Cognitive Wireless Networks: Concepts, Methodologies and Visions: Inspiring the Age of Enlightenment of Wireless Communications

Cognitive Wireless Networks Cognitive Wireless Networks Concepts, Methodologies and Visions Inspiring the Age of Enli...

Author: Frank H.P. Fitzek | Marcos D. Katz (Editors)

31 downloads 619 Views 79MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Cognitive Wireless Networks

Cognitive Wireless Networks Concepts, Methodologies and Visions Inspiring the Age of Enlightenment of Wireless Communications

Edited by

Frank H.P. Fitzek Aalborg University, Denmark and

Marcos D. Katz VTT, Finland

A C.I.P. Catalogue record for this book is available from the Library of Congress.

ISBN 978-1-4020-5978-0 (HB) ISBN 978-1-4020-5979-7 (e-book) Published by Springer, P.O. Box 17, 3300 AA Dordrecht, The Netherlands. www.springer.com

Printed on acid-free paper

© 2007 Springer No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work.

To our parents Eta-Marie and Werner

Fanny and Abraham (in memoriam) for their eternal support and loving.

What is Cognitive Radio and Cognitive Networks? Bernhard Walke RWTH Aachen University, Aachen, Germany

Cognitive Networks based on cognitive radio are addressing a revolutionary technology aiming, besides others, at remarkably improving efficiency of spectrum usage. When introduced, it will fundamentally change the way radio spectrum is regulated and used. Before this may happen, new enabling properties of radios are required such as sensing spectrum occupancy covering a wide range of spectrum and flexible spectrum access adapting to variable channel widths based on reasoning. Cognition has much to do with coexistence management: Coexistence of radio based systems operating in the same or in partly overlapping channels using the same or even different airinterfaces is the challenge to be solved. This appears to be especially difficult to achieve when coexistent radios operate different air interfaces and apply different transmit power levels, since the near-far and hidden terminal problems will apply then and would make coexistence management using de-central control hardly possible. The well-known pilot channel based control applied by an incumbent to control access to its licensed spectrum by third parties to manage coexistence of both appears workable only, if coexistence control is reduced to “yes or no” access permission for non-incumbents, depending on the needs of the primary user what is in fact a TDMA based resource sharing on a long time scale and not coexistence management. Much more sophisticated cooperation strategies appear necessary to enable small time-scale, location-specific coexistence management of radio systems, e.g. based on explicit information exchange between the radios involved or just being based on observations made, recently, by a radio following game theoretic reasoning, possibly, combined with well designed back-off algorithms. Prioritization of an incumbent might be part of the rules applied. As is visible from the assembly of contributions to this book, the state-of-the-art towards this is not much progressed, although a rich set of ideas exist potentially contributing to make efficient coexistence control of radios happen. It need not be mentioned that dynamic spectrum assignment can be achieved only if mobile terminals are able to re-configure on all layers of its protocol stacks. Cognitive radio aims to promote technologies as well as changes in radio regulation to overcome some

VIII

Bernhard Walke

existing barriers aiming to improve efficiency of spectrum utilization without scarifying highly reliable communication meeting high quality of service targets. Cognitive networks as a generic approach to exploit cognition in wireless networks: using cognitive principles to improve utilization of resources. In addition to spectral efficiency, for instance energy (power) efficiency can be also enhanced by exploiting cognition. Cognitive radio is just a particular instance of cognitive networks. However, Cognitive Networks / Cognitive Radio is a buzzword, too, since it umbrellas a number of more specific terms used to describe existent radio technology, namely • Mobile radio, a context-aware radio able to identify (based on cognitive capabilities like reasoning on measurement results) the best suited base station, supporting a given air interface standard, to serve a running session (called handover) or to associate to the base offering the strongest signal. • Multi-band adaptive radio to switch between distant channel groups, without changing the air-interface standard, like in use for GSM 900/1800/1900 MHz bands. Clearly, cognitive capabilities are required to decide on the band to use, based on measurement results. The related technique is dynamic spectrum assignment. • Multi-standard radios able to associate to one out of a number of different air interface standards like GSM/UMTS/CDMA2000/WLAN/others, or even switch air interface across standards during a running session (interstandards handover). Nobody would call this cognitive radio operation, although a lot of cognition related functions is required to operate a radio like this. • Multi-homing radios able to support different standards air interfaces at the same time. • Reconfigurable radio, e.g., – Hardware Defined Radio (HDR) comprising a set of radios housed in one box, each radio designed to serve a given air-interface, able to decide (based on cognitive capabilities), which radio to operate at a time. – Software Defined Radio (SDR) able to adapt transmission related or even protocol related parameters so that some (or many) properties of the radio are adapted to the needs. SDR would clearly need cognitive capabilities, too, to make senseful decisions. Wireless systems operated in license-exempt (ISM) bands, coexisting and sharing the spectrum according to a standard-specific set of common rules deciding on medium access, reflecting to own observations and based on measurement results, combined with reasoning (using cognitive capabilities). Spectrum etiquette and policies and open spectrum are the terms related to that operation.

What is Cognitive Radio and Cognitive Networks?

IX

Taking all of these already existent technologies into account, cognitive radios apparently need to go beyond in its aims and ambitions that currently are not well specified and expressed. There is a clear need for sorting things and differentiating known technologies from new ideas. Application of game theoretic models, vertical and horizontal spectrum sharing, overlay sharing (like UWB), reasoning and machineunderstandable regulatory rules - in general, feature detectors and cyclostationary detectors, spectrum opportunity identification and self-organization and cooperation in wireless networks are examples for new dimensions to be considered in the cognitive radios domain. Besides addressing cognitive networks, the book has a focus on user cooperation: The book advocates the concepts of wireless grids, which is an ad hoc cooperative cluster made of wireless terminals connected over short-range links, but at the same time, being connected to the cellular network. As behind each terminal there is a user who ultimately may decide to join or not a wireless grid, user decisions (individual and group) will have an impact on network operation and performance. Wireless networks enabling social networking and social networks shaping wireless networks are also discussed in the book. Cognitive networks and user cooperation are clearly related and the contributions presented by experts in the respective fields are really worth reading. I have found the themes of the articles invited to form the book very interesting and representing most recent research subjects. May this prove useful to you too.

Aachen, Germany, May 2007

Bernhard Walke

Preface

Sapere aude! Dare to know! Habe Mut, dich deines eigenen Verstandes zu bedienen! Sapere aude!, the emblematic motto associated with the Age of Enlightenment, is perhaps a rather eccentric expression to open a book on wireless communications. Alluding to that maxim, the German philosopher Immanuel Kant encouraged people to use their own minds as the basis for reasoning instead of following dogmatic rules. However, the expression has rich connotations, and we particularly see how dazzlingly inspires the current developments and future of wireless communication networks. This book basically deals with two complementary principles, cognition and cooperation, and how they are becoming essential for future wireless networks. Implicit cognition and cooperation have always been present in any wireless network as fundamental principles for ensuring basic network operation, as for instance the use of common protocols or signaling across the network, estimation of instantaneous channel conditions, etc. This book focuses on techniques exploiting cooperative and cognitive principles in an explicit manner, that is, purposely implemented by design, and aiming at enhancing the most relevant link and network performance figures as well as improving the efficiency in the use of resources. In general many of these techniques have just recently emerged and currently they are receiving increasing attention by the research community and industry. Cooperation in wireless networks is a well established and rather mature field, whereas cognitive principles, in their explicit way, are rapidly finding their way to the wireless world. The book presents a comprehensive cross-section of these promising fields, exploring these techniques in a highly

XII

Preface

motivating fashion. The goal is to describe key underlying concepts as well as their potentials and challenges. At the same time the book aims at providing mind sparkling discussions which will hopefully trigger further developments in this fertile research field. Thus, rather than being a conventional text book describing well established concepts and techniques, the spirit of the present volume is exploratory and highly motivational. The rationale for the recent extraordinary interest in exploiting cooperative techniques in wireless networks is clear, as demonstrated by the high number of new publications showing concrete advantages and potentials of cooperation. The collaborative interaction of different entities of a wireless network does pay off in basically any network, regardless of the involved access technology, architecture and operating scenarios. The cooperating entities include, among others, basic functional blocks, OSI layers, complete functionalities (e.g., access point, wireless devices), different networks. The benefits of cooperation are manifold, to name a few, increasing data throughput, extending the coverage area, enhancing quality of service and achieving higher efficiency in exploiting radio resources. An important remark is the fact that in addition to the aforementioned (purely technical) cooperating entities, the user, his or her eventual decision on joining a cooperative network and the preferred manner to cooperate are integral parts of the overall cooperative process. User decisions (individual and group) will have an impact on network operation and performance and therefore the social aspects of networking are of great importance. Wireless networks are increasingly enabling social networking, while social networks in turn are shaping the way in which wireless networks operate. Inter and intra network cooperation, presented here as the concept of cellular-controlled peer-to-peer communications will be particularly discussed and presented as a very promising composite access architecture for future wireless networks. Many important aspects of cooperation, together with novel emerging trends will be discussed in many of the upcoming chapters aiming at understanding the potentials and challenges behind the most relevant cooperative concepts suitable for wireless networks. Cooperation brings countless and unique opportunities to tackle many of the problems identified for future high performance wireless networks. From a wider perspective, the research challenges are not just technical but indeed multidisciplinary, involving for instance understanding individual and social behavioral patterns and their impact on the architecture, operation and ultimately performance of a wireless network. As users will become an important part of the cooperative equation, new aspects related to the social interaction between people will come into the scene, like incentives to foster user cooperation, users’ e-reputation and e-trust, security and privacy issues and others. For researchers and developing engineers the crossover between pure wireless communication techniques, group networking and social sciences is truly fascinating. Sapere aude!

Preface

XIII

Wireless communication networks have always had embedded some basic senses, like sensing the presence or activity of other users, monitoring the state of the channel, etc. In general sensing is followed by some action reacting in some way to the observation, like channel adaptation, signaling some information to the detected device, etc. The overall action can be understood as a simple cognitive process. In recent years many research initiatives have been focused on exploiting cognitive principles in an even more active fashion, by purposely deploying advanced sensory systems and intelligent processing able to interpret the observed wireless environment and adapt the system accordingly. The immense recent interest on cognitive radio can be considered the starting point triggering the research towards highly developed cognitive wireless networks. We certainly believe that senses of the wireless networks will further develop in accuracy and sensitivity, scrutinizing more and more parameters and activities of the surrounding wireless ecosystem. This will also trigger research targeting techniques for efficiently processing the increasing amount of data resulting from sensing. Moreover, and following the cognitive cycle, understanding the complex and dynamic surrounding scenario based on the on the acquired observations will require analysis tools making use of advanced reasoning techniques, creating a sound background for the following adaptation process. From the reigning spectrum oriented cognitive approach, we expect that additional figures will also monitored in future wireless networks, including other shared resources, configuration, capabilities and status of other surrounding (cooperating or competing) wireless devices, types of application being currently used, and others. A broad and deep understanding of the prevailing immediate wireless scene is paramount to attain high efficiency in the use of radio resources. Energy efficiency, vital for portable wireless devices, is in particular an important target for enhancing through cognition. Heterogeneous wireless networks with heterogeneous network entities, two key attributes of future wireless communication systems, need certainly to resort to cognitive principles in order to use efficiently their resources and ultimately, to support the use of cooperative principles within and across networks. A joint understanding of cognition and cooperation, two highly complementary principles, will render great benefits in terms of resource utilization and communication performance. Bringing cognition and cooperation into wireless networks in a harmonious and engineering attractive manner is a truly challenge for researchers and engineers. However, blending and embedding these principles into future wireless networks will be result in some of the most powerful techniques to cope with the clearly identified emerging problems. Eventually, we will witness the rising of wireless networks with such highly developed sensory, processing and reasoning capabilities, paving the way towards conscious wireless networks. Sapere aude!

XIV

Preface

Figure 1. Overview of the organization and chapters of the book.

Preface

XV

Structure of the Book This book consists of a collection of 35 chapters written by researchers from academia and research institutes from all over the globe. The chapters explore several technical and social aspects of cooperation and cognition in wireless networks, discussing also emerging trends and concepts in this field. Although this is a stand-alone volume, this book was conceived to complement our previous publication1 , in the same fashion that cooperation and cognition in wireless networks complement each other. Parts and corresponding chapters of the book are shown in Figure 1. Part I serves as an introductory and motivating overture of the book, presenting an overview of this emerging field, discussing promising trends and their challenges. Part II deals mainly with cooperative concepts applied to wireless networks, highlighting not only the communicational aspects but also the social and operational ones. In particular the social side of cooperative networking, where users, their interactions and decisions are also taken into account when designing a wireless network, is discussed in detail. Part III introduces cognitive networks, from a generic standpoint, and also considering the currently most studied aspect of them, namely cognitive radio. Part IV explores techniques exploiting both cognitive and cooperative principles, putting into evidence the natural synergy between these approaches. Part V discusses some methodologies and tools used to model, analyze and design cooperative and cognitive networks. Finally, some interesting visions, prospects and emerging technologies are presented in Part VI. We note that the book was deliberately planned to contain assorted chapters of different nature, including motivating discussions, overviews of technical existing technologies, ideas for new concepts, newly proposed and emerging techniques, performance analyses, description and applications of tools for modeling and analyzing cooperative and cognitive techniques as well as visions and prospects. With such a varied source of information we expect that reader will get not only a balanced view of the subject but also, and most importantly, an enlightening introduction to this enthralling field.

Intended Readership This book is intended to serve as a reference and stimulating source for scientists and research engineers developing ideas and concrete concepts for future wireless networks. Furthermore, the academic community engaged in research on advanced wireless networks will find in this book a useful source of ideas for further development. Students and self-study readers will be able to get a deep understanding to the latest ideas and future developments in mobile and wireless communications and motivating. 1

“Cooperation in Wireless Networks: Principles and Applications”, edited by Frank H.P. Fitzek and Marcos D. Katz, ISBN1-4020-4710-X. Springer, 2006, pp. 694

XVI

Preface

Contacting the Editors The editors welcome any suggestions, comments or constructive criticism on this book. Such a feedback would be used to improve forthcoming editions. Editors can be contacted at [email protected].

Aalborg, Denmark Oulu, Finland June 2007

Frank H.P. Fitzek Marcos Katz

Contents

What is Cognitive Radio and Cognitive Networks? Bernhard Walke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . VII Part I Introductory Chapter 1 Cooperative and Cognitive Networks: A Motivating Introduction Marcos D. Katz, Frank H.P. Fitzek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Ten Tenets Shaping Future Wireless Communications . . . . . . . . . . . 1.3 An Introduction to Cooperative Wireless Networks . . . . . . . . . . . . . 1.4 An Introduction to Cognitive Communication Systems . . . . . . . . . . 1.5 Towards Cooperative and Cognitive Wireless Communications . . . 1.6 Discussions and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 3 4 14 20 24 25 29

2 Cellular Controlled Peer to Peer Communications: Overview and Potentials Frank H.P. Fitzek, Marcos Katz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Challenges for Future Wireless Networks . . . . . . . . . . . . . . . . . . . . . . 2.2 Premises for Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Combining the Cellular and the P2P World . . . . . . . . . . . . . . . . . . . . 2.3.1 Cooperative Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Realization of the Cellular and the Short Range Link . . . . 2.3.3 The Importance of the Short-Range Communication Link 2.3.4 Somebody out There? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.5 Nature Inspired Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Cooperative Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Multicast and Broadcast Services . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Unicast Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31 31 36 39 40 41 42 44 44 46 47 53

XVIII

Contents

2.5 Service Discovery within Cooperative Cluster . . . . . . . . . . . . . . . . . . 2.6 Benefits of Cooperation in the Wireless World . . . . . . . . . . . . . . . . . 2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

54 56 58 58

Part II Cooperative Networks: Social, Operational and Communicational Aspects 3 Applying Evolutionary Approaches for Cooperation David Hales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 From Evolution to Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 The Prisoner’s Dilemma and Variants . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Tag-Based Cooperation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 The SLAC Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Possible Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63 63 64 66 67 68 69 72 73 73

4 The Social Qualities of Pervasive Wireless Networks Mark Pesce . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction: The Social Hormone . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Hyperintelligence and the End of Elites . . . . . . . . . . . . . . . . . . . . . . . 4.3 Pox Populi and the Collapse of the Mass Mind . . . . . . . . . . . . . . . . . 4.4 Read-Write Culture and the Restructuring of Institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Conclusion: Pervasive Wireless Networks and the Rise of Hyperpeople . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

84 86

5 Encouraging Cooperative Interaction among Network Entities Sonja Buchegger, John Chuang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Benefits of Network Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 The Cooperation Dilemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Solution Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 Reputation Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.2 Payment Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.3 Barter Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.4 Enforcement Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Selfish v. Malicious v. Faulty Behavior . . . . . . . . . . . . . . . . . 5.4.2 Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

87 87 88 91 91 93 94 94 96 96 96

75 75 76 79 82

Contents

XIX

5.4.3 Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 5.4.4 Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 5.4.5 Meta Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.4.6 Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 6 Competition and Cooperation in Wireless Multi-Access Networks Johan Hultell, Jens Zander, Jan Markendahl . . . . . . . . . . . . . . . . . . . . . . . . 109 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 6.2 Technology for dynamic cooperation and competition . . . . . . . . . . . 112 6.3 Cooperative Wireless Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 6.3.1 Benefits and Perils of Cooperation . . . . . . . . . . . . . . . . . . . . . 115 6.3.2 How Much Can Be Gained through Cooperation? . . . . . . . 116 6.3.3 Current Practice of Infrastructure Cooperation . . . . . . . . . . 118 6.4 Competitive Wireless Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 6.4.1 Benefits of Competitive Wireless Access . . . . . . . . . . . . . . . . 119 6.4.2 Feasibility of a Competitive Wireless Access Market . . . . . 120 6.5 Dynamic Cooperation and Competition . . . . . . . . . . . . . . . . . . . . . . . 126 6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 7 A Cooperative ID for 4G Simone Frattasi, Hanane Fathi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 7.2 The Fall of 3G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 7.3 The Raise of 4G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 7.3.1 Prophetic Visions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 7.3.2 A Pragmatic Methodology to Define 4G . . . . . . . . . . . . . . . . 135 7.4 Examples of User and Group Scenarios . . . . . . . . . . . . . . . . . . . . . . . . 137 7.4.1 Business on-the-Move . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 7.4.2 Smart Shopping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 7.4.3 Mobile Tourist Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 7.4.4 Personalization Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 7.5 Key Features of 4G from the User and Group Perspectives . . . . . . . 138 7.5.1 User Friendliness, User and Group Personalization . . . . . . . 139 7.5.2 Terminal and Network Heterogeneity . . . . . . . . . . . . . . . . . . 140 7.6 Technical Requirements and Expectations for 4G . . . . . . . . . . . . . . . 142 7.6.1 System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 7.6.2 Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 7.6.3 Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 7.7 Towards a Definition of 4G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 7.7.1 The Ad-Coop Network Model . . . . . . . . . . . . . . . . . . . . . . . . . 147 7.7.2 The Alchemy of Cooperation in 4G Wireless . . . . . . . . . . . . 148

XX

Contents

7.8

Enabling Wireless Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 7.8.1 Group Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 7.8.2 Cooperation Triggers and Types of Cooperation . . . . . . . . . 150 7.8.3 The User Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 7.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 8 Implementing Cooperative Wireless Networks Stefan Valentin, Hermann S. Lichte, Holger Karl, S´ebastien Simoens, Guillaume Vivier, Josep Vidal, Adrian Agustin . . . . . . . . . . . . . . . . . . . . . . 155 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 8.2 Approaches in User Cooperative Diversity . . . . . . . . . . . . . . . . . . . . . 157 8.2.1 From Relaying to User Cooperation Diversity . . . . . . . . . . . 158 8.2.2 Current Approaches – A Classification . . . . . . . . . . . . . . . . . 159 8.3 Designing Cooperative Systems – New Problems and Required Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 8.3.1 Mobile Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 8.3.2 Cooperation-Aware Resource Allocation . . . . . . . . . . . . . . . . 165 8.3.3 Medium Access Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 8.4 Towards Feasibility – Implementing Cooperative Systems . . . . . . . . 173 8.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks Peter Rost, Gerhard Fettweis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 9.2 Nomenclature and Relay Network Model . . . . . . . . . . . . . . . . . . . . . . 181 9.3 Protocols for Half-Duplex Relay Nodes . . . . . . . . . . . . . . . . . . . . . . . . 182 9.3.1 A Compress-And-Forward Based Approach . . . . . . . . . . . . . 182 9.3.2 A Decode-And-Forward Based Approach . . . . . . . . . . . . . . . 189 9.3.3 Mixed Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 9.4 Application to Wireless Communications . . . . . . . . . . . . . . . . . . . . . . 193 9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 10 Trigger Management and Mobile Node Cooperation Jukka M¨ akel¨ a, Kostas Pentikousis, Mikko Majanen, Jyrki Huusko . . . . . 199 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 10.2 Mobility Triggers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 10.2.1 TRG Producers and Consumers . . . . . . . . . . . . . . . . . . . . . . . 200 10.2.2 The Role of TRG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 10.3 A Trigger Management Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 202 10.3.1 Triggering Events Collection . . . . . . . . . . . . . . . . . . . . . . . . . . 203 10.3.2 Trigger Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 10.3.3 Trigger Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 10.3.4 TRG Policies and Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Contents

XXI

10.4

Routing Group Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 10.4.1 Routing Group Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 10.4.2 Stability-Based Multi-Hop Clustering Protocol . . . . . . . . . . 207 10.4.3 An Overview of the Gateway Selection Architecture . . . . . 209 10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 11 Cooperative Mobile Positioning in 4G Wireless Networks Simone Frattasi, Marco Monti . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 11.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 11.2.1 Hybrid Positioning Techniques . . . . . . . . . . . . . . . . . . . . . . . . 216 11.2.2 NLOS Error Mitigation Techniques . . . . . . . . . . . . . . . . . . . . 217 11.3 The Ad-Coop Positioning System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 11.3.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 11.3.2 Data Fusion Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 11.4 Simulation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 11.4.1 Statistical Models for TOA and AOA Estimation Errors . . 223 11.4.2 Statistical Channel Model for RSS Estimation . . . . . . . . . . 224 11.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 11.5.1 Performance Dependency on the Number of CMs . . . . . . . . 226 11.5.2 Performance Dependency on the Number of BSs . . . . . . . . 228 11.6 Localization, Cooperation and Cognition . . . . . . . . . . . . . . . . . . . . . . 228 11.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 12 Peer-to-Peer Information Retrieval Based on Fields of Interest Bertalan Forstner, Gergely Cs´ ucs, Imre Kel´enyi, Hassan Charaf . . . . . . . 235 12.1 Inspiration from Everyday Life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 12.2 Modeling Fields of Interests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 12.2.1 The Semantic Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 12.2.2 The Connection Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 12.2.3 The Reply Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 12.2.4 The Query Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 12.3 Protocol Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 12.4 Protocol Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 12.5 The Application of Our Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 12.5.1 Designing Symella . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 12.5.2 The Architecture of Symella . . . . . . . . . . . . . . . . . . . . . . . . . . 247 12.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

XXII

Contents

Part III Cognitive Networks 13 Introducing Cognitive Systems to the B3G Wireless World P. Demestichas, G. Dimitrakopoulos, K. Tsagkaris, V. Stavroulaki, and A. Katidiotis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 13.1.1 The Wireless World Today . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 13.1.2 Motivation: Cognitive Networks and their Management Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 13.2 Management Functionality for Cognitive Network Segments . . . . . . 257 13.2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 13.2.2 Cognitive Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 13.3 Management Functionality for Cognitive Access Points . . . . . . . . . . 260 13.3.1 The Autonomic Management of Access Points (AMAP) . . 260 13.3.2 Cognitive Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 13.4 Management Functionality for Cognitive Wireless Terminals . . . . . 262 13.4.1 The Cognitive Reconfigurable Equipment Management System (C REMS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 13.4.2 Cognitive Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 13.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 14 Architectures and Protocols for Next Generation Cognitive Networking B. S. Manoj, Ramesh R. Rao, Michele Zorzi . . . . . . . . . . . . . . . . . . . . . . . . 271 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 14.2 Definition of Cognitive Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . 272 14.3 Architectures for Cognitive Networking . . . . . . . . . . . . . . . . . . . . . . . . 272 14.3.1 Autonomous Cognitive Networking . . . . . . . . . . . . . . . . . . . . 273 14.3.2 Distributed Cognitive Networking . . . . . . . . . . . . . . . . . . . . . 277 14.4 CogNet: Cognitive Complete Knowledge Network . . . . . . . . . . . . . . . 277 14.4.1 CogPlane and CogBus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 14.4.2 Case Study: CogTCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 14.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 15 Scheduling in Cognitive Networks Chandrasekharan Raman, Jasvinder Singh, Roy D. Yates, and Narayan B. Mandayam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 15.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 15.3 Maximum Sum Rate Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 15.4 Fair Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 15.4.1 Max-Min Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 15.4.2 Proportional Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

Contents

XXIII

15.5

Distributed Dynamic Spectrum Access Policies . . . . . . . . . . . . . . . . . 293 15.5.1 Rate Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 15.5.2 Characterization of Rate Region for the Decentralized Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 15.5.3 Distributed Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297 15.6 Cross Layer Scheduling of End-to-End Flows . . . . . . . . . . . . . . . . . . . 299 15.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301 15.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 16 Design of Terminals and Infrastructure Components for Cognitive Wireless Networks Alexander Vießmann, Admir Burnic, Christoph Spiegel, Arjang Hessamian-Alinejad, Andreas Waadt, Guido H. Bruck, Peter Jung . . . . . 307 16.1 Ubiquitous Wireless Multimedia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 16.2 Reconfigurability and Cognitive Modes of Operation . . . . . . . . . . . . 309 16.3 Platform Based Design Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 16.4 PROMETHEUS Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 16.4.1 General Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 16.4.2 Transceiver Engines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317 16.4.3 HAWK Transceiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318 16.5 Future Proofness of the PROMETHEUS Platform . . . . . . . . . . . . . . 322 16.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325 17 Fundamental Limits of Cognitive Radio Networks Natasha Devroye, Vahid Tarokh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 17.1.1 Chapter Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 17.2 Fundamental Limits of Cognitive Radio Channels: Perfect CSI . . . 329 17.2.1 Gaussian Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 17.2.2 Discrete Memoryless Channel . . . . . . . . . . . . . . . . . . . . . . . . 337 17.2.3 Further Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342 17.3 Fundamental Limits of Cognitive Radio Channels: Imperfect CSI and Fading Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344 17.3.1 The Compound Gel’fand-Pinsker Channel . . . . . . . . . . . . . . 344 17.3.2 Carbon Copying onto Dirty Paper . . . . . . . . . . . . . . . . . . . . . 346 17.3.3 Gel’fand-Pinkser Coding with Unknown Phase . . . . . . . . . . 348 17.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

XXIV

Contents

18 Spectrum Awareness: Techniques and Challenges for Active Spectrum Sensing Marko H¨ oyhty¨ a, Atso Hekkala, Marcos Katz, Aarne M¨ ammel¨ a . . . . . . . . . 353 18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353 18.2 A Classification of Spectrum Awareness . . . . . . . . . . . . . . . . . . . . . . . 354 18.2.1 Passive Awareness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 18.2.2 Active Awareness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357 18.2.3 Response Time and Topology . . . . . . . . . . . . . . . . . . . . . . . . . 358 18.3 Spectrum Sensing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 18.3.1 Matched Filter Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360 18.3.2 Energy Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360 18.3.3 Feature Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 18.3.4 Interference Temperature Concept . . . . . . . . . . . . . . . . . . . . . 363 18.4 Spectrum Sensing Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365 18.5 To Cooperate or Not to Cooperate? . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 18.6 Emerging Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 18.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369 19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks Danijela Cabric, Robert Brodersen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373 19.1 Spectrum Sensing for Cognitive Radio Networks . . . . . . . . . . . . . . . 373 19.1.1 Requirements and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . 373 19.1.2 System Design Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374 19.2 Signal Processing Techniques for Spectrum Sensing . . . . . . . . . . . . . 375 19.2.1 Simple General Approach - Energy Detector . . . . . . . . . . . . 375 19.2.2 Exploiting Deterministic Signals - Coherent Processing . . . 377 19.2.3 Detecting Signal Features - Cyclostationary Processing . . . 381 19.3 Network Level Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388 19.3.1 Exploiting Diversity - Cooperative Sensing . . . . . . . . . . . . . 388 19.3.2 Limitations in Cooperative Sensing . . . . . . . . . . . . . . . . . . . . 391 19.4 System Design Guidelines for Spectrum Sensing . . . . . . . . . . . . . . . . 392 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 Part IV Marrying Cooperation and Cognition in Wireless Networks 20 Cognitive Resource Manager Marina Petrova, Petri M¨ ah¨ onen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 20.2 CRM Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399 20.3 Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 20.3.1 ULLA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 20.3.2 Common Application Requirements Interface . . . . . . . . . . . 405 20.3.3 Universal Network Interface . . . . . . . . . . . . . . . . . . . . . . . . . . 405

Contents

XXV

20.4

Core Unit Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 20.4.1 Learning and Reasoning with Genetic Algorithms . . . . . . . 407 20.4.2 Decision Making and Utility Functions . . . . . . . . . . . . . . . . . 410 20.4.3 Managing Time-Scales: CRM-core . . . . . . . . . . . . . . . . . . . . . 415 20.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 21 The C-Cube Concept - Combining Cross-Layer Protocol Design, Cognitive-, and Cooperative Network Concepts Thomas Arildsen, Frank H.P. Fitzek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423 21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423 21.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 21.3 Cognitive Networking in Cellular Networks . . . . . . . . . . . . . . . . . . . . 426 21.4 Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429 21.5 Cross-Layer and Cognition Combination . . . . . . . . . . . . . . . . . . . . . . . 431 21.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432 22 Cellular Controlled P2P Communication Using Software Defined Radio Jesper M. Kristensen, Frank H.P. Fitzek . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435 22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435 22.2 Realization Forms of the CCP2P Scenario . . . . . . . . . . . . . . . . . . . . . 436 22.2.1 Multi-Mode Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439 22.2.2 Combined Cellular and Short Range Air Interface . . . . . . . 439 22.3 SDR and SCR Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447 22.3.1 SCR Architecture Versus SDR Architecture . . . . . . . . . . . . . 449 22.3.2 SDR Receiver Architecture for Cooperating Terminal . . . . 451 22.4 CCP2P Testbed with the GNU Radio . . . . . . . . . . . . . . . . . . . . . . . . . 452 22.4.1 Introduction to GNU Radio . . . . . . . . . . . . . . . . . . . . . . . . . . 452 22.4.2 GNU Radio Setup for a CCP2P Cooperative Scenario . . . . 453 22.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454 23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks Chenguang Lu, Frank H.P. Fitzek, Patrick C.F. Eggers . . . . . . . . . . . . . . . 457 23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457 23.2 Description of the Proposed CSR Scheme . . . . . . . . . . . . . . . . . . . . . . 458 23.2.1 Cooperation Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460 23.2.2 CSR Capacity Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461 23.2.3 CSR Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462 23.3 CSR with Transmit Beamforming on MISO Links . . . . . . . . . . . . . . 463 23.3.1 Transmit Beamforming on MISO Links . . . . . . . . . . . . . . . . 463 23.3.2 MRC-TDMA Versus ZF-CSR . . . . . . . . . . . . . . . . . . . . . . . . . 463

XXVI

Contents

23.4

Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466 23.4.1 CSR Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466 23.4.2 Capacity Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467 23.4.3 Energy Efficiency Saving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467 23.5 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470 24 On the Energy Saving Potential in DVB-H Networks Exploiting Cooperation among Mobile Devices Qi Zhang, Frank H.P. Fitzek, Marcos Katz . . . . . . . . . . . . . . . . . . . . . . . . . . 473 24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473 24.2 Cooperative Strategy for IP-services over DVB-H . . . . . . . . . . . . . . . 475 24.3 Cooperative Short–Range Communication . . . . . . . . . . . . . . . . . . . . . 476 24.3.1 Topology Based Cooperative Algorithm . . . . . . . . . . . . . . . . 477 24.3.2 Signalling on the Short-Range Link . . . . . . . . . . . . . . . . . . . 478 24.4 Numerical Examples for Energy Consumption Analysis . . . . . . . . . . 479 24.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484 25 Cooperative Retransmission for Reliable Wireless Multicast Services Qi Zhang, Frank H.P. Fitzek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485 25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485 25.2 Non–Cooperative Error Recovery Strategies . . . . . . . . . . . . . . . . . . . . 487 25.2.1 ARQ Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487 25.2.2 FEC/HARQ Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488 25.3 Cooperative Retransmission Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 489 25.3.1 Frame Structure Design on Cellular Link with TDD Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 25.3.2 Design Cooperative Retransmission Scheme on the Short–Range Link . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490 25.3.3 Energy Consumption by Cooperative Retransmission Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 25.4 Comparison of Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . 494 25.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497 26 IP Header Compression for Cellular-Controlled P2P Networks Tatiana K. Madsen, Qi Zhang, Frank H.P. Fitzek . . . . . . . . . . . . . . . . . . . . 499 26.1 Introduction and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 26.2 Autonomous and Cooperative Header Compression in Cellular-Controlled P2P Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 502 26.3 Design of Information Exchange over Short Range Connections . . . 504 26.4 Evaluation of CCP2P Header Compression . . . . . . . . . . . . . . . . . . . . 506 26.4.1 Probability of Error Burst . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506 26.4.2 Bandwidth Savings and Energy Efficiency . . . . . . . . . . . . . . 507

Contents

XXVII

26.5 Discussion on Cooperation Strategies . . . . . . . . . . . . . . . . . . . . . . . . . 508 26.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510 27 Cluster Based Cooperative Uplink Access in Centralized Wireless Networks Qi Zhang, Frank H.P. Fitzek, Villy B. Iversen . . . . . . . . . . . . . . . . . . . . . . . 513 27.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513 27.2 CSMA/CA Based MAC Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514 27.2.1 RTS/CTS Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514 27.2.2 Packet Aggregation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 518 27.3 The One4all Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522 27.3.1 Throughput & Channel Access Delay Analysis . . . . . . . . . . 523 27.3.2 Energy Consumption Analysis . . . . . . . . . . . . . . . . . . . . . . . . 524 27.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524 27.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Part V Methodologies and Tools 28 Cooperation for Cognitive Networks: A Game Theoretic Perspective Cristina Comaniciu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 28.1 Future Generation of Wireless Networks: Opportunities and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 28.2 A Game Theoretic Framework for Cooperation . . . . . . . . . . . . . . . . . 534 28.2.1 Modeling the Cognition Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 534 28.2.2 Coalitional Game Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535 28.2.3 Non-Cooperative Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536 28.3 Cooperative Protocols for Cognitive Networks: Some Examples . . . 537 28.3.1 Implicit Cooperation in Protocol Design . . . . . . . . . . . . . . . . 537 28.3.2 Incentivizing Cooperation for Non-Cooperative Games . . . 539 28.4 Conclusions and Open Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550 29 Spectrum Sharing Games of Network Operators and Cognitive Radios Mohammad Hossein Manshaei, M´ ark F´elegyh´ azi, Julien Freudiger, Jean-Pierre Hubaux, Peter Marbach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555 29.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555 29.2 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556 29.2.1 Game Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557 29.2.2 Auction Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559 29.2.3 Graph Coloring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560

XXVIII

Contents

29.3

Network Operator Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561 29.3.1 WAN-WiFi Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561 29.3.2 National Border Spectrum Sharing . . . . . . . . . . . . . . . . . . . . 562 29.3.3 Network Operators Spectrum Sharing . . . . . . . . . . . . . . . . . . 564 29.4 Games in Unlicensed Bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567 29.4.1 Spectrum Sharing among Heterogeneous Wireless Systems 567 29.4.2 Spectrum Sharing among WiFi Operators . . . . . . . . . . . . . . 569 29.5 Cognitive Radio Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571 29.5.1 Opportunistic Spectrum Sharing . . . . . . . . . . . . . . . . . . . . . . 572 29.5.2 Auction Based Spectrum Sharing . . . . . . . . . . . . . . . . . . . . . . 573 29.5.3 Spectrum Sharing in OFDM Networks . . . . . . . . . . . . . . . . . 574 29.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577 30 Introduction to NetLogo Federico Albiero, Frank H.P. Fitzek, Marcos Katz . . . . . . . . . . . . . . . . . . . . 579 30.1 Why NetLogo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579 30.2 Main Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580 30.3 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581 30.4 A Model for Cooperation in Wireless Networks . . . . . . . . . . . . . . . . . 583 30.4.1 The PD Iterated Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583 30.4.2 Modeling Cooperation in Wireless Networks . . . . . . . . . . . . 586 30.5 NetLogo Libraries for Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587 30.5.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587 30.5.2 Cooperative Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595 30.5.3 Displaying Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599 30.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601 31 Analysis of Cooperative Power Saving Strategies with NetLogo Federico Albiero, Frank Fitzek, Marcos Katz . . . . . . . . . . . . . . . . . . . . . . . . . 603 31.1 Scenario of Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603 31.2 Theoretic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605 31.3 Strategies Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607 31.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610 31.4.1 Power Saving Gain of Cooperation . . . . . . . . . . . . . . . . . . . . 611 31.4.2 Mixed Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614 31.4.3 Two-Box Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615 31.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620

Contents

XXIX

Part VI Visions, Prospects and Emerging Technologies 32 Cooperation in Optical Wireless Communications Dominic O’Brien . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623 32.2 Overview of Optical Wireless Communications . . . . . . . . . . . . . . . . . 624 32.3 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625 32.3.1 Transmitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626 32.3.2 Receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626 32.4 Cooperation between RF and Optical Wireless (OW) Systems . . . 627 32.5 Potential Scenarios for Cooperative Working . . . . . . . . . . . . . . . . . . . 628 32.5.1 Hotspots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 628 32.5.2 Cooperative Transceivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 628 32.5.3 Enabling Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629 32.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631 33 Evolution of Digital Radios Friedrich K. Jondral, Volker Blaschke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635 33.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635 33.2 Transmission Physics and Standards . . . . . . . . . . . . . . . . . . . . . . . . . . 635 33.3 Radio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637 33.4 Digital Radio (DR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640 33.5 Software Defined Radio (SDR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643 33.5.1 SDR Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643 33.5.2 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644 33.5.3 Military SDR - The Software Communications Architecture (SCA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645 33.6 Cognitive Radio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647 33.6.1 Cognitive Radio Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648 33.6.2 Functional Enhancement of Cognitive Radios . . . . . . . . . . . 649 33.6.3 Cognitive Radio Features in Current and Future Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650 33.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653 34 CogMesh: A Cluster Based Cognitive Radio Mesh Network Tao Chen, Honggang Zhang, Xiaofei Zhou, Gian Mario Maggio, Imrich Chlamtac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657 34.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657 34.2 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659 34.3 Network Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661 34.4 MAC Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662

XXX

Contents

34.5 34.6

Spectrum Hole Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663 Neighbor Discovery and Cluster Formation . . . . . . . . . . . . . . . . . . . . 664 34.6.1 Analysis of Neighbor Discovery Approaches . . . . . . . . . . . . . 666 34.7 Inter-Cluster Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666 34.8 Topology Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668 34.8.1 Nodes Join Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668 34.8.2 Nodes Leave Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669 34.8.3 Spectrum Holes Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670 34.8.4 Cluster Shift Master Channel . . . . . . . . . . . . . . . . . . . . . . . . . 670 34.8.5 Merge Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671 34.9 Correctness of Network Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . 673 34.10 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675 34.11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677 35 Coordinating User and Device Behavior in Wireless Grids Lee W. McKnight, William Lehr, James Howison . . . . . . . . . . . . . . . . . . . . 679 35.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679 35.2 From Systems Management to Grid Coordination . . . . . . . . . . . . . . 680 35.3 Coordinating Strategic Behavior in Distributed Networks . . . . . . . . 683 35.3.1 Technical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683 35.3.2 Social . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685 35.3.3 Legal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687 35.3.4 Economic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688 35.4 Interactions and Dynamics in Regulation . . . . . . . . . . . . . . . . . . . . . . 691 35.4.1 Hardening Technical Regulation with Legal Enforcement . 691 35.4.2 New Legal Provisions and Their Surveillance Implications 692 35.5 Conclusion and Implications for Wireless Grids . . . . . . . . . . . . . . . . . 693 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709

List of Contributors

Marcos Katz VTT Technical Research Centre of Finland P.O. Box 1100 FI-90571 Oulu Finland [email protected] Frank H. P. Fitzek Aalborg University Department of Electronic Systems Niels Jernes Vej 12 9220 Aalborg [email protected] David Hales University of Bologna Dept. of Computer Science Mura Anteo Zamboni 7 40127 Bologna, Italy. [email protected] Mark Pesce Honorary Associate Digital Cultures Program University of Sydney [email protected] Sonja Buchegger Deutsche Telekom Laboratories Ernst-Reuter-Platz 7

D-10587 Berlin Germany [email protected] John Chuang School of Information University of California at Berkeley 102, South Hall Berkeley, CA 94720 [email protected] Johan Hultell Department of Communication Systems, Royal Institute of Technology Isafjordsgatan 30B, Electrum 418 SE-164 40 Kista [email protected] Jens Zander Department of Communication Systems, Royal Institute of Technology Isafjordsgatan 30B, Electrum 418 SE-164 40 Kista [email protected] Jan Markendahl Department of Communication Systems, Royal Institute of Technology

XXXII

List of Contributors

Isafjordsgatan 30B, Electrum 418 SE-164 40 Kista [email protected] Simone Frattasi Center for TeleInFrastruktur (CTIF) Dpt. of Electronic Systems Antennas, Propagation and Radio Networking (APNet) Group Aalborg University Niels Jernes Vej 12, 9220 Aalborg, Denmark [email protected] Hanane Fathi Research Center for Information Security (RCIS) National Institute of Advanced Industrial Science and Technology (AIST) Akihabara-Daibiru 1-18-13, 101-0021 Tokyo, Japan [email protected] Stefan Valentin University of Paderborn Warburger Straße 100 33098 Paderborn, Germany [email protected] Hermann S. Lichte University of Paderborn Warburger Straße 100 33098 Paderborn, Germany [email protected]

Guillaume Vivier Motorola Labs Parc Les Algorithmes St-Aubin 91193, France [email protected] Josep Vidal Technical University of Catalonia C/Jordi Girona 1-3 08034 Barcelona, Spain [email protected] Adrian Agustin Technical University of Catalonia C/Jordi Girona 1-3 08034 Barcelona, Spain [email protected] Peter Rost Technische Universit¨at Dresden Dresden, Germany [email protected] Gerhard Fettweis Technische Universit¨at Dresden Dresden, Germany [email protected] Jukka M¨ akel¨ a VTT Technical Research Centre of Finland Oulu, Finland [email protected]

Holger Karl University of Paderborn Warburger Straße 100 33098 Paderborn, Germany [email protected]

Mikko Majanen VTT Technical Research Centre of Finland Oulu, Finland [email protected]

S´ ebastien Simoens Motorola Labs Parc Les Algorithmes St-Aubin 91193, France [email protected]

Kostas Pentikousis VTT Technical Research Centre of Finland Oulu, Finland [email protected]

List of Contributors

Jyrki Huusko VTT Technical Research Centre of Finland Oulu, Finland [email protected] Marco Monti CTIF Italy Dpt. of Electronic Engineering University of Rome Tor Vergata Via del Politecnico 1, 00133 Rome, Italy [email protected] Bertalan Forstner Budapest University of Technology and Economics H-1111 Budapest, Goldmann Gy¨ orgy t´er 3. Hungary [email protected] Gergely Cs´ ucs Budapest University of Technology and Economics H-1111 Budapest, Goldmann Gy¨ orgy t´er 3. Hungary [email protected] Imre Kel´ enyi Budapest University of Technology and Economics H-1111 Budapest, Goldmann Gy¨ orgy t´er 3. Hungary [email protected] Hassan Charaf Budapest University of Technology and Economics H-1111 Budapest, Goldmann Gy¨ orgy t´er 3. Hungary [email protected] P. Demestichas University of Piraeus Department of Digital Systems 80 Karaoli Dimitriou str. Piraeus, 18534, GREECE [email protected]

XXXIII

G. Dimitrakopoulos University of Piraeus Department of Digital Systems 80 Karaoli Dimitriou str. Piraeus, 18534, GREECE [email protected] K. Tsagkaris University of Piraeus Department of Digital Systems 80 Karaoli Dimitriou str. Piraeus, 18534, GREECE [email protected] V.Stavroulaki University of Piraeus Department of Digital Systems 80 Karaoli Dimitriou str. Piraeus, 18534, GREECE [email protected] A. Katidiotis University of Piraeus Department of Digital Systems 80 Karaoli Dimitriou str. Piraeus, 18534, GREECE Tel: +30 210 414 2758, Fax: +30 210 414 2753 [email protected] B. S. Manoj Department of Electrical and Computer Engineering University of California San Diego CA 92093 [email protected] Ramesh R. Rao Department of Electrical and Computer Engineering University of California San Diego CA 92093 [email protected]

XXXIV

List of Contributors

Michele Zorzi Department of Information Engineering University of Padova, Italy [email protected] Chandrasekharan Raman WINLAB, Rutgers 671 US Route 1 South North Brunswick, NJ 08902 [email protected] Jasvinder Singh WINLAB, Rutgers 671 US Route 1 South North Brunswick, NJ 08902 [email protected] Roy D. Yates WINLAB, Rutgers 671 US Route 1 South North Brunswick, NJ 08902 [email protected] Narayan B. Mandayam WINLAB, Rutgers 671 US Route 1 South North Brunswick, NJ 08902 [email protected] Alexander Vießmann Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Alex.Viessmann@ KommunikationsTechnik.org Admir Burnic Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Admir.Burnic@ KommunikationsTechnik.org

Christoph Spiegel Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Christoph.Spiegel@ KommunikationsTechnik.org Arjang Hessamian-Alinejad Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Arjang.Hessamian@ KommunikationsTechnik.org Andreas Waadt Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Andreas.Waadt@ KommunikationsTechnik.org Guido H. Bruck Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Guido.Bruck@ KommunikationsTechnik.org Peter Jung Lehrstuhl f¨ ur Kommunikationstechnik Universit¨at Duisburg-Essen Germany Peter.Jung@ KommunikationsTechnik.org Natasha Devroye School of Engineering and Applied Sciences Harvard University Cambridge, MA, U.S.A. [email protected]

List of Contributors

Vahid Tarokh School of Engineering and Applied Sciences Harvard University Cambridge, MA, U.S.A. [email protected] Marko H¨ oyhty¨ a VTT Technical Research Centre of Finland P.O. Box 1100, FI-90571 Oulu Finland [email protected] Atso Hekkala VTT Technical Research Centre of Finland P.O. Box 1100, FI-90571 Oulu Finland [email protected] Aarne M¨ ammel¨ a VTT Technical Research Centre of Finland P.O. Box 1100, FI-90571 Oulu Finland [email protected] Danijela Cabric University of California, Berkeley [email protected] Robert Brodersen University of California, Berkeley [email protected]

XXXV

Thomas Arildsen Aalborg University Department of Electronic Systems Niels Jernes Vej 12 9220 Aalborg [email protected] Chenguang Lu Aalborg University Department of Electronic Systems, APNET Section Niels Jernes Vej 12, DK-9220 Aalborg East, Denmark [email protected] Patrick C.F. Eggers Aalborg University, Department of Electronic Systems, APNET Section Niels Jernes Vej 12, DK-9220 Aalborg East, Denmark [email protected] Qi Zhang Technical University of Denmark Department of Communication Optics & Materials Building 343, DK-2800 Kgs. Lyngby Denmark [email protected]

Marina Petrova RWTH Aachen University Department of Wireless Networks Kackertstrasse 9, 52072 Aachen Germany [email protected]

Tatiana K. Madsen Aalborg University Niels Jernes Vej 12, DK-9220 Aalborg, Denmark [email protected]

Petri M¨ ah¨ onen RWTH Aachen University Department of Wireless Networks Kackertstrasse 9, 52072 Aachen Germany [email protected]

Jesper M. Kristensen Aalborg University, Department of Electronic Systems Niels Jernes Vej 12 DK-9220 Aalborg, Denmark [email protected]

XXXVI

List of Contributors

Villy B. Iversen Technical University of Denmark Department of Communication Optics & Materials Building 343, DK-2800 Kgs. Lyngby Denmark [email protected] Cristina Comaniciu Stevens Institute of Technology Hoboken, NJ, USA [email protected] Mohammad Hossein Manshaei EPFL, Switzerland [email protected] M´ ark F´ elegyh´ azi EPFL, Switzerland [email protected] Julien Freudiger EPFL, Switzerland [email protected] Jean-Pierre Hubaux EPFL, Switzerland [email protected] Peter Marbach University of Toronto Canada [email protected] Federico Albiero VTT - Technical Research Centre of Finland Kaitov¨ayl¨a 1, P.O. Box 1100, 90571-FI Oulu, Finland [email protected] Dominic O’Brien Department of Engineering Science University of Oxford Parks Road, Oxford, OX1 3PJ United Kingdom [email protected] Friedrich K. Jondral Institut f¨ ur Nachrichtentechnik Universit¨at Karlsruhe (TH) 76128 Karlsruhe, Germany [email protected]

Volker Blaschke Institut f¨ ur Nachrichtentechnik Universit¨at Karlsruhe (TH) 76128 Karlsruhe, Germany [email protected] Tao Chen CREATE-NET Via Solteri 38, TN, Italy [email protected] Honggang Zhang CREATE-NET Via Solteri 38, TN, Italy [email protected] Xiaofei Zhou CREATE-NET Via Solteri 38, TN, Italy [email protected] Gian Mario Maggio CREATE-NET Via Solteri 38, TN, Italy gian-mario.maggio@create-net. org Imrich Chlamtac CREATE-NET Via Solteri 38, TN, Italy [email protected] Lee W. McKnight School of Information Studies Syracuse University [email protected] William Lehr Massachusetts Institute of Technology [email protected] James Howison School of Information Studies Syracuse University [email protected]

Part I

Introductory Chapter

1 Cooperative and Cognitive Networks: A Motivating Introduction Towards the Age of Enlightenment in Wireless Communication Networks All our knowledge begins with the senses, proceeds then to the understanding, and ends with reason. There is nothing higher than reason. Immanuel Kant

Marcos D. Katz1 and Frank H.P. Fitzek2 1 2

VTT, Finland, [email protected] Aalborg University, Denmark, [email protected]

Summary. In this introductory chapter, some key trends and emerging concepts for wireless communication networks are explored, highlighting chiefly those exploiting cooperative and cognitive principles. After discussing some key promising directions of development, a motivating overview of cooperative techniques in wireless networks in presented, followed by an overview of cognitive techniques and their use in wireless networks. The joint exploitation of these complementary principles are also considered in this chapter. Finally some visions on promising developments and possible evolutionary steps are included. The introductory discussions presented here serve as a an initial motivation to the rest of the book, where many of the concepts and techniques discussed here are explained in detail.

1.1 Introduction In the last two decades enormous efforts have been devoted to developing wireless communication technologies. Once affordable only to specific niche markets, these wireless communications are rapidly becoming everyone’s mainstream source of connectivity. Horizontal markets continue to grow in particular in developing regions of the globe, and it is expected that soon one third of the world population will use wireless devices for communication purposes. In many developed countries wireless voice-centric communications is replacing the well established wired counterpart. This trend is swiftly spreading into all regions of the world. Developments in the vertical markets are also significant, being propelled by the ever growing number of network access technologies as well as the incessant introduction of advanced terminals, mostly in developed countries. Behind schedule and with a currently 3 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 3–30. c 2007 Springer.

4

Marcos D. Katz and Frank H.P. Fitzek

smaller business size, data-centric communications appear to follow a similar pattern than voice, particularly driven by the industry push and growing user acceptance of wireless Internet. Even though in recent years the concept of the fourth generation (4G) wireless and mobile communication system has been intensively discussed by industry and academia, even today a common, clear and widespread understanding of its meaning is missing. In the most general sense, the term 4G is used to loosely describe advanced future highperformance wireless communication systems. Often enough, 4G is used as a synonymous of very high data throughput communications systems. We advocate here the former interpretation of 4G, but in an integrative manner, encompassing an eclectic array of different wireless networks which cover virtually every possible communication scenario. This leads to the concept of converging networks where heterogeneous networks harmonically coexist. Convergence is taking place also in other domains, noticeably in terminal and service. It is not the purpose of this chapter to present in detail views and visions on 4G - the interested reader is referred to [2] for that purpose - but to identify and discuss some emerging trends and concepts likely to make a profound impact on future wireless communications Figure 1.1 illustrates the current mosaic of wireless communication networks from the service coverage (range) standpoint. Two main network components are clearly distinguished, namely wide area networks on one hand, and short-range networks on the other hand. Curiously, range-wise, the development of wireless communication networks follows an ordered evolution from large to small networks, starting with very large distribution networks of up to hundred of kilometers wide down to sub-meter short-range networks. Several reasons can be attributed to the development of increasingly smaller wireless networks, including the pressure to move towards unused (and typically higher) frequency bands of the spectrum and the need to support higher data throughputs. In general these two component networks were developed independently of each other but aiming, by design, to coexist. As discussed in [2], such a coexistence may inevitably lead to competing situations between some networks. Different networks are most commonly seen as complementary, where each network is used in a given scenario or for a particular application. Heterogeneity and convergence of networks, terminals and services are perhaps the most distinctive characteristics of future mobile and wireless networks. They will certainly bring challenges for the technical development but also new opportunities to exploit.

1.2 Ten Tenets Shaping Future Wireless Communications In this section we identify and discuss some key emerging principles and trends likely to shape future wireless and mobile communication networks. We partic-

1 Cognition and Cooperation

5

Figure 1.1. Future wireless communications: An all–encompassing network of networks.

ularly highlight emerging ideas, concepts and developments where cooperative and cognitive techniques have leading roles or direct impact. 1. Exploiting Synergy in a World of Multiple Wireless Networks Future wireless communication networks will consist of a multiplicity of wireless networks of different capabilities, with the two main components being wireless wide area networks and short-range networks, for cellular mobile and local broadband access respectively. Such networks will be designed to exploit the complementarity between these two component networks. The coexistence of centralized architectures using licensed spectrum on one hand, and distributed (ad hoc) architectures using license exempt spectrum on the other hand brings unique opportunities as the advantages of both approaches can be combined. As measured in number of users or business size, cellular networks continue to be today the access network par excellence. Countless efforts have been put on research and development of ad hoc networks but the use of such networks remains far behind of their counterpart, cellular networks. In general network operators regard ad hoc networks as a threat to their core business. We argue that the relatively moderate business success of ad hoc networks can be explained by the fact that such networks were basically developed in isolation to other wireless networks, without a concrete mutual relationship. The obvious synergy between cellular and ad hoc networks was mostly ignored and thus left to a great extent unexploited. Given the increasing density of wireless devices, in most of the scenarios a given device could be regarded as a constituent node of a short-range ad hoc network (or cooperative cluster) including devices situated in the immediate neighborhood. The omnipresent

6

Marcos D. Katz and Frank H.P. Fitzek

closeness of peer devices is already changing the way that information will be delivered to and from a wireless device. Figure 1.2 illustrates the evolution of information delivery from the cellular access network to the mobile device and viceversa, starting with a) the conventional direct approach, b) using multihop techniques exploiting one or more relaying or repeating stations and c) through/from a short–range network . The composite architecture of the last approach, of increasing relevance in future wireless networks, will be discussed in detail in Chapter 2 under the designation cellular-controlled peerto-peer communications. The term short–range network or cooperative cluster, is used here to denote a variety of possible network configurations, with homogeneous or heterogeneous component nodes, like wireless grids, wireless sensor networks, ad hoc broadband networks, etc.

Figure 1.2. Evolutionary view of information delivery approaches through the cellular access network. a) conventional direct delivery, b) delivery through a repeater (or repeaters) and c) delivery through a short-range wireless network.

2. The Rising Role of Short-Range Communications Short-range wireless communications involve a very diverse array of air interface technologies, network architectures and standards. The most well known short-range wireless network technologies include wireless local area networks

1 Cognition and Cooperation

7

(WLAN), wireless personal area networks (WPAN), wireless body area networks (WBAN), wireless sensor networks (WSN) car–to–car communications (C2C), Radio Frequency Identification (RFID) and Near Field Communications (NFC). An overview of short-range communications from the standpoint of the typical range is presented in Figure 1.3. According to the Wireless World Research Forum (WWRF), by year 2017 seven trillion wireless devices will serve seven billion people. The overwhelming majority of these devices will be for short–range communications. This average figure of some 1000 devices per every person is even seen by some predictions as conservative. One of the evident consequences of the proliferation of wireless devices is that every wireless device will hardly be isolated of other other devices but, on the contrary, a given wireless device will always be surrounded by a considerable number of counterpart devices forming a potential cluster to interact with. As compared to long-range communications, short-range links require significantly lower energy per bits in order to establish a reliable link. They can achieve thus data troughputs of several orders of magnitude higher than the typical values for cellular networks. Usually short-range networks exploit distributed architectures, using unlicensed spectrum.

Figure 1.3. A classification of short-range communications according to the typical supported range.

3. The Pervasion of Cooperative Principles Cooperation is the basic principle of any communication system. In fact cooperation is deeply embedded at any wireless communication network, where tacitly all connected or interacting entities agree on using common signal formats, protocols and behavioral rules. Clearly, without this essential principle no communication would be possible. Such implicit form of collaboration, or

8

Marcos D. Katz and Frank H.P. Fitzek

passive or implicit cooperation, is the key underlying assumption when designing a communication network. However, more advanced forms of collaboration can be considered, where rich and dynamic interactions among entities are implemented by design. We refer in this case to active or explicit cooperation. The scope of active cooperation reaches in practice each and every OSI layer, and not only cooperative techniques are used within particular layers but also across them. But, why are cooperative techniques receiving increasing attention in the field of wireless communications? The answer is clear: cooperation has the potential to improve the most important link and network performance figures, including achievable data throughput, quality of service, network capacity and coverage. In addition, cooperation can in principle enhance the efficiency in the utilization of radio resources, a highly sought capability for spectrum-, power- and energy–limited systems. Furthermore, cooperation can be used in order to share and augment the capabilities of the interacting wireless devices, creating virtual devices with scalable architecture and enhanced features. An overview of the characteristics, scenarios and potentials of cooperative techniques applied in wireless networks is presented in Figure 1.4. The cooperative horizon covers three main cases, namely a) altruism, where cooperation focuses on supporting or benefiting a third party (e.g., a relaying station) and thus no pay-off is expected, b) cooperation, where the interactions are mainly driven by selfishness (e.g., cooperating with other entities in order to obtain a clear benefit, or pay-off driven cooperation), and c) non cooperation, characterized by autonomic operation. Cooperative concepts are rapidly emerging and even becoming the underlying principle in numerous engineering fields. We highlight first the emergence and impact of cooperation from a socio-technological perspective. The natural predisposition of people to cooperate aiming at common and self benefits is palpably being accentuated by the pervasive presence of communication networks. The realm of cooperation is reaching new frontiers due to the pervasive and far-reaching impact of communication networks. Cooperative principles are applied in Internet, a countless number of distributed ventures like eBay, Linux, OpenSource, Wikipedia, etc. Collaborative efforts are also present in file-sharing and distribution initiatives like BitTorrent and countless several other peer–to–peer file distribution protocols. Different forms of cooperative computing like grid and distributed computing are examples of joint interaction over a distributed network towards a common goal. We also mention the rapid emergence of the so called wireless communities, metropolitan- and rural-wide wireless networks connecting cooperatively computers and users over short-range broadband networks. A few hundred million people are already expected to be connected and share resources in such communities in some years from now. Interestingly, many of these initiatives can be approached as network-enabled user cooperation as well as user-enabled network cooperation. From a purely technological standpoint, research and development actively exploiting cooperative concepts are certainly booming in several areas of wireless communications, involving virtually all OSI protocol

1 Cognition and Cooperation

9

layers. In particular, cooperative principles in the physical (PHY) and medium access control (MAC) layers are extensively investigated today. Representative methods include, among others, conventional relaying (multi-hop) techniques, cooperative diversity,cooperative coding, distributed antennas, network coding. A detailed account of these techniques can be found in [2].

Figure 1.4. Cooperation in wireless networks: An overview of scope and benefits.

4. The Pervasion of Cognitive Principles Cognition is, according to the Encyclopedia Britannica, the process involved in knowing, or the act of knowing, which, in its completeness, includes awareness and judgement. Clearly, and as often pointed out, perception and reasoning are also closely related to the cognitive process. Cognition is not a new concept in wireless networks, in particular when approached from the awareness and knowledge viewpoint. In fact, in a wireless communication network information describing system state like channel condition, usage of available radio resources, etc., can be usually obtained by sensing the surrounding wireless environment from a received signal or be explicitly signalled to the receiving end. Techniques like channel estimation and sensing the presence of other users, can be seen as elementary forms of applying cognition in wireless networks. In general, acquisition of knowledge is followed by analysis and further reaction or adaptation, as selection of an appropriate modulation-coding combination

10

Marcos D. Katz and Frank H.P. Fitzek

in adaptive modulation and coding (AMC), computation and application of appropriate beamforming coefficients, etc. Link adaptation, scheduling and other techniques relying on link and network awareness are also examples of the use of cognitive principles in wireless networks. Cognitive principles are exploited in order to improve link and network performance as well as to attain a better utilization of radio resources. Notably, these are also the overall goals behind the use of cooperative principles. Like in the case of cooperation, techniques exploiting cognitive principles are starting to be explicitly used in the design of wireless communication systems. The interest on bringing cognition into the wireless world sparked in the late 90’s with the introduction of the basic concepts defining cognitive radio, by Mitola [7]. The scope and potential of cognitive radio was described in highly motivating paper by Haykin in 2005 [3]. The interest in cognitive radio grew explosively in the immediate aftermath of that paper, as we are witnessing it today. The main goal of cognitive radio is to make better use of the increasingly scarce radio spectrum. Spectral efficiency is of fundamental importance as not only the number of users of wireless systems increases steadily but also advanced broadband services become more popular. In recent years the research community is starting to look at cognition from a broader angle, considering wireless cognitive networks as networks taking advantage of cognitive principles to enhance efficiency in the use of radio resources as well as to improve basic performance figures of such networks. In that respect cognitive radio can be seen a particular instance of cognitive networks, where the goal is to increase spectral efficiency. Cognition is fundamental in heterogeneous networks as achieving end-to-end connectivity between different wireless devices operating in different networks requires that the ends are aware of the characteristics and capabilities of their counterparts. Cognition is also key at network level, whenever one targets an efficient exploitation of the shared radio resources. It is interesting to note the somewhat repeating patterns found in the evolution of wireless communications, in particular from the perspective of the exploitation of radio resources. Time, the most fundamental resource, was first exploited following rigid allocation strategies but the requirements for multiple access capabilities created the need of developing strategies exploiting more efficiently the time domain. Deterministic approaches, such as Time Division Multiple Access were then developed. The introduction of packetized transmission allowed a much better efficiency in the use of the time domain. Some systems were based on sensing the presence of transmissions at a given time to avoid collisions due to simultaneous transmissions, like in the Carrier Sense Multiple Access (CSMA). A similar development took place at the frequency domain, starting with fixed allocations, moving then towards deterministic approaches, such as Frequency Division Multiple Access (FDMA). Now we are moving towards the opportunistic use of the radio spectrum by exploiting the concept of cognitive radio.

1 Cognition and Cooperation

11

5. Spreading Out Source, Destination and Information Flow In general it is often assumed that in a communication system the information to be transmitted is originated in a point (or single) source and the destination is also punctual. An extension to this basic point–to–point communication paradigm is the point–to–multipoint model, as in for instance information distributing systems, typically broadcast and multicast delivery techniques. In interacting wireless systems such as cooperative clusters the information targeting the destination may also end up in, or could be deliberately directed to other, originally non-intended nodes, which, under certain conditions, may become actively interested in the content, with a clear benefit for all collaborating members. In such scenarios the notion of terminal becomes somewhat ambiguous, as the transmitted information is processed (forwarded and consumed) in several destinations. In cooperative scenarios the source can also be unfolded and as a result a multiplicity of nodes could become the source, each contributing with partial description of information, but all together providing the complete information. Multiple description coding (MDC)and scalable video coding (SVC) are representative examples where the source can be distributed. In cooperative scenarios the concept of routing also becomes distributed, as the same information propagates through a multiplicity of paths and connecting nodes, as done for instance with the recently introduced network coding techniques. 6. User–Centric Network Approach One of the underlying starting points when designing future wireless and mobile networks has been the user. Such user-centric approach is endorsed by the International Telecommunications Union (ITU) as well as the Wireless World Research Forum (WWRF) when referring to driving forces leading 4G development. But users will not only be source inspiring the key requirements for future networks but, more than ever, they will also play an active role within the wireless network. First, and as will be discussed below, the user will become an important source of content. Moreover, from the social cooperation standpoint, an important remark here is that the user will become an integral part of the collaborating chain and hence his decision will have a great impact on the operation and performance of the wireless network. Unlike communicational cooperation, where the collaborative interaction among entities is embedded by design in the system and transparent to the users (e.g., cooperative diversity, cooperative coding, distributed antenna, network coding, etc.), we envision users having a direct participation in the cooperative scene, as ultimately the user himself will decide whether to cooperate or not. The social aspect becomes fundamental here as several users come into play and both individual and collective decisions will have direct impact on the communicational cooperation. Today’s user predisposition to interact and cooperate through Internet is certainly phenomenal and such positive attitude

12

Marcos D. Katz and Frank H.P. Fitzek

is paving the way towards cooperative wireless communications among users in future networks. The sense of being connected has already emerged as a conscious awareness towards cooperation between users. As people are part of the social equation modelling cooperation, non technical aspects like trust, socio-geographical attitudes toward cooperation, incentives for cooperation, reputation and others will eventually have an effect on system performance for instance. The crossover of different fields make this research area not only challenging but also extremely interesting. 7. Balanced User Traffic Usually it is assumed that information transported by a wireless network predominantly moves towards the user, rather than from the user, like the typical downlink-uplink imbalance in cellular systems. This clear traffic mismatch is likely to continue in the future but with a less pronounced difference. One of the reasons for that is certainly the fact that mobile devices are more and more equipped with a mass storage and imaging devices, able to store and generate multimedia content. Thus, users will eventually become content providers, increasing the network-bound information flow. Furthermore, cooperating units will also tend to exhibit a more balanced traffic due to the rich exchange of information among the nodes of a cooperative cluster. 8. Learning from Nature Nature, with its evolutionary refining process spanning millions of years, has always been a rich source of inspiration for many fields of engineering. Cooperation is widely exploited by nature and examples abound in both microscopic and macroscopic worlds. Cooperative principles have been studied in biological, social and economic sciences, for instance. More recently, cooperative interactions and rules found in nature are starting to be considered as possible paradigms inspiring collaborative interactions in communication networks. The fact that in many ecological niches cooperative species out compete selfish ones motivates researchers to think that comparable results could be achieved if the collaborating entities are now parts of a communication network. Emulating ants behavior for Internet routing and the general concept of splitting are examples where nature fired up engineers’ minds. Promising cooperative and cognitive strategies to be applied within and across the layers of wireless networks can in principle be extracted from cooperative and cognitive patterns of behavior and rules found in nature. As decisions made by human beings can also have a direct impact on the way a network operates (e.g., a user joining or not a cooperative cluster, his decision on sharing certain resources of his mobile device, etc.), exploring social aspects of cooperative rules and cognitive process can shed some light on the process of developing cooperative and cognitive strategies for wireless networks. Many examples found in the animal kingdom give us some insights on how the knowledge of individual and group

1 Cognition and Cooperation

13

behavior of its members can be exploited to devise interacting mechanisms as well as cooperative rules and models for wireless networks. As a concrete and motivating example we mention the case of vampire bats, well known by their notorious cooperative strategies while eating and feeding other members of the colony. Indeed, very clear rules can be extracted from the behavior of these animals, including reciprocity, detection and punishing of cheaters and tolerance to pay-off delays. We can get more insights and extract additional rules from other gregarious species like monkeys. One interesting conduct is that monkeys may follow different behavior rules depending on their current counterpart in the tribe. For instance, they may accept unequal pay-offs or pay-off delays, depending if the interaction takes places with family members or an unknown member. Thus, additional rules include, to name some, groupmembership dependence of pay-off tolerance and knowledge (cognition) about group members. Rules such as the above discussed can be now exploited for instance in a cooperative cluster formed by interacting nodes of a network. Reciprocity (or the tit for tat rule) could be implemented by participating nodes in a fashion that nodes relay packets of information from nodes which also do the same job. A cheating node (e.g., refusing to relay but exploiting others) can be easily detected and disconnected of the cluster. For sure, more sophisticated rules can be extracted and applied. In 2 more discussions and examples are presented. 9. Air Interface Diversity There is no reason to believe that the current trend of equipping mobile devices with several air interfaces will continue and even strengthen in the future. The possible approaches to this form of air interface diversity are wireless devices exploiting multi-modality, where the air interfaces are integrated onboard using different chip solutions, and a more universal approach, flexible air interface, based on software defined radio (SDR). The availability of several air interfaces is essential for allowing a rich and dynamic cooperation among wireless devices. As of today, however, each air interface is used in an autonomic fashion, that is, in a particular scenario or for a given application. Concurrent use of different radio links is not a currently exploited feature. The implemented air interfaces represent both types of access networks, wide area cellular and short-range local access and as discussed previously, it is precisely interaction between these two networks one very promising concept for future networks. The main differences between the two air interface diversity approaches are in implementation complexity and cost, which favors the multi-modal approach, though the SDR concept is by definition much more flexible. Furthermore, from the standpoint of cooperation, multimodality transceiver naturally supports simultaneous parallel radio links while the SDR based transceivers better support a serial use of the air interfaces in a time division fashion.

14

Marcos D. Katz and Frank H.P. Fitzek

10. An Ideal Marriage: Cooperation and Cognition Cognitive and cooperative principles are complementary to each other and thus, it appears reasonable to exploit this natural synergy applying them jointly. Efficient cooperation relies on previously obtained knowledge, while effective acquisition knowledge and awareness can be achieved through cooperation. Future communication systems will consist of highly heterogeneous wireless ecosystems. In such a composite scenario, where different wireless devices and networks coexist, fruitful interaction cannot be devised without being aware of the surrounding wireless environment. Moreover, limitations and scarcity of resources will also put constrains in the design of future wireless networks. Power restrictions at access points (base stations), energy limitations in mobile devices and the scarcity of spectrum call for designs targeting high power, energy and spectrum efficiency. Complexity as well as power, energy and spectral efficiency are key resources that can be traded in different ways to achieve a desired level of performance and cooperation and cognition form a promising resource-trading framework for future wireless networks.

1.3 An Introduction to Cooperative Wireless Networks In this section we further introduce and discuss cooperative techniques for wireless networks. A vast and rapidly increasing body of literature already exists showing the potentials of cooperation basically in each and every OSI layer of a communication system. A basic taxonomy for cooperative techniques was recently introduced in [2], aiming at organizing and clarifying different approaches according to some particular characteristics and commonalities. The classification defines communicational cooperation, operational cooperation and social cooperation. These three concepts are discussed in this section. Figure 1.5 depicts this classification, highlighting the scope and goals as well. The shown correspondence between the types of cooperation and the OSI layers is mostly suggestive. Communicational Cooperation Communicational cooperation approaches are by far the most explored aspects of cooperation in wireless networks. Such methods encompass different techniques exploiting the joint collaborative efforts of multiple entities in the system aimed at bringing some advantages to the involved parts. In general communicational cooperation is inherently embedded in the wireless network and therefore invisible to the user. The mutually interacting entities include signals, algorithms, processing elements, building blocks and complete units. The goals of communicational cooperation are to enhance key performance figures at link and network level, to improve the utilization of basic radio resources and to create virtual augmented capabilities of the interacting entities

1 Cognition and Cooperation

15

Figure 1.5. Cooperation in wireless networks: classification, scope and goals.

of a cooperative cluster. A further classification of communicational cooperation includes implicit and explicit cooperation. The former refers to cases where interactions take place without any pre-established cooperative framework, such as protocols (e.g., all network nodes tacitly agree to use a common protocol like TCP, ALOHA, etc). The latter form of cooperation refers to interactions actively and purposely established through a given framework, that is, cooperative behavior is allowed and supported by design, allowing counterpart entities to actively interact directly with each other. The core of communicational cooperation mostly takes place in lower OSI layers, particularly in the physical (PHY) and MAC layers, though link and network layers may also be involved. We highlight here that in general these techniques mostly operate in an altruistic manner, where some the entities being active parts of the cooperative setup contribute with efforts or resources but do not gain anything in return. Such is the case of a typical relaying station. A great variety of techniques exploiting communicational cooperation exist. In the next paragraphs we briefly discuss the main principles involved in the most representative ones, including multi-hop techniques, cooperative diversity, cooperative antennas and network coding.

16

Marcos D. Katz and Frank H.P. Fitzek

Multi-Hop Techniques Of all cooperative approaches multi-hop techniques have perhaps received the largest share of attention, probably because of their inherent simplicity and rather direct application. The basic idea is to have one or more intermediate nodes whose main function is to repeat or retransmit in a convenient manner the received signal. In this way it is possible to extend coverage (range), resulting also in a more uniform provisioning of QoS within the service area. Multi-hop techniques, a basic concept of ad hoc networks, is also being considered for cellular networks, where, in addition to enlarging the coverage area, they will help guaranteeing high data throughput even at the edge of the cell. Multi-hop techniques usually exploit two basic approaches, namely amplify-and-forward (AF) [5] and decode-and forward (DF) [10]. AF is a simple method where, as its name implies, the signal is directly amplified and retransmitted to the next node in the chain. The processing station or device in the case of AF is usually denominated repeater. In the DF approach the signal is received and digitally regenerated before being transmitted forward, thus the noise is not amplified over the multi-hop chain as with the AF method. The processing node in this case is refereed to as the relay. Many other relaying (repeating) schemes exist such as estimate-and-forward, storeand-forward and hybrid combinations of them. In general, a two-hop solution (single repeater) is a good engineering compromise in cellular networks, while in distributed networks a generic multi-hop approach is usually considered. Multi-hop techniques are presented in detail in [2]. Note that particularly in cellular networks one of the two hops need not necessarily to be over a wireless link but it can be implemented on a wireline or optical fiber, as in the concepts of distributed base stations and radio over fiber. One important advantage of multi-hop techniques lies in the fact that the deployed repeating/relaying stations reduce the average link distance of the active terminals, resulting in a relaxed link budget and ultimately in extended duration of the battery. Cooperative Diversity Cooperative diversity refers to several techniques exploiting the presence of one or mode nodes between the source and destination. These distributed nodes cooperatively help the source to improve the overall capacity achieved between source and destination. The joint contribution of several nodes can be seen as a special form of spatial diversity and hence the end-to-end communication reliability can be also enhanced. These techniques are based on protocols allowing sequential reception and further retransmission in a time division fashion as simultaneous reception and transmission is not practically feasible (half-duplex constraint). Additional reading on cooperative diversity can be found in [4, 6] as well as in [2].

1 Cognition and Cooperation

17

Cooperative/Distributed Antennas Terminals in close proximity and forming a cooperative cluster can be used to form a virtual antenna array (VAA) by forming with single-antenna terminals a multiple input multiple output (MIMO) system. The VAA approach is also also known as distributed antenna or distributed MIMO. Each terminal contributes with a single antenna and by exploiting the short-range links these antennas can be interconnected making the cooperative cluster to appear either as a transmitting or receiving array. Such array, when communicating with a counterpart one in another cluster or in a base station form the basic MIMO structure. A comprehensive overview of cooperative/distributed antenna techniques is presented in [2]. The concept of space-time coding, that is the joint coding over temporal (e.g., repetition) and spatial (e.g., antennas) domains can be also applied in such a distributed antenna scenario, as showed in [11]. A different approach with distributed antennas is the concept of distributed beamforming where each antenna element of the terminals form a random array. Thus, assuming that nodes exchange their information and synchronization is achieved over the short-range links, it is possible to obtain array gain in both receiving and transmitting directions. In principle it is possible to achieve full array gain in a given desired direction as with conventional arrays (e.g., uniform linear arrays) but given the random distribution of nodes the side lobes cannot be controlled. Some insights of distributed beamforming are given in [2]. Network Coding Another form of cooperation which emerged rapidly in recent years is network coding. Network coding assumes also several repeating/relaying nodes between the source and destination. Source and destination may not necessarily be punctual but multiple sources and destinations can be considered. Network coding in general comprises the joint design of routing and coding in such multi-node scenario. A node will not just retransmit its received information but in principle will combine information received from many nodes to then forward that joint information to other nodes. From the packet standpoint, network coding exploits the fact that each wireless node broadcast its packets and a generic node thus receives and combine packets from many sources. Readers are referred to [2] for a in-depth discussions of network coding principles. Operational Cooperation An inherent characteristic of future wireless networks will be their heterogeneity, manifested by the presence of different access networks, a great variety of terminals with different capabilities and an large array of varied services. Clearly, providing connectivity in such heterogeneous environment is a far

18

Marcos D. Katz and Frank H.P. Fitzek

from trivial task. Operational cooperation can be defined as the interactional and negotiating procedures between entities required to establish and maintain communication between different networks. The main target is to ensure end-to-end seamless connectivity, where the main players could be terminals with different capabilities operating in different access networks. Research activities in this field are being carried out by academia and industry, we highlight here the work at the Wireless World Research Forum (WWRF) where network cooperation is mostly approached from the transport and network layers [9]. Also Ambient Networks project under the umbrella of EU’s Sixth Framework Programme addresses the problem of cooperation in heterogeneous networks, particularly where networks belong to different providers or exploit different access technologies [1, 8]. The ultimate goal of these projects is to ensure seamless operation regarding the type and associated networks of both source and destination. In addition to developing cooperative procedures, these initiatives also explore architectures required to support provision of end-to-end connectivity and quality of service. Social Cooperation One of the most interesting aspects of cooperation is perhaps its social perspective. In short, social cooperation can be defined as the dynamic process of establishing and maintaining a network of collaborative nodes, in our case, wireless devices. The process of node engagement is important as each node needs to decide on its participation in an (ad hoc) network, having each decision an individual and collective impact on performance. Unlike the communicational and operational approaches, where the cooperative mechanisms are embedded in the system, remaining therefore mostly invisible to the user, here the role of the user can be of fundamental importance as he will ultimately decide whether to cooperate or not. It is interesting to see that this process can be mapped into concrete communicational situations and therefore such a relationship could be exploited already during the system design phase. As an example, N interacting users may result in N possible diversity branches or they can be mapped into N interacting entities trading, sharing or combining their resources. The very first question to be answered in the realm of social cooperation is how to gather entities like wireless devices and form with them a cooperative cluster? As in general the user controls a mobile device, therefore it is primarily up to him to decide whether to join a cooperative cluster or not. There are many different reasons in support of joining a group, and for sure, also risks and doubts. A typical user will carefully ponder those before making his decision. Incentives need to be developed to stimulate users to support cooperative interactions. Clearly, since individuals are integral parts of the cooperative equation, non technical aspects like trust, socio-geographical attitudes toward cooperation, incentives for cooperation, reputation and others will eventually have an effect on system performance.

1 Cognition and Cooperation

19

Rich and dynamic user cooperation appears to be carried out more efficiently over short-range links, that is when devices are in close proximity. This is because, among others, the low energy (power) expenditure required to overcame short distances, the high achievable throughput and the typically unregulated spectrum used by these links. Already now, but even more accentuated in the future, a wireless device is always surrounded by counterpart units. In other words, we can assume that far and wide around us there are potential cooperative clusters. The nodes of such clusters can be, in addition to user-controlled wireless devices like mobile phones, a diverse array of wireless communications-enabled devices like home and office appliances, computers, repeaters, sensors and cars. The share of user-controlled nodes is likely to be significant or even highly predominant, depending on the scenario being considered. In principle, such ad hoc networking has been widely and deeply explored during the past, though the impact of individual and group decisions as well as the heterogeneity of components nodes have received less attention. This cooperative set-up, no doubt, is of fundamental importance as it has been predicted that short-range communications will be the dominant communication approach of the future, at least when measured by the number of wireless devices [12]. However, plain cooperative interactions over the short-range links, in isolation from the outside world, result in many technical and service limitations, which in practice have kept the penetration of such ad hoc networks relegated in comparison to that of cellular networks. Linking the short-range cluster to the cellular network and allowing thus cooperation between these network results in our view in a extremely powerful combination, exploiting the complementary characteristics of both distributed and centralized access architectures. The cellular network brings new possibilities to the cooperative cluster, being the service entry point for the cluster. Moreover, the cellular access is the natural link for the network operator to provide authorization, validation, authentication and security to certain transactions among users. In addition, the operator can promote cooperative services and even distribute incentives. Such a cellular-controlled short-range network approach is explored in detail in Chapter 2. Finally, we briefly mention that typical wireless scenarios exhibit favorable conditions for social cooperation. At home, potential cooperative nodes include wireless devices of other family members and wireless-enabled personal equipment (e.g., nodes of a personal network). The threshold for deciding upon cooperating with family members should be low or even non-existent as altruism is likely the dominant cooperative approach in such scenario. At the office, the devices of colleagues become the key nodes with which a cooperative cluster will be formed. In this environment cooperation may arise spontaneously or encouraged, or even required, by the employer. Finally, a public place, is perhaps the most demanding cooperative scenario, as the users behind the nodes are basically unknown to each other. In such a case cooperation can emerge also spontaneously but most importantly it can be encouraged by operators by different incentives.

20

Marcos D. Katz and Frank H.P. Fitzek

1.4 An Introduction to Cognitive Communication Systems Like cooperation, cognition is rapidly emerging as one of the underlying paradigms enabling future high-performance, high efficiency wireless networks. Exploiting cognition in increasingly complex wireless communication systems appears as one of the key evolutional steps towards the realization of cooperative networks. In this section the term cognition is purposefully positioned within the domain of a wireless network. In addition, the concept of Cognitive Communication Systems (CCS) is outlined. The relationships between CCS and Cooperative Wireless Networks are also sketched. Cognitive radio has been discussed considerably by the wireless communications research community lately as a new fundamental paradigm in the wireless world. However, cognition, interpreted as awareness, has always been present and exploited by wireless communications systems, as already discussed.

Figure 1.6. Cognitive cycle in wireless networks.

We first present an overview of the cognitive principles present in a typical wireless network. A typical cognitive cycle of a wireless network is depicted in Figure 1.6. This continuous cycle starts by sensing the surrounding wire-

1 Cognition and Cooperation

21

less ecosystem. We emphasize on the fact that an eclectic array of parameters and pertinent figures could be sensed, depending upon the type of scenario, wireless devices sharing the radio channel and relationship among the users. Sensing the status of shared resources (typically frequency) as well as radio channels and interference associated with wireless devices appear to be most important sensing targets. However, sensing representative figures such as configuration and capabilities of surrounding wireless devices, status of their particular resources, type of services/application being used, e-reputation of users owing the devices, etc., could certainly be highly significant when establishing cooperative clusters. Sensing can be carried out on centralized or distributed manners. Also cooperative and non-cooperative approaches could be used. We highlight here the broad meaning of the term sensing, focusing not only on measuring some physical magnitudes but also monitoring the status and behaviour of the wireless network and its entities. After the sensing process has been completed, the information is processed in order to figure out relevant features of prevailing wireless environment, aiming ultimately at understanding the wireless scene. The term understanding has also a broad connotation in this context, meaning, in one extreme, being aware of some basic network conditions, and, on the other extreme, having a deep understanding or knowledge of the overall network situation (acquired from the sensing information and a subsequent reasoning process). A sound understanding of the wireless network is a fundamental precondition for ensuring reliable outcomes of the decision making process. Decisions include the use of shared and device-particular resources, formation of cooperative clusters, change of communication parameters (e.g., link adaptation) and device configurations, use of a particular cooperative strategy or protocol, promotion of a service being used, etc. The final phase is adaptation, where active actions are taken, realising concretely the previous decisions. We define four cognitive scenarios, as depicted in Figure 1.7. The cognitive end-to-end link scenario, second box from the left, relies on direct signaling between counterpart protocol layers. For example at the link level the channel state is exploited for link adaptation (e.g., adaptive modulation and coding) and retransmission schemes. Similarly, information is exchanged at the application layer to build up knowledge about the underlying transportation medium and to adapt the coding rate accordingly. The generalization of this scenario leads to the cognitive network scenario, as shown in Figure 1.7 (second box from the right). Let us consider the case of multi-user communications. As long as the base station is aware of the channel state for any of the users in the cell and/or the user requirements in terms of delay, throughput or quality of service (QoS), it can change its scheduling policy and target specific design objectives, such as cell throughput or user fairness. Since the simultaneous achievement of conflicting objectives is not always possible, the design most frequently targets a trade-off among them. For any of these examples, crosslayer design is a methodology to build up knowledge within an entity (terminal

22

Marcos D. Katz and Frank H.P. Fitzek

or base station) across different protocol layers as given in Figure 1.7 (leftmost box) and is referred to as a cognitive terminal. The derived knowledge can be achieved by direct signaling between the protocol layers (breaking the ISO OSI model) or by deriving information of the packet flow.

Figure 1.7. Cognitive principles in wireless networks: present and future views.

Cognition is the first step to optimize a wireless communication system. However, cognitive capabilities per se are not effective unless the individual modules are flexible enough as to alter their behavior (i.e., adapt) responding to the acquired link/network knowledge. Therefore the enabling factors complementing cognition are flexibility and adaptability. In the domain of wireless communication, cross–layer design exploits cognition, while software defined radio, multi modality and novel air interfaces offer the required flexibility. Flexibility allows the possibility to implement more than one option while adaptability comes, by definition, with the additional task of selecting the adaptation policy. For the example of link adaptation, the terminals have to decide and agree on a common modulation order and coding rate, and should be able to encode/decode data in any of several possible ways. These two factors (decision and implementation) add complexity to parts of the device. The complexity can increase dramatically seen from a local point of view (e.g., protocol layer), but may pay off from a global point of view (e.g., the entire terminal or the end-to-end link). A general rule governing wireless system design is the linear relationship between the degrees of flexibility and

1 Cognition and Cooperation

23

complexity. One goal here is to break up this relationship, offering a larger degree of flexibility at a lower or at least the same cost in terms of complexity. The rightmost box of Figure 1.7 depicts a generalized cognitive communication system design approach, referred to as cognitive wireless system. The terminals are now allowed to be aware of each other state, and are able to communicate not only with the base station, but also directly with each other. The simplest form of cognition would facilitate the communication by limiting or avoiding interference. A more general case is that of a cooperative wireless network that exploits the link among terminals and improves the service to all of them. As the flexibility is not limited to the lower layers, the term Cognitive Communication System (CCS) is used instead of the well known term cognitive radio. A CSS is defined as the system being aware of its wireless environment by means of its own perception and/or by specific signaling protocols among its components. The components (such as the terminals, layers, etc.) are able to react by changing or adapting themselves according to the acquired knowledge while aiming at fulfilling certain pre-established goals. In that sense cognitive radio, link adaptation, scheduling and other techniques relying on awareness at local, link and network scales can then be considered as particular instances of CCS, as given in Figure 1.8. A CCS particularly optimized to a) enhance system spectral efficiency by sensing on a continuous basis the spectral utilization of surrounding wireless ecosystem and b) allocating opportunistically the unused bands is usually known as cognitive radio.

Figure 1.8. Classification of Cognitive Communication Systems.

24

Marcos D. Katz and Frank H.P. Fitzek

1.5 Towards Cooperative and Cognitive Wireless Communications In this section we explore the joint application of both cooperative and cognitive principles in wireless communication networks. To better illustrate the concepts we focus on cognitive wireless devices, though the discussion below can be straightforwardly extended to other network parts. Figure 1.9 depicts the main layers of generic cognitive wireless device, encompassing either a multi-modal or flexible air interface communication layer, a transport layer and a management layer. The last layer is a middle-ware needed to control and manage the cooperative communication network. The management layer gathers information from the cooperating cluster over the short-range communication links (e.g., how many devices would like to cooperate or already form the cluster?) as well as over the cellular ones (e.g., how does the network support cooperation?). The controlling unit takes care of managing the lower layers. The necessary flexibility can be achieved by the concepts of multimodality and flexible air interfaces, key for implementing a cognitive wireless device. We first briefly define these concepts.

Figure 1.9. Architecture of a cognitive wireless device.

Multi-Modality This approach uses different wireless technologies that are available on a single mobile device. Today, representative cellular communications (wide area) technologies are GPRS, EDGE and 3G, while for the short-range (local area, links among terminals) typical options include for instance IrDA, Bluetooth, or WLAN.

1 Cognition and Cooperation

25

Flexible Air Interfaces Different approaches have been explored to design flexible air interfaces, in particular space-time processing techniques have been extensively studied in the past decade due to their flexibility as well as potential to improve link and network performance. Here we approach flexible air interfaces from a different perspective, namely radio resource usage, in particular, but not limited to, spectrum. A flexible air interface allows for the opportunistic use of the spectrum. We distinguish two different approaches in the design of flexible air interfaces. The first method, referred to as closed-spectrum approach, focuses on a predefined (typically licensed) spectrum for the communication between the terminals and the base station as well as the communication among the terminals. On the other hand, the open-spectrum approach relies on the predefined spectrum only for the terminals to communicate with the base station, while the spectrum for short-range communication will be found by the terminals themselves by scanning for free spectrum. This is often referred to the cognitive radio approach. As this approach (after the scanning process) has to solve similar problems as the multi-modality approach and cognitive radio goes far beyond scanning for spectrum, we focus on the closed-spectrum approach to show that there is a need for cognitive radio even when using only the regulated spectrum. Cooperative wireless networks perform more efficiently as their flexibility to support a varying number of users grows. Moreover, cognitive capabilities are paramount to use this new degree of flexibility in an optimal manner. In general, cognition can be considered as the fundamental capability enabling cooperation in wireless networks. This condition, already compelling in homogeneous networks, becomes vital in heterogeneous networks, where a priori information on likely different cooperating partners needs to be known to all interacting entities. Inherently cross layer protocol concepts are applied at both the terminal and network sides. Cognition in combination with cooperation are principles with high potential to tackle some of the key problems identified for future heterogeneous wireless networks, including spectral and power efficiency and complexity issues, among others. In summary, Cognitive Communication Systems are the foundation of Cooperative Wireless Networks. The former acquire knowledge from the wireless ecosystem (locally and widely) and react in a rational fashion (locally and widely), and the latter allowing mutual interactions within and across network components aiming at fulfilling well established (global) objectives. Figure 1.10 illustrates the concept of cognitive wireless device

1.6 Discussions and Conclusion In this introductory chapter we explored several approaches to cooperation and cognition in wireless networks. The use of cooperative and cognitive principles in wireless networks was motivated, highlighting the key approaches to

26

Marcos D. Katz and Frank H.P. Fitzek

Figure 1.10. A cognitive wireless device.

cooperation, namely communicational, operational and social. The role of cognition, a highly complementary principle to cooperation, was also discussed here. Cognition is a key principle in wireless communications as networks and their elements become more and more heterogeneous. In order to ensure connectivity in such eclectic systems it is needed to acquire knowledge about the surrounding environment and participating network entities. Cognitive and cooperative principles are also essential in order to exploit efficiently the available resources. We highlight here that one of the the most important aspects of future communications will be the clever and efficient exploitation of resources, in many cases distributed along interacting entities. Resources can be defined from different standpoints. We consider mainly radio resources, built-in resources, user interface resources and social resources. • Radio resources - essentially time, frequency, space and power/energy - are fundamental when defining the communicational aspects of cooperation and cognition. The gradual introduction of multi-carrier and multi-antenna techniques as well as the increasingly higher supported sampling rates give to the system designer an unprecedent degree of granularity in the temporal, spectral and spatial domains. Such a fine grain availability of resources is a vital prerequisite supporting the emergence of cooperative frameworks exploiting complex interactions among entities (e.g., layers, networks, functionalities, algorithms, etc.)

1 Cognition and Cooperation

27

• Built-in resources encompass different hardware-related assets distributed over the network. These include devices for mass storage (e.g., memory devices), energy sources (e.g., batteries) and processing units (e.g., CPU, DSP). In the social context of communications, probably the most challenging set-up is the case when these resources are distributed among the wireless devices of a collaborative cluster. • User interface resources refer to typical capabilities integrated in the wireless devices, like speakers, microphones, keyboard, display, imaging devices (cameras) and other sensors. Both built-in and user interface resources can be shared, for instance, over the cooperative cluster aiming at creating augmented or enhanced virtual capabilities. These resources can be also transferred to a particular device of the cluster for a specific use. • Social resources are the individuals controlling their own wireless devices and deciding in which manner and to want extent to cooperate. Social resources should be considered in isolation - single users - as well as collectively. Both, a user and a group have particular requirements and exhibit specific patterns of behavior. Figure 1.11 illustrates the above defined wireless networks resources. In general it is convenient to consider these resource categories as being part of a resource pool distributed across the cooperating entities. The interacting entities share and trade these resources in order to use them efficiently, to enhance link and network performance and to share and augment capabilities of the wireless devices. Of course, users, at the end of the chain, are ultimately the most important resources, who individually and collectively will exploit and enjoy social networking. Cooperation and cognition are the underlying resource-trading principles defining a resource-trading framework. Resources are traded over several possible domains, basically short-range (local) networks, cellular (wide/metropolitan area) networks, and the OSI layers. These resource-trading domains are also illustrated in Figure 1.11. The assumption that wireless communication devices will be found in virtually every substantial entity that we deal with (i.e., man and machine) becomes more and more realistic as time goes by. The vision of an hyperconnected world is rapidly becoming a tangible reality, where wireless connectivity is more than ever taking a leading role. As discussed, it is not hard to imagine that regardless of our current location there will always be nodes around us with which we can form cooperative clusters. A cluster connected over short-range links makes sense because the typical energy and spectral efficiency associated with links over short distances significantly outperform those of wide-area networks. However, cellular networks have also a key complementary role as the counterpart network in the cooperative set-up. Centralized architectures will continue to be the dominant access approach, though the information to the target wireless device will not be delivered directly but through the short-range cooperative cluster that the target device belongs to. The nodes of the cluster could be autonomous like a repeater station or

28

Marcos D. Katz and Frank H.P. Fitzek

Figure 1.11. Resorces, trading domains and principles in wireless networks.

1 Cognition and Cooperation

29

a wireless communications-enabled appliance but more important, a human being could be controlling the device and ultimately deciding to collaborate or not. Users are integral parts of the cooperative arrangement and hence we emphasize on their importance. The challenge for network designers is to map social aspects into the technical domain. Expectations, behavior and requirements of users need to be well understood in order to devise appropriate cooperative techniques, while at the same time, users themselves need to understand their role and impact as well as achievable advantages. Except cases of altruistic behavior, in general users will cooperate when individual and group benefits are clearly realistic and well defined. Joining a cooperative cluster depends upon the attainable pay-off (e.g., type and amount) but also the delay involved in getting the pay-off shapes the user decision. Concurrently to cooperation, and complementing it, cognitive processes need to be implemented in the wireless network, bringing awareness and knowledge on the composition of the surrounding environment, capabilities of the interacting entities, requirements of each connected user as well as current usage of resources. In a somewhat similar manner to the cognitive process in human beings, cognitive wireless networks initiate the cognitive process by sensing the environment, continuing by understanding the prevailing conditions and ultimately making clever decisions. These decisions basically will try to fulfill the requirements of all connected users by properly allocating the available resources, taking into account relevant network information obtained through the cognitive process. We are at the doorstep of the Age of Enlightenment of wireless networks as already today we are conceiving future wireless communication systems with locally and widely deployed sensory systems analysis tools to interpret the observations and flexibility to adapt the network to the dynamics of the system. The upcoming evolutionary steps will likely direct the developments towards networks with advanced capacity of reasoning, exploiting sensor data, rich signalling across network components, and experience gained in previous interactions to understand consciously and autonomically the current network conditions. Cognitive radio is the very first attempt to bring explicit cognition to wireless networks, but this is just the starting point of developments, paving the way towards cognitive wireless networks. Ultimately, though far-off in the future conscious wireless networks, are expected to emerge.

References 1. B. Ahlgren, L. Eggert, B. Ohlman, and A. Schieder. Ambient networks: Bridging heterogeneous network domains. In International Symposium on Personal Indoor and Mobile Radio Communications (PIMRC 2005), 2005. 2. F.H.P. Fitzek and M. D. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-4020-4710-X. Springer, April 2006.

30

Marcos D. Katz and Frank H.P. Fitzek

3. S. Haykin. Cognitive Radio: Brain-Empowered Wireless Communications. In IEEE Journal on Selected Areas in Communications, Paris, France, 2005. 4. J. N. Laneman, D. Tse, and G.W. Wornell. Cooperative diversity in wireless networks: Efficient protocols and outage behavior. In IEEE Trans. Inf. Theory, 2004. 5. J. N. Laneman and G. W. Wornell. Energy-efficient antenna sharing and relaying for wireless networks. In IEEE WCNC, Chicago, IL, 2000. 6. J. N. Laneman and G.W. Wornell. Distributed space-time-coded protocols for exploiting cooperative diversity in wireless networks. In IEEE Trans. Inf. Theory, 2003. 7. J. Mitola. Cognitive Radio. In PhD thesis, Royal Institute of Technology (KTH), Stockholm, Sweden, 2000. 8. N. Niebert, A. Schieder, H. Abramowicz, G. Malmgren, J. Sachs, U. Horn, C. Prehofer, and H. Karl. Ambient networks: An architecture for communication networks beyond 3G. In IEEE Communications Magazine, 2004. 9. C. Politis, T. Oda, S. Dixit, A. Schieder, K. Y. Lach, M. Smirnov, S. Uskela, and R. Tafazolli. Cooperative networks for the future wireless world. In IEEE Communications Magazine, 2004. 10. A. Sendonaris and E. Erkip B. Aazhang. Increasing uplink capacity via user cooperation diversity. In IEEE ISIT, Cambridge, MA, 1998. 11. A. Stefanov and E. Erkip. Cooperative space-time coding for wireless networks. In Proc. IEEE ITW, 2003. 12. WWRF. Wireless World Research Forum. In WWRF, 2007.

2 Cellular Controlled Peer to Peer Communications: Overview and Potentials Coming together is a beginning keeping together is progress working together is success. Henry Ford

Frank H.P. Fitzek1 and Marcos Katz2 1 2

Aalborg University [email protected] VTT [email protected]

Summary. In this chapter we present a dynamic approach to bridge cellular and peer-to-peer network architectures, referred to as Cellular Controlled Peer-to-Peer (CCP2P) communication. This approach goes beyond the concepts used in composite networks, which focus mainly on coverage extension and data relaying. In CCP2P networks, besides being connected to an outside world using cellular links, a group of mobile devices in close proximity form a cooperative cluster contributing their onboard capabilities and resources to exploit them a more efficient way. Using peerto-peer technology in combination with cellular networks, CCP2P has the potential to overcome many important limitations of current cellular networks. The main driver behind such a concept is the short-range communication among mobile devices.

2.1 Challenges for Future Wireless Networks Currently the commercial cellular communication scene is dominated by the transition from the second (2G) to the third generation (3G). Some of the key problems of those communication systems include the limited achievable data rates over the air interface and the lack of new and appealing services motivating the customers to shift their 2G mobile devices to 3G ones. While designing the architecture of future wireless networks, also labeled as the fourth generation (4G), another challenge becomes evident: the ever increasing power consumption of the mobile device. While Moore’s law dangles even larger processing power in the future, the battery capacity does not keep pace with the increasing demand for energy to handle the upcoming tasks. In [1] it is claimed that the battery capacity has only increased by 80% within the last ten years, while the processor performance doubles every 18 month following Moore’s law. By cramming mobile devices with a large number of placebo functionalities and capabilities like advanced imaging features (camera, 31 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 31–59. c 2007 Springer.

32

Frank H.P. Fitzek and Marcos Katz

high-definition display, etc.) as well as versatile and high-performance wireless connectivity including short-range communication (Bluetooth, WLAN, etc.), and higher data rates over the cellular interface to support new services, manufacturers face a serious problem as the mobile devices are using more and more energy. Figure 2.1 shows the energy consumption for different parts of the mobile device of the Nokia 6630 mobile phone (Figure taken from [6]).

Figure 2.1. Power consumption for video conferencing in the Nokia 6630.

That energy consumption of mobile devices is already a problem in our daily life can be derived by the fact of an increasing number of public cell phone chargers around the world as given in Figure 2.2. Such as service is highly appreciated if the mobile device is out of battery, but on the other side the mobile user wastes time to charge its device plus she/he has to pay money for that services (3 US dollar for 30 minutes or 6 US dollars for 60 minutes for the public cell phone charger given in Figure 2.2; note, for the same amout of 3 US dollar a user can call 30 minutes on a public phone booth next to the public cell phone charger from the US to any country in Europe). So public cell phone charger can be very helpful, but if future devices will even consume more energy, we will spend more time and money on these machines. The increased need for energy has two major impacts: first, the absolute energy consumption could make active cooling of the mobile device necessary, and secondly, the operational time of that device decreases, as the develop-

2 Cellular Controlled Peer to Peer Communications

33

Figure 2.2. Cell phone charger by smartcarte (tm).

ments in battery capacity and efficiency cannot cope with the steep energy consumption increase. The operational time, including the time where the mobile device is actively used and the standby time, of mobile devices has been identified by Taylor Nelson Sofres (TNS) [11] as the number one criteria of the majority of the customers purchasing a mobile device. Therefore, one of the biggest impediments to future wireless communication systems is the need to limit the energy consumption of the battery–driven devices so as to prolong the operational times and to avoid active cooling. Thus, solutions have to be engineered aiming at decoupling the problem of increasingly more complex mobile devices on one hand, and the need for supporting advanced capabilities like high data rates as the enablers for new services on the other side. Before we introduce a potential solution to this problem, let us review the current communication architecture of omnipresent cellular communication systems. They are characterized by the communication between a base station and mobile devices as given in Figure 2.3. The base station is part of the network and the only access point for mobile devices to the network services. The mobile devices do not communicate directly with each other as the main services requested by them are accessed from the core network through the cellular system. Therefore, in state–of–the–art cellular networks, the mobile

34

Frank H.P. Fitzek and Marcos Katz

Figure 2.3. Conventional Cellular Communication Architecture.

device is also referred to as terminal as all services terminate in the device. This paradigm encourages a monolithic design for mobile devices. Hardware and software solutions are more or less static, they do not change in response to the current situation or user request. In the past this was acceptable as the dimension of different services was quite small with the focus mainly on voice centric services. Now, with an increasing number of different services, each part of the mobile device needs to be designed for the worst–case situation, or, in other words, to serve a huge variety of services. In Figure 2.3 those parts are, among others, the cellular air interface, the short–range communication link, the battery, the storage unit, and the processing unit. To overcome the aforementioned drawbacks, the very first idea is to enable cooperation among several mobile devices with their different capabilities forming wireless grids. The wireless grid uses the short range communication technology to communicate among the devices. Furthermore, as given in Figure 2.4, the cellular link is used to connect to the cellular world. This is a clear departure of today’s use where a given air interface and corresponding architecture is used in a given scenario or for a particular service. In such conventional approach dynamic interaction between and across air interfaces are not possible. Clearly cooperation will enable resource sharing or augmentation among the collaborating mobile devices, with the cellular network being basically the service entry point and the short–range links the gluing element allowing such a wireless grid to have a wirelessly scalable architecture. Such a wireless grid opens up a rich array of possibilities for distributed processing, sharing and scaling of functionalities, better utilization of resources, enhancement of performance, etc. As the communication link and even services do not necessarily terminate in a given mobile device but infor-

2 Cellular Controlled Peer to Peer Communications

35

Figure 2.4. Cellular Controlled Peer–to–Peer Network.

mation is shared, distributed and processed by several wireless devices, the term terminal may appear in such situation as misleading and its use should be avoided.

Figure 2.5. Different approaches to cooperation in wireless communication.

How mobile devices might cooperate with each other depends heavily on the ownership of the devices. Figure 2.5 shows three different users with their mobile devices. Each user may have one or more devices. It is quite easy

36

Frank H.P. Fitzek and Marcos Katz

to understand that cooperation might pay off in case the user’s own devices form a wireless grid or Personal Wireless Grid. This form of cooperation is dominated by altruistic behavior. This grid might be expanded with users known to each other, like family members, friends or colleagues, forming a Private/Professional Wireless Grid. But we will show throughout this chapter that even a Public Wireless Grid, where users do not know each other, can lead to an overall benefit for all participating users. Such a form of cooperation will be realized if and only if all participating entities are gaining at the same time. In [5] we have shown that egoistic behavior is the main driving force to build up cooperation and it is based on non altruistic behavior. Quoting Kurt Edwin: “Real egostic behavior is to cooperate!”

2.2 Premises for Cooperation After introducing the cooperative architecture, this section deals with the question of which capabilities and functionalities of the mobile devices can exploit cooperation. As given in Figure 2.6 a mobile device has many different capabilities and functionalities that can be used in a cooperative way. We group those of a mobile (smart) phone into three classes, namely user interfaces, the built-in resources and the communication interfaces. The user interfaces comprise the speaker, microphone, camera, display, built-in sensors, and keyboard capabilities. The built-in mobile device resources are the battery, the central processing unit and the data storage. We highlight particularly the communication interfaces, which typically include cellular and short–range capabilities. Instead of having one mobile device hosting all functionalities with the best possible quality, the concept presented here considers cleverly using the capabilities of several, in principle different, mobile devices. Indeed, in a heterogeneous mobile device scenario each device could specialize in one service such as a simple voice-centric mobile phone or a communication device with a music profile featuring an mp3 player. In general terms, such a concept is not totally new. Bill Joy, the visionary driving force behind Jini Technology and co–founder of Sun Microsystems, argued for cooperative software in 1999: “Jini technology is a dislocation in the context of cooperation of subspecies of devices. Once you have lots of different kinds of devices combining in different ways, you can not do monolithic software anymore. Each of these devices has a certain set of functions, and if we have to assume when we build them what they are going to be used for, it is not very flexible.” – Bill Joy The main idea behind this is to identify core functionalities of mobile devices and re–assemble them in a new way each time a certain service has to be provided; without holding the full set of functionalities in one mobile device. Figure 2.7 illustrates this concept with an example of two cooperative wireless devices. As given in Figure 2.5 one user may have two different mobile devices,

2 Cellular Controlled Peer to Peer Communications

37

Figure 2.6. A mobile device broken up into capabilities grouped into user interface, communication interface and the built-in resources.

each with a subset of different capabilities (e.g., cellular communication facilities for the mobile phone) and a subset of the same functionality but with a different realization level (e.g., the display of the tablet PC has a bigger size and higher resolution). To support a wirelessly scalable architecture the devices need to locally communicate with each other, which is done by the short–range communication links. In the future the role of the short–range links is expected to become far more important than it is today. Numerous cooperative applications can be devised, like extending the capabilities’ or functionalities of a given device with those of remote devices, for instance for storage, imaging (camera), and audio (microphone) purposes. This kind of cooperation is mainly envisioned for mobile devices owned by the same customer or by people with some relation to him/her, and therefore referred to as “altruistic cooperation” as we know it from bees or ants. Such cooperation can be applied to personal or private/professional wireless grids. In Figure 2.8 a comparison between monolithic (i.e., autonomous) and cooperative communication systems is presented from the requested service and related costs viewpoint. For the monolithic case the cellular communication link has to provide a given data rate to support the requested service (height of the cellular box). This may result in a complex architecture design. In the

38

Frank H.P. Fitzek and Marcos Katz

Figure 2.7. Example of a mobile phone and a tablet PC as two cooperative devices of the same user.

cooperative case several wireless devices collaborate by each receiving a lowerrate component stream of the whole signal and exchanging these streams over the short–range links. The same data rate is obtained, but through a less complex cellular air interface augmented by the short–range links (same height of cellular and short range box). As the short–range link supports higher data rates and exhibits a better energy efficiency per bit, the overall costs are less than for the monolithic case. On the other hand, the cooperative case requires cooperative peer devices. By combining the capabilities of different composable mobile devices a cooperative cluster, sometimes referred to wireless grid, is formed. As given in Figure 2.9 we see two different ways of combining the capability sets. On the left side three devices combine their capabilities having a larger set of capabilities, where as on the right side the three devices have the same capabilities but by cooperation they augment the performance of the capabilities. We define the first example as capability extension and the second one as capability augmentation. An example for the capability extension is given in Figure 2.7 by two devices with different capability sets, but even if the mobile devices would have exactly the same capabilities, using them in a different context, we still refer to capability extension. An example for the later one is the idea of

2 Cellular Controlled Peer to Peer Communications

39

Figure 2.8. Comparison of monolithic and cooperative communication systems with respect to requested services and the related costs.

relaying or multi-hopping. In this example the capability sets of the mobile devices are exactly the same, but the role of being source/sink or relay is different. From the cooperation’s point of view the pay off depends on the role (relay or source/sink). The relay tends to invest into the cooperation, while the source/sink is benefiting from it. An example for the capability augmentation is the cooperative communication architecture presented in Figure 2.4. Here the mobile devices bring in the same capabilities once more. But from the cooperation’s point of view both devices benefit equally cooperating with each other. This characteristic of cooperation is important to convince each customer to join the cooperation. Remember, we need to get them onboard by their egoistic way of thinking.

2.3 Combining the Cellular and the P2P World In this section we first propose a cooperative architecture to then consider possible realization forms. Some cooperation rules, in some cases inspired from nature, are also derived and discussed.

40

Frank H.P. Fitzek and Marcos Katz

Figure 2.9. Classification of Grids: Grid Extension (left) and Grid Augmentation (right).

2.3.1 Cooperative Architecture The previously introduced private and personal/professional wireless grids are the most obvious cooperative clusters that we can think of. However, public wireless grids constitute an environment that is more generic, challenging, and rich in business potential, as cooperation is envisioned also for mobile devices of users that might not know each other. These grids are in principle much more interesting as the probability of finding such a situation is much larger than for the other two grid forms. As each user is basically following egoistic interests and customers may not know about the neighboring mobile device capabilities, this kind of heterogeneous cooperation appears to be more challenging to establish. The second difference is that capabilities are not simply accumulated but negotiated beforehand among the devices which then dynamically adapt/change their capabilities according to the upcoming needs. One example of such adaptation is the air interface that needs to support cellular as well as short–range communication. As the pure accumulation is linearly summing up the benefits of different entities, the adaptation will lead to even further gains. The proposed wirelessly scalable architecture is an extension of a cellular network, where the mobile devices communicate with the base station coexisting with other mobile devices. This architecture is shown in Figure 2.4.

2 Cellular Controlled Peer to Peer Communications

41

Now mobile devices are able to create cooperative clusters with neighboring devices in their proximity. While the distance between the base station and the mobile device can typically be of up to several kilometers, the short-range link is usually limited to several tens of meters at the most. Each mobile device is then contributing to the cooperative cluster with its cellular link, energy of its battery, and some of its processing power. The grouped members can then be used as virtual entities in a cooperative manner aiming at overcoming the above described problems. This hybrid network combining a centralized and distributed architecture is referred to as composite network. The cellular link has a key role with the proposed architecture as it is needed as the service entry point and to perform administrative tasks such as authentication, billing, management, etc. Therefore the architecture is referred to as cellular controlled peer–to–peer network. The combination of cellular and ad hoc networks has been proposed before. However, the considered solutions approach such a system from a rather static standpoint, targeting usually range extension and data throughput enhancements by using multi-hop techniques between the two networks. The proposed architecture and associated cooperative framework dynamically exploits cooperation in two domains, namely a) inter-network cooperation, encompassing the collaborative interaction between the centralized and distributed networks, and b) intra-network cooperation, exploiting the interaction taking place within the distributed short-range network (i.e., cluster). This collaborative framework constitutes the key underlying principle associated with the composite architecture. The cooperation is described as a non altruistic form of collaboration in which each mobile device joins the cooperative group as long as it gains instantaneously. This cooperative behavior is based on egoistic reasoning. The fundamental idea behind the considered network and associated collaborative framework is exploiting the resulting synergy of having a component network with highly complementary characteristics, combining features such as licensed and license-extempt spectrum usage as well as high power / wide area / low data rate together with low power / short-range and high data rate. Moreover, the considered strategy brings the two main component networks of current and future wireless communication systems in a closer and amicable relationship, rather than regarding them just a coexistent networks or even competing ones in the worst case. This brings, in addition to the technical advantages, a universal convergence platform as well as a politically attractive solution embracing the two prevailing models of wireless communications. Note that in principle any type and air interface technology for distributed short-range networks can be employed. 2.3.2 Realization of the Cellular and the Short Range Link There are two main directions to implement the envisioned architecture, namely through a multi modality or through a common air interface approach. The multi modality approach can be realized already with 2G or 3G mobile

42

Frank H.P. Fitzek and Marcos Katz

devices which are equipped with cellular and short-range communication capabilities. One example employing today’s technology could be a mobile device with dual GPRS and Bluetooth air interfaces. A first implementation of this scenario has been carried out at Aalborg University on the Symbian OS platform using commercially available mobile devices (Nokia N70). Two cooperating phones agree on splitting a file to download and start to receive it over the GPRS link. Simultaneously the received data is exchanged over the short-range link. Measurements have shown a download time reduction of 50% and energy consumption reduction of 44% for only two cooperating mobile devices. Figure 2.10 depicts a screenshot of one of the cooperating mobile devices where the contributions of the cellular, short-range and virtual (cooperatively combined) links are displayed while cooperation takes place. Note that additional gain can be achieved if more devices collaboratively interact.

Figure 2.10. Mobile device screenshot (Nokia N70) for the two cooperating entities case. The data rate over the cellular link (1st bar on the left), the short-range link (2nd bar/in, 3rd bar/out) and the virtual link (4th bar) is given.

The second direction focuses on a new air interface design. Indeed, while the multi modality concept is based on two orthogonal spectra, the common air interface is focusing on flexibly splitting the spectrum for cellular and short-range communications. Such a future solution is discussed in Chapter 22. 2.3.3 The Importance of the Short-Range Communication Link The short-range communication capability is a must for functional cooperative networking in a CCP2P. Furthermore, the short-range technology dependent system parameters determine how large the benefit of CCP2P can be in practice. These parameters include the data rate of the short-range technology,

2 Cellular Controlled Peer to Peer Communications

43

the power consumption (sending, receiving, idle), and medium access control (MAC) operations. While the data rate and the energy values have a clear impact on the energy per bit ratio (EpBR), the impact of the MAC may not necessarily be evident. Therefore, as an example we refer to wireless local area network (WLAN) and Bluetooth as representative candidates for distributed and centralized short-range communication approaches, respectively. Certainly, wide-area cellular networks could also be considered in the example. While WLAN supports the mobile device providing the cellular information received over the WLAN in a multicast fashion, Bluetooth does not. In a Bluetooth-based communication network, the master device is able to control the communication of up to seven active slaves and some more inactive slaves. No direct communication between slaves is possible and therefore the transparent energy per bit ratio (TEpBR), defined as the overall energy needed to convey information from a particular slave to another one (via the base station), is larger than the EpBR. To illustrate this concept, we consider the WLAN scenario, with J cooperative devices. In such as scenario J multicast packets are exchanged among the mobile devices. Each mobile device sends one packet to the cooperating device and expects to receive J − 1 packets as given in Table 2.1. The calculation becomes slightly more complex when the number of packets sent over a Bluetooth enabled cooperative cluster is estimated: First the master can multicast its own packet to the J − 1 mobile slave devices. The J −1 mobile slave devices cannot multicast their packets directly to the neighboring slaves and need to relay them through the master. This ends up in an overall number of packets sent equal to 2J − 1. It is important to note, as given in Table 2.1, that the master is more active than the individual slaves. While the power consumption of the slave is equal to a cooperating mobile device in the WLAN scenario, the master takes all the burden as he is sending J times more packets than the slaves. Because of this a round robin of the master role would be fair. In this case the mean value of sent packets is 2 − 1/J. For large values of J, we can say that we need double the number of packets to exchange, which also doubles the TEpBR values in contrast to the EpBR. As one conclusion on the designing principles for future short-range wireless networks, we can highlight the importance of the three key factors, data rate, power values, and the choice of MAC. Table 2.1. Number of packet exchanges for the WLAN and the Bluetooth case (master and slave separated). Sending Receiving WLAN 1 J −1 BT Master Slave Master Slave individual J 1 J −1 J −1 mean 2 − 1/J J −1

44

Frank H.P. Fitzek and Marcos Katz

A lesson learned from last section is that the MAC should support direct communication between cooperating entities in a point-to-multipoint fashion. The reason for this is to avoid relaying through a central entity as in the Bluetooth communication case. Furthermore, to minimize the idle time as well as to reduce collision times, the channel should be slotted in time or frequency. This is not viable in IEEE based WLAN systems. For the power saving it is crucial to be able to switch on and off the RF/baseband chain in time domain or to dynamically enable/disable frequency bands reducing the complexity, a feature that can be found in DVB-H technology. 2.3.4 Somebody out There? Cooperation among mobile devices depends highly on the number of devices that open their short range communication. There are certainly places where such cooperation can be established more easily than in others. At Aalborg University, students have carried out measurements of Bluetooth connectivity among mobile devices in buses, in the stadium and at Aalborg International Airport. One mobile device has been placed in the waiting room of one gate scanning all open Bluetooth connections. Such a measurement can be seen as a first estimation as people tend to switch off the Bluetooth device to save energy and to not get any virus on the device. The number should be increasing with upcoming social networks using Bluetooth technology and the proposed cooperative networking. Instead of the battery and security issue, the social and cooperative networks bring a clear benefit to have switched on the Bluetooth technology. In Figure 2.11 the number of short range connection is shown for Aalborg International Airport. The observing mobile device is scanning every minute the neighborhood. It filters out mobile phone devices only and write the MAC address of the found entities in a file. Postprocessing log data the connection to individual mobile devices can be displayed (see Figure 2.11 top) and the overall number of connection per scan is known (see Figure 2.11 bottom). In addition to that even the number of new devices from one scan to the previous one is illustrated. The overall number is interesting to understand if cooperation can be established, the number of new device from one scan to the next will tell us something about the amount of signalling to keep the cooperative cluster alive. 2.3.5 Nature Inspired Cooperation In any kind of peer-to-peer network the communication ultimately depends on the willingness of the peers to cooperate. Such cooperation can only be established and maintained if fairness is guaranteed among these peers. Now fairness in the context of the relationship can have different meanings to the participants. To understand the ground rules for cooperation, a short overview on cooperation in nature is given. Next, as two representative cases of cooperation in nature, we refer to vampire bats and monkeys. Inspired by these

2 Cellular Controlled Peer to Peer Communications

45

Figure 2.11. Number of short range connections at Aalborg International Airport in the waiting room of one gate.

cooperative rules analogous principles will be developed and applied to wireless communication systems. Vampire bats live together in groups, where the group size is around 150 animals. Vampire bats leave the dwelling at night time on their own to gather food. Food in this case means typically mammals with lacerations giving the vampire bat the possibility to sip blood from the existing open wounds. Each vampire bat needs blood feeding every approximately sixty hours to not suffer starvation. However, not all vampire bats will be successful in gathering food. Fortunately the successful vampire bats carry more blood to the dwelling than what they need for feeding themselves and thus, they are able to share their extra food with others. Interesting to note, the sharing procedure is not based on altruism (which could be based on family membership), but on non altruistic cooperation. Vampire bats have a remarkable good memory, remembering which individuals they have already exchanged blood with previously. Thus the feeding process involves a reciprocal behavior that prevents other vampire bats from staying in the dwelling and exploiting those successful ones returning with food in excess. We identify here the first rule of cooperation as reciprocal behavior. The second rule of cooperation is detection of cheaters. In fact, vampire bats are able to detect cheating individuals. In case a vampire bat of their cooperative group is unwilling to share some food, the hungry animal will check out the stomach of the non-cooperative bat to verify if there is some blood at all. This is because the denying action could be based on the fact that the bat does not have any blood to share or it just does not want to share with the requesting bat. If the bat was cheating (having blood

46

Frank H.P. Fitzek and Marcos Katz

but not sharing it), the requesting bat will remove the requested bat from its cooperative group and ignore it afterwards. The next rule derived from the vampire bats is an inherent one, meaning that the pay off for cooperation should be received within a pay off cycle or within a reasonable delay. In the case of vampire bats, the pay off cycle is the lifetime of the animal. Frans de Waal and his group have worked with monkeys to investigate cooperation and fairness among these mammals [2, 3]. Among the very interesting findings, one result refers to the tolerance for unequal pay-offs among monkeys. The experiments were made with monkeys trained to exchange a token for cucumber. Cucumber is a respected exchange value for the token by the monkeys. In an experimental setup, two monkeys were put together in a cage separated by a grid. The two monkeys were either from the same social group or from different ones. Independent from the settings, all monkeys exchanged the token happily for a cucumber. However, if one of the monkeys was given a grape instead of the cucumber, while the other monkey still exchanged the token for a cucumber, the results changed. The grape has a much higher value for the monkeys than a cucumber. Now the behavior of the monkey depends on the group membership. If the monkeys are from different groups, the exchange of the token for a cucumber is refused. The monkey understood that the token may have a much higher price. Even though the token by itself has no value at all for the monkey and is even lower than the cucumber, the monkey will not trade. From his point of view the exchange is not fair. Even though the behavior seems to be irrational, its fits with the findings of Kahneman and Smith (see e.g., [8, 9]) stating that all markets are dominated by behavioral economics instead of rational maximizing strategies. The results changed if two monkeys from the same group were tested. In this case the token was exchanged for a cucumber even though the partner received a grape for the same token. Thus, as suggested by de Waal, the tolerance of unequal pay offs depends on the group membership. This leads to the fourth rule of cooperation. The tolerance to the pay off delay is inversely proportional to the degree of closeness of the involved group members. In other words, the better I know a partner entity, the longer I may accept waiting to get revenue back. Clearly, if I do not know a partner entity I should foresee an instantaneous gain in order to encourage cooperation with that partner. These findings are not restricted to monkeys but can also be found in human interactions.

2.4 Cooperative Services This section gives a first overview of possible cooperative services. It is claimed that cooperative services can support unicast as well as multicast and they are available on uplink as well as on downlink.

2 Cellular Controlled Peer to Peer Communications

47

2.4.1 Multicast and Broadcast Services Once the cooperative cluster is established different services can be provided. Obviously multicast or broadcast services are well suited for cooperative clusters. For such services the cooperative cluster can accumulate their cellular bandwidth to open up a virtual “big pipe”. Each mobile device is receiving only a part of the content which it makes available over the short–range communication link to all cooperative partners. Also on the short–range communication link every collaborating mobile device would like to receive the disjoint information, thus the missing parts. Such a scenario works perfectly for file downloading or video reception. As an example of video services, the benefit of cooperation for DVB-H services is investigated in Chapter 24. The main reasons why DVB-H was introduced even though DVB-T was already available are the video format and the battery consumption. In DVB-H the air interface is only active a fraction of the receiving time to receive high data rate bursts. Battery savings are available as the air interface is switched off from time to time. The main idea in Chapter 24 is to virtually increase the times the RF/BB is switched off. This is done as one mobile device of the group receives the broadcast data and forwards it to its local neighborhood, which has deactivated the cellular air interface and listens only on the short–range link. As the short– range link is much more energy efficient than the cellular link, energy can be saved for all remote devices. The forwarding device is investing slightly more energy as it has to activate two interfaces. Therefore for fairness reasons the role of the forwarder is changed within the group from time to time. In this scenario each cooperating entity is more than welcome as it helps to decrease the power consumption while viewing the video service. Here the benefit for the customer is clearly visible. An extension to this scenario is the usage of multiple description coded (MDC) video. Here the network provider and the service provider agree to transmit a broadcast video stream in small sub streams, where each sub stream is decodable. The more sub streams available at the mobile device the better the video quality becomes. Thus, to improve the video quality the customer needs a high data rate connection. In general, this would imply using a high class mobile device. This device will not only cost much more than a basic device but also it will also consume considerably more energy as the realization of a high data rate over the cellular link is quite complex. Another option would be to exploit cooperation. Here, even simple devices will receive one or a few of the sub streams and they will exchange them cooperatively within the group. The customers will gain in terms of power consumption reduction and also in a lower service price, if supported by the network and service provider. The reason why the network and service provider should do so is the fact that they increase their market share and revenues. High class services can only be used in a stand alone fashion by high class mobile devices. Those wireless devices are largely outnumbered by the omnipresent basic devices. To sell such

48

Frank H.P. Fitzek and Marcos Katz

high class services even to larger groups of mobile devices, the network and service provider should agree that the grouping will pay off in some way. In the following we assume that a service provider offers multiple description coded content, which is distributed by the network provider over a certain access point in multiple IP streams. Mobile devices within the coverage of the access point have now three possibilities: 1. Autonomous full-signal reception: The mobile devices try to receive all sub streams by themselves, which would result in high service costs and high energy consumption. As this may require support for a high data rate, not all mobile device may be able to perform this operation. In this case only the sophisticated and high priced mobile devices subscribe to the service. 2. Autonomous partial-signal reception: The second possibility is to receive a smaller number of sub streams, which saves money but results in a service of lower quality. 3. Cooperative full-signal reception: The third possibility is an extension to the second one but based on cooperation. In this case the mobile devices use their short–range communication link (such as Bluetooth or WLAN) to search for other mobile devices in their proximity also willing to receive the same broadcast service. Once they have found additional partners they can exchange their received descriptors over the short–range link. As learned in the section above, the descriptors should be disjoint to increase the quality of the broadcast service. This grouping of mobile device is referred as micro cooperation in [5]. The third option is the focus of this chapter, while we take the first option as a baseline reference case to illustrate the gain attained by cooperation. To illustrate this technique, Figure 2.12 shows two different video qualities for a non cooperative mobile device T 4 and the cooperative group of T 1, T 2, and T 3. By exchanging the video descriptors within the cooperative group over the short–range link, the video quality becomes better. In this example we assumed that all mobile devices can receive the same amount of descriptors over the cellular air interface. The non cooperative device T 4 could achieve the same video quality if and only if its cellular air interface could support a three-times larger data rate. But this will obviously increase the complexity of the mobile device and, as we will show later, come at the price of a larger battery drain. The feasibility of the concepts illustrated in Figure 2.12 were demonstrated by implementing some testbeds. This was done within the EDWIN (EnhanceD cooperative WIreless Networks) project, under the umbrella of the C3 initiative, at Aalborg University, Denmark. Three different demonstrators are introduced here to show the benefit of cooperation in terms of video quality and energy consumption on different platforms. The first testbed architecture is depicted in Figure 2.13 and focuses on the service quality improvement by cooperation. A video server conveys multiple

2 Cellular Controlled Peer to Peer Communications

49

Figure 2.12. Example for multiple description quantizers showing one noncooperative mobile device and a group of three cooperative mobile devices with the resulting video quality.

description coded sub streams towards the 802.11g access point operating in the 2.4 GHz band. This technology was meant to represent the cellular communication. The access point conveys the sub streams to four laptop computers. Each laptop has a built–in 802.11g interface as well as an interface card for 802.11a in the 5 GHz band, which is used for communication between the laptops. In the first step multiple description coding was implemented in the VLC [4] running on the video server. The video service conveys the sub streams over the access point (802.11g) to the mobile devices. An IP switcher was implemented on the mobile device side. The incoming IP stream was forwarded to the local VLC client and a copied version sent out over the short–range link (in this case 802.11a). Furthermore, the received central stream was merged with the incoming cooperative streams to be displayed within the VLC client. On each mobile device a user interface as given in Figure 2.14 was running. With this interface the user can specify whether to cooperate or not, start the receiving process, and monitor the incoming streams. Furthermore, information about each of the active wireless air interfaces is given. Figure 2.14 illustrates how the virtual data rate increases with each additional cooperating entity. Each substream has a mean data rate of 300 kbit/s. Therefore, three cooperating entities have a virtual data rate of 900 kbit/s. The data rate plot fluctuates as we used variable bit rate (VBR) encoded videos.

50

Frank H.P. Fitzek and Marcos Katz

Figure 2.13. Demonstrator setup for a cooperative cluster exploiting MDC.

It can be demonstrated that the video quality increases with each additional cooperative device joining the group. Moreover, the transmission is robust to the dynamics of the network, in particular when devices leave or join the network. The user experiences a loss in quality if mobile devices are disconnected, but the degradation is graceful and at least a minimum service can be guaranteed even thoughhen no other cooperative device can be found. Finally, we would like to note that services from different wireless communication systems can be merged in the cooperative cluster as given in Figure 2.15. From the perspective of the service provider and network operator, cooperation among mobile devices is directly related to an increased revenue. As an example we refer again to the example of MDC video distribution, but can be extended to more examples easily. As given in Figure 2.16 (revenues for the operator are given in a circles and those of the service provider in a box), we assume the case of i.) one low class device receiving one MDC stream, ii.) one high class device receiving three MDC streams, and iii.) a cooperative cluster of three low class devices (low class in terms of cellular data rate reception but with a sufficient display). We assume a linear cost model for the network and service provider. Thus the user of the low class device pays 1 unit to the service as well as the network operator (2 units in total). The user of the low class device may like to receive a better service but she/he cannot receive it due to the technical limitations. This changes for the user of the high class device, who receives three times more data rate and even a better services, therefore paying two times 3 units (6 units in total). Each device of the cooperative cluster receives the same data rate a the stand alone low class device and therefore each device pays 1 unit to the

2 Cellular Controlled Peer to Peer Communications

51

Figure 2.14. The client GUI of the MDC testbed.

network operator. As the devices exchange the descriptions, they increase the service quality to the maximum equal to that of the high class terminal and therefore they pay 3 units each. Thus, by cooperation the cooperative cluster gains 2 units compared to the high class terminal. From the customers point of view the benefit of cooperation is clear. For the service provider the gain of cooperative behavior is also clear as he has trebled its revenue in this example compared to the low class device. The network operator also increases its revenue, if: • service provider and network operator are the same (mostly the case in Europe and US). • the low class terminals would not even consider to use the service as one MDC stream may not provide sufficient quality. • the revenue and the supported data rate are not related linearly. Regarding the last point, it has been proven over the last years that the pricing strategy can be described as regressive, thus less money per bit for high data rate services. While SMS services with some cents for a couple of bits were very interesting for the network operator, video services requiring a larger data rate, are not achieving the same financial efficiency anymore. Interestingly, the service providers have a different pricing strategy. The relationship between service quality and costs can be described by a progressive

52

Frank H.P. Fitzek and Marcos Katz

Figure 2.15. Four cooperative mobile devices getting the information streams from different wireless communication system such as 2G and DVB-H.

Figure 2.16. Service and operator costs for stand alone and cooperative mobile devices.

2 Cellular Controlled Peer to Peer Communications

53

behavior. E.g. mp3 files can be bought on the Internet with a different quality levels following a certain pricing strategy as given in Table 2.2. Table 2.2. Cost of Grace Kelly - Mika mp3 song and related album on allofmp3 (2007 within TOP10 in Europe). Song Full Album encoding rate price amount of data price amount of data kbits cent Mbit cent Mbit 128 10 2.9 145 43.3 192 14 4.4 200 65.0 320 21 7.3 306 108.3

Therefore it will be very interesting to see how network operators and service providers will setup their pricing strategy for cooperative services. Furthermore, going a little bit further, the example shows that the network operator has to be careful to not degrade to a pure bit delivery entity as the real benefit lies in the service. The service provider can decode its content in small portions such that the costs for the network delivery are small and the mobile users are combining their portions in a cooperative manner paying the service providers the full price. 2.4.2 Unicast Services Cooperation is not restricted to multicast services. Unicast services can also be exploited in a cooperative manner. In Chapter 24 the DVB-H scenario is extended to unicast IP services. As the unicast IP services are conveyed in a broadcast fashion within the DVB-H cell, the concepts derived above for video services are still valid. In Chapter 26 the possibility of cooperative header compression is introduced. Here the mobile devices cooperate to exploit the diversity of the cellular links making the service support more robust and reliable. On top of this any application can be used, including unicast services. And further examples are available. Here we note only the possibility of cooperative web surfing. We take the example of two mobile devices with a given cellular data rate and once again we assume that the supported data rate over the short–range link is much larger than that of the cellular link. For the ease of understanding we assume a GPRS connection for the cellular and a Bluetooth connection for the short–range. The communication architecture is now assumed to be as given in Figure 2.4. If both customers start to surf the web, their traffic can be modelled as ON OFF traffic. If they do not cooperate each cellular link will be idle for most of the time and active when a new page is requested. The time they have to wait for the full page to be displayed depends on the content size and the available data rate. In the cooperation

54

Frank H.P. Fitzek and Marcos Katz

case the traffic can be split in two parts. One part is the direct cellular part and the second part is relayed by the neighboring mobile device. We note that each web page is already divided into many IP sub streams and therefore this scenario can be easily applied. In the cooperative scenario with two mobile devices the data rate is nearly doubled and therefore the waiting time should be cut in half. It may happen that both stations request the medium at the same time. In such a case the performance is only as bad as in the non cooperative scenario. With larger numbers of cooperating mobile devices the probability that all request new pages simultaneously is reduced even further. How the system deals with unequal cooperation awards (if one mobile device uses more resources than it contributes) depends on the network costs and the relationship of those mobile devices to each other. As one candidate for cooperative uplink we refer to Chapter 27.

2.5 Service Discovery within Cooperative Cluster This section briefly discusses ideas and concepts for service announcement within a cooperative cluster. We believe that the announcement should be carried out by the entities themselves instead of the cellular core network. Certainly the base station may enhance the service announcement, but the main task should be done by the interacting mobile devices in part because they can best determine what their ranges and capabilities are. We assume that running on each mobile device is an application showing which cooperative services are currently active. Figure 2.17 depicts such an application. It gives the user an overview of ongoing cooperative services. Furthermore, the application has to inform the user about the possible costs in case he or she joins the group. The overall cost is composed of network and the service provider costs. Charge information needs to be given by the network and/or service provider. Such an application obviously benefits users, but it benefits the network operator and the service provider as well. Knowing that there is an ongoing cooperative group may encourage other users to join. Figure 2.18 shows the market potential increase due to the announcement of the ongoing services. This kind of advertisement is different from those that could be launched by the network operator. Here the surrounding customers are using the service and indicate that it might be interesting for them. The customer may spontaneously join the service. This will lead to a significant increase of the market potential. We stress that cooperation increases market potential due to the following facts. First, cooperative clusters advertise their services and may convince other customers to join spontaneously. A good service or a service that is used by others intensively may be frequented also by customers that would not have considered such a service beforehand. The second reason why cooperation increases the market potential lies in the fact that more mobile devices can

2 Cellular Controlled Peer to Peer Communications

55

Figure 2.17. Possible application of service announcement within a cooperative group.

Figure 2.18. Increased market potential for cooperative services.

now enjoy high class services. While high class services can be utilized by high class devices in a stand alone fashion, the same service can be used by basic devices exploiting cooperation.

56

Frank H.P. Fitzek and Marcos Katz

2.6 Benefits of Cooperation in the Wireless World In this section we summarize and discuss the potential of cooperation, taking into account in particular customers, network operators, service providers, and mobile device manufactures. Why should end users be interested in cooperation? • Virtual Capacity: Even though the cellular link capacity is low such as for CSD or GPRS, cooperation offers a virtual high capacity link available with the current infrastructure. The bundling of low capacity links is especially interesting for broadcast services as all users are interested in the same content. The exchange over the short range technology is not bound to any capacity constraints as they offer data rates in the order of Mbps. This benefit can only be achieved if the maximum data rate on the mobile phone is smaller than the maximum data rate that the network provider offers at the base station or access point. • Energy Saving: If data is shifted from the cellular air interface to the short–range air interface, the overall energy consumption is reduced as the energy per bit ratio is much larger on the cellular than on the short–range [5]. • Low Service Costs: As we only get a partial service of the cellular link, the cost for the network operator and service provider are reduced. • Robustness: As the service is provided over different wireless paths, diversity gain can be expected. Why should the network provider be interested in cooperation? • Increased Revenue: The network operator faces problems introducing services such as video as they need much more bandwidth. The resulting cost is not acceptable for most of the consumers. By splitting the cost among different users the price decreases and more customers are willing to join. Note, the costs need not be split in a linear fashion, but can be set by the network operator. The prices should be set in such a way that the network operator and user are encouraged to cooperate. • Larger Market Penetration: The biggest advantage for the network provider is that heterogeneous devices may cooperate. In case a 3G and a 2G phone cooperate, they contribute with a different cellular capacity. The 3G phone can contribute with a data rate ten times larger than the 2G phone. Without cooperation the 2G phone may not be able to receive a high bandwidth service but through cooperation it can. This increases the number of potential customers. • New services: On top of the broadcast service, the network provider can offer some services to further stimulate the cooperation. As in the aforementioned example of the 2G and 3G phones, the network provider may charge the 2G phone a higher rate than the cooperating 3G phone. The 2G phone would essentially be compensating the 3G phone for the

2 Cellular Controlled Peer to Peer Communications

57

higher energy consumption the 3G phone incurs when it forwards data to the 2G phone. This differential pricing has to be done in accordance with the service provider. Why should the service provider be interested in cooperation? • •

Higher Revenue: As we have pointed out clearly before, the service provider will achieve higher revenues. Service Scalability: One of the characteristics of future wireless communication systems is that the mobile devices become more and more heterogeneous in terms of cellular data rate, battery, display, etc. Providing the appropriate content for all device classes may result in the need for huge storage and data pipe to the mobile device. To overcome this problem multiple description coding is offering an inherent solution [7] supported by cooperation. Why should the device manufacture be interested in cooperation?

•

Low cost and energy saving devices: The biggest problems of mobile device manufactures are the device price and the energy consumption. As the first one is pretty obvious, the later one can be divided into two sub problems. The first part is related to the operating time of the mobile device. The more energy that is wasted, the shorter the operating time becomes, which has been reported as important to the end user [11]. The second part is that more energy consumption may result in the need for active cooling, which increases the cost and size even more and is not appealing to the end consumer. Therefore, the manufactures are interested in cooperative devices, as they can offer high data rate services without supporting high data rate in a stand-alone fashion. Figure 2.19 shows the reduction in complexity for a cooperative mobile device. While in state of the art mobile devices the air interface complexity and the display size are coupled directly, cooperative devices are equipped with a less complex air interface and a huge display. The later is needed as the cooperative device should be able to display the same services as a high class device. As shown in the figure, cooperative devices can offer high class services with low class air interfaces.

As outlined, the advantage of cooperation is manifold, but still there exist some barriers to tackle. The most important one is the digital right management (DRM), which might be used to protect the content. Therefore a solution has to be found to use DRM for cooperative devices or new ways to protect the ownership, such as digital ownership management (DOM) presented in [10], have to be found.

58

Frank H.P. Fitzek and Marcos Katz

Figure 2.19. Linear increase in complexity from low to high class mobile devices and the breakthrough with cooperative mobile devices.

2.7 Conclusion This chapter gave an overview of the cellular controlled peer–to–peer network architecture. References to the following chapters are given introducing cooperative applications more in detail. Acknowledgement. The authors would like to thank all researchers around the world for the interesting discussions about this topic, which helped us to improve the chapter. Special thanks to Professor Frank Reichert for the discussion about the pricing strategy for network and service providers. We would like to thank also the students of Aalborg University, namely Beatrice Pietrarca and Giovanni Sasso, for making the measurements at the airport of Aalborg represented in Figure 2.11.

References 1. C. Andersson, D. Freeman, I. James, A. Johnston, and S. Ljung. Mobile Media and Applications, From Concept to Cash: Successful Service Creation and Launch. Wiley, 2006. 2. S.F. Brosnan and F. de Waal. Monkeys reject unequal pay. Nature, 425:297–299, 2003.

2 Cellular Controlled Peer to Peer Communications

59

3. S.F. Brosnan, H.C. Schiff, and F. de Waal. Tolerance for Inequity may increase with Social Closeness in Chimpanzees. Proceedings of the royal society, B-Biological Sciences, 272:253–258, 2004. 4. VLC developer team. Videolan client (vlc). 5. F.H.P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-40204710-X. Springer, April 2006. 6. F.H.P. Fitzek and F. Reichert, editors. Mobile Phone Programming and its Application to Wireless Networking. ISBN 978-1-4020-5968-1. Springer, 2007. 7. F.H.P. Fitzek, H. Yomo, P. Popovski, R. Prasad, and M. Katz. Source Descriptor Selection Schemes for Multiple Description Coded Services in 4G Wireless Communication Systems. In The First IEEE International Workshop on Multimedia Systems and Networking (WMSN05) in conjunction with The 24th IEEE International Performance Computing and Communications Conference (IPCCC 2005), Phoenix, Arizona, USA, April 2005. 8. D. Kahneman. A Psychological Perspective on Economics. American Economic Review, 93(2):162–168, may 2003. 9. V. Smith, E. Hoffman, and K. McCabe. Reciprocity: The Behavioral Foundations of Socio-Economic Games. Springer-Verlag – Understanding Strategic Interaction - Essays in Honor of Reinhard Selten, pages 328–344, 1997. 10. M. Stini, M. Mauve, and F.H.P. Fitzek. Digital Ownership: From Content Consumers to Owners and Traders. IEEE Multimedia-IEEE Computer Society, 13(5):4–6, Oct-Dec 2006. 11. TNS. Two-day batter life tops wish list for future all-in-one phone device. Technical report, Taylor Nelson Sofres, September 2005.

Part II

Cooperative Networks: Social, Operational and Communicational Aspects

3 Applying Evolutionary Approaches for Cooperation A General Method and a Specific Example

David Hales Bologna University [email protected] Summary. In this chapter we describe a simple general method by which existing evolutionary algorithms originating in the biological or social sciences can be translated into always-on protocols that adapt at run time. We then discuss how this approach has been applied to import a novel cooperation producing algorithm into a simulated peer-to-peer network. Finally we discuss possible applications and open issues.

3.1 Introduction Increasingly, within biological and social sciences, models of behavior are expressed in the form of evolutionary algorithms. That is, individual entities such as cells, animals or human agents are represented as interacting, mutable and reproducing entities which are modeled computationally. Computer simulation is used to specify and analyze such models because their behavior is complex and often produces emergent results that are not easily tractable analytically. Such models are typically co-evolutionary in nature in which the performance of individual entities is a result of some kind of interaction with other evolving entities in the population. This can be contrasted with evolutionary optimization algorithms, such as traditional Genetic Algorithms [5, 13] which aim to optimize an a priori objective fitness function. Such algorithms generally specify some rule by which entities interact, gaining some reward (often termed utility) and, then differentially reproduce based on utility. This differential reproduction process (the evolutionary bit) often requires that some entities “die” - they are removed from the population - and other entities produce “offspring” - they produce copies of themselves. In the context of biological models the interpretation is clear: survival of the fittest and death to the weakest. In the context of sociological models the interpretation is less clear. The assumption here is that some imitation process is occurring that favors entities with high utility. Entities are seen as behaviors or ideas that can replicate horizontally, between peers within a generation, in a population. Such culturally replicating entities are sometimes termed “memes” [4].

63 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 63–74. c 2007 Springer.

64

David Hales

This latter cultural interpretation gives us a clue as to how evolutionary models can be accommodated within information systems composed of distributed processing entities sharing some communications network. Rather than requiring the entities themselves to die and reproduce - which is obviously not currently viable - we can implement the differential imitation of behaviors between entities. To put it more directly, computational entities can transmit executable codes to each other. The target infrastructures we have in mind for the application of evolutionary algorithms are unstructured peer-to-peer (P2P) overlay networks. In a P2P overlay network there is a population of nodes, typically processes situated within a physical network, which maintain symbolic links to other nodes (often called their neighbors). P2P applications, like Skype1 or BitTorrent2 implement these to provide services. A valuable property of the overlay net abstraction is that rewiring nodes or changing the topology of the network is a logical process in which nodes simply drop, copy or exchange symbolic links. It is therefore feasible to maintain highly dynamic network topologies at the overlay layer. We make use of this property when we translate a novel tag-based algorithm into a P2P protocol (see Section 3.5). This chapter is structured as follows, firstly we provide a general method for translating evolutionary algorithms into P2P protocols (in Section 3.2), then we discuss the basic problem of cooperation and formulate it as a Prisoner’s Dilemma (Sections 3.3 and 3.4). We then apply the general technique to the specific case of a novel tag-based cooperation algorithm (Sections 3.5 and 3.6). Finally we conclude with a brief discussion of open issues (Section 3.8).

3.2 From Evolution to Protocols Translating evolutionary algorithms into parallel distributed P2P protocols is a relatively simple process. Generally this involves a parallel and asynchronous copying of application behaviors between pairs of processing entities based on a utility measure. In order to illustrate this translation process we give a set of pseudo-code template algorithms, starting with the kind of evolutionary algorithms given in biological and social simulation work and ending with an outline of a set of threads that could be the basis of a protocol design for a distributed P2P system. Listing 3.1 shows an outline of a typical co-evolutionary algorithm. We do not show here the particular way that entities interact to gain fitness (utility) or the specific reproduction method used. The reproduction phase may be implemented in many ways. A common approach in biological models is to use so-called “Roulette Wheel” selection [5]. This is a probabilistic approach requiring access to all the finesses of the the entire population. A simpler approach, from the point of distributed implementation, is to use a Tournament Selection approach. Listing 3.2 shows the same general outline algorithm but with the reproduction phase expanded with a simple tournament selection approach. Some number of reproductions are performed in which random pairs of entities are selected, utilities are compared and the entity with the lower utility copies the behavior of the node with the higher utility - meaning the behavior of the higher utility node is effectively 1 2

http://www.skype.com http://www.bittorrent.com

3 Applying Evolutionary Approaches for Cooperation

65

Listing 3.1. A generalized synchronous evolutionary algorithm. Similar to the kind of algorithms used in biological and social simulation work. I n i t i a l i z e some p o p u l a t i o n P o f N e n t i t i e s l o o p f o r some number o f g e n e r a t i o n s e n t i t i e s i n P i n t e r a c t i n some way and o b t a i n a f i t n e s s ( utility ) r e p r o d u c e a new p o p u l a t i o n P2 by r e p l i c a t i n g e n t i t i e s from P i n p r o p o r t i o n a l t o f i t n e s s ap p ly mutation t o each e n t i t y i n P2 with some low probability P = P2 end l o o p

Listing 3.2. A generalized asynchronous evolutionary algorithm using Tournament Selection method during reproduction. I n i t i a l i z e some p o p u l a t i o n P o f N e n t i t i e s l o o p f o r some number o f c y c l e s s e l e c t some e n t i t i e s i n P t o i n t e r a c t i n some way and obtain u t i l i t y l o o p f o r some number o f r e p r o d u c t i o n s s e l e c t a random p a i r ( i , j ) o f e n t i t i e s from P i f U t i l i t y ( i ) > u t i l i t y ( j ) then copy b e h a v i o r o f i t o j ap p ly mutation t o j with low p r o b a b i l i t y utility ( j ) = 0 end i f end l o o p end l o o p

reproduced. After reproduction and with low probability some “mutation” is applied to the behavior, meaning that some randomized change is made in behavior. The benefit of working with evolutionary algorithms is that, although algorithms are often presented as sequential and synchronous which aids simulation and analysis on a single machine, by their nature they should be easily translatable into distributed implementations because evolutionary processes are fundamentally distributed. Sequential evolutionary algorithms are a simulation abstraction of a parallel process. Listing 3.3 gives a simple example of a set of threads that would need to be executed by each node in a peer-to-peer system such that it would implement the same tournament based selection.

66

David Hales

Listing 3.3. A generalized Tournament Selection approach represented as three concurrent threads assumed to be running in each node over a population of nodes forming a peer-to-peer system. Here Si representing an application behavior (or strategy) of node i and Ui represents the utility. A c t i v e a p p l i c a t i o n t h r e a d f o r node i : do f o r e v e r : Engage i n a p p l i c a t i o n l e v e l i n t e r a c t i o n with o t h e r nodes using Si Update u t i l i t y v a l u e Ui A c t i v e r e p r o d u c t i o n t h r e a d f o r node i : do f o r e v e r : wait ( d e l t a ) j = selectRandomNode ( ) r e c e i v e ( Uj , S j ) from node j i f Uj > Ui then Si = Sj with low p r o b a b i l i t y Mutate ( S i ) Ui = 0 end i f P a s s i v e r e p r o d u c t i o n t h r e a d f o r node i : do f o r e v e r : send ( Ui , S i ) t o r e q u e s t i n g node j

3.3 Cooperation It is well known that the maintenance of cooperation between entities within distributed open systems is a major issue for a successful protocol design. Consider, for example, a file-sharing system in which nodes in a network may download files without uploading. If all nodes behaved in this selfish way then no files would be shared at all and the network would have no value to any node. Another example might be the broadcasting of a message through the entire network where each node relays the message to its immediate neighbors. If a substantial number of nodes choose not to pass on the message then the broadcast would not be received by all nodes. A further example could involve the sharing of load and cooperative replication of content between servers responding to client queries. All these are particular manifestations of so-called Commons Tragedies [12]. These are widely found in biology and human societies and are hence well studied in biological and social sciences. Given the problem of designing cooperative protocols several general and interrelated approaches have been proposed including: incentives, mechanism design, micro-payments and evolutionary approaches. Each have their strengths and weaknesses, but here we will focus on a novel tag-based evolutionary approach which maintains cooperative behavior through a form of group-level selection [6, 16]. To test the novel approach we use a canonical game (the Prisoner’s Dilemma or PD) which captures the dilemma of cooperation, reflected in the practical examples

3 Applying Evolutionary Approaches for Cooperation

67

we have given. Within both biological and social sciences the PD game (or variants) are often used to test proposed cooperation producing mechanisms. Firstly we introduce the PD game and then the original tag-based evolutionary algorithm. We then translate this into a P2P protocol, following essentially the process we have discussed above. Finally we give some simulation results and discuss some of our on-going work applying the protocol to application domains within P2P.

3.4 The Prisoner’s Dilemma and Variants The two player single-round Prisoner’s Dilemma (PD) game captures a situation in which there is a contradiction between collective and individual self-interest [3, 18]. Two players interact by selecting one of two choices: to “cooperate” (C) or “defect” (D). For the four possible outcomes of the game, players receive specified payoffs. Both players receive a reward payoff (R) and a punishment payoff (P) for mutual cooperation and defection respectively. However, when individuals select different moves, different payoffs of temptation (T) and sucker (S) are awarded to the defector and the cooperator respectively (see Figure 3.1a). Assuming that neither player can know in advance which move the other will make and wishes to maximize her own payoff, the dilemma is evident in the ranking of payoffs: T > R > P > S and the constraint that 2R > T + S. Although both players would prefer T , because it’s the highest payoff, only one can attain it in a single game. No player wants S because it’s the lowest payoff. No matter what the other player does, by selecting a D-move a player always gets a higher score than it would have obtained if it had selected C. D is therefore the dominant strategy – hence an ideally rational player would always choose D.

Figure 3.1. Payoff matrix (a) shows the PD payoff structure, (b) shows the values used, (c) shows alternative values from an asymmetric generalized PD (GPD) where a single player determines payoff values. It is claimed this captures more realistically interactions between clients and servers [15].

Therefore, the dilemma is that if both players select a cooperative (C) move they are jointly better off (getting R each) than if they both select D, but selfish players will select mutual defection, getting only P each, because of the individual

68

David Hales Listing 3.4. The tag-based cooperation algorithm as given in [6].

Loop some number o f g e n e r a t i o n s Loop f o r each a g e n t ( a ) i n t h e p o p u l a t i o n S e l e c t a game p a r t n e r a g e n t ( b ) with same t a g ( i f possible ) Agents a and b i n v o k e t h e i r s t r a t e g i e s and g e t p a y o f f s End l o o p Reproduce a g e n t s i n p r o p o r t i o n t o t h e i r a v e r a g e p a y o f f Apply mutation t o t a g and s t r a t e g y o f each r e p r o d u c e d a g e n t with low probability End l o o p

incentive to select defection. We select this game as a minimal test that captures a range of possible application tasks in which nodes need to establish cooperation and trust with their neighbors but without central authority or external mechanisms that enforce it. Many variants of the PD are possible. For example, the Generalized PD (GPD) as used by Feldman et al [15] to model client / server interactions where payoffs are asymmetric (shown in Figure 3.1c). Additionally, players may use probabilistic strategies where the move selected is determined by a real value indicating a probability to cooperate. Also, of course, payoffs can be varied, specifically the size of the temptation (the T payoff) inside the interval [2R..R]3 .

3.5 Tag-Based Cooperation Algorithm Listing 3.4 shows outline pseudocode for a cooperation producing tag-based algorithm [6]. Agents play the PD in pairs. The model is composed of very simple agents. Each agent is represented by a small string of bits. On-going interaction involves pairs of randomly selected agents, with matching tags, playing a single round of PD. Agent bits are initialized uniformly at random. One bit is designated as the PD strategy bit: agents possessing a “1” bit play C but those possessing a “0” bit play D. The other (L) bits represent the agents’ tag – a binary string. Tag bits do not affect the PD strategy played by the agent but they are observable by all other agents. Each agent is selected in turn to play a single-round of PD. Agents do not selected an opponent randomly but selectively based on the tag string. The opponent is selected randomly from the subset of the population sharing the same tag string as the agent. If this subset is empty, because no other agents have an identical tag, the agent plays against some randomly chosen partner from the entire population – whatever their tag values. After each pair of agents plays a game of PD the payoffs are accumulated against each agent. When all agents have been selected in turn, and played a game, agents 3

When T > 2R or T < R then there is no longer a dilemma.

3 Applying Evolutionary Approaches for Cooperation

69

are reproduced probabilistically in proportion to the average payoff they received (using a “roulette wheel” selection algorithm). With a small probability, each bit of each reproduced agent is mutated (i.e., flipped). There is no topological structure since agents are not situated in a space – such as a lattice or a ring – interaction is only structured using tag similarity and random selection. The tag algorithm leads to very high levels of cooperation - even when the population of agents is initially set to all defect strategies. The key to understanding the tag process is to realize that agents with identical tags can be seen as forming an “interaction group” or “tribe”. The population can be considered as partitioned into a set of such groups. If a group happens to be entirely composed of agents selecting action C (a cooperative group) then the agents within the group will outperform agents in a group composed entirely of agents selecting action D (a selfish group). This means that individuals in cooperative groups will tend to reproduce more than agents in selfish groups because they will obtain higher average payoffs. If an agent happens to select action D within a cooperative group then it will individually outperform any C acting agent in that group and, initially at least, any other C acting agent in the population – remember the T payoff is 1.9 but the best payoff a C acting agent can get is R = 1. However, due to its high payoff such a D acting agent will tend to reproduce many copies of itself and then the group to which it belongs becomes very quickly dominated by the newly reproduced D acting agents. The group then becomes a selfish group and the relative advantage of the lone D acting agent is lost – the group becomes unsustainable due to the interaction being kept within the group. So by selecting the D action an agent destroys its group very quickly (remember groups are agents all sharing an identical tag). A similar algorithm of tag-based cooperation was presented in [16]. Initial work on the tag ideas was proposed by John Holland (the “father” of genetic algorithms) [14].

3.6 The SLAC Protocol We translated the tag-based algorithm in the manor discussed perviously. This involved introducing a tournament selection process and application level behavior (in this case playing the PD with randomly selected neighbors) as a node level protocol composed of three threads. However, the strategy here, is not just the PD strategy of cooperate or defect, but also the tag. The tag needs to be a copyable feature that specifies the possible interaction partners of the node. Essentially it needs to specify some kind of group membership. The trick we used was to translate the tag into a neighbor list. That is, each node stores a list of it’s immediate neighbors and this neighbor list (or view) determines which other nodes from the population can be selected for interaction - it performs the function of a tag. The combination of the PD strategy and the View (neighbor list) comprises the entire composite strategy. We called the new protocol SLAC (Selfish Link-based Adaption for Cooperation) because we no longer use tags as such but rather adapt links between nodes. Interestingly this approach can be closely compared to previously proposed link-based incentive schemes [19]. Listing 3.5 shows the outline pseudocode for the SLAC protocol. Figure 3.2 shows the typical evolution of a SLAC network. Notice the quick formation of components

70

David Hales

Listing 3.5. The SLAC protocol. Each node in a P2P network runs the above threads. Here Si stores both PD strategy and the View (links to neighbors in the network) of node i and Ui represents the utility. We expand the mutation step to include the way mutation is applied to both PD strategy and View. The node View represents the function of the tag from the previous tag-based algorithm. A c t i v e a p p l i c a t i o n t h r e a d f o r node i : do f o r e v e r : S e l e c t a random node from c u r r e n t n e i g h b o r l i s t Play PD with n e i g h b o r and g e t p a y o f f Update u t i l i t y v a l u e Ui a s p a y o f f r o l l i n g a v e r a g e A c t i v e r e p r o d u c t i o n t h r e a d f o r node i : do f o r e v e r : wait ( d e l t a ) j = selectRandomNode ( ) r e c e i v e ( Uj , S j ) from node j i f Uj > Ui then Si = Sj with low p r o b a b i l i t y Mutate ( S i ) : mutation o f PDstrategy = f l i p s t r a t e g y mutation o f View = drop a l l l i n k s and l i n k t o random node Ui = 0 end i f P a s s i v e r e p r o d u c t i o n t h r e a d f o r node i : do f o r e v e r : send ( Ui , S i ) t o r e q u e s t i n g node j where : S i = { PDstrategy f o r node i , View f o r node i } PDstrategy = {C | D} View = l i s t o f immediate n e i g h b o r l i n k s ( up t o some max . = 20)

and then the rapid spread of cooperation over the nodes. Figure 3.3 gives some results from computer simulations showing the time to attain high-levels of cooperation when starting from a random network with all nodes following defect strategies. SLAC networks are highly robust because the evolutionary process will always push the network towards a cooperative state even if many nodes fail or leave the system or new nodes enter the system. This is one of the major benefits of using evolutionary approaches - they are inherently robust to noise. In fact, they rely on noise, via mutation, to function. We have only given an overview of SLAC here, more detail can be found in previous publications [7, 9, 10].

3 Applying Evolutionary Approaches for Cooperation

71

˜ Figure 3.2. Evolution of a SLAC network with nodes playing the Prisoner Os Dilemma. From an initially random topology composed of all nodes playing the defect strategy (dark shaded nodes), compo- nents quickly evolve, still containing all defect nodes (a). Then a large cooperative component emerges in which all nodes cooperate (b). Subsequently the large component begins to break apart as defect nodes invade the large cooperative component and make it less desirable for cooperative nodes (c). Finally an ecology of cooperative components dynamically persists as new components form and old components die (d). Note: the cooperative status of a node is indicated by a light shade.

72

David Hales

Figure 3.3. Results of simulation experiments with the SLAC protocol playing the PD game. Each bar shows the time taken to attain high levels of cooperation. All nodes are initialized to play the defect strategy within a random topology. The results are averages over 10 independent runs, variances are low and not shown. Notice the slight reverse-scaling property which improves performance as network size is increased. Figure 3.2 shows snapshots from a typical single run.

3.7 Possible Applications We have adapted the SLAC protocol for application in a number of simulated task domains4 . In [9] we added probabilistic link copying producing fully connected cooperative networks following a small world topology. We then adapted and applied this for a broadcasting task, where randomly selected nodes need to spread a message to the entire network [2]. In [8] we applied SLAC to a task sharing scenario in which nodes receive jobs that require skills - this requires specialization within the group. We recently applied a modified form of SLAC to a distributed replica management domain [11]. However, these simulated application domains are still represented at a highly abstract level and further work is needed before implementations can be produced. Although we have, what appears to be, a general mechanism for sustaining cooperation between nodes in an adaptive network there are several open issues that need to be addressed to produce a deployable protocol. The SLAC protocol creates an incentive for cooperative behavior because non-cooperative nodes become quickly surrounded by others of the same type. This relies on nodes communicating and copying both links and behaviors honestly. This mechanism can be subverted by nodes lying about strategies and links and other malicious behaviors. Interestingly, we performed some experiments with certain classes of malicious behavior and found that in some cases this can improve performance [1]. However, we also found that a small number of highly malicious nodes can degrade performance of the entire network significantly. 4

The code for SLAC and related simulations can be found on the Peersim webpage: http://peersim.sourceforge.net

3 Applying Evolutionary Approaches for Cooperation

73

3.8 Discussion and Conclusion We aimed in this chapter to practically illustrate how to take an evolutionary algorithm and translate it into a P2P network protocol in general. We also presented a specific example and gave some results. We have briefly discussed subsequent work in which we have adapted and applied the developed protocol to simulated task domains. There are a number of open issues that need to be addressed before practical implementations of these protocols can be deployed. These mainly related to malicious behavior. Currently the protocols require some degree of honesty in the nodes and this can not be assumed in open systems. In some recent work we have proposed not to use utility and link reporting between nodes but rather to rely on a binary node satisfaction function and randomized linking. In this case each node needs to set a desired utility level which it attempts to obtain by changing links randomly [11]. Finally, in the context of wireless systems, the current proposed protocols rely on the ability to randomly sample from the entire population of nodes. This would appear highly implausible or costly in the real world. However, recently evolutionary models from biology have shown that this may not be a requirement of such link based approaches. It has been shown, for example, that it is only necessary for nodes in a network to have access to their neighbors neighbors (two hops) in order to support cooperative evolution [17]. Acknowledgement. Thanks go to Stefano Arteconi, from the Dept. of Computer Science, University of Bologna, for the picture shown in Figure 3.2 and also for the Peersim implementations of SLAC.

References 1. S. Arteconi and D. Hales. Greedy cheating liars and the fools who believe them. Technical Report UBLCS-2005-21, University of Bologna, Dept. of Computer Science, December 2005. 2. S. Arteconi and D. Hales. Broadcasting at the critcial threshold. Technical Report UBLCS-2006-22, University of Bologna, Dept. of Computer Science, October 2006. 3. R. Axelrod. The Evolution of Cooperation. Basic Books, 1984. 4. R. Dawkins. The Selfish Gene: Second Edition. Oxford University Press, 1989. 5. D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley, 1989. 6. D. Hales. Cooperation without Space or Memory: Tags, Groups and the Prisoner’s Dilemma. In Multi-Agent-Based Simulation, Lecture Notes in Artificial Intelligence 1979, Springer, 2004. 7. D. Hales. From Selfish Nodes to Cooperative Networks: Emergent Link-based Incentives in Peer-to-Peer Networks. In proceedings of The Fourth IEEE International Conference on Peer-to-Peer Computing (p2p2004), 25-27 August 2004, Zurich, Switzerland. IEEE Computer Society Press, August 2004. 8. D. Hales. Emergent group-level selection in a peer-to-peer network. Complexus, 2006(3):108–118, 2006.

74

David Hales

9. D. Hales and S. Arteconi. Slacer: A self-organizing protocol for coordination in p2p networks. IEEE Intelligent Systems, 21(2):39–35, 2006. 10. D. Hales and B. Edmonds. Applying a socially-inspired technique (tags) to improve cooperation in p2p networks. Transactions in Systems, Man and Cybernetics - Part A: Systems and Humans, 35(3):385–395, 2005. 11. D. Hales, A. Marcozzi, and G. Cortese. Towards cooperative, self-organised replica management. Technical Report UBLCS-2007-02, University of Bologna, Dept. of Computer Science, January 2007. 12. G. Hardin. The tragedy of the commons. Science, 162:1243–1248, 1968. 13. J. Holland. Adaptation in Natural and Artificial Systems. University of Michigan Press, 1975. 14. J. Holland. The effect of labels (tags) on social interactions. Working Paper 93-10-064, Santa Fe Institute, 1993. 15. J. Chuang M. Feldman, K. Lai and I. Stoica. Robust Incentive Techniques for Peer-to-Peer Networks. In ACM Conference on Electronic Commerce, ACM Press, 2004. 16. R. L. Riolo, M. D. Cohen, and R. Axelrod. Evolution of cooperation without reciprocity. Nature, 414:441–443, 2001. 17. F. C. Santos, J. M. Pacheco, and T. Lenaerts. Evolution of cooperation without reciprocity. PLoS Comput. Biol., 2(10), 2006. 18. J. M. Smith. Evolution and the Theory of Games. Cambridge University Press, 1982. 19. Q. Sun and H. Garcia-Molina. SLIC: A Selfish Link-based Incentive Mechanism for Unstructured Peer-to-Peer Networks. In Proc. of 24th IEEE Int. Conf. on Distributed Systems, IEEE Computer Society, 2004.

4 The Social Qualities of Pervasive Wireless Networks Mark Pesce Honorary Associate, Digital Cultures Program, University of Sydney [email protected]

4.1 Introduction: The Social Hormone In 1997 half of humanity had never made a telephone call; by the end of 2007, half of us will own a mobile telephone. It took seven years to get the first billion mobile handsets into use, two years for the second billion, and about a year for the third billion. Everyone who can afford a handset will soon own one. In the third world, inexpensive pre-paid services bring mobile telephony to individuals whose incomes are just a few US dollars a day. Why would such poor people make an investment that might cost them a month’s income? Recent studies (Economist 12 May 2007) indicate that people use mobile phones to create their own arbitrage networks; fishermen in Kerala, in Southern India, make a few calls when still out at sea - their GSM service is available 5km offshore - to determine where their catch will net the best prices. In this way, the cost of a mobile handset is recouped in about a month, and the markets are never short of fish. Everyone is satisfied because the connective intelligence [8] within the human network is effectively maximized when the individuals within the network are empowered with wireless communications. First-world individuals have lived with pervasive wireless communication for years, enough so that it has become an incorporated environment [3]; that is, we have ontologically incorporated the existence of the network into our understanding of how the world works [13]. No one is ever late any more; a simple voice call or SMS takes care of that. Instead, we are better coordinated than ever before; the social arbitrage of pervasive wireless networks has both smoothed and accelerated the pace of social interactions. Considered in this way, pervasive wireless networks function as a social hormone [11], instantaneously signaling to parts distant. The pervasive nature of the wireless network has inverted the human relationship to the physical layers of the network. Until the last decade, we had to present ourselves at the network’s point of presence. Today, the human has become the point of presence, and the network comes to us. Where a land-line ties the network to a place, a mobile handset is bound to a person, irrespective of location. The collapse of space [11] reaches completion; not only has the electrification of communication collapsed distance, it has also freed us from any particular position. Proximity is now measured in response time, both the sub-second response time of the global network, and the enhanced ability of human beings to respond quickly to any message

75 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 75–86. c 2007 Springer.

76

Mark Pesce

sent through that network. This “telecosm” [7] has engendered substantial cultural transformations; the “global village” predicted by McLuhan seemed to arrive first with television, but, in truth, awaited an age of pervasive wireless networks. Only when every individual is empowered both to transmit and receive could the feedback mechanisms of networked intelligence begin to emerge [1]. The qualities of networked intelligence are difficult to predict in advance of their emergence. The poor in India - who nonetheless own mobile handsets - have perfected a system of missed calls as out-of-band signaling. When a relative goes on a journey - perhaps only to the next village - a missed call to home indicates a safe arrival. Although tariff rates in India are perhaps the lowest in the world, even a cent or two for a call is too much for these very poorest users of the pervasive wireless network; instead, they use the their own human intelligence to encrypt and decrypt messages cast up by the mobile network. (Do the low-cost Indian carriers mind this abuse of their network? Arguably no, because these individuals will grow to see the mobile as an indispensable item and the carriers will eventually get tariffs from voice and text message traffic as their customers grow in wealth.) “The street finds its own use for things,” William Gibson wrote in 1989, “uses the manufacturers never imagined.” This sums up the current state of affairs; we have manufactured a pervasive, multimodal, multi-band, multi-purpose global wireless network. It comes in many flavors, from UWB connectivity for high-speed, short-range connectivity, through to WiMax broadband, spread over tens of square kilometers; all of it, now handed over to an enthusiastic humanity, will permute and evolve in unpredictable ways. Three trends can already be identified, emerging from the newly empowered condition of man as network node. Their emergence will challenge the basic assumptions of modern, globalized civilization.

4.2 Hyperintelligence and the End of Elites In December 2005 the prestigious journal Nature published the results of a study of the relative accuracy of two encyclopedias: the 250 year-old definitive reference in the English language, Encyclopedia Britannica, and the four year-old “peer-produced” reference, Wikipedia. A random survey of articles showed that each entry in Britannica had an average of three errors, while analogous articles in Wikipedia had an average of four errors per entry. While most of the public seemed more surprised that Britannica was itself riddled with misinformation, perhaps the more shocking fact was that an encyclopedia of information freely contributed to by the public could nearly rival Britannica in accuracy. Wikipedia has become the archetype for an ancient mode of human behavior knowledge sharing - that has been amplified and accelerated by pervasive wireless networks. Knowledge sharing is a common feature among all the members of biological subfamily Hominidae. Bonobos, chimpanzees, gorillas and human beings, all very social animals, practice knowledge sharing as an innate behavior. The obvious selection advantages which benefit any animal who can share the product of their experiences with their relations have made this a conspicuous feature of all these species, and recently chimpanzees have been observed in the wild instructing one another in the use of tools - a behavior that dates back at least to the Paleolithic (Proceedings of the National Academy of Sciences, February 2007). We know that infants and toddlers practice highly mimetic behaviors; they closely observe the

4 The Social Qualities of Pervasive Wireless Networks

77

behavior of their parents, and put those observations into practice. Since the development of language (which likely evolved from a gestural “sign language” before the emergence of speech), human beings have communicated their knowledge through symbology. Each familial-tribal grouping of prehistoric humans had developed own rich symbolic systems which expressed the shared knowledge of the group. When human culture made the transition from oral to written forms, some 5500 years ago, the human world underwent its second “knowledge explosion” (the first arguably happening with the birth of language). By the time of the Alexandrine Library (3rd century BCE), an enormous reservoir of human experience, observation, and knowledge had been gathered into a form where it could be freely shared among an elite of scholars, rulers and clerics. The tens of thousands of works (no exact count or index remains) represented the Wikipedia of the ancient world; any question could be answered, if one only knew where to look - and provided one could read the text. The flowering of art and science associated with Alexandria (including the first steam engine, and a wealth of medicine and mathematics rediscovered under the Caliphate) represented a pre-modern version of a “knowledge economy.” McLuhan argued that the Roman Empire fell not because of barbarian invasions, but because the supply of Egyptian papyrus - the information medium of the ancient era - dried up. An empire can not be run on expensive parchment; without a constant information flow, the empire soon found itself fragmenting into the fiefdoms of the Early Modern Era. With the advent of the printing press in the mid-15th century, the modern world began its rise, built on the availability of inexpensive, high-quality information. Individual libraries of hundreds of volumes of “essential” knowledge became a common feature in upper-class households, and universities boasted thousands, then tens of thousands, then millions of volumes. The scientific revolution embodied the principle of knowledge sharing as its basic tenet; researchers from all over Europe (and eventually, all over Earth) pooled their hypotheses and results, leading to an exponential growth in knowledge. Even so, the creation of knowledge was an asymmetric process; elites in laboratories, universities and monasteries created the knowledge, then shared this knowledge within their elites. Only occasionally did the mass of humanity have any need for newly-created knowledge - generally where it concerned agricultural practices [5] - and that knowledge typically spread slowly and horizontally. These two knowledge-production systems, elite/top-down and plebian/peer-to-peer remained essentially isolated from one another until the early 21st century. Intersecting at Wikipedia’s launch in January 2001, the collision of elite and peer-to-peer systems of knowledge sharing paradoxically produced more light than heat. Where it might have been expected that the mass of humanity, collectively contributing to a writeable “wiki” (Hawai’ian for “quick”), would only create a mass of inaccuracies and urban legends, precisely the opposite happened: Wikipedia rapidly became a high-quality factual source. This has often been attributed to the “wisdom of crowds,” that is, the median estimation of the truth from a large sample of the population. In other words, with a large enough sample inaccuracies will cancel themselves out. But this is not what has happened within Wikipedia; every one of the millions of contributors who have edited Wikipedia have not edited every single page. Only a very few of Wikipedia’s entries have had enough editors to qualify as crowd wisdom, and these few are more like ontological battlegrounds for socially provocative issues. Instead, Wikipedia embraces a new phenomenon: hyperintelligence.

78

Mark Pesce

When any group of human beings establishes a certain level of connectivity proximal and embodied, or networked and virtual - knowledge sharing emerges. This we know from an analysis of zoology and anthropology. When the number of people connected reaches into the hundreds of millions, individuals seek to form connections with others who share their interests. These “amateurs”, who may lack professional qualifications, but who nonetheless possess deep knowledge of a particular area of interest, share this knowledge freely within their own self-defined social networks. Each individual’s position within these emergent social networks is based solely upon expertise; one is judged by one’s peers by the depth of contribution made to the community. This was as true for Paleolithic communities as it is for the virtual communities of the 21st century - the same very ancient biological and anthropological forces are in play. An emergent meritocracy quickly forms, and humans being social actors on a constant quest for dominance [2] - each actor in the social network is constantly challenged to maintain and improve their social standing within the network by increasing their contributions to it. In this way wellconnected individuals bound by a common interest quickly learn how to mine and reap the benefits of the collective intelligence within their community. This process, repeated countless thousands of times, is the engine that powers Wikipedia. Wikipedia, with its potentially infinite range of subject material - everything known to humans, factual and cultural, is relevant to Wikipedia - has become an attractor for these social networks of hyperintelligence. No network is all-knowing, but within a specific area of expertise, it can easily come to outclass certified “experts”. A Wikipedia article, as a product of hyperintelligence, can easily equal the quality of a Britannica article written by an expert, and peer-reviewed. This is one of the major disruptive transformations of the 21st century, one specifically wrought by pervasive networking: the transition from a culture of elite professionals to a culture of hyperintelligent amateurs. The published attacks of elites, raging against the teeming masses who contribute to Wikipedia, as typified by the strong negative response to the Nature article, represent the opening salvos in a drawn-out war, as the cultural torch is passed to those amateurs who harness hyperintelligence to out-think, out-maneuver, and out-produce their professional peers. In the era of pervasive wireless networking, each of us brings the network of hyperintelligence with us. As we develop systems to harness hyperintelligence Wikipedia is simply the first such example, important because it has sensitized us to the power of hyperintelligence latent in all our social networks - the pervasive wireless network will ensure that we are always fully connected to those networks, constantly contributing to them and receiving the full value of our participation within them. Like the Kerala fishermen using mobile telephony to create their own arbitrage networks (one form of hyperintelligence), we are already becoming closely bound to these networks, because they improve our effectiveness. Effectiveness is a very seductive quality, touching on the psychological underpinnings of will; the more we benefit from hyperintelligence, the more likely we are to continue to use it, and benefit from it. That the mobile telephone has so quickly become the one indispensable electronic accessory provides the proof of this; mobile telephony provides baseline connectivity for the emergence of hyperintelligence, but much more will come as we learn how to create the fertile conditions for the rapid and effective emergence of higher levels of hyperintelligence.

4 The Social Qualities of Pervasive Wireless Networks

79

4.3 Pox Populi and the Collapse of the Mass Mind The presence of the pervasive wireless network carries one quality innately its own - hyperconnectivity. Where networks in the pre-Internet era were characterized by a top-down centralized distribution of information, as typified by the mass media, present day networks allow for a fully horizontal and rhizomatic [4] distribution of information. Whereas in the earliest days of pervasive networks this had been mostly a one-to-one arrangement (such as in the case of electronic mail), we have recently seen the emergence of individuated, one-to-many forms of “hypercasting” (Pesce, 2006), where an individual, employing peer-to-peer technologies such as BitTorrent, or Web2.0 tools such as Flickr or YouTube, can effectively reach an audience as large as or larger than any which could be reached by a corresponding broadcasting technique. Hyperconnectivity facilitates more than just the lateral transfer of information; it creates the necessary preconditions for two qualities of emergent behavior: the interrelated phenomena of “swarms” and “storms”. In a pervasively networked environment, individuals rapidly learn about new or newly-exciting topics of interest to them; this is an innate function of hyperintelligence. These individuals then tend to cluster around the resources (in this case, network resources) which allow them to explore the topic at hand. Such an individual is rarely ever alone in their interest; instead a “swarm” of individuals with similar interests maintains “co-presence”, constantly interrogating and being interrogated by the swarm of peers. This swarm could be entirely based around the exchange of a specific set of information: file-sharing networks fall into this category. Users in file-sharing networks create a “swarm” of network resources to provide a resilient system for the sharing of large audiovisual files, much to the consternation of copyright holders, whose commercial distribution rights are often violated by these swarms. These swarms could be entirely social: Mizuko Ito, in Personal, Portable, Pedestrian: Mobile Phones in Japanese Life (MIT Press, 2005) noted that teenagers in Japan often maintain a list of four or five peers whom they are constantly in touch with, via SMS, during all waking hours, simply reinforcing the social network which binds them together. Swarms and their associated emergent hyperintelligence can be highly focused and task-driven. Perhaps the most significant case of a task-driven swarm (outside of Wikipedia, which has already been described) began early in 2007, on a blog known as Talking Points Memo, or TPM. Josh Marshall, the editor-in-chief of TPM, noted that a US Attorney had been fired, without explanation, and asked his readership (which at that time numbered in the tens of thousands) for any light they could shine on the matter. The swarm around TPM went to work and soon it became clear that at least six US Attorneys had been fired simultaneously, all without explanation. This discovery quickly escalated into a full-scale scandal and a US Congressional investigation. As part of this investigation, the US Department of Justice made regular “document dumps” - numbering in the thousands of pages late on successive Friday evenings. This is a common political technique to avoid the fast-paced weekday news cycle, and has the additional benefit of ruining the weekends of congressional investigators. Marshall, beginning to sense the power of the swarm around TPM, invited his readership to download some or all of the documents from the Department of Justice’s website, asking them to perform their own forensic investigations. Very quickly - in minutes to hours - TPM was able to unearth the kinds of factual information which would have taken congressional investigators

80

Mark Pesce

days or weeks to uncover. Moreover, with each document dump, the swarm at TPM became more perceptive and more effective at sifting relevant facts from the morass of information. Where it was once possible to drown investigators in paperwork, the distributed hyperintelligence of the swarm has effectively evolved a response which renders that technique impotent. The swarm moves faster, learns faster, and responds faster than any top-down hierarchy working against its aims. Copyright holders constantly attempt to frustrate file-sharers, only to find themselves outpaced, both technologically and creatively, by the swarm of file-sharing users and sites. Politicians seek to jam the system with paperwork, encountering a swarm of dedicated amateur investigators only to happy to distribute the workload amongst themselves, pooling their results, and sharing their insights through their broader networks of affiliation. This is only possible because of the network effects of hyperconnectivity; where individuals can freely and instantaneously form swarms within pervasive networks, these kinds of emergent behaviors are the norm, rather than the exception. Alongside the swarm, the “storm” typifies another new form of hyperconnected behavior. It is now possible for a single piece of media - a photograph, or a video, or a clip of audio - to be nearly instantaneously shared around the entire world. The defining case in point came on 7 July 2005, the day of the bombings in the London subway system. The first photos of the wreckage - from within the subway tunnels were available on the photo-sharing site Flickr, not through the mainstream media. Why? An individual with a camera-equipped mobile telephone snapped an image and, using the pervasive GPRS network, immediately transmitted that image to Flickr, where it could be seen, tagged and shared globally - and eventually broadcast by the BBC. Storms are another phenomenon which emerges directly from the innate knowledge sharing capabilities of human beings. The behavior of knowledge sharing can be broken down into three distinct phases: first, some new information is found by an individual; next, the individual filters this information against his social network, weighing which actors in the network would most benefit from this information (leading to the largest increase in social standing within that network); finally, the information is forwarded to those actors. These actors, reacting to the forwarded knowledge, also repeat these same steps, and so it continues until the natural limit of forwarding is reached. For information which is highly relevant - such as those photos of bombed-out subway trains - that natural limit might extend to a large proportion of all individuals within the pervasive wireless network. We have storms before; JibJab’s Flash video “This Land” - a comedic take on the 2004 Presidential election in the United States - garnered some seventy million viewers in the space of four months; the YouTube video “Free Hugs Campaign” had an audience of thirteen million viewers in just six weeks. This so-called “viral” distribution of media was first restricted to entertainment, but now, as with Wikipedia, storms have begun to collide with the mainstream news media, and are consequently having a profound influence on politics. In August 2006, Virginia’s incumbent US Senator, George Allen, campaigning for re-election, made a campaign stop at a park in the city of Breaks. There, Allen made a speech to assembled supporters, a speech videotaped by S.R. Sidarth, an American of Indian ancestry. Sidarth worked for Allen’s opponent, Democratic candidate Jim Webb, and - as is standard practice in tight elections - shadowed Allen’s every step along the campaign trail, collecting an audiovisual record of Allen’s state-

4 The Social Qualities of Pervasive Wireless Networks

81

ments. At this particular stop Allen - perhaps exasperated by Sidarth’s omnipresence - pointed Sidarth out to the crowd, saying, “This fellow here, over here, with the yellow shirt, macaca, or whatever his name is?” The word in question, “macaca,” a declension of macaque, or “monkey,” became the defining utterance of Allen’s re-election campaign. Within hours, Sidarth had posted the video to YouTube, and informed political bloggers of the footage. Those bloggers reported it on their blogs, and forwarded the video along to people who forwarded it along to people who forwarded it along. Within a day a casual (and racist) utterance by Allen had become the central issue of the campaign - both on the political blogs and within the mainstream media - a campaign which Allen unexpectedly lost, a loss which equally unexpectedly handed control of the US Senate to the opposition Democratic Party. Swarms and storms represent newly emergent techniques for political, cultural and economic self-organization, techniques impossible before the advent of pervasive networking. Proximal self-organization, at the community-level, is an ancient feature of human culture, arguably its defining feature. Now that greater than two hundred million people carry still and video cameras, integrated into their mobile handsets, each has been empowered with the necessary capability to participate in a swarm of hyperintelligence, or produce a storm of forwarded media. Once again, this amateur activity trumps the professionals; a reporter is no longer first on the scene of some event, gathering footage and getting the story. The members of the community proximal to the news event are themselves fully capable to provide coverage of it, and, equally, are empowered to distribute their own version of events. No one need rely on the mass media to reach a massive audience; pervasive wireless networking has supplanted the central role of broadcasting as the shaper of public opinion, the creator of mass mind. Instead, we see a disintegration into a million polities, each with their own particular point of view, and each sharing their own point-of-view within their swarm. The 21st century is already seeing the “collapse of consensus” [15], because the instrumentality of consensus building - the mass media - has been obsolesced by the pervasively networked swarm. The mass media have neither the manpower nor the bandwidth to confront the emergence of a multitude of points of view. Yet the individuals who participate in swarms demand both a depth of detail and a deference to the deeply-held views of the community, only receiving their full satisfaction within the swarm. Thus, the more one uses the swarm to form a picture of the world, the more likely one is to rely on the swarm as a source of information, until it becomes the primary source. Individuals, connected within their own social networks, and participating within emergent systems of hyperintelligence, tend to seek reinforcement for the commonly-held views of the group. There is no longer any common space for discussion or debate; the unified polity of the Industrial Era has dissolved into the “Pox Populi” - the poisoning voice of the people, where every individual, swarming as both storm broadcaster and information processor, will eventually form a moiety of one. Of all the dangers inherent in pervasive wireless networking, this is perhaps the greatest. The empowerment engendered by these networks allows us to work together, but it is corrosive in equal measure. There is no separating the distributed, collective hyperintelligence efforts of Wikipedia from the “open-source warfare” of Al Qaeda [14]; both developed along similar principles of knowledge sharing, and use similar techniques to achieve their ends. A political video of George Allen distributed in a storm via YouTube and a video portraying the beheading of American journalist

82

Mark Pesce

Daniel Pearl (which was widely distributed through file-sharing networks) are two sides of the same phenomenon. Humanity is speaking to itself, not in a unified voice, but in a confusion of tongues. The center can not hold.

4.4 Read-Write Culture and the Restructuring of Institutions Recently, an Adelaide science teacher, discussing some of the current developments in physics with his high-school class, found himself interrupted by a voice from the back of classroom, correcting him. One of the experiments the instructor cited had recently been contested - and arguably disproved - by another group of researchers. This voice from the back of the room was a student, wirelessly connected to Wikipedia, reading its article on the subject. Who, at that moment, was teaching whom? With the end of elites in an era of swarming hyperintelligence, culture opens the door to something else, a new form of organization which directly challenges both the authority of elites and the institutions which support them. Nowhere is this more immediately evident than in our educational institutions, which persist in an industrial mode of pedagogy well into the pervasively networked era [16]. Children, who have grown up with pervasive wireless networks as an accepted fact (much as an older generation unquestioningly accepts pervasively available electric current) are placing informational pressures on all the institutions which seek to perpetuate obsolete elites. This pressure is analogous to the informational pressure faced by the mass media when confronting a swarm of peer-producers generating storms of media forwarding. Everywhere, because of the existence of pervasive wireless networks, the informational pressure of culture is steadily increasing, and growing exponentially [9]. Institutions formed during pre-Industrial or Industrial Era systems of knowledge production have fallen out of synchronization with the culture at large, yet they still command the resources and attention of governments and other institutions. Nearly all institutions - political, educational, commercial and cultural - fall into this category. Like the mass media, every institution faces an obsolescence brought on not by its own actions, but because the elite/plebian relationship is no longer enforceable; knowledge can not be withheld, constrained, or protected in any meaningful way [6]. Instead, the entire human universe has been transformed into a palimpsest; that which is written can be re-written, freely, by all. This “read/write culture” was the original promise of the Internet [12], yet it has taken nearly thirty years for this transformation to leave its fingerprints on culture. The autopoeic qualities [10] of pervasively networked human beings emerged only after the network had been in place for some years; those participating in the network collectively learned the “rules of engagement” before hyperintelligence and swarming could appear. However, once those forms appeared, they rapidly disseminated through the network. Every new emergence is quickly shared and copied; any innovation that works for one group is quickly copied by others. Wikipedia, for example, has served as the template for countless other wikis, each attempting to codify the hyperintelligence within a particular flavor of social network. Informational pressure facing institutions resembles a form of natural selection. When institutions resist this selection pressure they only refine the capabilities of

4 The Social Qualities of Pervasive Wireless Networks

83

the swarms seeking to obsolesce it. Since these swarms disseminate and share their refined capabilities, they evolve those capabilities much more rapidly than any institution. Within a space of just two years, broadcasting has transitioned from the dominant force in shaping the mass mind to one (louder) voice among many. This escalating desynchronization between institutions and the constituencies they serve frames the first decades of the 21st century. What, then, is to be done? The pervasive wireless networks are here, and are so beneficial to those who use them - even if one only considers them only in the purely utilitarian context of increased efficiency - that it is unlikely any institutional force, or collection of institutional forces, could overwhelm our increasing need to remain constantly connected, continuously fed by and feeding our networks of hyperintelligence. It is not the intention of this author to be polemical, but rather, to be prescriptive; the paradox of power in the 21st century is that its redistribution, away from institutions and into moieties and micro-polities, both disrupts and enables. Institutions can adapt to both aspects, if they incorporate the ontological lessons which have already become second nature to their constituencies. The best of these institutions will adapt to the selection pressures that confront them with obsolescence; the rest will find themselves trapped in an evolutionary cul-de-sac.

Capability over authority Swarming hyperintelligence has effectively and permanently obsolesced the role of self-appointed elites. Institutions, as the private preserve of elites, can no longer trade on their expertise. The functional difference between an institution and a swarm lies not in expertise, but in capability. An institution commands physical resources well beyond the reach of any swarm; its collection of things has become more valuable than any knowledge which amplifies the value of those physical resources. Swarming hyperintelligence will always know how to deploy those resources more wisely than any institution. For this reason, institutions must transform themselves from loci of expertise to loci of capability. Swarms will decide what is to be done in the real world. Institutions will do it.

Transparency over obscurity Institutions tend to preserve their elite status either by deliberately obscuring their internal operations or by constructing a private language, known only to an institutional elite, to functionally produce the same result. The informational pressure experienced by all institutions in the era of pervasive wireless networks is dammed by this wall of obscurity. The longer the institution persists in its preservationby-obscurity, the stronger this pressure becomes, until, in the worst-case scenario, the institution is swept away in an informational overload. To avoid this situation (while admitting that some institutions which thrive on obscurity will be unable to make such a fundamental adaptation), institutions must listen closely to their constituencies and build the appropriate feedback and feed-forward mechanisms into the institutional structures so that information can seamlessly flow between the institution and the swarms which emerge around it. This is not simply “customer service” or “listening to the voters”; transparency requires an intensive restructuring of the dynamics of the institution. Top-down information flows, to some degree

84

Mark Pesce

necessary within any institution, must be subordinated to the horizontal and rhizomatic information flows which are already permeating the defined boundaries of the institution.

Flexibility over tradition Institutions are established to promote a set of established principles through time; in this way they transcend the influence of any single person - indeed, institutions set themselves up as superior in motivation to the actions of the individual. Principles naturally accrete into traditions; the institution behaves in a particular way because of what it believes to be true. While this allows an institution to be internally self-consistent (a necessary pre-condition for the hierarchical organization of large numbers of individuals), this also means that when conditions outside the institution change, the institution normally reacts to conserve its principles. The inherent conservatism of an institution is a feature essential to its nature, but one which is now wildly at variance with the world within which it acts. This means that institutional responses become increasingly less viable over time; they are being selected against by the environment at large. This represents the paradox at the core of all institutions in the 21st century: to maintain integrity, they stick to their principles, yet this adherence to principle produces a corresponding selection pressure which ends in their disillusion. No human institution has ever proven to be eternal: no government, no religion, no system of commerce, no language, no civilization. Each are eventually overwhelmed by time. This process has heretofore been a gradual one: even the collapse of the Roman Empire played out over millennia. Today, because of pervasive wireless networks, these processes are no longer gradual, but catastrophic. Perhaps the only valid institutional response to this impending catastrophe is to abandon traditional behavior, informed by principle, in favor of flexibility, informed by the information flows passing transparently through the institution. This strikes at the heart of the meaning of the institution, but unless everything is sacrificed, nothing at all will remain.

4.5 Conclusion: Pervasive Wireless Networks and the Rise of Hyperpeople At the start of the 21st century humanity finds itself suddenly empowered with a global voice, and access to a global mind. Like a child given an exciting but complex new toy, we alternate between frustration and enlightenment as we press the buttons, twist the dials and learn that - above all else - this toy connects us to others. The toy itself is ephemeral, and has mutated from the heavy, fixed-point access device of the Internet’s first decades into the portable, pervasive, ubiquitous field of information which constantly bathes us. The device is unimportant - even though this is where much of the commercial activity concerns itself. The connective and collective intelligence - hyperintelligence - for which the device serves as portal has created a new social milieu, a new form of cultural organization, unbounded by proximity, tethered together by lines of salience. Individuals who share similar views about the importance of particular topics of interest now quickly find each other,

4 The Social Qualities of Pervasive Wireless Networks

85

self-organizing into swarms who eagerly share their knowledge for the benefit of the all. This much has already happened. But now, with the widespread deployment of pervasive wireless networks, we bring that swarm with us, everywhere, as an externalized cognitive and sensory presence which increasingly acts to inform us, increasing our capability and efficiency in all human realms: social, cultural, political and economic. As we incorporate the lessons of networked hyperintelligence into our ontological frameworks, these networks have already begun to change the way we behave. We “google” to find information, expecting it come to hand immediately; we scour Wikipedia for an overview of knowledge on a particular subject. As our interests deepen, we reach out, through the networks, searching for those who will gather us into their own networks of expertise, granting us the full benefit of their knowledge, and - if we choose to put in the necessary effort - transforming ourselves into experts. These are old behaviors, wrought in new forms. Where once a tribal community would pool their extensive reservoir of local knowledge to the benefit and increased selection fitness of all, we now work in swarms, networked tribes, sharing knowledge and continually increasing our expertise and effectiveness. We already rely on these networks more than we realize. Individuals who spend a substantial amount of time participating in networked culture have a significant investment in it, and have likely made a large contribution to social networks that lie well outside the antique definition of community. Interaction with these social networks changes the way these individuals think, what they believe, and the language they use to speak of what they know. This transformation is undeniably value-neutral; just as swarms can collaborate to create Wikipedia, they can also collaborate to create Conservapedia, a “Trustworthy Encyclopedia” which presents articles filtered through a politically conservative and evangelical Christian mind set. A child who reads Conservapedia will not grow into the same world-view as a child who uses Wikipedia as a resource, yet both are internally self-consistent. The 21st century will see these two swarms - among many, many others - engage in ontological warfare as they battle toward some absolute, external definition of truth. Partisans in both swarms, continuously connected to their swarming hyperintelligence via pervasive wireless networks, will use their reservoirs of shared knowledge to out-compete the other. What follows - and it is happening right now, in the political and cultural spheres - looks like rational debate. It is not. It is, instead, the continuous testing of each swarming hyperintelligence for selection fitness in an environment of increasing informational pressure. As the selection pressure increases, individuals become ever-more-reliant on their swarming hyperintelligence, until this exteriorized capability undergoes a thorough ontological integration. The boundary between knowledge “out there” (communicated knowledge) and knowledge “in here” (experiential knowledge) collapses, and those individuals effectively exhibit a shared consciousness drawn from their interactions within the swarm. These hyperpeople use swarming hyperintelligence to constantly increase their capability and effectiveness; the more these hyperpeople increase in capability and effectiveness, the more they come to rely upon these qualities, until it becomes inconceivable to do without them. We, who rely on Google and Wikipedia and countless other sources of information and knowledge, who keep a list of contacts in our head, ready to make a call whenever we need some specific expertise, who travel everywhere with a fullycharged mobile in our pocket, are already hyperpeople. We’ve already crossed the

86

Mark Pesce

threshold into swarming hyperintelligence. The signs are already plainly visible: media, commerce, culture, politics - each of them now bears the imprint of the swarm. The swarms have arrived, and we are already bound to them in so many ways that it has become nearly inconceivable to imagine a world without pervasive wireless networks. What we make of them remains an open question. Emergence is not predictable; we must walk the path to know what lies along it [17]. We know this much: because of pervasive wireless networks, the informational pressure is constantly increasing, and this pressure is forcing us, collectively, into a new mode of being.

References 1. Alexander Bard and Jan Soderqvist. Netocracy: The New Power Elite and Life After Capitalism. FT Press, 2002. 2. Howard Bloom. The Lucifer Principle: A Scientific Expedition into the Forces of History. Atlantic Monthly Press, 1997. 3. Michael Crichton. The Terminal Man. Ballantine Books, 1973. 4. Gilles Deleuze and Felix Guattari. A Thousand Plateaus: Capitalism and Schizophrenia. University of Minnesota Press, 1987. 5. Jared Diamond. Guns, Germs, and Steel: The Fates of Human Societies. W. W. Norton, 1997. 6. Cory Doctorow. Microsoft research drm talk. http://www.craphound.com/ msftdrm.txt, June 2004. This talk was originally given to Microsoft’s Research Group and other interested parties from within the company at their Redmond offices on June 17, 2004. 7. George Gilder. Telecosm: How Infinite Bandwidth Will Revolutionize the World. Free Press, 2000. 8. Derrick De Kerckhove. Connected Intelligence. Somerville Press, 1997. 9. Ray Kurzweil. The Singularity is Near: When Humans Transcend Biology. Viking Adult, 2005. 10. Humberto Maturana and Francisco Varela. The Tree of Knowledge: The Biological Roots of Human Understanding. Shambala Press, 1980. 11. Marshal McLuhan. Understanding Media: The Extensions of Man. New American Library, 1964. 12. Theodor Holm Nelson. Literary Machines. Mindful Press, 1982. 13. Mark Pesce. Final amputation. In Proceedings of CyberConf3, the Third International Conference on Cyberspace, 1993. 14. John Robb. Brave New War: The Next Stage of Terrorism and the End of Globalization. John Wiley, 2007. 15. Alvin Toffler. The Third Wave. Bantam Books, 1980. 16. Alvin Toffler and Heidi Toffler. Revolutionary Wealth. Knopf, 2006. 17. Stephen Wolfram. A New Kind of Science. Wolfram Media, 2002.

5 Encouraging Cooperative Interaction among Network Entities Incentives and Challenges

Sonja Buchegger1 and John Chuang2 1 2

Deutsche Telekom Laboratories [email protected] UC Berkeley [email protected]

Summary. The research on cooperation in networks has focused on types of networks that depend on cooperation among the nodes just to enable communication at all: in self-organized networks such as peer-to-peer and mobile ad-hoc networks that have neither central authorities nor infrastructure to allow cooperation-less functioning. This focus is now being extended to networks that potentially have such infrastructure, such as cellular, wireless mesh or sensor networks, because even in such networks, nodes can benefit from cooperation to obtain services or performance gains not provided otherwise. In this chapter, we explain why cooperative interaction does not just occur spontaneously but needs changes in the network environment that allow nodes to benefit from their own cooperation. We then give an overview over such incentives devised to elicit cooperation from network nodes, such as reputation, payment, barter or enforcement systems and mechanism design and discuss challenges arising from the distributed and transient nature of non-fixed networks.

5.1 Benefits of Network Cooperation Networks without central authorities and infrastructure depend on cooperative interaction between the network entities to make communications happen. In multi-hop wireless networks such as mobile ad-hoc networks, nodes can only send messages to other nodes beyond their own wireless range when using intermediate nodes as relays. The same is true for wireless sensor networks where the sink is more than a hop away from the sources. Peer-to-peer file-sharing networks need the cooperation of other participants to forward and return queries and provide data. Without cooperation between network entities such networks cannot function, there is no communication without cooperation. Fixed or centrally operated networks such as cellular networks do not depend on cooperation to such a fundamental extent for basic communication. Yet also in such networks where cooperation is not strictly necessary, it can nevertheless be beneficial to enhance the functionality of the existing network or facilitate the transition to new services and protocols.

87 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 87–108. c 2007 Springer.

88

Sonja Buchegger and John Chuang

We can imagine examples for added value by cooperation on the network layer and above to include the following, as applied to cellular networks and illustrated in Figure 5.1. Extended coverage. Instead of needing to talk to an access point for every communication, multi-hop cellular networks give some control to the end-systems to self-organize and communicate beyond the range of the access point, thereby extending the coverage area spatially. Local exchange. If nodes cooperate to exchange data locally, they can do so without having to have an access point nearby or when it is cheaper to use short-range communication. This extends the coverage area both spatially and temporally. Speed-up of data dissemination. For applications such as broadcasting or file sharing, cooperative down- and upload can speed up the dissemination of data by reducing redundancy, as compared to every end-system communicating separately with the access point. Load balancing. Besides the advantage of accelerated data dissemination, the system load can be both reduced and better balanced by allowing cooperative interaction between network entities. Statistical multiplexing can be used at a higher level of aggregation. The load balancing can be both from access point to end systems as well as from licensed to unlicensed spectrum, for example. In a theoretical analysis of how much cooperation mechanisms can help by increasing the probability of a successful forward in a wireless multi-hop network, Lamparter, Plaggemeier, and Westhoff [24] find that increased cooperation superproportionally increases the performance for small networks (i.e., fairly short routes). Cooperation increases more if the initial probability e (the probability to cooperate by forwarding) is fairly acceptable (above 0.6). Even small increases in e as given by δi, the change of the probability to cooperate in the presence of an incentive mechanism such as a reputation system, can have a dramatic improvement. They find, however, that the benefit is much more pronounced in small networks with fairly short routes than in medium to large scale networks. Networks such as the examples above can benefit substantially from cooperation between entities within the network, or even depend on it. Cooperation between different networks, in terms of peering and routing also has potential to improve outcomes overall. In this chapter we focus on intra-network cooperation rather than inter-network cooperation.

5.2 The Cooperation Dilemma As motivated in the previous section, nodes in self-organized networks need to cooperate in order to communicate. Even in networks with central authorities and infrastructure, nodes can benefit from cooperation to increase the performance they perceive. Cooperation in networks, however, usually comes at a cost in the form of battery consumption, computation cycles, bandwidth or storage. Network nodes can have strategic behavior and are not necessarily obediently cooperating by making their resources available without the prospect of rewards for their good behavior. Unreciprocated, there is no inherent value of cooperation for a node. A lone cooperating node draws no benefit from its cooperation, even if the rest of the

5 Encouraging Cooperative Interaction

(a) Cellular Network

(b) Extended Coverage

(c) Local Exchange

(d) Data Dissemination

89

(e) Load Balancing Figure 5.1. Examples of cooperation benefit.

network does. Guaranteed cost paired with uncertainty or even lack of any resulting benefit does not induce cooperation in a rational, utility-maximizing node. Rational nodes therefore would not cooperate in such an environment. A network of rational nodes would thus not cooperate and all be worse off than if they cooperated. Not to cooperate is the dominant strategy, regardless of what other nodes choose to do. The collection of individually chosen, dominant, strategies leads to an outcome where everyone is worse off than had they chosen a different strategy. This is a dilemma. Ostrom [29] discusses the role of trust in a dilemma: “Because the less valued payoff is at an equilibrium, no one is independently motivated to change their choice, given the choice of other participants. These situations are considered to be dilemmas because at least one outcome exists that would yield higher returns for all participants. To get to this outcome, however, individuals have to trust one another.

90

Sonja Buchegger and John Chuang

Rational participants making independent choices are not predicted to realize this Pareto-optimal outcome. A conflict is thereby posed between acting from individual rationality and gaining sufficient trust to achieve the optimal outcomes for a group. The problem of collective action raised by social dilemmas is finding a way to avoid Pareto-inferior equilibria and to move closer to the optimum. Those who find ways to coordinate strategies in some fashion receive a ‘cooperator’s dividend’ equal to the difference between the worst outcome and the outcome achieved.” Dilemmas have been studied in Game Theory [28], the best-known example being the prisoner’s dilemma, where the dilemma lies in the prisoner’s choice between defecting, i.e., confessing and telling on his partner in crime to escape the harshest sentence and thereby, possibly, if the the other prisoner remains silent, getting away with the shortest sentence (the dominant strategy) and cooperating with the other prisoner, i.e., remaining silent when questioned, the consequences of which are the longest sentence if the other prisoner confesses or a shorter sentence for both otherwise. When both prisoners cooperate they are both better off than when they defect. Regardless of what the other prisoner chooses, each prisoner has to choose to defect to minimize his own sentence, this is the Nash equilibrium. This is the case at least if the choice arises only once, when it is a one-shot game. The situation changes when a dilemma arises repeatedly, as in the iterated version of the prisoner’s dilemma, where strategies like tit-for-tat (TFT), i.e., initiating with cooperation and continuing with direct reciprocity, are successful at eliciting cooperation and show a way out of the dilemma, avoiding the attraction of detrimental equilibria of dominating strategies of the one-shot game [2]. Table 5.1 shows the payoff a player gets at each combination of actions, from player A’s perspective. For example, if player A defects and player B cooperates, player A gets the so-called temptation payoff. A game is a dilemma when the following relation for payoffs hold: Temptation > Cooperation > Mutual Defection > Sucker’s Payoff.

Table 5.1. Payoff Matrix: Player A (vertical reading) vs. Player B (horizontal reading). PLAYER A/B COOPERATE DEFECT COOPERATE mutual cooperation sucker’s payoff DEFECT temptation punishment

Applying insights from well-studied dilemmas such as the prisoner’s dilemma to communication systems, it becomes clear that encouraging cooperative interaction between network nodes, the players in the communication game, can lead to an outcome preferred by all nodes and increase the overall performance of a network. Several incentives for cooperation have been suggested [23]. Yet it is not straightforward to do that. In networks, the decisions are more complicated than a simple binary cooperate/defect resulting in a richer strategy space, as noted in [3]: “Peers make strategic decisions concerning the revelation of private information, such as local resource availability, workload, contribution cost, or willingness to pay. Peers decide on the amount of exerted effort given the nonobservability of their hidden actions. Peers may adjust their spatial engagement

5 Encouraging Cooperative Interaction

91

with the network through strategic network formation, and temporal engagement through strategic churning (arrivals and departures). Finally, peers may choose to manage their own identities and treat the identities of others differently given the availability of cheap pseudonyms.” In summary, in a dilemma, there is a fundamental tension between individual rationality and collective welfare. The goal of incentive mechanisms is to align these and make cooperation pay off and thus to establish rules of the game to reduce or eliminate the gap between social optimum and worst-case Nash equilibrium. The ratio of the social optimum to the worst-case Nash equilibrium is called the price of anarchy. If we take into account strategic behavior when designing networks, we can encourage cooperative interaction between network entities and reduce the price of anarchy. We next look at solution concepts that have been proposed to that end.

5.3 Solution Approaches Several approaches to elicit cooperation between network nodes have been proposed in the last few years. They can be broadly classified into reputation, payment, barter, and rules-based enforcement systems.

5.3.1 Reputation Systems Reputation systems can provide a way of distinguishing cooperative from uncooperative entities, to enable an informed decision about which entities to choose from for future cooperation and which ones to avoid. They provide a basis for risk assessment of different types of behavior toward other entities. In the context of networks, these entities are nodes, but reputation systems have been used and developed for a wide range of applications, such as buyers and sellers in in online auctioning sites such as eBay, reader feedback and recommendations for book sites such as Amazon, restaurant critiques, and entities in networked systems such as peer-to-peer file-sharing, mobile ad-hoc or sensor networks, and providers in wireless networks. The basic premise for reputation systems is that one can predict future behavior by looking at past behavior. We could call this “the shadow of the past” in analogy to “the shadow of the future” [2] which describes an increase in willingness to cooperate when future interactions are anticipated. Reputation systems combine both the shadow of the past and the future to elicit cooperation. The shadow of the past here means that past behavior has consequences for the present standing of a network entity and how it will be treated in the future, independent from whether future interaction will be with the same entities. The shadow of the future refers to the expected repeated interaction with the same entity. Combined, these two shadows provide an incentive for cooperation, in that they potentially elicit cooperative behavior from other network entities. Reputation systems provide a mechanism of service differentiation [3], where network entities are treated differently depending on their reputation rating. The difference in treatment can be binary, resulting in the effective exclusion of network entities perceived as uncooperative [4] or provide flexible service-level agreements [21]. In general, reputation systems work as shown in Figure 5.2 and explained as follows. First, information about the behavior about other network nodes is gathered. This information consists of own observations, reports from third parties or a

92

Sonja Buchegger and John Chuang

combination thereof. The gathered information is then transformed into a reputation rating, by way of aggregation, weighting or other functions. There are several factors that can be considered for a function that generates reputation ratings. One can look at the absolute number of cooperative interactions, the number of defects, the ratio of cooperate to defect decisions, or any combination and weighting of these factors.

Figure 5.2. Reputation System. Nodes evaluate the reputation a another node before allowing a transaction.

The reputation rating is either global, i.e., everyone sees the same rating of an entity, or local, where nodes do not have to agree on a single value of a reputation rating for another entity. These reputation ratings can be stored at a central location, if possible, or in a distributed way otherwise. Taking the reputation rating as a basis, network entities are then classified according to their eligibility for future interaction. The classification can consist of simple thresholds which in turn can depend on the network situation and the distribution of network entities to choose from for cooperation. Reputation systems apply to a broader range of desired behavior as long as it is observable and classifiable. They can, if they use second-hand information and have means to cope with false accusations or false praise, partially prevent misbehavior by excluding misbehaved nodes [5]. This way, nodes can protect themselves before encountering the misbehaved node. If the reputation systems rely exclusively on first-hand experience to build reputation ratings, they can only prevent more of the misbehavior experienced by a node after it occurred. Reputation systems that do use second-hand information provide a means for indirect reciprocity via third party observations in addition to direct reciprocity provided by one’s own observations. The reputation rating provides not only an abstract notion of the “goodness” of a network entity, but a network entity can derive several kinds of more tangible values from its own reputation rating: The operating value denotes the present value of the expected future increase in profits thanks to reputation, i.e., the advantages

5 Encouraging Cooperative Interaction

93

gained for eliciting cooperation from others. The throw-away value measures the potential profit the network entity can make by cheating, capitalizing on its previous reputation. The replacement cost is interesting to so-called whitewashers (see Section 5.4.3) and means expected cost of recreating reputation after having discarded the reputation. The values listed above are derived from a good reputation, which in turn depends on the degree of cooperation perceived. It it therefore in the interest of network entities to be perceived as cooperative to get a good reputation. Barring easily duped reputation systems, there is an incentive to actually cooperate.

5.3.2 Payment Systems

(a) Direct Payment.

(b) Indirect Payment.

Figure 5.3. Payment System. With and without banks for billing and money transfer. Payment systems, e.g., Sprite [35], serve as an incentive to provide a well-defined service, such as packet forwarding, to others for remuneration. The payment, in the form of a currency or token, can then be used to pay for one’s own service, e.g., for forwarding one’s own traffic, using other network entities as relays. The payment can be direct or via a central authority that serves as a bank, as shown in Figure 5.3. Some payment schemes offer a fixed price per service, e.g., one unit of payment per hop forwarded [7, 8], others use variable pricing depending on functions of cost such as bandwidth or power [1, 13]. The payment has to be unforgeable. To ensure this, tamper-proof hardware and trusted third parties have been suggested. With payment systems, the issue of pricing and other economic questions, such as how to deal with lost payment, arise. Since they assume rationality, payment systems can prevent selfish uncooperative behavior, however, they do not address malicious or faulty behavior. Mechanism design [14], [17] provides ways of eliciting truthful revelation of cost and enable efficient pricing for payment systems. This is important for heterogeneous networks, where simple tit-for-tat meets with different capabilities of the participating entities. Mechanism design can help to make cooperation incentive-compatible by making it favorable for network entities to truthfully report their local resources

94

Sonja Buchegger and John Chuang

and cost. If truth revelation is a dominant strategy, i.e., each agent has a bestresponse strategy no matter what strategy the other agents select, the mechanism is strategyproof.

5.3.3 Barter Systems Systems such as Bittorrent [12] for peer-to-peer file-sharing build on direct reciprocity among network entities and form a barter system as shown in Figure 5.4. Bits and pieces of files are exchanged directly following a TFT strategy that has been shown to be successful in similar dilemmas such as the iterated prisoner’s dilemma, and the interaction is split into a large number of smaller interactions. This enables the system to escape some of the pitfalls in one-shot games or interaction with strangers in peer-to-peer system. As the number of interactions increases, the so-called shadow of the future is longer and nodes cannot afford to defect in view of the many interactions with the same network entity that are needed to complete the ongoing transaction. This way, even with an increased network size, interactions have consequences and therefore elicit cooperation.

Figure 5.4. Barter System. Nodes exchange services directly.

5.3.4 Enforcement Systems In the previous sections we have discussed incentive mechanisms for cooperation. Inherent in the expression is that they provide an incentive to behave in a cooperative way, so rational network entities would choose to cooperate as the incentive mechanism is such that it is advantageous for them to do so in order to maximize their utility. Incentives thus encourage cooperative behavior on a voluntary basis, but do not enforce anything. In contrast to this model of rational entities, there can be network entities with bounded rationality or even seemingly irrational behaviors when

5 Encouraging Cooperative Interaction

95

Figure 5.5. Enforcement System. Nodes check the validity of a service before allowing a transaction.

considering utility functions. This is the case for malicious uncooperative behavior, or simply misbehavior. In the case of malicious intent, incentives for cooperation are not enough, but cooperation has to be enforced. Enforcement can be done by setting rules of behavior, that if broken, nullify the transaction. For example, several secure routing protocols for mobile ad-hoc networks have been suggested, such as [20, 30] that only return routes to the requesting source, if all the nodes on the route comply with their requirements and that has been verified, e.g., by checking the validity of hash chains, detecting tampering with message headers, as shown in Figure 5.5. A method for thwarting attacks is prevention. According to Schneier [32], a prevention-only strategy only works if the prevention mechanisms are perfect; otherwise, someone will find out how to get around them. Most of the attacks and vulnerabilities have been the result of bypassing prevention mechanisms. Given this reality, detection and response are essential. Secure protocols prevent preconceived deviations from specific protocol functions. They do, however, not aim at serving as incentives for cooperation or dealing with novel types of misbehavior that occur by going around the protected functions. Preventive schemes can only protect what they set out to protect from the start. There can, however, be unanticipated attacks that circumvent the prevention. It is vital that this misbehavior be detected and prevented from happening again in the future. Self-policing schemes are only as limited as their intrusion detection component regarding detected attacks. The schemes themselves are flexible and can accommodate an evolving intrusion detection component. If the detection of a new attack is conceived of, the detection component can be changed to reflect this added knowledge. This does not in any way change the protocol. If a preventive scheme needs to be extended to accommodate the advent of a new attack, a new version of the routing protocol is required. Yet, prevention or enforcement mechanisms can be a complement to incentive mechanisms.

96

Sonja Buchegger and John Chuang

The question of a tamper-proof security module remains controversial [31], but might prove inevitable to implement enforcement mechanisms and protection such as authentication.

5.4 Challenges The goal of incentives for cooperation is to align the individual strategic behavior with the collective welfare of all network entities. When devising such incentives, networks, and in particular wireless networks pose several challenges due to the transient nature of the medium and communication system.

5.4.1 Selfish v. Malicious v. Faulty Behavior Lack of cooperation can arise from different reasons besides rational utilitymaximization by selfish entities. Devices could just be faulty and act erratically or the behavior could stem from maximizing a different utility function, namely the goal of denying service to another node, disrupting the whole network, or other malicious intents. Designing mechanisms to encourage cooperation among network entities that are, at least seemingly, not rational is by definition quite difficult if not infeasible. If behavior modification cannot be elicited, the next best thing is to recognize its occurrence and reign in the consequences of uncooperative behavior. Economic systems assume a rational node that aims at maximizing its utility expressed in energy savings or payment units. The node misbehavior targeted by payment systems is thus selfish concerning utility but it is not malicious. A malicious node is not necessarily aiming at a economizing on its resources. Its interest lies in mounting attacks on others. Secure routing protocols aim at preventing malicious nodes from mounting attacks. Although some reactive systems focus on selfish (watchdog [25]) or malicious misbehavior (intrusion detection [34]), this is not an intrinsic limitation. Self-policing networks by reputation systems enables coping with both selfish and malicious, and, in addition, with non intentional faulty misbehavior, the only requirement being that such misbehavior be detectable, i.e., observable and classifiable. We deem the consideration of non intentional misbehavior such as bugs of high importance, and we think it is vital to protect the network against misbehaved nodes regardless the nature of their intentions. Non intentional misbehavior can result from a node being unable to perform correctly due to a lack of resources, due to its particular location in the network, or simply because of the node being faulty. Self-policing misbehavior detection, reputation, and response systems can be applied irrespective of the actual cause of the misbehavior, be it intentional or not. When a node is classified as misbehaved it simply means that the node performs badly at routing or forwarding. No moral judgment is implied.

5.4.2 Observability Incentives for cooperation benefit from, or even depend on, the ability to identify cooperative behavior and distinguish it from uncooperative behavior exhibited by network entities.This ability is needed to rate the reputation of a node, to prove

5 Encouraging Cooperative Interaction

97

the basis for payment or that a contract has been fulfilled, that a service has been rendered that merits reciprocity, or to detect breaches of or conformity with rules of the game, such as security. In networks, however, this transparency is not always achievable, due to both hidden action and hidden information. The information might be asymmetric. In economics, information asymmetry occurs when one party to a transaction has more or better information than the other party. In a network, it is typically the service provider that knows more than the service consumer, e.g., a node requested to relay packets for another knows the local conditions of the network and its own cost better than the initiator of the communication. Information asymmetry models assume that at least one party to a transaction has relevant information whereas the other does not. Some asymmetric information models can also be used in situations where at least one party can enforce, or effectively retaliate for breaches of, certain parts of an agreement whereas the other cannot. These models are studied in Principal-Agent Theory. In adverse selection models the party with less information does not know about the properties of the other party and thus selects the wrong partner for a transaction. In moral hazard the party with more information can exploit the situation by acting differently than expected, the party with less information cannot determine that this is the case and cannot retaliate for a breach of the agreement.

Hidden Action In a wireless medium, intentional packet drops are hard to distinguish from congestion or link breaks; mobility can look like misbehavior, and the wireless range limits the observability of nodes further along the path. Intentional, strategic behavior can seem like the result of network conditions and vice versa. Figure 5.6 shows how the limited wireless range lets nodes hide misbehavior, e.g., dropping packets.

Figure 5.6. Hidden Action. The source of a packet cannot see the misbehavior along the path beyond its own range.

98

Sonja Buchegger and John Chuang

Hidden Information It is not obvious from the outside what properties nodes have, what utility function they have, how rational they are, what there cost for cooperation is and what capabilities they have in terms of battery, computational power, storage or bandwidth. The collection of these properties along with their utility function constitute the type of a network entity, which is not necessarily revealed to other network entities. Furthermore, as shown in Figure 5.7, nodes can lie about their properties, i.e., their type.

Figure 5.7. Hidden Information. A node does not reveal its true resources.

5.4.3 Identity After establishing what a network entity has done, it is important for incentives for cooperation to work, to find out who did it and whether the node who seems to have behaved in a certain way actually is the actor in question. Reputation ratings have to be assigned to the correct network entity, payments or reciprocal behavior returned to the correct initiator and consequences of security enforcement need to be born by the node involved in the behavior. There are several obstacles to the precise mapping between behavior and identity.

Shadow of the Future From analyzing good strategies for the iterated prisoner’s dilemma we know that the anticipation of repeated interaction and opportunities for cooperation in the future, the so-called shadow of the future [2], lead nodes to increase cooperation for fear of retaliation or hope of reciprocal cooperation. To cast a shadow of the future, an identity must persist long enough to allow for repeated interaction. Short-lived

5 Encouraging Cooperative Interaction

99

identities reduce the cooperation dilemma to a one-shot game, where cooperation is unlikely the dominant strategy for network entities. Identity persistence is precluded by a node’s ability to change its identity either while in the network or by leaving and entering the network again. Identities that persist over a long term cast a longer shadow of the future and thus can offset some of the transient nature of networks. This transient nature arises from several factors including short-lived transactions, a large network population, a volatile medium, user mobility, or any combination thereof. The impact of the shadow of the future can be increased not only by longer persistence of identities to enable repeated transactions, but also by increasing the frequency of interaction during a given time period. Interactions can be split into several sub-interactions that each require a modicum of cooperation. This way, reciprocity is immediate and cooperation incentives are strengthened even under conditions that result in short games. Immediate reciprocity simultaneously addresses another factor that typically shortens the shadow of the future. Even if the identity persists long enough and the nodes have repeated interactions, it is not a given that the nodes take turns as in an idealized game, where one player moves after the other. In networks, a node does not automatically take turns being a client or a server, a sender or an intermediate node or a destination. Finding ways to reciprocate directly and immediately provides proper role reversals.

Sybil Attacks Nodes can claim to have more than one identity at the same time and pretend to be several nodes. The requirement of distinct identities is the target of the so-called Sybil attack analyzed by Douceur [16], where nodes generate several identities for themselves to be used at the same time. This property does not so much concern a reputation system, since those identities that exhibit misbehavior will be excluded, while other identities stemming from the same node will remain in the network as long as they behave well. The Sybil attack can, however, influence public opinion by having a network entity’s rating considered more than once. To prevent the Sybil attack, impersonation, and guaranteeing minimum identity persistence, nodes could be required to register with a certification and pseudonym authority that does not hand out more than one identity to a node at a time and requires a minimum time to have elapsed before changing an identity. In the scenario where the network is not completely cut off the Internet, one can make use of certification authorities. An example for such a scenario are publicly accessible wireless LANs with Internet connection. The detection and isolation of uncooperative and misbehaved nodes as achieved by a distributed reputation system are still necessary, even in the presence of network operators. For the case of a pure ad-hoc network without Internet connectivity or secure hardware, Weimerskirch and Westhoff [33] propose zero-common knowledge authentication which provides recognition of nodes that have been dealt with before, without requiring geographical proximity. When the infrastructure allows, solutions based on public key infrastructure are possible. There are also reputation systems that attempt to deal with the sybil attack directly, e.g., [9]. The sybil attack can also affect payment systems, for example in the case when newcomers are endowed with a specified amount of currency, or when it seems

100

Sonja Buchegger and John Chuang

beneficial for a node to pretend to be several nodes to gather more requests for service, thereby gaining money-earning opportunities.

Whitewashing To get rid of disadvantages accumulated under an identity, such as a bad reputation or payment debt, or to reap the benefits of a system that encourages node participation and rewards nodes new to the system, nodes may decide to strategically leave and reenter the network, and thereby not only contribute to overall network churn and instability but decreasing the efficacy of incentives for cooperation in the network. Whitewashers are indistinguishable from newcomers if identities are not persistent over a longer time, so whitewashers repeatedly get the benefits of a blank slate without being detected. One way of preventing whitewashing attacks [18] is to have new nodes acquire a good standing in the network by cooperating for a while before they are granted much benefit from other cooperating nodes until they have been in the network long enough to receive the same privileges as regular nodes in the network. While this decreases the incentive of leaving the system to start anew, it also decreases the incentive to participate in the system at all, especially for purposes of small transactions.

Anonymity and Pseudonymity There is a tradeoff between anonymity and the privacy it provides to the user of a network device, and the needs of a stable identity imposed by reliable mechanisms for cooperation and non-repudiation Identity persistence, even if linked to a pseudonym and not a straightforward identity of the user and her device, nevertheless provides continued data on a specific network entity and can enable location tracing, inferences from who talks to whom, and other correlations of identity and access services or information. When this information is aggregated and used to link the pseudonym to the true identity, anonymity is gone. Even when only linking it to a pseudonym, privacy may be compromised.

5.4.4 Fairness Going back as far as Plato, there are four principles for distributive justice [27], that are incompatible yet equally valid depending on the situation. They are exogenous rights, compensation (for other inequities already inflicted), reward (for good behavior or effort), and fitness (i.e., who can make the best use of a resource). In networks, we distribute resources such as bandwidth or payoffs gained from incentives for cooperation, and it is not straightforward to decide how such resources should be distributed. Self-organized networks depend on cooperation, yet selfish behavior leads to noncooperation and thus potentially to a non-functional network which does not benefit even selfish rational nodes. To address this dilemma, incentives for cooperation (such as payment and reputation) have been proposed. These solutions, however, have not taken into account the different opportunities for cooperation of nodes due to their location. The same behavior leads to payoff inequities in different locations.

5 Encouraging Cooperative Interaction

101

Incentives for cooperation aim at rewarding cooperative behavior or punishing non-cooperative behavior. Yet as a consequence of neglecting cooperation opportunities (e.g., due to node location), the desirable and the actually rewarded behavior might not be the same. Nodes lacking opportunities for cooperation can starve when their payoff depends solely on them exhibiting cooperative behavior. When there is no opportunity for cooperation, rewards for cooperation cannot be collected. In the extreme case, nodes can have negative payoff even if they are willing to cooperate. As an example, if nodes remain at the edge of the network that has a payment system in place where nodes are paid for each packet they forward and have to pay other nodes to forward the traffic they generate, they can cease to afford their own traffic and starve if monetary resources are limited. Since such a node might never be called to forward for others as it does not lie on any path for other nodes, it cannot afford to send any traffic despite its willingness to cooperate. With reputation systems that punishes a denial of cooperation, nodes in the center can get such a bad reputation that they become isolated from the network, if they experience network congestion. These cases show that both the willingness and the opportunity to cooperate need to be taken into account for incentive mechanisms. Another consequence of payoffs from incentive mechanisms varying by location and not only behavior is that it enables nodes with favorable locations in terms of cooperation opportunities to erect entry barriers for new nodes or keeping competitors at the outside. As an example, peering agreements can be denied to new Internet Service Providers (ISPs) to limit their position in the logical network to a location with less payoff. Taking into account location information and keeping in mind payoff inequities of incentive mechanisms that depend on cooperation opportunities as well as cooperative behavior raise the need for effective evaluation models to uncover such inequities. Uniformity assumptions about location can mask such inequities and thus bring about misleading evaluations of the effects of incentive mechanisms for cooperation in networks. As an example, incentive schemes for self-organized networks such mobile ad-hoc network often implicitly assume that nodes will move about enough so that they, over time, do not spend dis-proportionally much time at either the center or the edge of the network. It is common practice to evaluate the performance of incentive schemes in mobile ad-hoc networks using the random way-point mobility model. More realistic mobility models, such as those based on trajectory traces of people on a campus, for instance, do not show this uniform distribution. Payoff inequities due to location therefore may lead to starvation scenarios as described above. Even without incentive schemes, there are inequities in performance metrics attributable to location [10] that can also lead to strategic network formation [11]. Incentive schemes networks aim at fostering cooperative behavior by rewarding it and withholding rewards for (or penalizing) selfish non-cooperative behavior. Another dimension of cooperative behavior is fairness: How often do nodes have the opportunity to exhibit cooperative behavior and are therefore eligible for rewards or penalties? Depending on whether the incentive mechanism in place emphasizes rewards of cooperative behavior or penalties for non-cooperative behavior, the number of cooperation opportunities relative to other nodes have different consequences. If rewards for cooperative behavior are emphasized, a node benefits from having many opportunities for cooperation, e.g., if a forwarding action is remunerated, the opportunity

102

Sonja Buchegger and John Chuang

to earn money is proportional to the load. Conversely, if discouraging of noncooperative behavior is emphasized, a node benefits from having fewer instances of behavior opportunities that demand cooperation, e.g., a node will not get a bad reputation for denying forwarding when it is never asked to cooperate. If we take the example of packet forwarding, the opportunities for cooperation are more numerous in the center than at the edge of a network.

5.4.5 Meta Cooperation When the incentives for cooperation themselves depend on the cooperation of the entities by participating and contributing, a cooperation dilemma can arise at this meta level of cooperation, for instance participating in a reputation system by evaluating other entities and publishing corresponding reputation ratings. Devising incentives for collaboration in an incentive system only defers the problem to the next higher level of cooperation. Ideally, the participation in an incentive system brings immediate benefit. When this alignment is not possible, as is the case in a dilemma, one can at the minimum contain the consequences of such noncooperation, ranging from free-riding to more harmful manipulation of the incentive system.

Telling the Truth What is the incentive to tell the truth when rating the reputation of another network entity? The benefit of doing so is not obvious, whereas there are immediate benefits for lying for a node that wants to manipulate the reputation for another node: it can lead to having better access to a service by denying it to others, weakening competitors, loading forwarding work off to others, attracting or deviating traffic, etc. There are several strategic behaviors that lead to spurious ratings: Network entities can deliberately worsen another node’s reputation by false accusations reflected in a bad reputation (blackmailing). Nodes can boost each other’s reputation by giving good feedback repeatedly (ballot stuffing), or collude against a third entity. Besides such deliberate lying for reasons outside the transaction they are evaluating, nodes can give spurious reputation ratings out of retaliation or simply reciprocity when they have been given a false rating themselves. False ratings can also occur for fear of such reciprocity, anticipating retaliation for giving a negative rating, even if that accurately reflects their experience. Several solutions have been proposed to counter such spurious ratings, by allowing only positive ratings for instance, thereby eliminating at least the blackmailing attacks [26], or by detecting liars and discounting their opinion [6], limiting both types of spurious rating: too good and too bad ones. Telling the truth about their cost is only beneficial to nodes when the right incentive mechanisms are in place. Truth revelation is therefore one of the main goals for mechanism design.

Participation We have discussed in earlier sections that network entities need to have an incentive to cooperate. Yet also on a meta level, what is the reason for a node to cooperate

5 Encouraging Cooperative Interaction

103

in the incentive mechanism, to provide reputation ratings for other nodes, exerting effort in observing their behavior? There needs to be a benefit nodes can gain from participating, otherwise there is a reluctance to participate as encountered for cooperation in the underlying network protocol. This reluctance is exacerbated when there is negative reputation and a therefore a fear of retaliation. Another hurdle to cooperation in the incentive mechanism itself such as a detection and reputation system in mobile ad-hoc networks is that when a node enforces compliance with an incentive system by, for example refusing to forward messages for a node with a bad reputation, this can be misinterpreted as a lack of cooperation itself, even though it is not only conforming to the protocol but also helping the efficiency of the incentive mechanism by punishing uncooperative behavior of other nodes, thereby increasing the quality of the network performance. Applying insights from sociology, where it was found that in human behavior, observed punishing misbehavior (at a cost) is seen as cooperation and increases the willingness to cooperate, incentives need to be aligned accordingly. To entice participation in the layer of the incentive mechanism, it has been proposed to provide side-payments for honest feedback in reputation systems [22]. To determine whether feedback is honest is not straightforward due to its subjectivity and behavior dynamics, but by using incentive-compatible mechanisms that provide an incentive for truthful revelation it can be encouraged.

Banking Payment systems need a way of billing the service consumer getting the payment to the service provider. Most research on payment systems for cooperation in communication networks assumes that the micro-payments can be made somehow and focus on pricing issues. In principle, there are two types of banking systems envisioned: direct payments in a distributed system by means of tamper-proof hardware (e.g., [7]) and central authorities providing a banking infrastructure (e.g., [35]). Neither of these two might be realistic, depending on the environment. While cellular networks already have billing and payment infrastructures in place, it is not straightforward to include facilities for micro-payments between network entities and not only between end-systems and central service providers.

5.4.6 Time Many analyses of cooperation emphasize what happens at equilibrium, yet some challenges occur in transient phases, or due to the mere passing of time. The following examples illustrate such transient behavior.

Money Supply With payment systems for packet forwarding, it is usually the initiating nodes who pay intermediate nodes for relaying packets. The price is based on some definition of accumulated forwarding cost, be it per hop or taking into account different cost at the intermediate nodes. Budget balance here means that the amount of money (or money equivalents) circulating in the system remains constant and no

104

Sonja Buchegger and John Chuang

money has to be inserted from the outside of the network. It is a goal of mechanism design to achieve budget balance along with other criteria such as efficiency, strategy-proofness, individual-rationality, and pareto-optimality, yet all goals cannot be achieved simultaneously. Using VCG auctions for forwarding prices achieve truthful revelations but not budget balance. In Ad-hoc VCG [1], for instance, the source has to pay the price even if it is exceeding the sum of cost for the intermediate forwarding nodes (the source was chosen and not the destination to prevent the case when recipients of spam would have to pay for it). When payments are added to packets for forwarding in systems like nuglets [7] (much like stamps on envelopes), in the so called packet-purse-model, nuglets are lost when the actual route taken turns out to be shorter than estimated (additional payment is not used and lost) or longer (packets do not reach the destination, have to be re-sent) or when packets are dropped. If the money supply in the network is suboptimal, a payment system cannot work effectively. If the supply is too small, not everyone can afford to send their own traffic. Conversely, if it is too large, network entities have a reduced incentive to behave in a way that earns them money, since they already have enough. An economy created by a payment system can suffer from the same inflictions that real economies deal with, such as inflation or crashes. A system to control incentive efficiency (social welfare) by adapting the money supply in the network was proposed in [19]. Newcomers are assumed to have no money, and the price of service is adjusted (equivalent to adjusting the money supply) to maintain the ratio that maximizes efficiency.

Inconsistent Behavior The basic premise of a reputation system is that one can predict future behavior by looking at past behavior. This does not hold for all cases, since there can be erratic behavior that is completely inconsistent with past behavior, as in the case of sudden failure, for instance, but the assumption is that such cases are the exception and not the norm and that past behavior can be used as a basis for the prediction of future behavior. To provide this basis, the reputation system has to keep track of past behavior. This can be done in several ways. Here are some decision points to guide the design process of a reputation system. What information is kept? About whom? Where? For how long? When is information added? How is information from others considered? How is it integrated? What does this information look like over time? What has to happen to change this information? In summary, a reputation system needs a way of keeping information about the entity of interest, of updating it and of incorporating the information about that entity obtained from others. Also it has to keep that relevant as time passes. This provides the basis of decision making. Then the decision making itself has to take place to allow nodes to chose other nodes for cooperation. Humans can look at graphical representations of reputation such as the number and color of stars on eBay, and glance over some qualitative information given in feedback comments. In self-organized networks, we want the reputation system to

5 Encouraging Cooperative Interaction

105

not only be able to present information about reputation to the user, but to make automatic and autonomous decisions. For example, the system can estimate the most likely behavior in the future [15], but that does not automatically lead to a decision. The reputation system therefore has to have a mechanism to make decision and classifications. These decisions can depend on the network conditions and should be adaptive to the situation encountered, such as congestion, availability of alternate routes, number and perceived quality of service providers, importance and urgency of the service, etc. As time passes, the importance of parts of the reputation data collected can change. For instance, recent steady behavior is probably a better predictor of future behavior than behavior observed a long time ago. On the other hand, looking only at the most recent behavior can yield a distorted picture of past behavior as one instance observed is not enough to measure a trend. Reputation systems need to have a way of factoring in time in a reasonable manner that would either conform to the user’s expectation or be proven to work well in the system environment. Giving more weight to recent behavior and discounting past behavior as time passes achieves two objectives: better correlation to future behavior and allowing for node redemption: When past behavior is discounted, nodes cannot capitalize on previous good behavior but have to consistently behave well to maintain a good reputation. Information about nodes has to be constantly reinforced to stay current. Node redemption allows for a node to regain at least a neutral reputation after a specified time period (determined by the discount rate) without bad behavior. This is crucial for example for dealing with formerly faulty nodes that have been repaired and useful in general to adapt to behavior changes of nodes regardless of the reason.

5.5 Conclusions Encouraging cooperative interaction among network entities is a task that is not only interesting and necessary in self-organized networks such as mobile ad-hoc and peerto-peer networks, where cooperation is needed for basic network functionality. It also gains importance in infrastructure-based networks to extend their functionality or even enable new services. Typically, cooperation does not come naturally and automatically in networks, as there are costs resulting from cooperation and often their is no immediate benefit to offset, let alone exceed those costs. Network conditions are such that they embody in a modern form the Tragedy of the Commons of days past, when farmers would let as many as possible of their sheep and cattle graze on the commons, thereby rendering the commons unsuitable for grazing of all, such that the farmers were worse off than had they restricted the grazing of their animals. The tragedy lies in that everyone maximizes their own utility at the expense of the overall utility of everyone. Analogies to commons in networks are unlicensed spectrum and bandwidth. Since cooperation does not come for free, incentives for cooperation are being devised, in the form of reputation or payment system, barter or enforcement, and in some cases, combinations of these systems. In this chapter we gave an introduction to these approaches to elicit cooperation and identified challenges that still need to be overcome in order to make those incentives effective, efficient, and robust.

106

Sonja Buchegger and John Chuang

References 1. Luzi Anderegg and Stephan Eidenbenz. Ad hoc-vcg: a truthful and cost-efficient routing protocol for mobile ad hoc networks with selfish agents. In MobiCom ’03: Proceedings of the 9th annual international conference on Mobile computing and networking, pages 245–259, New York, NY, USA, 2003. ACM Press. 2. Robert Axelrod. The Evolution Of Cooperation. Basic Books, 1984. 3. Moshe Babaioff, Michal Feldman, John Chuang, and Michal Feldman. Chapter on: Incentives in peer-to-peer systems. In Algorithmic Game Theory, to appear 2007. 4. Sonja Buchegger and Jean-Yves Le Boudec. Performance Analysis of the CONFIDANT Protocol: Cooperation Of Nodes — Fairness In Dynamic Ad-hoc NeTworks. In Proceedings of IEEE/ACM Symposium on Mobile Ad Hoc Networking and Computing (MobiHOC), Lausanne, CH, June 2002. 5. Sonja Buchegger and Jean-Yves Le Boudec. The effect of rumor spreading in reputation systems in mobile ad-hoc networks. Wiopt’03, Sofia-Antipolis, March 2003. 6. Sonja Buchegger and Jean-Yves Le Boudec. A robust reputation system for peerto-peer and mobile ad-hoc networks. P2PEcon, Harvard University, Cambridge, MA, USA, June 2004. 7. Levente Butty´ an and Jean-Pierre Hubaux. Enforcing service availability in mobile ad-hoc wans. In Proceedings of IEEE/ACM Workshop on Mobile Ad Hoc Networking and Computing (MobiHOC), Boston, MA, USA, August 2000. 8. Levente Butty´ an and Jean-Pierre Hubaux. Stimulating cooperation in selforganizing mobile ad hoc networks. ACM/Kluwer Mobile Networks and Applications, 8(5), October 2003. 9. Alice Cheng and Eric Friedman. Sybilproof reputation mechanisms. In P2PECON’05: Proceeding of the 2005 ACM SIGCOMM workshop on Economics of peer-to-peer systems, pages 128–132, New York, NY, USA, 2005. ACM Press. 10. N. Christin and J. Chuang. A cost-based analysis of overlay routing geometries. In Proceedings of IEEE INFOCOM’05, Miami, FL, March 2005. 11. Byung-Gon Chun, Rodrigo Fonseca, Ion Stoica, and John Kubiatowicz. Characterizing selfishly constructed overlay routing networks. In INFOCOM, 2004. 12. Bram Cohen. Incentives build robustness in bittorrent. In Proceedings of the First Workshop on the Economics of Peer-to-Peer Systems, Berkeley, CA, June 2003. 13. Jon Crowcroft, Richard Gibbens, Frank Kelly, and Sven Ostring. Modelling incentives for collaboration in mobile ad hoc networks. In Proceedings of WiOpt’03: Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks, Sofia Antipolis, France, March 2003. 14. Rajdeep K. Dash, Micholas R. Jennings, and David C. Parkes. Computational mechanism design: A call to arms. In IEEE Intelligent Systems, November/ December 2003. 15. Zoran Despotovic and Karl Aberer. Maximum likelihood estimation of peers’ performance in p2p networks. In The Second Workshop on the Economics of Peer-to-Peer Systems, 2004. 16. John R. Douceur. The sybil attack. In Proc. of the IPTPS02 Workshop, Cambridge, MA (USA), March 2002.

5 Encouraging Cooperative Interaction

107

17. J. Feigenbaum and S. Shenker. Distributed algorithmic mechanism design: Recent results and future directions. In Proceedings of the 6th International Workshop on Discrete Algorithms and Methods for Mobile Computing and Communications, pages 1–13. ACM Press, New York, 2002. 18. Michal Feldman, Christos Papadimitriou, John Chuang, and Ion Stoica. Freeriding and whitewashing in peer-to-peer systems. In PINS ’04: Proceedings of the ACM SIGCOMM workshop on Practice and theory of incentives in networked systems, pages 228–236, New York, NY, USA, 2004. ACM Press. 19. Eric J. Friedman, Joseph Y. Halpern, and Ian Kash. Efficiency and nash equilibria in a scrip system for p2p networks. In EC ’06: Proceedings of the 7th ACM conference on Electronic commerce, pages 140–149, New York, NY, USA, 2006. ACM Press. 20. Yih-Chun Hu, Adrian Perrig, and David B. Johnson. Ariadne: A secure on-demand routing protocol for ad hoc networks. Technical Report Technical Report TR01-383, Department of Computer Science, Rice University, December 2001. 21. R Jurca and B. Faltings. Reputation-based pricing of p2p services. In Proceedings of the Wokshop on Economics of P2P Systems, Philadelphia, USA, 2005. 22. Radu Jurca and Boi Faltings. Minimum payments that reward honest reputation feedback. In Proceedings of the ACM Conference on Electronic Commerce, pages 190–199, Ann Arbor, Michigan, USA, June 11-15 2006. 23. Kevin Lai, Michal Feldman, John Chuang, and Ion Stoica. Incentives for cooperation in peer-to-peer networks. Workshop on Economics of Peer-to-Peer Systems, Berkeley, CA, June 2003. 24. Bernd Lamparter, Marc Plaggemeier, and Dirk Westhoff. Estimating the value of co-operation approaches for multihop ad hoc networks. In Elsevier Journal of Ad Hoc Networks, Elsevier Science, January 2004. 25. Sergio Marti, T.J. Giuli, Kevin Lai, and Mary Baker. Mitigating routing misbehavior in mobile ad hoc networks. In Proceedings of MOBICOM 2000, pages 255–265, 2000. 26. Pietro Michiardi and Refik Molva. CORE: A collaborative reputation mechanism to enforce node cooperation in mobile ad hoc networks. Sixth IFIP conference on security communications, and multimedia (CMS 2002), Portoroz, Slovenia., 2002. 27. Herve J. Moulin. Fair Division and Collective Welfare. The MIT Press, 2003. 28. Martin J. Osborne and Ariel Rubinstein. A Course in Game Theory. The MIT Press, 2002. 29. Elinor Ostrom and James Walker. Trust and Reciprocity. Interdisciplinary Lessons from Experimental Research. Russel Sage Foundation, New York, 2002. 30. Panos Papadimitratos and Zygmunt J. Haas. Secure routing for mobile ad hoc networks. In Proceedings of SCS Communication Networks and Distributed Systems Modeling and Simulation Conference (CNDS 2002), San Antonio, TX. IEEE, January 27-31, 2002. 31. Andreas Pfitzmann, Birgit Pfitzmann, Matthias Schunter, and Michael Waidner. Trusting mobile user devices and security modules. In Computer, pages 61–68. IEEE, February 1997. 32. Bruce Schneier. Secrets and Lies. Digital Security in a Networked World. John Wiley & Sons, Inc, 1 edition, 2000.

108

Sonja Buchegger and John Chuang

33. Andr´e Weimerskirch and Dirk Westhoff. Zero-common knowledge authentication for pervasive networks. In Proceedings of Selected Areas in Cryptography, SAC’03, Ottawa, Ontario, Canada, August 2003. 34. Yongguang Zhang and Wenke Lee. Intrusion detection in wireless ad-hoc networks. In Proceedings of MOBICOM 2000, pages 275–283, 2000. 35. Sheng Zhong, Jiang Chen, and Yang Richard Yang. Sprite: A simple, cheatproof, credit-based system for mobile ad hoc networks. Proceedings of IEEE Infocom, 2003.

6 Competition and Cooperation in Wireless Multi-Access Networks Johan Hultell, Jens Zander, and Jan Markendahl Department of Communication Systems, Royal Institute of Technology [johan.hultell|jens.zander|jan.markendahl]@radio.kth.se

Summary. Providing affordable wireless access for all user-needs is one of the remaining great challenges for the telecommunications industry. It has over the years become evident that neither a single wireless access technology nor a single business actor is capable of solving this problem alone. The appearance of multimode terminals, however, opens up for new, more flexible architectures where adaptation to emerging user-needs can be done in a faster, and more cost efficient way than by replacing the “entire” infrastructure. In addition to these technical advantages, multimode terminals have several rather far-reaching business implications. The fact that operators no longer need to provide “full coverage” with their preferred wireless access technology is probably the key factor. It means that infrastructure can be incrementally deployed (where needed) and shared in other locations, resulting in large cost reductions, as well as lowered entry barriers for new specialized actors focusing on a “niche market”, e.g., indoor coverage. To exploit these potential advantages to the full extent, however, new technology that enables a user and network, or alternatively networks, to form a dynamic relation “on the fly” has to be developed. In this chapter we examine how “infrastructure sharing” can be performed by means of such new technology and what consequences it may have on retail prices. We start by studying the archetypical cases of “pure” cooperation and competition and both are shown to be feasible from a technical as well as business standpoint. Furthermore they are shown to result in cost reductions, and in the competitive scenario these may also transfer to lowered end-user price. Although the cases provide valuable insight into the mechanisms of cost reduction and retail pricing, they are note realistic as such and our observation is that future business models may neither be categorized as competitive nor cooperative. Two emerging examples are implicit access cooperation (“coopetition”) and “free” access concepts. The chapter is concluded by a discussion concerning some implications of these scenarios and an outline of some new related research topics are presented.

6.1 Introduction The success of wireless technologies is likely to continue and even accelerate if wireless applications and services become pervasive; used by everyone and everywhere.

109 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 109–132. c 2007 Springer.

110

Johan Hultell et al.

If we are able to position our technology properly, this will not only be limited to the traditional vision of personal communication, entertainment and information sharing applications. We will also see that computation and wireless communication capabilities become integrated in a great variety of “everyday things”. Examples range from simple sensors and interactive appliances (cards, rings, eyeglasses...), via pocket and lap-sized devices to wall or table screen working areas. The key driver for wireless technology is here (unlike systems deployed for personal communication services) not mobility, but rather the convenience of increased functionality and the ability of avoiding cables and connectors for rapid and flexible deployment. In this process the technology undergoes a transformation; from expensive and highly visible (cf. early cellular phones) towards an “invisible technology” that is present everywhere in massive volumes, easy to use and deploy (“self-configuring”), extremely reliable (“self-healing”) and affordable to everyone. We have recently seen this development in electronic and computing equipment and we are now about to see networking and wireless communication heading down the same road as also infrastructure components are becoming disposable.

Supply-Side Economics of Wireless Broadband Access It is well known that wireless access systems are confined to the range/rate tradeoff defined by the “Shannon bound” of communication theory. This states that, for reliable communication, each transferred bit has to have a minimum amount of energy at the receiver. As higher data rates result in less available energy per transferred bit (given a fixed amount of power available at the transmitter) increased peak-data rates reduce the ranges of the radios. Providing broadband access over a wide area will therefore inevitably require a large number of access point (“base stations”), regardless of the advances in technology. As a matter of fact, it has been shown; see e.g., [24]; that all else equal, the number of required access points grows about linearly with the bandwidth provided. This means that the marginal production cost per transferred bit, second and Hertz is constant. An immediate consequence of this is that in order to provide several orders of magnitude higher data rates to the same customers (with the same amount of money available for spending on telecommunication services), the price per transmitted bit will have to drop radically. This observation has during the last decade generated a considerable amount of research in alternative system architectures. From a business point of view, the problem of providing broadband wireless data access is quite clear: For every deployed access point, operators need a sufficiently large number of users in order to recover investments without claiming “unacceptable” prices. This means that for rural areas, where the user density is low, only a low density of access points can be supported. In order to cover sufficiently many users, the cell radii of these access points furthermore has to be large, which reduces the feasible peak-data rates. In locations where the user density is high, on the other hand, offering high data rates is not a problem.

The Traditional View: One All-Inclusive Wireless System The traditional solution to the problem of varying system requirements has been to design access systems with flexible air interfaces capable of providing both

6 Competition and Cooperation in Wireless Multi-Access Networks

111

high rate/short range and low rate/long range communication. A leading paradigm behind this approach has been that the user-terminal can handle only one (albeit very complex) air interface and that a single worldwide standard is necessary for commercial success. An obvious drawback with utilizing one system, reasonably suited for all purposes, is the complexity and inherent lack of flexibility (as it must meet the most extreme requirement in every possible dimension). This have resulted in that we today have broadband wireless data systems capable of reliable wide area coverage and high speed handover everywhere, but pedestrian (lap-top) users in city-centers will only very rarely need those capabilities. Besides this, there exists a considerable risk that system’s built-in flexibility is insufficient for meeting future user-needs. Wide-area infrastructure deployment is a matter of decades, whereas user-needs often change more rapidly.

Network of Networks: Multi-Access Wireless Systems Combined with Dynamic Network Interworking A more attractive scenario would be triggered by the appearance of a flexible multimode terminal. Such a terminal will be capable of communicating by means of several wireless access standards and adapt to varying access network quality. This scenario would allow a plethora of specialized access systems to coexist, each optimized to provide cost efficient access for its “niche” market1 without any requirement (on each and every system) for service ubiquity. As terminals furthermore are consumer products with a life-cycle of 2-3 years and generally perceived by the buyer, to be intimately connected with the applications of choice, choosing the proper set of air interfaces for a particular terminal is not a problem and each time a user buys a new terminal it is likely to contain a different set. The advent of multimode terminals has also the possibility radically lower the entry barriers to the wireless access domain where new business opportunities could be created [4]. Facility, shop and restaurant owners or even private persons would be capable of offering wireless access to global services as well as value-added localized services. Together the infrastructure components can form integrated parts of a “wireless grid”, accessible to the public [18]. In the prolonging large diversity and efficient competition between providers of networks, service elements or combinations thereof could provide seamless service according to user preferences. This vision challenges many of the current paradigms in mobile communication; including that a single worldwide wireless access standard and a single operator model are necessary for commercial success; and poses a number of new problems. The potential economic advantages of the described multi-access, multi-provider system scenario where each provider specialize seems significant to all actors. The key question is if (and how) these advantages can be realized in a business context and this is the topic treated. The remaining part of this chapter is organized as follows: We start by briefly reviewing recent technology advancements that support automated short-term, dynamic relations between various actors. Thereafter we analyze the archetypical cases of operator cooperation and competition in Section 6.3 and Section 6.4, respectively. 1

This niche can for example be a geographical area, support for mobility, quality of service, etc..

112

Johan Hultell et al.

With full operator cooperation there is no duplication of access network resources and the total infrastructure cost is reduced. The problem is, however, that under such a oligopolistic regime the achievable cost reductions may not be reflected in retail price reductions. In a competitive scenario, on the contrary, the access resources are not minimized but end-user prices are controlled by market forces. Analyzing under which conditions a competitive access market is feasible and to which extent end-user price can be lowered (compared to a cooperative regime) is one of the objectives of this chapter. The two archetypical cases provide valuable insight in the mechanisms of cost reduction and end-user pricing, but they may not be that realistic as such. Section 6.5 therefore extends the analysis by introducing some more realistic use-cases that neither can categorized as competitive or cooperative; implicit access cooperation (“coopetition”) and “free” access concepts. The chapter is concluded with a discussion concerning some implications of these cases and an outline of a few new related research topics are presented.

6.2 Ambient Network: Technology for Dynamic Cooperation and Competition An environment populated by numerous devices, pervasive applications and where a multitude of wireless access technologies; deployed by traditional mobile operators, stores (e.g., airports, malls, restaurants, etc.), and even private persons; coexist faces several challenges. In this section we discuss new technology capable of establishing dynamic, short-term relations between various actors (e.g., between a user and network owner) in an automatic fashion. Such technology would enable users to drive competition and triggering the cooperation. To describe the technology we rely on concepts from the ambient network projects2 [13].

The Vision of Ambient Networks: Ubiquitous Interworking Ambient network technology offers scalable, automated control plane interworking (mobility management, quality of service control, security, media delivery, context provisioning, etc.) between heterogeneous networks managed by, potentially competing, economic and administrative domains [13]. The essence of the technology is network composition, a process during which two networks, autonomously, determine to cooperate and settle the related conditions (see Figure 6.1).3 Besides 2

3

The ambient network project partially funded by the European Commission under its sixth framework program focuses on devising technology for self-configuring (“plug-and-play”) networks, efficient use of infrastructure investments, and access competition. Amongst other, the project has developed a framework that enables increased competition and cooperation in environments with numerous wireless access technologies, operators and other business actors. The framework is based on “dynamic composition” of networks so as to provide access through other networks by means of instant and automatic establishment of inter-network agreements. For more information see www.ambient-networks.org. It should be highlighted that also user-terminals are encompassed by the term ‘network’ within the ambient network paradigm.

6 Competition and Cooperation in Wireless Multi-Access Networks

113

more efficient usage of available network resources, the developed technology allows, at least in principle, users to connect to any available network and not only the infrastructure managed by the operator with whom they subscribe.

Figure 6.1. Cooperating network appear as a single network to external users. Notice that network composition, which will be further discussed in the following, allows a user-terminal to access any available infrastructure.

Compared to current technologies for operator interworking, the perhaps largest advantage is the capability to support automated and dynamic interworking. This means that relations between actors can be formed on a demand basis, without any pre-negotiated agreements [13]. Besides ensuring a scalable and self-configuring network architecture, ambient network simplifies cooperation between infrastructure owners as it: 1. Radically reduces the transaction cost of entering sharing agreements. This is because the conditions for cooperation are settled “on the spot” by the involved networks (as opposed to pre-negotiated agreements that involves the network owners). 2. Enables networks to dynamically renegotiate and exit agreements as their conditions (e.g., load, interference) change. Although also this is an effect of lowered transaction cost it is important to note that it, implicitly, increases the inclination for cooperation as the associated risk reduces. Besides easing network management; including incremental deployment of new wireless access technologies; for operators with multi-access systems the developed technology allows cooperation that involves local operators, or even end-users operating a single access point network. This is important, both since an increasing amount

114

Johan Hultell et al.

of network nodes will be user-deployed and because end-users or other local access operators usually deploy their infrastructure indoors where a considerable proportion of the traffic is generated and where cellular technology has inherent difficulties as it is deployed outdoors. For end-users with multimode terminals ambient network enables: 3. The ability to be always best connected [7] since users with ambient network technology can utilize the network, and wireless access technology, that currently offers best experience. This may act as a catalyzer for access network competition also in the shorter run. It should be noted that although ambient network targets the provisioning of wireless access services (“transport layer”) it may have far-reaching industrial implications as it reduces entry barriers for new actors focusing on a specific market segment. This suggests that we, besides traditional operators with vertically integrated value chains, also may see new niche actors emerge. An access market that consists of a larger number of access operators may, furthermore, in the prolonging, result in that more content providers emerge since they do not have to act in a monopsony.

Technology for Dynamic Cooperation and Competition Self-organizing architectures wherein subsystems and users determine with whom they should connect and compose brings about several challenges; both technical and economical. These include: Discovery available networks. One technical problem, particularly relevant for user-terminals with limited energy, is the design of protocols that allow them to discover the presence of subsystems in their proximity. To be efficient the protocol also needs to allow subsystems to “advertise” their current capabilities (e.g., price models, available resources, etc.). This information can for example be used when networks determine whether they should compose, and by terminals for prioritizing between available access services. Note that even the discovery of network candidates is a non-trivial problem since there may be many frequency bands to search in and numerous options to evaluate. Networks, advertising their presence on several modes and directory based method (accessed over a wide-area network) seem to be promising solutions currently under investigation [25]. Decentralized resource management. When the capabilities and characteristics of the available networks are known, a prioritization can be done. Two key problems here are that the decision has to be taken on input data, e.g., signal quality, price, etc., that may change in the near future, and that decentralized control, without the proper design, can result in an inefficient resource usage (cf. the “tragedy of commons”) [9]. Automated formation of financial agreement. Once the selection has been completed, a financial agreement has to be reached. An important issue here is to keep the cost of such transactions low since the number of transactions may be high. The ultimate challenge is to provide access anonymously, without pre-arranged subscriptions and with full protection of privacy [25].

6 Competition and Cooperation in Wireless Multi-Access Networks

115

Realizing automated negotiations between actors will require the introduction of autonomous agents that act on behalf of the users and networks. In addition to discovering available networks, the agents have to, continuously, evaluate the available alternatives so as to maximize their stakeholder’s utility (“value for money”). Thus agents are not merely a resource manager responsible for evaluating the technical performance (peak-data rates, quality of services, etc.), but also a trade-agent entering agreements and performing electronic financial transactions. Since the terminals will have to take a more active part in the resource management they will have a larger battery consumption. In fact, it is likely that many of the procedures, e.g., how to search for available networks, implicitly, are determined by the terminals’ battery consumption. Hence an important research issue for this type of systems is the to design energy efficient scanning procedures.

6.3 Cooperative Wireless Access Cooperation between infrastructure providers has the ability to lower the social cost of an infrastructure. This can, for example, be realized by avoiding duplication of assets, from network specialization and thereby increased production efficiency, to more subtle advantages such as lowered financial risk [6]. This section starts by illuminating the inherent advantages that access operators can exploit through cooperation. Then we describe a few existing use-cases in which infrastructure cooperation has been employed.

6.3.1 Benefits and Perils of Cooperation Cooperative wireless access enables network operators to lower their production cost for deploying, operating and maintenance of wireless infrastructure. The direct advantages are: 1. Avoiding duplication of assets and utilizing economies of scale. Duplication of assets (i.e., parallel infrastructures) leads to higher costs without necessarily resulting in larger social value. This argument holds for many networked infrastructure systems, e.g., railroads where operators avoid the cost of running separate railroad lines to the same destination. It is fairly obvious that large cost reductions can be achieved in rural areas where demand is so low that a single infrastructure provides sufficient capacity4 . Such markets are usually referred to as natural monopolies, a situation where one firm can produce a desired output at a lower social cost5 than two or more firms [20]. Cost reduction gains achievable are often gigantic, but somewhat fictitious. Even though, say 5 operators would like to operate in a certain area (cost reduction of 80 percent with shared 4

5

Often the wireless systems are designed so that a minimum data rate can be supported with “reasonable” availability. In rural areas, where demand is low, the required access point density is set so that users at the cell border experience a peak-data rate equal to this minimum data rate. As the system is limited by range and not load (cf. capacity-limited scenarios) it often has a low load and could, thus, carry more traffic. Social costs within economic is the total costs associated with an activity [19].

116

Johan Hultell et al.

infrastructure), these operators would never actually contemplate building their own (parallel) infrastructure. In capacity-limited scenarios, i.e., in areas with high user densities the advantages from cooperation are less obvious. Although it can be argued that a single (cooperative) network would make best use of resources, these gains are diminishing as the traffic load increases.6 2. The ability to specialize, which refers to that each access operator can focus on the market segment (e.g., geographical area, wireless access technology) were they are, relatively, most efficient. Within economics, this is referred to as the theory of comparative advantage (Ricardo’s law). It states that two actors capable to trade should specialize and produce the service on which they are relatively most efficient. Under conditions where the costs associated with trade is low, this may result in a lowered social cost. As ambient network technology offers automated interworking, it may adopt the role of a catalyzer towards increased specialization. An obvious example where specialization would be advantageous is indoor coverage. This is very hard (costly) to provide with cellular systems since they are deployed outdoors. Automated, dynamic control plane interworking opens up for the possibility of cooperation between traditional cellular operators that can provide efficient outdoor coverage, and local indoor operators (facility owners, end-users, etc) that can reuse existing broadband connections and offer WLAN coverage in a cost-efficient manner [4]. Other secondary (indirect) benefits of cooperative wireless access include a broader range of access services and lower financial investment risk. Beside the obvious benefits of cooperation, there is also a risk that cooperation between access actors creates new monopoly (full cooperation of all access provider in a certain market) or oligopoly situations (groups of access provider acting in collusion) where the “aggregate” market actors have less incentives to share the achieved cost reductions with their customers.

6.3.2 How Much Can Be Gained through Cooperation? One immediate benefit from cooperation between access networks is that users can connect to more access points which will provide better coverage and/or higher enduser quality (e.g., higher data rates). An additional advantage with the (aggregate) system is that in can handle more user traffic since the load can be distributed over more access points. As we will see in our example, these two factor seldom come in to play simultaneously. In a scenario where access points of the cooperating networks are placed in (almost) the same locations (almost “co-sited”) the “coverage gain” is small, whereas the gain from load balancing (often referred to as improving the “trunking efficiency”) is large. If the access points in the cooperating networks, on the contrary, are spatially separated, the gain associated with load balancing diminishes whereas the “coverage gain” increases. 6

This is usually called “trunking efficiency”, which refers to the very general queueing theory result that pooling all resources, results in higher capacity/higher availability. With low average traffic load per network, we have to provide a large relative overload margin to guarantee a certain resource availability. As the traffic load increases these gains, however, decrease.

6 Competition and Cooperation in Wireless Multi-Access Networks

117

This is illustrated in Figure 6.2, which shows the gain in uplink throughput (data rate) experienced in a cellular data communication scenario where three networks cooperate. Here we study the 10th percentile of the user throughput, i.e., the minimum data rate that more than 90 percent of the users achieve.7 It is evident that the coverage gains are 70–180 percent and increase with the spatial separation of the cooperating operators’ network and that the gain is rather insensitive to traffic (average number of users per access point). Although the presented gains are lowered in scenarios where more sophisticated radio resource management techniques are used it should be noted the gains with shared infrastructure are still substantial [10].8

Figure 6.2. Gain in asymptotic user throughput for the 10th user percentile (“coverage”) for uplink transmission in a case where three operators cooperate.

Figure 6.3a shows the corresponding gain for a case two WLAN operators cooperate in a traffic hot-spot. The composed WLAN system is assumed to form a noise-limited (high reuse) system where a total of 19 access points have been randomly deployed in a 250 m×250 m area [1]. It is apparent that overall gain in system throughput is constant and about 65 percent although, the gain associated with 7

8

Each individual subsystem have a cell radius equal to 360 m. All access points utilize HSPA technology, three-sector antennas and each sector reuse all bandwidth (although the subsystems does share spectrum). The access points are equipped with two receiving antennas and a channel oblivious round robin scheduler is used. A more detailed description of the conditions surrounding this study can be found in [10]. Systems employing opportunistic scheduling, more receiving antennas or any other technique that exploits microscopic diversity are associated with higher SIR and, thus, also user throughput levels.

118

Johan Hultell et al.

macroscopic diversity order (that is attained if users always are assigned to the access point with lowest path-loss) diminishes with increasing access point density.9

Figure 6.3. Gain in asymptotic system throughput for a geographically limited traffic hot-spot where two WLAN (IEEE 802.11a) networks cooperate. As the combined system is noise limited, additional gains can be realized by load aware user assignment.

6.3.3 Current Practice of Infrastructure Cooperation Interconnection of networks The perhaps most rudimentary form of infrastructure cooperation between operators is the interconnection of networks. Its primary purpose is to increase the value of the interconnected network and since it is socially desirable but not always beneficial for individual operators it has often been mandatory for operators (under conditions stipulated by national regulatory authorities) [14]. Examples of current obligations imposed on incumbent American and European operators can be found in the US Telecom Act of 1996 [21] and the regulatory framework issued by the European Union [23]. Both impose extensive regulations on the wholesale market for access pricing and interconnection regulation. At large, they require that incumbent operators 1. Interconnect for the purpose of originating and terminating traffic. 2. Sell capacity at regulated wholesale prices. 9

In interference-limited scenarios the gain from balancing the load across access points reduces however [9].

6 Competition and Cooperation in Wireless Multi-Access Networks

119

3. Allow competitors to use parts of their network so as to provide competing services. The traditional type of regulation associated with interconnection is important for voice services where it is essential for a subscriber to be able of reaching other users, independently of their operator. However for future data services the key issue is instead the ability to initiate data session from any network using the same identity or user account.

Network Sharing Network sharing received a significant amount of attention during the early phases of deployment of 3G system. Its main motivation was potential cost savings (lowered deployment, as well as operation and maintenance costs) that could be attained in situations where license requirements included coverage obligations of rural areas with low traffic demand, as in Sweden [15]. The regulation concerning network sharing have guidelines that stipulate how network can be shared in most European countries. These have, however, been settled on national level and therefore the situation may vary greatly between countries [16]. Some countries have adopted a “liberal” approach and allow most forms of network sharing and mobile virtual network operator operation. Other countries, have instead taken a more “conservative” standpoint. In those, sharing of core network elements is prohibited, while sharing of elements in the radio access networks only is allowed under premises that access network and network management functionality are logically separated. In some other countries, mobile virtual network operation in 3G networks is not allowed at all. Similar for the cooperation models above is that they originate from prenegotiated agreements, specifying how revenue, costs, and responsibilities should be divided between the participating parties. Notice that since this is associated with considerable transaction and monitoring costs cooperation has traditionally been restricted to nationwide mobile operators.

6.4 Competitive Wireless Access This section focuses on open systems where the different networks are managed by providers that compete also on the shorter time-perspective. Users can gain connectivity through any of the available subsystem, e.g., with the ambient network technology described in Section 6.2. Note that this scenario resembles a common market place where consumers can purchase a good or service from multiple providers that compete with each other for earning as large revenue as possible. We start by describing some of the generic advantages that arise with a competitive wireless access market. This is followed by an example, in which we illustrate how the end-user price, and service quality may differ between a competitive and cooperative (oligopolistic) access market. The example is based on work that previously, in parts, have been published in [2, 3, 9].

6.4.1 Benefits of Competitive Wireless Access From previous section it is evident that cooperative wireless access, in which operators manage part of their assets jointly can reduce costs for wireless access. It is

120

Johan Hultell et al.

however unclear to what extent these cost reductions would accrue the end-users and in fact it could be argued that cooperation between access operators even could increase prices since there are less competition, and thus incentive, for maintaining low prices. One way to ensure lowered retail prices would be through establishing a competitive access market where access providers also compete in the shorter perspective (per session, second, etc.). This could for example be established through the technology discussed in Section 6.2. Beside this obvious benefit of a lower retail price, a competitive access market is, at least in theory, also associated with several more subtle advantages such as increased innovation (as a consequence of lower entry barriers), higher flexibility to emerging needs, and a broader range of services. Whether these gains can be realized or not is outside the scope of this chapter but we may note that they are generally dependent on that users have a low experienced “cost” when switching from a provider to another; something explicitly targeted by ambient network.

6.4.2 Feasibility of a Competitive Wireless Access Market In the example, we study an agent-based architecture where trade agents can connect to multiple access points; see Figure 6.4. To divide resource (i.e., downlink “air time” in a time division multiple access system) between competing trade agents, each access point employs a proportional fair divisible auction [11]. One property of such these is that all participating trade-agents obtain a share of the resource. This is in contrast to, e.g., a Vickrey auction where the highest bidder claims the entire resource at a price determined by the second highest bidder. Assuming that there are N competing users at, say access point m, and letting sm i,j ∈ [0, smax ] (where 0 < smax < ∞) denote the bid agent j place in auction i at access point m. The proportion of transmission time (“air time”) allocated to trade-agent j is then determined by the following predefined, and publicly known, allocation rule sm sm i,j i,j xm = m ∈ [0, 1] (6.1) i,j = P m si,j + sm i,−j k6=j si,k + εm Here εm is the reservation price and it corresponds to a nonzero bid placed by the network that can be interpreted as a price floor below which the resource is not sold. Both sm i,j and εm are counted in monetary units (mu) and it should be stressed that we throughout the presented example assume that the resource (transmission time) is infinitesimally divisible. We envision that access points (networks) will change their reservation price on a rather slow time-basis and they are here assumed to vary on the same time-scale as the average offered load (cf. “time-of-day” pricing). Faster variations are instead handled implicitly through the structure of the auction which, all else equal, results in higher prices when the access point is congested.

Acceptance Probability Since the price at which the transmission time is sold (through the auction) depends on the demand, it will vary dynamically as users enter and leave the system. We assume that each user has a fixed, and individual, price threshold pth (measured

6 Competition and Cooperation in Wireless Multi-Access Networks

121

Figure 6.4. System model for the studied example. User trade-agents try to minimize their “cost” (disutility) by determining how large bids they should place at the networks, while the access points concurrently try to maximize their individual revenue by choosing an appropriate reservation (minimum) price for the resource.

in monetary units per transferred MByte) that is unknown to the access points as well as to other trade-agents. If the expected price exceeds this threshold the tradeagent will simply refrain from transmission. For simplicity, the expected price is assumed to be known to all trade-agents and it could, for example, be advertised by the access points (networks) via the ‘network advertisement’, which is part of the ambient network technology outlined in Section 6.2. To account for that users may have varying price sensitivity we assume that the price threshold pth for each user is drawn from a probability density function fp (p). Although the price thresholds are private information, access points are assumed to have perfect knowledge about its distribution fp (p). Utilizing the probability density function the acceptance probability, PA (p), which describes the probability that a randomly selected user will enter the system at a certain expected price p is defined as Z p PA (p) , fp (ξ) dξ. (6.2) 0

122

Johan Hultell et al.

The acceptance probability allows us to express the total demand (at an access point) as D (p) = D0 PA (p) . (6.3) Here D0 is the potentially offered load per access point and it is corresponds to the offered load if the price of the access point would be set to zero. Three different demand functions are studied; one convex, one concave, and one linear; and each of them has been normalized so that Z ∞ 1 PA (p) dp ≡ (6.4) β 0 where β is demand-related parameter.

Demand Models The studied demand models as well as their associated price elasticity of demand are illustrated in Figure 6.5. The latter is defined as ED (p) ,

∂D (p) p % change in quantity = ∂p D (p) % change in price

(6.5)

and it is used within economics to measure the rate at which demand varies due to a price change [19]. If ED (p) < 1 the demand is said to be price-inelastic while ED (p) > 1 instead is referred to as price-elastic. The special case where ED (p) = 1 corresponds to the price that a revenue-maximizing AP would utilize in a monopoly situation [19]. We highlight that the expected price threshold E[pth ] = β1 for each of them [9]. Moreover we stress that the variance associated with the users’ price thresholds for the convex demand function exceeds the variance of the linear, which in turn is larger than the one associated with the concave function. A more extensive description of the demand functions are given in [3, 9].

A “Persistent” Trade-Agent Strategy (“The User Problem”) Trade agents that decide to enter the system try to minimize a weighted sum of the monetary expenditure and the file transfer delay. The latter is (for user j with a file of size qj ) defined as n X M X m xm Tj , TA · min (n) : i,j Ri,j TA ≥ qj ,

n ∈ N = {1, 2 . . . }.

(6.6)

i=1 m=1

Here TA is the time-duration between two consecutive auctions, M =|M| is the m number of available access points, and Ri,j is the peak-data rate that user j experience to access point m during auction i. xm i,j is the proportion of the resource allocated to user j at access point m as a result of auction i (an example where a user with a file to transfer is shown in Figure 6.6). In a similar manner the monetary expenditure associated with the file transfer is given as Ej ,

z X M X i=1 m=1

sm i,j ,

(6.7)

6 Competition and Cooperation in Wireless Multi-Access Networks

123

Figure 6.5. Illustration of the studied normalized demand functions (acceptance 1 probabilities). In the studied example β = 150 .

Figure 6.6. Illustration of the auction procedure associated with a file transfer. In the example, trade-agent j initiates a file transfer in auction 1. In auction i user j is allocated a portion xi,j of the total available transmission time TA and depending on its current peak-data rate Ri,j the trade-agent will be able to transfer a total of xi,j Ri,j TA bits. After participating in z auctions the file transfer is completed and consequently the file transfer delay becomes Tj = zTA in this case.

124

Johan Hultell et al.

where z denotes the number of auctions in which that the trade-agent has to participate in order to complete the file transfer (see Figure 6.6). Combining Equation 6.6 and Equation 6.7, the disutility (“cost”) may be written as cj , qj (Ej + αj Tj ) =

z X M X

cm i,j (si,j , si,−j ) ,

(6.8)

i=1 m=1

where αj is parameter that describes how sensitive user j is to delays in comparison to increased monetary expenses. In our example trade-agents are assumed to be selfish and only interested in minimizing their individual disutility for transferring the file. Hence their problem can be formulated as min

1 1 ,...sM ,sM ...,sM s1 1,j ,s2,j ...,sz 1,j 2,j ˜,j z ˜,j

E

z X M hX

cm i,j (si,j , si,−j )

i

∀ j.

(6.9)

i=1 m=1

Since trade-agents’ actions are intertwined, e.g., through Equation 6.1, they are involved in an interactive decision process (or a noncooperative game). As they furthermore typically have to participate in multiple auctions to complete their transfer, agents are in fact involved in a repeated noncooperative game where the strategy-space is given by their bids in the different networks, the set of players is dynamically changing between auctions (as users enter and leave the system), and their objective function is the expected cost for transferring a file. Here we focus on a “persistent” agent-strategy. The strategy departs from Equation 6.9 and it is based on the assumptions that 1. The peak-data rate of the particular user, on whose behalf trade-agent j is acting m m remains unchanged during the entire file transfer, i.e., Rk,j = Rl,j ∀ k, l, m, and that 2. The total demand associated withP the other trade-agents remains unchanged P m during the entire file transfer, i.e., i6=j sm k,i = i6=j sl,i ∀ k, l, m. Under these two assumptions trade-agent j would, for each network, place identical bids in all auctions and we will therefore heron forth drop the subscript i (denoting the auction number). The corresponding transfer delay and monetary expenditure would be P m qj M qj m=1 sj Tj = PM , E (6.10) P j = M m m m TA m=1 xm j Rj m=1 xj Rj Inserting this into Equation 6.9, the repeated game simplifies to ! PM PM m m qj m=1 sj + αj TA m=1 sj + αj TA min ⇐⇒ min . P P M M m m m m 2 M TA 2 M s1 s1 j ,sj ,...sj j ,sj ,...sj m=1 xj Rj m=1 xj Rj

(6.11)

Hence with the persistent strategy, the set of trade-agents J = {1, 2, . . . , J} are involved in several sequential noncooperative games. Each game is described by n o Gu = J , S, {cj } (6.12) and we will refer to Equation 6.12 as the “user game”. Here S = S1 × · · · × SJ is the strategy-space, Sj = {s1j , s2j , . . . sM j }, and cj is the cost function associated with player j and it is described by the objective function in Equation 6.11.

6 Competition and Cooperation in Wireless Multi-Access Networks

125

The Access Point Problem: Revenue Maximization Just as user trade-agents want to minimize their disutility, access points try to maximize their revenue by selecting an appropriate reservation price ε. Here we analyze the problem for the cases where access points 1. Compete for users while trying to maximize their individual revenue, or 2. Cooperate with each other and maximize their total revenue. While the former case are modeled through noncooperative game-theory, the latter is treated by means of cooperative game theory. Competitive access points. In the studied context, trade-agents tend to associate with the access point with lowest price. This can be influenced by the access point through the reservation price εm . The appropriate choice for an access point depends on the reservation prices used by the other access points (as in any marketplace). Since each access point, further, want to maximize its own expected revenue, the available access points M = {1, 2, . . . , M } are involved in a noncooperative game defined by n o Ga = M, ε, {Λm } . (6.13) where ε = ε1 × · · · × εM is the strategy-space, εm the reservation price that access point m employ, and Λm the utility function (i.e., expected revenue) characterizing access points m. We will refer to Equation 6.13 as the “access point game”. Let Λm (εm , ε−m ) depict the expected revenue per second that access point m earns if it employs a reservation price εm concurrently as the other access points utilize reservation prices ε−m = {εi , ∀ i ∈ M\m}. Then the best response function for access point m, ϕ(ε−m ), can be written as ϕ(ε−m ) = arg max Λm (εm , ε−m ) .

(6.14)

εm

Similarly as for the user game, the Nash equilibrium point is obtained when εm = ϕ(ε−m ), ∀ m ∈ M

(6.15)

and it is characterized by that no access point, unilaterally, can increase its expected revenue by deviating from its current reservation price εm . As access points in our example are assumed to be identical, we confine ourselves to symmetric solutions where εk = εl , ∀k, l ∈ M. It should be stressed that even though we have not been able to show that there exist a Nash equilibrium point analytically, all our simulation experiments indicate that this, indeed is the case [3]. Cooperative access points (oligopolistic regime). An alternative to competing for the wireless users would be for access points to cooperate so that higher expected revenues can be earned. To model this case, which corresponds to an oligopolistic regime, we utilize the Nash bargaining game where the Nash equilibrium point is used as disagreement point. The disagreement point corresponds to the expected revenue that access points would earn if the “cooperating” access points fail to reach a agreement.

126

Johan Hultell et al. It is well known; see e.g., [22]; that the solution to the Nash bargaining game coincides with the solution to the optimization problem arg max ε

M Y

(Λm (ε) − Λm,dis )

(6.16)

m=1

where Λm,dis is the disagreement point. For the special case where all access points are identical; i.e., εk = εl , ∀ k, l (Λk,dis = Λl,dis , ∀ k, l); it can readily shown [9] that arg max ε0

M Y

(Λm (ε) − Λm,dis ) ⇐⇒ arg max

m=1

ε0

M X

Λk ε0 ,

(6.17)

k=1

where ε = {ε1 , . . . , εM } and ε0k ∈ R+ .

Numerical Example This section applies the developed framework to an example with two access points that either compete or cooperate. To quantify the performance experienced by users in the two regimes we utilize the average user throughput and monetary expenditure per MByte. For access points we instead use the average revenue per second. Figure 6.7 and Figure 6.8 depict the average cost per transferred MByte and the average user session throughput, respectively. Note that we, in both figures, have omitted the simulated results for the case where the access points cooperate, i.e., the Nash bargaining solution, and instead relied on validated analytical results; see [9]. It is evident that an architecture where access points compete for users, and consequently share their infrastructure implicitly, in combination with trade-agents can reduce end-user price and increase session throughput considerably, especially at low loads. The difference between the two regimes reduces with the “concavity” of user demand. This is an effect of that the price elasticity reduces with concavity. As a low elasticity suggests that an increase in price only affects demand marginally, this motivates both the higher price, and lower user throughput (since more users have to share the channel). Figure 6.9 shows the relative revenue-gain between the oligopolistic and competitive regime. As expected, competing access points earn less than those who cooperate although the difference reduces with D0 . Already for medium loads (an average channel occupancy between 0.40-0.55), the relative revenue loss is smaller than 50 percent for all of the studied demand functions and for the concave demand function the difference is diminutive. In general, it is evident that the loss in revenue is less pronounced than the users’ “quality” improvement. This can be ascribed to the fact that scenarios where access points compete for wireless users will result in a lower price per transferred MByte p (than the monopoly price), which in turn results in a higher entering probability.

6.5 Dynamic Cooperation and Competition The two archetypical cases of “pure” cooperation and competition provide valuable insight into the mechanisms of cost reduction and end-user pricing, but they may

6 Competition and Cooperation in Wireless Multi-Access Networks

127

Figure 6.7. Average monetary expenditure per transferred MByte as a function of the potentially offered load D0 (file arrival density λ) for the studied demand functions.

Figure 6.8. Average throughput experienced by users as a function of the potentially offered load D0 .

128

Johan Hultell et al.

Figure 6.9. Relative revenue gain for cooperating access points as a function of the potentially offered load D0 . We see that although access points always will benefit from cooperation the relative gain decreases rather quickly with D0 . This means that, as long as the system is not overprovisioned (and resource management is not a problem), an architecture where access points compete for wireless users is feasible. not be that realistic as such. In the following section a few interesting, upcoming access provisioning scenarios that contain elements of cooperation and competition; and therefore can not be mapped on our two basic cases; are discussed.

Coopetition: Competition where Feasible and Implicit Cooperation where Needed One key advantage with a technology like ambient network is that it allows operators to specialize and find the niche where they can offer the access service that makes them most efficient (“profitable”). As discussed in Section 6.2, the chosen niche can be a geographical region, a certain wireless access technology, a certain quality of service offering, or a combination thereof. The analogy with other lines of business is obvious; not everyone need to be “department stores”. As a matter of fact, the theory of comparative advantage, suggests that markets where providers specialize, and trade goods with each other, may be more efficient [19]. An inherent problem is of course that users need to know in which store a certain good is sold. The same holds true for geographical specialization where one operator may choose to run an access network only in a limited geographical region and rely on roaming elsewhere. Here access providers compete locally while they, implicitly, cooperate as their customers can use other access provider in locations where they are not present.10 This is an example of coopetition [5], which is used to denote 10

Gas station chains are an obvious analogy from another infrastructure business domain. A chain of gas stations (“Brand X”) does not need to have a presence

6 Competition and Cooperation in Wireless Multi-Access Networks

129

situations where business actors compete in some market segments and cooperate in others. Coopetitive wireless access can be realized in several ways, although we primarily would think of two cases; decentralized (“user centric”) resource management and resource management by means of access brokers. In the first case; user centric resource management most of the resource control stays with the user, or rather with an agent residing in the terminal and act on behalf of its user. Notice that this could be implemented means of ambient network technology and would result in a situation similar to the competitive access market discussed in Section 6.4. One significant new complication is that in competitive access, the user agent has to, continuously, evaluate access offerings from a technical as well as economical standpoint so as to maximize the “value for money”. The user agent is not just a resource manager in the traditional sense evaluating technical performance criteria, but also a trade-agent, entering agreements and performing electronic financial transactions. In conventional business models these choices are done on beforehand when a user subscribes to the services of a certain operator, or when a operator make agreements with roaming partners. In the second case, a virtual operators, or access broker, act on behalf of many and purchase the access services from the available providers either “on the spot” or in wholesale. This case is very similar to the user-centric case. The difference is that the access broker uses his trade-agent as proxy for the end-user, evaluating advertisements, performing all negotiations and access selection decisions described above. This solutions might, for example, be preferred in situations where the computational power of terminals is limited, or necessary energy consumption for supporting a fully decentralized too high. When access services are purchased in wholesale, the negotiations are done on beforehand, e.g., in futures, options, and other derivatives on bandwidth. The competitive advantage of the access broker that allow him to successfully cut in in between the physical operator and the user is superior information about access availability, volume discounts and the decreasing the systematic risk (as an insurance company) [8]. An example of the access broker model used in reality, is the case where physical operators act as mobile virtual network operators in other geographical areas, e.g., countries, on a wholesale basis. Another example where geographical roaming has been used, is the 2G operator SpringMobil, who in Sweden used its licence to offer private indoor coverage coverage to business users.11 Outdoors, their subscribers roamed into a nationwide 2G network managed by another operator. It should however be noted that for cellular systems, where “coverage” often is used a means for differentiating the access services towards other operators, geographical infrastructure sharing has been rare; see e.g., [17]. Important research questions are, thus, methods for service quality differentiation and for dividing costs between involved parties.

11

everywhere - if a customer preferring Brand X cannot find such a station in some location, he can use Brand Y. The two brands implicitly cooperate in this case, although brand X and Y stations are direct competitors if they are located at the same freeway exit. See www.SpringMobil.se for more information.

130

Johan Hultell et al.

Geographical infrastructure sharing has also started to emerge in smaller scale between WLAN operators. Examples range from access aggregators12 interconnecting different hotspot operators, to user deployed community based networks. Often the latter type is based on barter trade where a user can gain access to other community members’ WLAN access points such as FON.13

Wireless Access as an Integrated Part of Bundled Services Following the success of short-range unlicensed WLAN technology, new business models for providing public broadband wireless access has emerged. Some actors have originated from the operation mode used in cellular systems and typically these actors sell wireless data access at a fixed charge. Usually these actors have employed price-skimming14 , explicitly targeted towards business users and sold wireless broadband access around e10/h and e10-20/24h [4].15 To a large extent the strategy has been used in contexts where a single actor can exert a local monopoly. [10] Concurrently other actors, e.g., hotels and restaurant owners, have chosen to bundle the wireless broadband access with products and thus seemingly give it away for free. The rationale for this strategy is to exploit that “free” wireless data access can attract consumers so that they can sell more of their primary product [12] and a prerequisite for the strategy is that the production cost for wireless access lower than the cost of their main product. If sufficiently many local operators (hotels, malls, etc.) start to offer cheap (or even “free”) wireless access this may influence the business case of mobile operators since their, relatively, high price margins are undermined.

6.6 Conclusions In this chapter we have examined how new technologies such as multimode terminals and a novel networking architecture offer new means to provide wireless access in 12 13

14

15

See www.thecloud.net. An example is FON, which is a community-based network with more than 160,000 members worldwide that was launched in 2006. Members of the community (“Foneros”) allow each other to access their respective access points and typically the network available to users, or consumers, will consist of high-capacity pockets. According to their mission statement; FON is a value-added service for existing fixed network operators since it enables subscribers of the fixed network operator a larger footprint. Also users who are not members of the community can utilize the network, although at a a cost of e3/24h (see www.fon.com). Price-skimming involves charging a relatively high price policy and it is often used, for a period of time, when a new service is introduced. The ability and possible benefit of using price-skimming is highly dependent on the inelasticity of demand (for the product as a whole or by a certain market segment, e.g., business users) and potential entry barriers. Note that the latter can be viewed as a type of Goldilocks pricing, where a one alternative of the service is offered at a very high price in order to make the nextlower price option look more reasonably (than when viewed alone). The strategy exploits general bias of aversion of the extreme.

6 Competition and Cooperation in Wireless Multi-Access Networks

131

a more cost efficient way. The archetypical cases of operator cooperation and competition were shown to be feasible, both from a technical and business standpoint. Both cases lead to cost reductions, and in the competitive scenario these may also transfer into lowered retail prices. Although the archetypical cases provide valuable insight into the mechanisms of cost reduction and retail pricing, they are not realistic as such and our observation is that future business models may neither be categorized as competitive nor cooperative. We studied two examples that are already emerging: implicit access cooperation (“coopetition”) where operators specialize and compete in certain niches and cooperate in other markets. The other treated example consisted of “free” access concepts, where local operators (e.g., restaurants) are using wireless access, seemingly free of charge, either as portal to their services or as complement to attract customers to other (non-electronic) services they provide.

References 1. M. Berg and J. Hultell. Access Selection in Partially Backhaul-Limited MultiOperator IEEE 802.11 Networks. In Proceedings of the IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06), Helsinki, Finland, September 2006. 2. M. Blomgren and J. Hultell. Decentralized Market-Based Radio Resource Management. In Proceedings of the IEEE Vehicular Technology Conference (VTC), Dublin, Ireland, May 2007. 3. M. Blomgren and J. Hultell. Demand Responsive Pricing in Open Wireless Access Markets. In Proceedings of the IEEE Vehicular Technology Conference (VTC), Dublin, Ireland, May 2007. 4. M. Blomgren, J. Hultell, J. Markendahl, P. Valiente, B. Thorngren, J. Werding, ¨ M¨ and O. akitalo. Novel Access Provisioning - Final report. Technical report, available at: www.wireless.kth.se/projects/NAP/, January 2007. 5. A. Brandenburger and B. Nalebuff. Co-Opetition: A Revolution Mindset that Combines Competition and Cooperation. ISBN-038 5479 506. Doubleday Publishing, New York, NY, December 1996. 6. M. Ganslandt. Konkurrens vid en reglerad utbyggnad av 3G-n¨ at i Sverige. Technical report, Rapport till Post-och Telestyrelsen, available at: www.pts.se, October 2005. 7. E. Gustafsson and A. Jonsson. Always best connected. Wireless Communications, IEEE [see also IEEE Personal Communications], 10(1):49–55, 2003. 8. J. C. Hull. Options, Futures, and Other Derivatives. ISBN 0-13-046592-5. Prentice Hall, International Edition, Dec 2003. 9. J. Hultell. Access Selection in Multi-Access Architechtures: Cooperative and Competitive Contexts. PhD thesis, KTH Communication Systems, Stockholm, Sweden, March 2007. 10. J. Hultell and K. Johansson. An Estimation of the Achievable User Throughput with National Roaming. Technical report, The Royal Institute of Technology, available at: www.wireless.kth.se, June 2006. 11. I. Kremer and K. G. Nyborg. Divisible-Good Auctions: The Role of Allocation Rules. The RAND Journal of Economics, 35(1):147–159, April 2004.

132

Johan Hultell et al.

12. F. J. Mulhern and R. P. Leone. Implicit Price Bundling of Retail Products: A Multiproduct Approach to Maximizing Store Profitability. Journal of Marketing, 55:63–76, 1991 1991. 13. N. Niebert, A. Schieder, H. Abramowicz, G. Malmgren, J. Sachs, U. Horn, C. Prehofer, and H. Karl. Ambient networks: an architecture for communication networks beyond 3g. Wireless Communications, IEEE [see also IEEE Personal Communications], 11(2):14–22, 2004. 14. E. Noam. Interconnecting the Networks of Networks. ISBN-10: 026 2140 721. The MIT Press, June 2001. 15. Northstream. Network Sharing - Savings and Competitive Effects. Technical report, The Swedish National Post and Telecom Agency, available at: www.pts. se/Dokument, September 2001. 16. Northstream. 3G rollout status - a report about the 3G status in Europe. Technical report, The Swedish National Post and Telecom Agency, available at: http://www.pts.se/Dokument/, October 2002. 17. Office of Communications. National roaming - A further consultation. Technical report, Office of Communications, available at: http://www.ofcom.org.uk/ accessibility/rtfs/consultations/, 2004. 18. J. M. Pereira. Fourth generation: Now it is personal! In Proceedings of the IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’02), Lisbon, Portugal, September 2002. 19. R. S. Pindyck and D. L Rubinfeld. Microeconomics. ISBN-10: 013 0165 832. Prentice Hall, 5th edition, July 2000. 20. W. Sharkey. The Theory of Natural Monopoly. ISBN-10: 0521271940. Cambridge University Press, November 1982. 21. H. Shelanski. Inter-modal competition and telecommunication policy in the United States. Communication & Strategies, 60:15–37, 2005. 22. C. Touati, E. Altman, and J. Galtier. Generalized Nash Bargaining Solution for Bandwidth Allocation. In Elsevier, Computer Networks, 2006. 23. H. Wittig, J. Dellis, R. Sinha, and D. Wright. European Telecom Services: Agents of Deflation - Disruptive Technologies in Mobile Europe. Technical report, August 2006. 24. J. Zander. On the Cost Structure of Future Wideband Wireless Access. In Proceedings of IEEE Vehicular Technology Conference (VTC Spring), Phoenix, Arizona, May 1997. 25. J. Zander. Competitive Wireless Multi-Access. In Proceedings of IEEE International Symposium on Indoor and Mobile Radio Communications (PIMRC), Helsinki, June 2006.

7 A Cooperative ID for 4G Simone Frattasi1 and Hanane Fathi2 1

2

Center for TeleInFrastruktur (CTIF), Dpt. of Electronic Systems, Antennas, Propagation and Radio Networking (APNet) Group, Aalborg University [email protected] Research Center for Information Security (RCIS), National Institute of Advanced Industrial Science and Technology (AIST) [email protected]

Summary. The time for reflections and visions about 4G is getting closer to the X hour, therefore the scientific community has to finally declare how 4G will really look like. Besides various prophetic visions, we have to have a look into the “4G crystal ball” and attempt to pragmatically define the forthcoming system by fusing the sociological and the technological perspectives. In this chapter, we have adopted a top-down methodological approach, whose focal point stands in a user-group-centric vision of the wireless world. In this way, we have been able to elaborate some examples of usage scenarios and consequently to derive and interrelate the main key features of 4G. The latter have been then mapped into technical requirements and expectations in terms of system, services and devices, which has finally resulted in a pragmatic definition of the forthcoming generation. At this point, we have identified in cooperation the most promising enabling paradigm of such a definition. In particular, we have listed the main technical benefits of cooperation in wireless and presented a hybrid wireless network model in which cooperation can be easily exploited and successfully started up. Furthermore, we have discussed more practical issues connected to the establishment of cooperation, such as group formation through profile- and location-based attributes, how to trigger cooperation and the consequent types of cooperation established, and what is the user experience with respect to the interactions involved while setting up cooperation.

7.1 Introduction The ever-increasing growth of user demands, the limitations of 3G and the emergence of new mobile broadband technologies on the market have prompted researchers and industries to a thorough reflection on 4G. Many prophetic visions have appeared in the literature, which present the forthcoming “G” as the ultimate boundary of wireless communications without any limit in its technical potentials. However, since the failures of 3G were not only related to technical limitations but also to the cultural and social settings of its launching, a stronger intertwining of technology and society would be beneficial to develop 4G in a way that will maximize its acceptance and penetration in the market, while minimizing the down-side risk of its flop. Hence, in this chapter we characterize the next generation by presenting a combined

133 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 133–153. c 2007 Springer.

134

Simone Frattasi and Hanane Fathi

user-group-centric vision of the wireless world, which considers the user as the “cornerstone” of the generational design and emphasizes the potential of the possible social networks connected to him. Consequently, we elaborate examples of user and group scenarios from plausible sketches of people’s everyday life, which implicitly reveals the key features and the technical requirements and expectations coming along with 4G in terms of system, services and devices. While on the one hand this results in a more pragmatic and motivated definition of 4G, on the other hand it recognizes in cooperation a promising way to achieve such goals. This is illustrated in a wireless network setting by introducing the ad-coop network model, where cooperation can be enabled and exploited to achieve many technical and social benefits. Note that the cooperation we refer to does not embrace only users, as the innate human cooperative behavior can also trigger and be triggered by cooperation among terminals and networks. Indeed, cooperation is seen not only as a mean to enhance one’s social capital, but also to obtain services that would not be available otherwise, to enhance the quality of the services available, or to improve one’s reputation record in a reputation-based system. Finally, in this chapter we also list the attributes that are salient when defining group distinctions and the interactions among users when establishing cooperation; this implies new challenges and possibilities in terms of user experience. The rest of the chapter is organized as follows: Section 7.2 presents the “3G failures”; Section 7.3 introduces the prophetic visions of 4G and our methodological approach; Section 7.4 outlines examples of user and group scenarios; Section 7.5 extrapolates, interrelates and describes the key features of 4G; Section 7.6 identifies the real technical step-up of 4G with respect to 3G; Section 7.7 gives the resulting definition of 4G, introduces the ad-coop network model and highlights the potentials of cooperation; and Section 7.8 shows how wireless cooperation can be established. Finally, our concluding remarks are given in Section 7.9.

7.2 The Fall of 3G 2G was a huge success story because of its revolutionary technology and the services that it brought to its users. Besides high quality speech service, global mobility was a strong and convincing reason for users to buy 2G terminals. 3G has been launched in several parts of the world, but its low penetration in the market suggests that the success story of 2G is hard to repeat. The major reason for this assertion stands in the tremendous worldwide downturn in the economy that led to a starvation of new investments and thus to delays in technological development, service roll out, etc. In some countries with high penetration of mobile services, this was also combined with a reward policy for releasing the spectrum usage rights that did not favor the investments in the IMT-2000 infrastructure. However, besides these difficulties, the evolution from 2G towards 3G has brought only few novel additional services, which leave the business model largely unchanged and may not be enough to encourage the customers to change their equipment. This deficiency was considered too late by 3GPP, the 3G standardization body, which attempted to incorporate in the latest standards some advanced services, such as the MBMS in combination with the IMS. Nevertheless, these improvements were made without the possibility to adjust the access technology properly. Ultimately, it has to be emphasized that the limited

7 A Cooperative ID for 4G

135

success of 3G has also depended on the cultural and social settings in which the new system has been deployed. Indeed, 3G has been more accepted in Asian countries than in Europe. In Japan, teenagers share videos and books on their mobile phones in public places, whereas in Europe they experience the same exchange indoors, in their own rooms with their own TV sets or computers [13]. The Japanese success mainly lies in the ability of SPs to deploy many different services for various market segments and then to adjust their offerings according to what their subscribers found attractive. Europe’s 3G problem is, instead, a matter of “Now why would I need 3G?”. The fact is that the average European still lacks the knowledge and understanding of the services that 3G can provide and its overall capabilities in the wireless market. This is then translated into a lack of the enthusiasm that could aid its success and survival in Europe [21].

7.3 The Raise of 4G 7.3.1 Prophetic Visions Following the paradigm of generational changes, it was originally expected that 4G would follow sequentially after 3G and emerge between 2010 and 2015 as an ultra-high-speed broadband wireless network [5]. In Asia, for example, the Japanese operator NTT DoCoMo defines 4G by introducing the concept of Mobile Multimedia; Anytime, Anywhere, Anyone; Global Mobility Support; Integrated Wireless Solution; Customized Personal Service (MAGIC) [16], which mainly concentrates on public systems and envisions 4G as the extension of 3G cellular service. This view is referred to as the linear 4G vision and, in essence, focuses on a future 4G network that will generally have a cellular structure and will provide very high data rates (exceeding 100 Mb/s). In general, the latter is also the main tendency in China and South Korea [4]. Nevertheless, even if 4G is named as the successor of the previous generations, the future might not be limited to cellular systems and thus 4G should not be seen exclusively as a linear extension of 3G. In Europe, for example, the European Commission envisions that 4G will ensure seamless service provisioning across a multitude of wireless systems and provide an optimum delivery via the most appropriate (i.e., efficient) network available [14]. This view is referred to as the concurrent 4G vision. However, it does not give us the underlying methodology that could justify such a broad definition.

7.3.2 A Pragmatic Methodology to Define 4G Today, communication technologies have become something that people live with, an integral part of everyone’s life. In fact, their use cannot be separated from the rest of peoples’ lives or examined under a microscope as an isolated object. Indeed, in a broader context, developing technology for technology’s sake is limiting as it does not primarily target the final users. This is especially true for telecom industries, since they will most likely not recover their initial investments. Therefore, it would be more logical and less risky to set a goal of developing technologies that will fulfill users’ expectations by providing (and selling) new appealing services. From this point of view, users are the main actors playing on the stage of the wireless

136

Simone Frattasi and Hanane Fathi

world, although they are often unaware of and indifferent to which technology to use in order to obtain a desired service. Hence, if we consider users’ requirements secondary with respect to the technological issues, we risk unanticipated failure (e.g., WAP). Indeed, without a broad horizon obtained through an extended overview of the general problem, or with just the limited and narrow viewpoint of the technology, no one is able to predict the level of acceptance and penetration in the market of a given technology or product. Needless to say, if users are not fully considered in the development process, huge investments and enormous efforts by industry and academia may eventually be wasted. Thus, it becomes crucial to understand users, their expectations and needs, and to consider them as the “cornerstone” in the design of 4G in order to potentially turn a new technology into a big success; here also cultural and societal differences become noticeable. Furthermore, it has to be taken into consideration that novel technologies may have a significant (and unpredictable) influence on users’ behavior and, consequently, their usage may change the emerging products. Understanding the users means also understanding how the users change as the society around them changes in general and, specifically, how they change through interaction with the products that are introduced. Neither technological determinism nor societal determinism give full recognition to the complicated and dynamic interplay between individuals, society, and communications technologies. If technological developers start from understanding human needs and their societal background, they are more likely to accelerate evolutionary development of useful technology. The payoff from technological innovation is that it supports some human needs while minimizing down-side risks. Therefore, responsible analysis of technology opportunities will consider positive and negative outcomes, thus amplifying the potential benefits for society [3]. Clearly, there is a need for a new approach, there is a need for contextual understanding, and there is a major methodological challenge in the design of 4G. The methodology we propose in this chapter is a top-down approach that focuses on a combined user-group-centric vision of the wireless world, and consists of the following four steps: (1) Consideration of the user as a socio-cultural person with subjective preferences and motivations, cultural background, customs, and habits. This leads to the identification of the user’s functional needs and expectations in terms of services and products. However, to interrelate socio-cultural values and habits with functional needs is a sociological problem that is not described in this chapter [1]; (2) Reflection of the functional needs and expectations derived from Step 1 in exemplary user and group scenarios derived from sketches of people’s everyday life; (3) Extrapolation and interrelation of the key features of 4G from the user and group scenarios assessed in Step 2; they represent the basic pillars for a very relevant and pragmatic definition of the forthcoming technology; and (4) Identification of the real technical step-up of 4G with respect to 3G by mapping the key features described in Step 3 into advances in terms of system, services and devices. These technological developments are necessary to support the requirements of the different user and group scenarios defined in Step 2.

7 A Cooperative ID for 4G

137

7.4 Examples of User and Group Scenarios 7.4.1 Business on-the-Move Even before leaving home to reach the place of a work appointment, users would like to receive information about train / subway schedules, door-to-door delays, and so forth, as well as more personalized information, such as knowing how long it takes to walking to be on schedule in order to eventually wait for the next train. According to the users’ decisions, their time-plan must consequently be scheduled in the most efficient way. During their stay on the train, users would like to download e-mails, listen to radio, watch TV, and so on. Finally, before they get off the last planned train, the most time-saving exit and way to reach their final destination must be known and available in multimedia format.

7.4.2 Smart Shopping Users would like to receive pop-up advertisements informing them of an offer not only when passing by or through a shopping mall, but also anywhere else (e.g., while relaxing at home, or while on the bus / subway), where they can start thinking about their spare time. With such a service, based on users’ preferences and hobbies, the targeted advertisements become precious information: they are not as annoying as massive ones, because they result from a user request and thus they answer a real need. Hence, users can utilize those inputs to get more detailed information regarding the route and the overall cost of their planned activity.

7.4.3 Mobile Tourist Guide Tourists walking in Paris would like to use their personal devices to receive not only directions to a sightseeing place but also last-minute alerts when alternative (e.g., less congested) routes become available, or detours to other sites of interest appear en route to the sightseeing place. They can also avoid the problem of long queues at the famous museums by buying tickets via their terminals or by signing up online on the waiting list, which sends back the approximate waiting time to them. Inside the museum, instead of buying the brochure or renting an electronic guide, all they need is to download a package in their language and enjoy their tour listening to the audio guidance. For each artifact in the exhibition, they can automatically listen to the comments and explanations, without any effort of browsing through the guide. Their personal devices can also provide information about the culinary specialties of the city / region; for example, advising them about the location of a typical restaurant situated nearby.

7.4.4 Personalization Transfer In a music festival or during a concert, users would like to take pictures and record special moments with their friends by using the multimedia capabilities of their recently purchased hand-held devices. On the way back, the pleasure of watching pictures and video clips is not limited by the small screens of their devices, since they can transfer any content to publicly available larger displays – on the bus, on the train, at the airport – and fully enjoy together with their friends and the other people who were present at the event.

138

Simone Frattasi and Hanane Fathi

7.5 Key Features of 4G from the User and Group Perspectives

Figure 7.1. The user-centric system. Inspired by the Helios-centric Copernican theory, the user is located in the center of the system and the different key features defining 4G rotate around him on orbits with a distance dependent on a user-sensitive scale (see Figure 7.1): the further the planet is from the center of the system the less the user is sensitive to it. The minimum is reached when the user-centric system is translated into the technocentric system, where network heterogeneity has a much stronger impact than user friendliness. This kind of representation shows also the interdependency between key features, e.g., service personalization is a satellite of terminal heterogeneity. The user-centric system demonstrates that it is mandatory in the design of 4G to focus on the upper layers (max user-sensitivity) while tailoring the lower ones to them. For example, without user friendliness the user cannot exploit his device and have access to other features, such as user personalization. Although we are used to think that stars come as individuals, as in the Helioscentric system, this is not the norm. The evidence, instead, is that more than 85% of them are parts of multiple star systems, where each of them revolves around a common center of mass under the influence of their mutual gravitational force. Users, like stars, are seldom found in isolation. A group of users can be formed by a number of users having certain characteristics in common (e.g., hobbies, ethnicity, interests, musical preferences, etc.) or belonging to the same social group, such as family, friends and colleagues. As a consequence, the user-centric system can be extended to a group-centric system, where each “sun” comprises its own “planets” (see Figure 7.2). Moreover, as well as multiple star systems when the stars on their orbits are getting closer to each other raising up their velocity accordingly to their increasing mutual gravitational force (violet curve in Figure 7.2), the group-sensitivity, i.e., the socio-technological possibilities relative to a group of users, increases the closer the users get to each other.

7 A Cooperative ID for 4G

139

Figure 7.2. The group-centric system.

7.5.1 User Friendliness, User and Group Personalization In order to encourage people to move towards a new technology, which is a process that usually takes a long time and a great effort from the operators’ side, the combination of user friendliness and user personalization appears to be the winning concept. User friendliness exemplifies and minimizes the interaction between applications and users thanks to well designed interfaces that allow users and terminals to naturally interact (e.g., speech interfaces). For instance, in Scenario A, users can get traveling information in the most user-friendly way: text, audio, or video format. User personalization refers to the way users can configure the operational mode of their devices and pre-select the content of the services chosen according to their preferences. Since every new technology is designed having in mind as the principal aim to penetrate the mass market and to strongly impact the people’s lifestyle, the new concepts introduced by 4G are based on the assumption that each user wants to be considered as a distinct, valued customer demanding special treatment to satisfy his exclusive needs. Therefore, in order to embrace a large spectrum of customers, user personalization must be provided with high granularity, so that the huge amount of context- and location-dependent information is filtered according to the users’ flavors. This is illustrated in Scenario B, where users can receive targeted pop-up advertisements. Virtual information agents, portals, and information brokers will assist users in finding, filtering, and personalizing information. Two examples will be personal assistants that keep track of users’ meetings, tasks, alarms, or address books; and virtual guides for tourists in unfamiliar cities or museums [17]. In conclusion, the combination between user personalization and user friendliness will give to the users the idea of an easy management of the overall features of their devices and the maximum exploitation of all the possible applications, thus conferring the right value to their expense. User personalization can be easily mapped into group personalization, where services

140

Simone Frattasi and Hanane Fathi

and contents can be targeted to a specific and declared group of users according to their features and their social connections. In Scenario C, for example, a group of tourists that would like to visit a museum receives information on the current expositions and the tickets availability.

7.5.2 Terminal and Network Heterogeneity

Figure 7.3. Heterogeneous terminals.

In order to be a step ahead 3G, 4G must not only provide higher data rates but also a clear and tangible advantage in people’s everyday life. Therefore, we believe that the success of 4G will consist in the combination of terminal and network heterogeneity. Terminal heterogeneity refers to the different types of terminals in terms of screen size, weight / portability, performance, etc. (see Figure 7.3). Network heterogeneity is related to the increasing heterogeneity of wireless networks due to the proliferation in the number of access technologies and wireless infrastructures available (e.g., HAP, DAB / DVB, UMTS, WiMAX, Wi-Fi, Bluetooth, etc.). Most of these heterogeneous wireless access networks (see Figure 7.4) are packet switched, based on enhanced versions of IP (e.g., IPv6) and typically differ in terms of coverage, data rate, latency and loss rate. Therefore, each of them is practically designed to support a different set of specific services and devices. As a consequence of terminal and network heterogeneity, 4G will encompass various types of terminals, which may have to provide common services independently of their capabilities. Therefore, tailoring contents for end-user devices will be necessary to optimize the service presentation. Furthermore, the capabilities of the terminal in use will determine whether new services are to be provisioned or not, in order to offer the best enjoyment to the user and prevent declining interest and elimination of a service offering. This concept is referred to as service personalization (see Figure 7.1) and implicitly constrains the number of access technologies supportable by the user’s personal device. However, this limitation may be solved in the following ways: (1) By the development of devices with “evolutionary design”. A naive example can clarify this concept: in case the user has a watch-phone on which he would like to see a football match, just by pressing a button on the watch’s side a

7 A Cooperative ID for 4G

Figure 7.4. Heterogeneous networks.

141

142

Simone Frattasi and Hanane Fathi

self-extracting monitor with a bigger display can come out. Therefore, having the most adaptable device in terms of design can provide the customers with the most complete application package, thus maximizing the number of services supported; and (2) By mean of a “personalization transfer”. An example extracted from Scenario D can clarify this concept: in case the user has a watch-phone on which he would like to see a video, he does not need to possess larger display terminals as all the publicly available terminals can be borrowed for the displaying time. Therefore, the advantage for the customers is to buy a device on which they have the potential to get the right presentation for each service, freeing it from its intrinsic restrictions. Furthermore, in a private environment users can optimize the service presentation by exploiting the multiple terminals they have at disposal. The several levels of dependency highlighted by the user-centric system definitely stress the fact that it is not feasible to design 4G starting from the access technology in order to satisfy the user’s requirements, something that instead is widely done nowadays. A contextual understanding and a strong preliminary consideration of the user are a more relevant and pragmatic approach to the design. As previously mentioned, a user can benefit from terminal and network heterogeneity due to the availability of various networks and terminals with heterogeneous characteristics. It is straightforward to suppose that such condition is better satisfied in case of a group of users, since they can likely have access to more heterogeneous devices and networks than a single user. Indeed, the gathering of users in socialor interests-oriented groups represents an agglomeration of highly diverse terminals and increases the opportunities to have access to several types of networks. This leads each user to a virtual possession of numerous and various terminals and to an indirect access to unsubscribed networks without investing major efforts. Yet, the mobility of a group of users carrying their personal devices implies the mobility of a whole established network, like it is in case of transportation systems. This is an issue to be tackled in 4G and it is referred to as NeMo. In order to preserve service personalization, the services targeted to a group of users should be tailored to the different capabilities of the end-user devices, so that the best service presentation is delivered to each user. Hence, the same information can be delivered in multiple media formats. Finally, when a group of users is gathered together in a certain location, personalization transfer is highly relevant because it can allow users carrying devices with small screens to borrow the bigger screens of their fellows’ devices.

7.6 Technical Requirements and Expectations for 4G To cope up with the growing demand in downlink data rate, the UMTS specifications recently defined the HSDPA architecture for 3G cellular networks based on WCDMA. HSDPA is designed to provide up to 10 Mbps downlink data rate to mobile users. However, the poor coverage of a cell – caused by path-loss, shadowing and multi-path fading – is the main reason why a user may not be able to get this 10 Mbps. Indeed, since HSDPA does not employ power control, each user will always experience a location-dependent downlink data rate [23]. Hence, the only way for improving the performance of a 3G cellular network is to significantly increase the density of base stations, which nevertheless results in more severe inter-cell interference and considerably higher deployment costs. Under the working assumption

7 A Cooperative ID for 4G

143

that subscribers will not be willing to pay the same amount per data bit as for voice bits, this solution does not seem economically justifiable [20]. It is obvious from the above discussion that more fundamental enhancements are necessary for the very ambitious requirements of 4G in terms of system, services, and devices.

7.6.1 System Data rate. 4G systems are expected to support much higher data rate services compared to evolving 3G systems (up to 100 Mbps in outdoor environments and up to 1 Gbps in indoor environments). To reach such targets, multi-carrier technologies (e.g., OFDM), link layer techniques (e.g., AMC and HARQ), and multi-antenna technologies (e.g., MIMO) are the most promising. Coverage. In Figure 7.5, we show the shift in paradigm: while 2G was focused on full coverage for cellular systems offering only one technology and 3G provides its services only in dedicated areas and introduces the concept of vertical handover through the coupling with WLAN systems, 4G will be a convergence platform extended to all the network layers. Hence, the user will be connected ‘almost’ anytime and anywhere thanks to a widespread coverage due to the exploitation of the various networks available. The system or the terminal will always choose to access the best network available by considering bandwidth demand, time, cost, or any other QoS parameter. In particular, the service provision will be granted with at least the same level of QoS when passing from one network to another. The common IP-based infrastructure will facilitate this seamless mobility across networks. Finally, global roaming should also be enabled for users traveling across the world. Energy consumption. The increasing need for power to drive computers, routers, and battery-powered devices is an environmental problem on a global scale. In particular, 4G users will demand long usage time both in standby mode and while using their terminals. Unfortunately, the battery drain is a chronic problem of wireless devices and the battery technology is not progressing at comparative pace. This is evident when considering that whereas 2G phones were shipped out with one battery, 3G phones are shipped out with two batteries. If we follow this escalation, the energy consumption will then increase proportionally to the more advanced services. Hence, 4G needs to modify the 3G rule E3G ∝ QoS3G into E4G ∝ 1/QoS4G , where E represents the energy consumption. The strong environmental concern among the public will also push the industry to radically decrease the transmitted power and thus the energy consumption, due to the possibly real and perceived risks concerning electromagnetic radiation. Therefore, all BSs and APs will be designed to meet very low limits on radiation levels. In urban areas these solutions will involve low-power, wide-band, and short-range communications inside a pico- and micro-cells instead of the old macro-cellular approach [17]. Spectrum usage. The resource sharing among the various networks available will smooth the problem related to the spectrum limitations relative to 3G [24], where the interworking between networks operating in different bands will result in a larger spectrum usage. Nevertheless, 4G should also use the different spectrum bands efficiently in order to optimize the capacity of the system – this can be obtained, for example, via the use of cognitive radio techniques [6].

144

Simone Frattasi and Hanane Fathi

Figure 7.5. Generational evolution from 2G to 4G.

7 A Cooperative ID for 4G

145

Location estimation accuracy. Even though the GPS is the most popular solution on the market, the introduction of handsets with built-in GPS receivers leads to an increased cost, size, battery consumption, and a long time for a full market penetration [22]. Furthermore, it is sometimes unfeasible in dense urban environments to obtain any sort of location information due to the impossibility of having a clear view of at least four satellites, or due to signal blocking and multipath. Terrestrialbased technologies have the same drawback of the GPS in multipath environments and in NLOS conditions, when no accurate environmental information is available. Hence, investigations have started in connection with 4G in order to define a solution – probably integrated with the GPS –, which would be able to provide location information with a high level of accuracy anywhere and anytime.

7.6.2 Services Services are heterogeneous in nature (e.g., different types of services such as audio, video, pop-up advertisements, etc.), quality, and accessibility. In fact, at a certain time and place, the quality of and the accessibility to a service may not be the same due to the intrinsic heterogeneity of the network. For instance, in Scenario B, users in proximity of the shopping mall but outside the coverage of a WLAN can still receive pop-up advertisements by exploiting a possible multi-hop ad-hoc network in their surroundings. Therefore, thanks to the dynamics of the network environment, which can change in topology, number of users and terminals, 4G can maximize the probability to provide users with the requested connectivity. Therefore, contrarily to the previous generations, where the service is either constant S2G ∝ const or only place-dependent S3G ∝ f (place), 4G will provide services that will be function of time, place, terminal, user, and number of users: S4G ∝ f (time, place, terminal, user, number of users), where the dependency on terminal is due to terminal heterogeneity and service personalization, and the dependency on user and number of users is due to user and group personalization. Apart from some soft additional emerging services, such as fast Internet connection and pop-up advertisements, there is still a lack of really distinct services that will enable new applications with tangible benefits for the users. The real advantage that 4G will bring in terms of services will be due to the integration of technologies designed to match the needs of different market segments. For example: (1) Short-range wireless technologies, such as Wi-Fi and Bluetooth, will enable M2M communications, such as in Scenario C, where users sign up online on the waiting list, which sends them back the approximate waiting time, or in Scenario D, where users can transfer any content to publicly available larger displays. In particular, from the sociological point of view, in the latter case the private and the public sphere are definitely mixed. This re-combination can result in the enhancement of the public access, where the access to displays will be as common as the access to public telephone booths is nowadays [9]; (2) Since 3G networks are not able to deliver multicast services efficiently or at a decent quality, broadcasting technologies, such as DAB and DVB, will open the possibility to provide mobile users with interactive or on-demand services – so called IP datacasting –, and audio and video streaming in a much more efficient way than using the point-to-point switch network [2]; and (3) The embedding in the user terminal of a global positioning solution will offer

146

Simone Frattasi and Hanane Fathi

the essential feature of location-awareness necessary to provide users with the most comprehensive and extensive level of information, thus bringing the real revolution in terms of personalized services. Hence, the user terminal will offer not only location information, such as maps and directions, to reach a specific place, but also useful information relevant in time and space, such as pop-up advertisements concerning offers in shops nearby (see Scenario B) or multimedia data aiding guided tours in museums (see scenario C). Finally, it is worth highlighting that though users are attracted by high data rates, they would certainly be even more attracted by useful services exploiting high data rates. The support of imaging and video as well as high quality audio gives SPs a myriad of possibilities for developing appealing applications. These features, blended with the support of high data rates, result in a particularly attractive combination. Indeed, in addition to an explosive increase in the data traffic, we can expect changes on the typically assumed downlink-uplink traffic imbalance. Data transfer in the uplink direction is expected to increase considerably and ultimately, as a result of these trends, the mobile user will become a content provider (CP). In future wireless networks, this concept will broaden to encompass not only the conventional small- or middle-size business-oriented service companies, but also any single or group of users (e.g., shops / cinemas / theaters in Scenario B, museums in Scenario C, users themselves in Scenario D). Mobile CPs will open up a new chapter in service and content provision.

7.6.3 Devices As it is illustrated in Figure 7.3, 4G is characterized by the support of heterogeneous terminals, ranging from pen-phones to cars. However, due to its wide acceptance and usage in the past ten years, we expect that the mobile phone will be still in the next future “on the edge of the wave” of the mass market. Indeed, while the penetration of other devices will still occupy a restricted niche role (e.g., PDAs, watch-phones and pen-phones will continue to be designed for an elite of tech-savvy people), the mobile phone will still have no competitor, thanks to its size and weight that guarantees high portability. Moreover, due to the casual and informal feeling it gives, people will pay more attention to the pop-up advertisements / news / events they receive on it than on any other device. In general, due to the increasing network heterogeneity, the future trend of wireless terminals will move towards: •

•

Modularized devices: terminals with clearly defined interfaces between modules, for example radio unit, input / output unit, screen, and keyboard. The different modules will be even sold as separate components and then used to create individual device systems. Many of these cheap and service-dedicated wireless components will be integrated into fashion items like clothes or accessories or into portable computers and PDAs, and will be managed as BANs [17]; Multi-mode / reconfigurable devices: terminals able to access the core network by choosing one of the several access networks available and to initiate the handoff between them without the need for interworking devices (see next bullet). This leads either to (1) Multi-modality: integration of different access technologies in

7 A Cooperative ID for 4G

•

147

the same device; or (2) Reconfigurability: capability of dynamically adjusting the hardware features via software, in response to dynamic changes in the wireless environment (e.g., number and type of access networks available). SDR is the key enabling technology of such a reconfigurability [12], which would also make 4G as much adaptable as possible to the various worldwide markets; Interworking devices: terminals able to bridge information from one access network to another, while guaranteeing end-to-end QoS. For example, an integrated AP performing the interworking between a WMAN and a WLAN will reduce the hardware embedded in the user terminal and the software complexity. Indeed, instead of integrating both technologies, the user terminal will only incorporate a Wi-Fi card, where the price to be paid for this relief is an increased system (infrastructure) complexity [8].

Most terminals will feature several functions that will greatly facilitate communication and use. Advance solutions will be common for voice control, touch screens, and interactive control. The same service will be provided on different input and output devices, such as a big screen on a wall, a desktop, or a laptop computer, and a PDA for people on the move. Advanced display technologies will allow the virtual size of the display to be much larger than the physical size of the screen. Self-learning devices will help the user to personalize the interface and to filter and organize information coming from various service and content providers, or the Internet. Advanced voice interfaces can eliminate buttons, so the size of the terminal can be very small [17].

7.7 Towards a Definition of 4G 7.7.1 The Ad-Coop Network Model According to all the previous considerations, we can derive that: 4G is expected to supply the increasing population of mobile users with a various range of appealing services (from pop-up advertisements to location-based and interactive or on-demand services – so called IP datacasting), which will definitely require a technological improvement in terms of data rate, coverage, energy consumption, spectrum usage and location estimation accuracy. All these characteristics will be supported by modularized, multi-mode / reconfigurable, and interworking devices. Toward this end, in addition to novel air-interface technologies and collocated antenna technologies, some major modifications in the wireless network architecture itself are required. The most promising architectural upgrade relies on the use of a combination of the cellular network model with the P2P one, which is usually used only in a special class of wireless networks called ad-hoc networks. Whereas in conventional cellular networks mobile hosts operate in a purely peer-agnostic fashion, in ad-hoc networks, they act cooperatively as routers or relays for other hosts, where communications are enabled through multi-hopping without the need for a centralized base station. By using transmission powers that are just large enough to ensure network connectivity, the ad-hoc network model achieves several performance benefits over the cellular one, including better spatial reuse characteristics and lower energy consumption [11]. It is straightforward to realize that a hybrid network model, such as the

148

Simone Frattasi and Hanane Fathi

cellular ad-hoc one, is the most natural type of environment in which cooperation not only between users or terminals, but also between networks can be established and best flourish. Such network model is thus referred to as ad-coop network model.

7.7.2 The Alchemy of Cooperation in 4G Wireless Cooperation is a raising alchemic paradigm in wireless communications, which gives the designers the potentials to achieve enhancements in terms of data rate, coverage, energy consumption, spectrum usage and location estimation accuracy [7]- [10]. These goals can be achieved either by exploiting exclusive cooperative stations (e.g., fixed or mobile RSs / APs) or short-range communications among neighboring MSs. In general, cooperation through relays has shown to have the main following benefits [20]: •

•

•

•

While a cellular base station is assumed to cover a region of diameter [2-5] km, a relay is only be expected to cover a region of diameter [200-500] m. This means that the transmit power requirements for a relay are significantly reduced compared to those for a base station, which permits an economical design of the amplifier. Furthermore, the mast on which a relay is placed does not need to be as high as for a base station, thus reducing operating expenses such as tower leasing and maintenance costs for the SP; A relay does not have a wired connection to the backhaul. Instead, it stores the data received wirelessly from the base station and forwards them to the mobiles in its range, and vice versa. Thus, the costs of the backplane that serves as the interface between the base station and the wired backhaul network can be eliminated; If the density of relays in a cell is moderately high, most mobiles are significantly closer to one or more RSs than to the BS. This means that the propagation loss from the BS to such mobiles is larger than from a nearby RS. This results in higher data rates on RS-MS links than BS-MS links, thereby potentially solving the coverage problem for high data rates in large cells, but also reducing the energy consumption of the terminals served by relays; RS-MS links can use a different (unlicensed) spectrum (e.g., IEEE 802.11x) than BS-MS links (licensed spectrum), yielding significant gains in terms of spectrum usage and load balancing. Moreover, the delivery cost of the service can be reduced.

Most of these benefits can be also achieved by cooperation through mobiles. In this case, [10] shows that P2P communications can be exploited in a mesh fashion within established clusters for cooperative localization purposes, which gives the potential to enhance the location estimation accuracy with respect to traditional wireless location techniques in cellular networks. In conclusion, all the aforementioned benefits can be summarized in the motto “better services at lower prices”.

7 A Cooperative ID for 4G

149

7.8 Enabling Wireless Cooperation 7.8.1 Group Formation Group formation is one of the fundamental pillars to establish cooperation in the society. It might be performed based on several attributes with the aim of supporting interpersonal awareness, enable communication, and identify previously unknown affinities. Specifically, when it is performed by using a combination of profile- and location-based attributes, it results in an opportunistic cooperation, which is the most suitable for wireless cooperation purposes, as it can only take place at a certain time and place. Note that the collection of these various selective attributes may raise privacy issues. Therefore, users have to previously agree on making their personal information partially or fully publicly available.

Profile-Based Attributes The user profile may include a wide range of attributes, which reflect personality, background, ethnicity, general attitudes, choices, etc. Specifically, they can be grouped as (1) User intrinsic characteristics: age, sex, marital status and job. They can also be related to the user background, such as family origins, education and work experience. These attributes do not vary much in time and, therefore, are quite reliable but very basic ones; (2) User preferences and interests: music, movies, hobbies, sports, cultural activities, food preferences, etc. They can be represented by a list of preferred genres or keywords, and they can be stated explicitly by the user or inferred by information collected by the terminal on the users’ habits and customs (this is possible if the terminal is equipped with a cognitive functionality). These attributes do vary in time and, therefore, updates are to be performed rather frequently to be a good ground for group formation; and (3) User history: duration and frequency of actions and activities in which the user has taken part. For example, the user has been attending dance classes in a specific place, twice a week for a year.

Location-Based Attributes The location attribute can include absolute location, proximity, and crowding. The absolute location represents the geographical position in terms of latitude and longitude of a certain user, and can be used for group formation between individuals that have strong social ties. Proximity allows the creation of new social ties based on physical collocation of users having matching profiles. Finally, crowding is the term used to refer to the effect of people gathering at the same place due to a common attribute in their profile (e.g., music preferences or food preferences) [15]. This information can also be used to let others know about the occupancy of a certain place or to help users to coordinate interactions that reinforce existing ties. In general, the location attribute may also consider the user location history, which keeps track of places visited by the user (e.g., restaurants, shops, museums, neighborhoods, parks, etc.). This can be exploited to form groups of users with high personal affinities and common geographical routines.

150

Simone Frattasi and Hanane Fathi

7.8.2 Cooperation Triggers and Types of Cooperation The first trigger is to consider cooperation as a mean to improve one’s social capital1 . Users accept cooperation with other users, because this permits them to create new social ties by meeting new acquaintances with whom they share common interests. This is actually one of the most important motivations behind P3 systems [15], which strengthen the relationship between social networks and physical places. We refer to this type of cooperation as ad-lib cooperation. The second trigger is to consider cooperation as a mean to obtain services that are not available otherwise or to improve the quality of the services offered (from the SP point of view) or requested (from the user point of view). This trigger implies a cooperation rather technical than social, where proximate terminals exploit the benefits of short-range communications. We refer to this type of cooperation as on-demand cooperation. The third trigger is to consider cooperation as a mean to improve the reputation record in a reputation-based system, where reputation is defined as the collaborative level of a node in participating in a certain protocol (e.g., routing and forwarding in mobile ad-hoc networks). Thus, reputation provides a basis for the choice of prospective transaction partners. With a good reputation record, a certain terminal can expect the surrounding terminals to relay its information or inbound messages, thanks to its publicly available reputation record. In this way, it is also possible to find out about selfish nodes and isolate them by denying them service. This isolation has the purpose to stimulate and encourage cooperation in general, and to reduce selfish behavior by depriving the non-cooperative node of the opportunity to participate in the network when highly needed. The isolation can be done by each node autonomously, without consensus or human intervention. As a consequence, in order to avoid isolation, a node is motivated to cooperate with others to improve its reputation record even if the current cooperation is not immediately beneficial for it [18]. We refer to this type of cooperation as reputation-based cooperation.

7.8.3 The User Experience According to the type of cooperation to be established, the level of the interactions involved may vary. As a consequence, the user experience may differ from case to case, and we can mainly refer to: •

1

Perceived cooperation: it involves both users and their terminals. Terminals help identifying previously unknown affinities between users in the same location and enabling communication between them, where potential cooperative types can be made aware of the presence of a matching profile through pop-up advertisements. Hence, an ad-lib cooperation can be established, where new social ties can be created by supporting interpersonal awareness.

Social capital can be defined as the aggregate of the actual or potential resources, which are linked to the possession of a durable network of relationships of mutual acquaintance and recognition [19].

7 A Cooperative ID for 4G •

151

Transparent cooperation: it involves only terminals. Users are not made explicitly aware of the cooperation, as there is no generation of social capital, but it only takes place to provide some technical benefits. Hence, an on-demand or reputation-based cooperation can be established. Note that this type of user experience can be exploited by vendors, unless any user-perceived parameter (e.g., energy consumption) is evidently varying during the cooperation time.

7.9 Conclusions In the first part of this chapter, we have adopted a top-down methodological approach, whose focal point stands in a user-group-centric vision of the wireless world. In this way, we have been able to elaborate some examples of usage scenarios and consequently to derive and interrelate the main key features of 4G: user friendliness, user and group personalization, and terminal and network heterogeneity. The latter have been then mapped into technical requirements and expectations in terms of system, services and devices, which has finally resulted in a pragmatic definition of the forthcoming generation: 4G is expected to supply the increasing population of mobile users with a various range of appealing services (from pop-up advertisements to location-based and interactive or on-demand services – so called IP datacasting), which will definitely require a technological improvement in terms of data rate, coverage, energy consumption, spectrum usage and location estimation accuracy. All these characteristics will be supported by modularized, multi-mode / reconfigurable, and interworking devices. In the second part of this chapter, we have identified in cooperation the most promising enabling paradigm of the aforementioned definition. In particular, we have listed the main technical benefits of cooperation in wireless and presented a hybrid wireless network model in which cooperation can be easily exploited and successfully started up. Furthermore, we have discussed more practical issues connected to the establishment of cooperation, such as group formation through profile- and location-based attributes, how to trigger cooperation and the consequent types of cooperation established, and what is the user experience with respect to the interactions involved while setting up cooperation. Contrarily to the allegedly soft dimensions of culture and society that had proved to be hard facts in the case of 3G, the socio-technical approach that we have adopted hopefully indicates a more successful method of technological development, where an increase in the social cooperative behavior of the users may be translated in a regained enthusiasm that could aid 4G success and survival. The innate human tendency to cooperate and operate in groups is shaping the future wireless communications systems. It is therefore relevant to wonder if there are other human behaviors that can inspire the forthcoming technologies. The human faculties, which we recognize in cognition and contextual awareness, seem to be promising concepts to be extrapolated into terminals and systems. An open question remains: How far can technology get inspired by human behaviors and faculties?

References 1. Gimmler A. 4g – social aspects of the next generation of communications technologies – an analytical-explorative report. Deliverable D1.4, JADE, July 2004.

152

Simone Frattasi and Hanane Fathi

2. Fernandes B. Combining the best of dvb-t and umts: Potentials and opportunities for interactive mobile multimedia. Proceedings of Workshop on Multiradio Multimedia Communications (MMC), November 2000. 3. Shneiderman B. Leonardo’s Laptop – Human Needs and the New Computing Technologies. MIT Press, October 2002. 4. Bauer F and et al. Synthesis report on worldwide research on 4g systems. Deliverable D7.1, IST MATRICE, September 2003. 5. Bohlin E, Lindmark S, Bjrkdahl J, Weber A, Wingert B, and Ballon P. The future of mobile communications in the eu: Assessing the potential of 4g. ESTO Publications, February 2004. ˆ 6. Cabri´ c D, Mishra SM, Willkomm D, Brodersen R, and Wolisz A. A cognitive radio approach for usage of virtual unlicensed spectrum. Proceedings of the 14th IST Mobile and Wireless Communications Summit, June 2005. 7. Fitzek FHP and Katz MD. Cooperation in Wireless Networks: Principles and Applications. Springer, 2006. 8. Frattasi S, Cianca E, and Prasad R. An integrated ap for seamless interworking of existing wman and wlan standards. Springer – Wireless Personal Communications, Special Issue: Increasing Efficiency in Broadband Fixed Wireless Access Systems: from Physical to Network Layer Solutions, 36(4):445–459, March 2006. 9. Frattasi S, Fathi H, Gimmler A, Fitzek FHP, and Prasad R. Designing socially robust 4g services. IEEE Technology and Society Magazine, 25(2):51–64, Summer 2006. 10. Frattasi S and Monti M. Ad-Coop positioning system (acps): Positioning for cooperative users in hybrid cellular Ad-Hoc networks. Wiley – European Transactions on Telecommunications, to appear in 2007. 11. Hsieh HY and Sivakumar R. On using peer-to-peer communication in cellular wireless data networks. IEEE Transactions on Mobile Computing, 3(1):57–72, January-February 2004. 12. Hui SY and Yeung KH. Challenges in the migration to 4g mobile systems. IEEE Communications Magazine, 41(12):54–59, December 2003. 13. Ito M and Daisuke O. Mobile Phones, Japanese Youth, and the Re-Placement of Social Contact. Front Stage – Back Stage: Mobile Communication and the Renegotiation of the Social Sphere Conference, June 2003. 14. Pereira JM. Fourth generation: Now, it is personal. Proceedings of the 11th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2:1009–1016, September 2000. 15. Jones Q and Grandhi SA. P3 systems: Putting the place back into social networks. IEEE Internet Computing, 9(5):38–46, September-October 2005. 16. Murota K. Mobile communications trends in japan and docomo’s activities towards 21st century. Proceedings of the 4th ACTS Mobile Summit’99 (AMOS), June 1999. 17. Karlson B, Bria A, Lind J, L¨ onnqvist P, and Norlin C. Wireless Foresight: Scenarios of the Mobile World in 2015. Wiley, August 2003. 18. Mandalas K, Flitzanis D, Marias GF, and Georgiadis P. A survey of several cooperation enforcement schemes for manets. IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), pages 466–471, December 2005. 19. Bourdieu P. The forms of capital. In Handbook of Theory and Research for the Sociology of Education, Ed. J.G. Richardson, Greenwood Press, pages 241–258, 1986.

7 A Cooperative ID for 4G

153

20. Pabst R, Walke BH, Schultz DC, Herhold P, Yanikomeroglu H, Mukherjee S, Viswanathan H, Lott M, Zirwas W, Dohler M, Aghvami H, Falconer DD, and Fettweis GP. Relay-based deployment concepts for wireless and mobile broadband radio. IEEE Communications Magazine, 42(9):80–89, September 2004. 21. Forbes S. 3g services have taken japan by storm. Press Release, August 2002. 22. Sayed AH, Tarighat A, and Khajehnouri N. Network-based wireless location: Challenges faced in developing techniques for accurate wireless location information. IEEE Signal Processing Magazine, 22(4):24–40, July 2005. 23. Wei HY, Ganguly S, and Izmailov R. Routing and scheduling for adhocell downlink data capacity enhancement. Proceedings of IEEE Vehicular Technology Conference (VTC-Fall), 4:2917–2921, September 2004. 24. Zhen L, Wenan Z, Junde S, and Chunping H. Consideration and research issues for the future generation of mobile communication. Proceedings of IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), 3:1276– 1281, May 2002.

8 Implementing Cooperative Wireless Networks Towards Feasibility and Deployment

Stefan Valentin1 , Hermann S. Lichte1 , Holger Karl1 , S´ebastien Simoens2 , Guillaume Vivier2 , Josep Vidal3 , and Adrian Agustin3 1 2 3

University of Paderborn [stefanv|hermann.lichte|holger.karl]@upb.de Motorola Labs [simoens|guillaume.vivier]@motorola.com Technical University of Catalonia [pepe|agustin]@gps.tsc.upc.edu

Summary. Theory has shown that letting nodes cooperate to construct a virtual multiple-antenna array provides significant performance gains in many scenarios. Many cooperative relaying schemes were proposed using different codes and protocols. While each of these schemes has its individual benefits and employs different methods, all of them are based on common fundamental principles and characteristics. In this chapter we, firstly, provide a discussion and classification of typical state-of-the-art cooperative relaying schemes. Secondly, we focus on putting cooperative relaying into practice. In theory and practice fundamental problems have to be solved to let nodes benefit from cooperation. So far it is unclear how in mobile scenarios – e.g., cellular, mesh, WMANs, and WLANs – the optimal relaying scheme, partners, and cooperation level can be selected. Furthermore, cooperative relaying requires more complex multiplexing on the Medium Access Control (MAC), which can be realized by different cooperative patterns. We discuss several design paradigms along with their practical advantages and disadvantages, contrast current approaches, and show open issues for cooperation on the MAC sublayer. Based on this discussion we, finally, focus on implementing user cooperative systems. We discuss the state of the art and derive guidelines for implementing future wireless networks that let users benefit from cooperation.

8.1 Introduction To cope with fading channels, interference, and mobility in future wireless networks, nodes may cooperate with each other. Nodes turn into partners by cooperatively forwarding each other’s data. This approach, called cooperative relaying, takes advantage of the broadcast nature of the radio channel. Partners can overhear each other’s data and by cooperation they create a distributed multiple-antenna array. Such an antenna array utilizes several independent physical channels to form a single cooperative link. This introduces diversity in time and space making the link more robust against transmission errors caused by fading and interference. This diversity, only introduced by cooperation is called user cooperation diversity. Exploiting

155 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 155–178. c 2007 Springer.

156

Stefan Valentin et al.

user cooperation diversity requires no new antennas per node and can increase robustness, data rate, or coverage in mesh, cellular, WMAN, and WLAN scenarios. Cooperative relaying goes back to the early work of Van der Meulen [29], Cover and El Gamal [4], and Gallager [6] and currently heavily attracts the research community [5, 21]. Many cooperative relaying schemes and protocols were proposed [1, 9, 14, 19, 22, 33] and extensively studied in theory [10, 13, 15, 23], general performance bounds were derived for various scenarios [8, 34], and first implementations have been made [2, 12]. However, many fundamental questions are still open whose solution is crucial to put cooperative relaying systems into practice. In this chapter, we provide a detailed discussion of these open questions and a survey of the state of the art in solving them. More precisely, we focus on the following problems which are specific for implementing cooperative wireless networks: Choosing a cooperation scheme and its parameters: Many cooperative relaying schemes were proposed, achieving cooperation diversity by forwarding signals, symbols, code words, or packets to distribute a user’s data among several nodes and antennas. Each of these schemes provides individual benefits in specific scenarios. For example, while forwarding data packets reaches high performance with good channels between the cooperating partners, simply forwarding modulation symbols achieves higher performance if this situation reverses [26]. Furthermore, cooperative relaying schemes may introduce new parameters, e.g., defining when to transmit the partner’s data or which amount of transmission time or power to reserve for the own or the partner’s data. So which cooperation scheme should one choose for optimal performance and how should it be parameterized? Even in theory, answering these questions is not trivial since a single cooperative transmission involves multiple physical channels and nodes. Compared to a direct point-to-point transmission, this requires to study many new factors since each node and channel may be individually affected by fading, interference, mobility, or traffic characteristics. Selecting a cooperation scheme for a specific scenario requires to characterize its performance and to find optimal operating points for several schemes. Selecting partners and cooperation-aware resource allocation: The performance of the cooperative link depends on the states of the utilized channels which are inherent to the chosen partners. Hence, the link’s performance can be optimized by selecting a set of partners and the number of partners to cooperate with (which includes direct transmission as a special case). This choice may be adapted if the channel states change, e.g., due to mobility or node failures. Seen from a network-wide perspective this introduces the question which physical channels between the nodes to allocate to form a cooperative link? This question is basically a resource allocation problem. It is specific for cooperative networks that its solution requires to evaluate information which is distributed among several nodes. For example, it does not suffice to choose partners with perfect channels to the destination if these nodes cannot “hear” each other. Unlike a point-to-point link where only source and destination are involved, with cooperation the CSI between source, destination, and partners has to be considered jointly. Another problem specific to cooperative networks is the allocation of resources among the partners. How the amount of shared resources, e.g., transmission power, OFDM subcarriers, and transmission time is distributed among the partners is a further unsolved question.

8 Implementing Cooperative Wireless Networks

157

All these questions, how optimal partner sets can be found, how many partners are beneficial, how resources are efficiently distributed among the partners, and based on which information and metrics this is decided, are currently actively discussed in the research community. Cooperative MAC: Based on the current partner selection and parameters of the cooperation scheme, a cooperative MAC has to make sure that during a cooperative transmission all resources forming the cooperative link are reserved. There are several new paradigms for designing MAC protocols for cooperative networks. For example, the intense usage of broadcast for transmitting data and control information, the required spatial separation of the partners to profit from spatial diversity, the distributed information relevant for resource allocation, and time-variant selection of partners and the proportion between own and partner’s data. These and further specifics of cooperative relaying have to be taken into account when designing efficient MAC protocols for cooperative wireless networks. Implementing cooperation: Deploying cooperative relaying schemes requires to integrate them into future mesh, WMAN, WLAN, or cellular network standards. These amendments to the standard’s PHY and MAC have to be designed with respect to the limitations of the specified systems and to legacy issues. However, before cooperative networks get into everyday life a fundamental question has to be answered: What is the performance gain in a realistic scenario with limited development and manufacturing cost? Prior to production this question can only be studied by prototyping, i.e., the construction of testbeds for cooperative wireless networks. The more these testbeds are based on widespread communication standards, e.g., IEEE 802.11 or UMTS, the more relevant are the answers of these studies. The above problems show that distributing the data among multiple nodes is not sufficient for implementing cooperative wireless networks. In addition, new cooperative functions at multiple layers of the protocol stack are required. To provide a systematic survey on these issues and show early approaches this chapter is structured as follows. In Section 8.2 we introduce and classify state-of-the-art cooperative relaying schemes. In Section 8.3 we discuss open issues and current approaches in implementing the functionality required with cooperation under practical considerations. Finally, in Section 8.4, we focus on prototyping and introduce current cooperative wireless testbeds.

8.2 Approaches in User Cooperative Diversity Based on the simple relaying approach, many user cooperative relaying schemes were developed to increase robustness, range, or data rate by user cooperation diversity. While each of these schemes has its individual benefits and employs different methods, all schemes are based on common fundamental principles and characteristics. In this section these fundamentals are discussed and a classification of typical cooperative relaying schemes is provided.

158

Stefan Valentin et al.

(a) Simple relaying

(unicast)

(b) Broadcast relaying

(c) User cooperative relaying

Figure 8.1. Three basic 3-terminal relay scenarios: Unicast relaying and, specific for the wireless case, broadcast relaying and user cooperative relaying. Each figure shows the utilized half-duplex channels in the first (solid line) and second phase (dashed line).

8.2.1 From Relaying to User Cooperation Diversity Figure 8.1(a) illustrates simple relaying as the most basic cooperative scenario in the unicast case. Here, a nearby terminal, called relay r, forwards the data of source s to the destination d. As illustrated, this is done by transmitting the data between three nodes via two half-duplex channels. Although this scenario is rather simple, it includes two basic elements of more complex cooperative relaying schemes. At first, relaying requires two time phases. In the first phase (solid line in Figure 8.1) the relay has to receive the data from the source, then, it forwards the source’s data to the destination (dashed line). The second basic characteristic is that a relay permanently or temporarily lends its channel to other nodes. In this example, the relay lends its channel to the source during the second phase. Channels to the destination are called uplink channels while we will call the channels between the source and relay inter-node channels. However, simple relaying ignores one specific attribute of the radio channel – its broadcast nature. This was taken into account by Van der Meulen [29], Cover and El Gamal [4], and Gallager [6]. In their early work, they extended the above simple unicast relaying by a broadcast transmission (Figure 8.1(b)). Assuming that relay and destination are in range, in the first phase the source’s data equally reaches the relay and the destination before it is conventionally relayed in the second phase. Compared to simple relaying, this broadcast introduces a redundant transmission in the first phase via the so-far unutilized (s, d) channel. If this transmission is affected differently by fading then diversity is introduced. Here, this is the case if the two uplink channels (s, d) and (r, d) fade independently in both phases. Based on broadcast relaying, Sendonaris et al. proposed the concept of user cooperation diversity [22] for wireless scenarios. User cooperation diversity refers to spatial and temporal diversity achieved only by cooperative relaying between multiple users. As with broadcast relaying, here nodes broadcast and relay data via independently fading channels. However, cooperative relaying introduces two fundamental differences. Firstly, there are no dedicated relays. As opposed to conventional relaying, a node’s channel is not exclusively used by its own or the source’s data. Each of the nodes temporarily acts as source or relay thereby turning an asymmetric source-relay pair into a symmetric pair of partners. This leads to the second fundamental difference: With user cooperative relaying, each of the partners makes use of a broadcast channel and, hence, lets its partner profit from user cooperation

8 Implementing Cooperative Wireless Networks

159

diversity. Compared to broadcast relaying (Figure 8.1(b)) this enables to use a further inter-node channel, i.e., channel (2, 1) in Figure 8.1(c), introducing additional diversity.

8.2.2 Current Approaches – A Classification The concept of user cooperation diversity has inspired the development of many cooperative relaying schemes. Such a cooperative relaying scheme consists of a protocol defining its basic relaying procedure and makes use of further functions, e.g., modulation, coding, or FEC schemes. In their seminal papers [14,15] Laneman et al. proposed several cooperative relaying protocols and provided first capacity bounds. Based on this pioneer work and on the current state of the art [3,10,17,26] we derive the following criteria to classify cooperative relaying schemes: Regenerative/non-regenerative: Cooperative relaying can be done with or without regenerating the partner’s bits. With non-regenerative schemes partners only work at signal or symbol level and do not estimate the source’s data bits prior to forwarding. Such a non-regenerative scheme is Amplify-and-Forward (AaF) originally defined in reference [14]. During the first phase of AaF, a node receives a degraded version, e.g., by fading and noise, of the partner’s signal, and amplifies and retransmits it in the second relaying phase. At the destination the two received signals are combined and the source’s data is regenerated. The Compress-and-Forward (CaF) approach, initially suggested in Theorem 6 of [4], works similarly. CaF detects the modulation symbols from the partner’s signal. In addition to AaF it performs some processing one the partner’s message. Prior to forwarding the detected symbols are compressed in order to reduce redundancy to an amount specified by the phase length. The benefit of these two non-regenerative schemes comes from their simplicity. Since no regeneration of bits is done it is only left to the destination to estimate the bit values and to decide if these values are correct. Due to relaying, the destination can take two versions of the source signal or symbol into account. If these versions were transferred via independent (r, d) and (s, d) channels user cooperation diversity can be exploited and the probability of wrong bit estimation decreases. This may lead to correct detection of the source’s data at the destination even if this is not possible at the relay, which can consider only the single (s, r) signal or symbol version. Non-regenerative schemes perform best with bad (s, r) channels [26]. However, their performance suffers from error-propagation in other scenarios. For example, in scenarios with a sporadically erroneous (s, r) channel the signal extracted from a bad (s, d) channel may not help the destination to correctly decode the data. However, using a regenerative scheme in this case the relay may be able to estimate the original bit values or correct several bit errors by FEC decoding. With a medium quality (s, r) channel, the resulting packet may contain fewer errors than prior to detection and decoding. Finally, the relay forwards this packet to the destination which prevents error-propagation. Assuming a good (r, d) channel at the destination this packet can be reconstructed and the erroneous copy extracted from (s, d) may still provide additional diversity gain. In the simplest regenerative scheme, Decode-and-Forward (DaF) [14] the regenerated data is decoded, localized errors may be corrected, and, finally, this

160

Stefan Valentin et al.

data is repeated by forwarding it using the same code as the source. In such a fixed DaF scheme the success of the error correction is not checked, e.g., based on a FCS. Hence, even erroneous data may be forwarded. Similar to the non-regenerative schemes this leaves it up to the destination to consider two independently faded versions of the message. However, with DaF one of the received packets has passed additional estimation and FEC decoding at the relay. In several scenarios, this may reduce the final error rate at the destination, even if the relay does not react to remaining errors in the decoded bit stream. Adaptive/non-adaptive: Instead of always forwarding the regenerated data as with fixed DaF, prior to relaying a partner could check the decoding result for errors and react to it. If decoding has failed it may decide to avoid error propagation by remaining silent or transmitting its own data. Further, it could switch to a more robust modulation or code, a different cooperative relaying scheme (e.g., AaF or CaF), or even let the source or another relay take over its second phase. This functionality is defined in the protocol part of the cooperation scheme and is called selection relaying [15] or relay-adaptive cooperation. The type of knowledge used for this adaptation is for the complete cooperation scheme. For example, if the protocol adapts to the decoding result of the data the cooperation scheme is based on DaF since the relay has to interpret the data. Adaptive cooperation based on the decoding result relies on implicit CSI in terms of decoding decisions. However, even considering explicit CSI is reasonable, e.g., directly measured assuming symmetric channels or obtained via feedback at PHY or MAC level. Implicit or explicit CSI may not only come from the partner. It can even be provided by the destination, e.g., as feedback via hybrid ARQ [15]. In addition to CSI, adaptive relaying may also take routing or traffic information into account. Finally, the relay may not always be the ideal place to make this decision. Although so-far only relay-adaptive protocols are studied for cellular scenarios, CSI may be easier available at the destination, i.e., the BS, which could signal a decision to the partners. In summary, we can classify adaptive relaying schemes according to (i) the place where the decision is made, which may be source, relay, or destination, (ii) the information on which this decision is based, and (iii) how it is obtained. Applied code: The question of the applied codes arises only for DaF schemes where the data is regenerated. In the simplest case, repetition coding, the same FEC code is employed at the relay and the source. However, the relay may use even a weaker, stronger, or different code. Adaptive cooperative relaying schemes make use of this, by either letting the relay decide [15] or let both partners agree on their code and rate [9]. Typical adaptive schemes employ RCPC codes [7] as assumed in [10], or a combination of several codes, for example RCPC and STC as in [11]. Re-encoding at the relay also enables to interleave the source’s and relay’s data prior to forwarding. Hence, the relay can forward a packet including a “mixture” of both partner’s data. For example this is performed in Distributed Turbo Coding (DTC) schemes [17,32] or with Network Coding (NC)-based strategies [3,30]. Cooperation level: Relay-adaptive cooperation schemes like Coded Cooperation (CC) [9] may adjust the resources among the two phases of a cooperative transmission. The proportion of resources allocated to a partner is called its cooperation

8 Implementing Cooperative Wireless Networks

161

level [10]. It may be defined in terms of bits, transmission time, transmission power, or even OFDM subcarriers [16]. For example, with CC we can define the cooperation level as α = N1 /(N1 + N2 ), where N1 stands for the total bits (including redundancy) of the first and N2 for the bits of the second node. CC adjusts α by controlling the amount of FEC redundancy and, hence, N1 and N2 separately for both partners. The cooperation level allows to optimize the required redundancy, transmission power, or other resources separately for the two phases and related channels. This optimization is based on scenario observations, e.g., on the decoding decision at the relay. We will discuss further aspects and current approaches in optimizing the cooperation level in the next section. We summarize the above discussion of schemes and criteria in Table 8.1. Here we provide a classification of common cooperative relaying schemes according to the level of regeneration, whether and on which node the relaying protocol adapts, which code is employed, and which resource is used to adjust the cooperation level. The resources for cooperation level adaptation are bit or symbol time T and transmission power P . If a resource is placed in brackets this means that its adaptation is only implicitly considered in the original scheme. Please note that this overview is only based on the cited basic variants of the cooperation schemes. There may be similarly named improved versions with different characteristics.

Table 8.1. Classification of some cooperative relaying schemes. Relaying scheme AaF [14] CaF [4] DaF [14] Cooperation diversity [23] CC [9] Space-time cooperation [11] DTC [32] Network coded diversity [3]

Regeneration

Adaptive

Code

Cooperation level

Signal Symbols Bits

fixed/relay fixed/relay fixed/relay

— — Repetition

(P ) (P, T ) (P, T )

Bits

fixed

Repetition

P, T

Bits

relay

FEC (RCPC)

T , (P )

Bits

relay

FEC & STC

T, P

Bits

fixed

PCCC

(P )

Bits

relay

Rep. & NC

(P )

Based on these criteria and on the performance evaluation in the cited literature, a system engineer may choose the most suitable scheme for given system and scenario parameters, e.g., based on the terminal’s processing power and expected channel conditions. However, in varying scenarios, e.g., with mobility, it may be even considerable to switch between several cooperative relaying schemes depending on current scenario observations.

162

Stefan Valentin et al.

8.3 Designing Cooperative Systems – New Problems and Required Functionality User cooperation diversity gains do come at a cost. The additional nodes and channels employed by the relaying scheme introduce new control problems to a network. For example, in mobile scenarios, where channel and node states change frequently, not a single scheme or partner selection can achieve optimal performance. The new problems and dependencies introduced by cooperative relaying are discussed in the first part of this section. Then, we focus on new resource allocation and Medium Access Control (MAC) functionality, required to solve these problems. We discuss current approaches and open issues in designing cooperation-aware resource allocation and MAC.

8.3.1 Mobile Cooperation In a conventional spatial diversity system, where multiple antennas are placed on a single node, the inter-node channels can be assumed to be perfect (they are, in essence, short-distance copper wires). Obviously, this is not the case with cooperative networks where spatial diversity is provided by connecting the partner’s antennas via radio channels. Here, the multiple antennas are distributed among the nodes and the radio transmission between them may suffer from noise, path loss, or fading. Compared to a direct transmission these additional channels introduce new factors which may severely affect the overall performance of the cooperative network. For example, if we consider the simplest cooperation scenario in Figure 1(c), a direct transmission only depends on the uplink channel (1, d). With cooperative relaying the state of the inter-node channels (1, 2) and (2, 1) define whether cooperation using both uplink channels is possible. Hence, even in this simple example, the performance of cooperative diversity depends on the states of three additional channels, namely (1, 2), (2, 1) and (2, d). This situation gets worse if at least one of the nodes starts to move. Mobility introduces time-selective fading and path loss, both causing the channel states to vary over time. While this aspect is not specific for cooperative networks, here it is even more relevant due to the multiple channels and nodes involved in a single transmission. For example, with independently moving partners, it is not clear whether a “good” partner remains “good” in the future. Hence, for practical relevant performance evaluation and, finally, implementing cooperative relaying systems it is highly relevant to take mobility and its parameters, e.g., the speed of a node, into account. In this section we start with a rather simple mobility scenario with fixed relays which is very relevant for cellular networks and WMANs. Later, we generalize to mobility scenarios with arbitrarily moving nodes, which may occur in WLANs, cellular, or mesh networks.

Fixed Relays and Mobile Terminals in Cellular Networks Let us consider a cellular WMAN (Wireless Metropolitan Area Network) scenario where a MIMO-OFDM system operates in 10 MHz bandwidth around 2.5 GHz. The Base Station (BS) is equipped with 3 sectors and 4 antenna elements per sector.

8 Implementing Cooperative Wireless Networks

(a) Cooperation gain (downlink)

163

(b) Non-regeneration gain (uplink)

Figure 8.2. Effective throughput ratio of regenerative cooperation vs. direct transmission for the downlink (left) and ratio of regenerative vs. non-regenerative cooperation in uplink (right) for a single cell of a typical MIMO-OFDM WMAN scenario.

A fixed cooperative Relay Station (RS) is deployed in each sector and operates in Time Division Duplex (TDD). Each RS may be located on lamp poles or roof tops, and benefits from LOS propagation to the BS and consequently from a high SNR. The mobile terminals are assumed to be not in LOS of each other, which is typical in urban and sub-urban environments. To study the benefit of cooperation we investigate the gain of the space-time cooperation [11] scheme vs. non-cooperative techniques, i.e., direct transmission and non-cooperative D&F. For the downlink, i.e., mobile destinations, the result is shown in Figure 8.2(a). Here, the cooperation gain is plotted as a function of the terminal location within the cell. We define the cooperation gain as the throughput ratio of the best cooperative strategy to the best non-cooperative strategy. As shown, cooperative relaying always outperforms the other techniques, but the gain is only significant (up to 40 %) in areas which are far away from the BS and RS. When looking at the uplink and mobile sources, the situation changes because now the most robust link is between the relay (RS) and destination (BS). Due to the weak internode link (mobile terminal to RS) we have to take non-regenerative schemes into account due to their superior performance in such situations. Figure 8.2(b) depicts the ratio of throughput obtained with non-regenerative vs. regenerative cooperation schemes. The regenerative strategy, e.g., D&F, is optimal around the RS, i.e., when the source-to-relay link is good enough. However, non-regenerative C&F becomes optimal in other parts of the cell with gains up to 50 %. This highlights the need for implementing multi-mode relays with adaptive strategies in order to maximize capacity in both the downlink and uplink.

Independently Moving Nodes Even in mobility scenarios with independently moving partners and destination cooperation may provide significant performance gains. We analyze this based on three mobility scenarios. In the first scenario, all stations move independently with a certain relative speed and distance to each other. Here, all 4 related channels

164

Stefan Valentin et al.

(a) Uplink channel SNR 10 dB

(b) Uplink channel SNR 20 dB

Figure 8.3. Outage probabilities vs. mean inter-node SNR Γi for the direct and coded cooperative (CC) transmission of two nodes n1 and n2 to the destination d. Shown for 2 different values of the mean uplink SNR Γu (plot), 3 mobility scenarios (line style) and 2 cooperation levels α (marker type). Simulated for overall code rate R = 1/4 and node speed v = 10 m/s (Results from [28]).

(Figure 1(c)) are independently affected by fading. In the second scenario the destination d moves. Here, only the uplink channels fade which is inspired by the following example: Consider two cooperating nodes in the same moving train. Both nodes are relatively fixed to each other but move relatively fast with respect to the destination. Hence, the channels between the nodes are good, while the uplink channels (1, d) and (2, d) may be severely faded. In the third scenario only the partner n1 moves, which is the above fixed relaying case. Here only one node n1 moves and the other (n2 ) is fixed resulting in the good uplink channel (2, d). The resulting error performance is shown in terms of outage probability Pout in Figure 8.3. We show results for all three mobility scenarios of two cooperating terminals to a single destination. We consider the Coded Cooperation (CC) relaying scheme and direct transmission of n1 to d via the faded uplink channel (1, d). The results are plotted vs. mean inter-node SNR Γi and shown for a “good” (high SNR) and “medium” (lower SNR) uplink channel. Since we assume CC as cooperation scheme we also show the effect of adjusting the cooperation level α. CC handles both partners symmetrically. Hence, it is sufficient to only show the results for one partner which we consider to always move. With considerable inter-node SNR in all shown scenarios CC performs better than direct transmission. The results for low uplink SNR (10 dB) show that the above train scenario, i.e., two fixed partners and a moving destination, performs best with equal phase lengths (α = 1/2), while increasing the length of the first phase (higher α) leads to best performance if only n1 moves. If all nodes move the optimal value of α depends on the inter-node SNR. For a high uplink SNR (20 dB) CC performs better than with low uplink SNR. This shows that cooperative diversity even helps with good channels to the destination where the error probability is low. The cooperation level α shows no effect if only n1 moves. In the other cooperative scenarios α has still an impact on the

8 Implementing Cooperative Wireless Networks

165

performance. However, with high uplink SNR it does not depend on the mobility scenario. As illustrated, choosing the optimal α depends on the uplink SNR, the inter-node SNR, and on the mobility scenario. This demands for resource allocation schemes dynamically adjusting α according to these scenario factors.

Multiple Factors and Asymmetric Mobility Scenarios While the above studies illustrate the general performance of cooperative relaying in mobile scenarios, only a small amount of scenario factors is studied so-far. However, cooperative relaying introduces further factors which may significantly affect the performance. These factors have to be studied to decide on their importance for resource allocation. Even more limiting is that the above studies, and the most studies introduced in Section 8.2, are based on symmetrical scenarios. For the simple two partner scenario this means that although the nodes may move independently both inter-node channels and both uplink channels are equally parameterized. This is done in terms of mean SNR or further parameters, e.g., node speed. For this reason, we studied the effect on the error performance (outage probability) of CC in asymmetrical scenarios, where multiple factors are varied independently. The results are shown in Figure 8.4 for the outage probability Pout of one of two partners. It turns out that although mobility is a significant factor, interestingly, the relative speed has no effect. Spatially correlated inter-node channels do not affect Pout ; correlated uplink channels result in a small performance decrease. Naturally, correlated uplinks equally affect only mobility Scenarios 1 and 2 where both uplink channels fade. Increasing the mean inter-node SNR increases the probability that cooperation is possible and node cooperation diversity may be achieved. Hence, increasing the SNR for channel (1, 2), i.e., Γ1,2 , decreases the outage probability for n1 (negative effect) due to the more frequent cooperation with its partner. This situation reverses with a better channel (2, 1) for n2 since for n1 this case is unfair. Here, a relay-adaptive CC scheme utilizes the second phase for n2 ’s data since it was not able to correctly decode the data of node 1 during the first phase. Increasing the mean uplink SNR for both nodes significantly decreases Pout . The coding parameters show only slight effects due to the small available range for the R and α levels in this scenario. Naturally, increasing the overall code rate R increases Pout due to the decreased redundancy. The effect of α is interesting. While mobility Scenario 1 and 3 profit from a longer first phase, i.e., increasing α and the redundancy in this phase, a slight performance decrease is shown if only d moves. Due to the large amount of additional factors in cooperative networks the above studies show only a fraction of the current state of the art in performance characterization. The interested reader can find further detailed studies in the references [8, 27, 28, 34].

8.3.2 Cooperation-Aware Resource Allocation The above results clearly show that with time-variant factors in mobile scenarios a single cooperative relaying scheme or parameter configuration cannot constantly

166

Stefan Valentin et al.

Figure 8.4. Effects on the outage probability Pout if n1 transmits directly or using Coded Cooperation (CC). For each of the 3 mobility scenarios the mean effect µ and the effects of the factors mean SNR Γ , node speed v, spatial channel correlation coefficient c, overall code rate R, and cooperation level α are shown. If possible the factors are separately configured per channel (Results from [27]).

Figure 8.5. The optimization function fOpt maps observable scenario factors to controllable system parameters in order to reach a given optimization objective.

reach optimal performance. Here, dynamic schemes are required which adapt parameters to the current state. In this section, we focus on factors and parameters which are specific for cooperative relaying, e.g., the cooperation partner, relaying scheme, code, or cooperation level. As illustrated in Figure 8.5, a general resource allocation scheme maps controllable system parameters to observable scenario factors. This is done by an optimization function fOpt according to a given optimization objective. Hence, in general a resource allocation scheme consists of an optimization function or algorithm and methods to define the optimization goal, to measure observables, and to adjust controllables. Which observables and controllables are employed and how accurate they have to be measured or controlled, firstly, depends on the optimization objective. Typical objectives are, for example, maximizing the effective throughput of a node, minimizing the latency of transmitting a packet, and maximizing the number of supported nodes per cell at given throughput or latency. Secondly, not every scenario factor can be measured and not every system parameter can be controlled at the required accuracy. Hence, observables and controllables are only subsets of the existing scenario factors and system parameters (Figure 8.5). For example, without accurate

8 Implementing Cooperative Wireless Networks

167

node locations it may not be easy to always choose the closest neighbor as a cooperating partner. In this case an observable which is easier to obtain may serve as metric, e.g., the mean SNR of the channels to the neighbors. Designing such optimization schemes requires to analyze the effect of the observable factors and the adjusted parameters. While studying scenario factors has been discussed in the earlier parts of this section, we now focus on the question how observations of these factors can be used in resource allocation schemes. Based on the observables in cooperative networks we can classify approaches for resource allocation schemes as follows: Channel State Information (CSI)-based allocation: As discussed, factors introduced by the channel have an enormous effect on the performance of cooperative diversity schemes. Due to its frequent changes, in fading channels the instantaneous channel state cannot be directly considered as a decisive metric for selecting the appropriate cooperation scheme or parameters. Hence, practical resource allocation schemes have to rely on long-term observations of the channel, e.g., the mean SNR. Since in most systems the mean SNR of a channel can be measured easily, e.g., via the preamble of a MAC frame, this provides an important metric for the optimization decision. A simple approach for cooperation level adaptation is proposed in reference [28]. This scheme relies only on mean SNR and provides a simple method to select partner, cooperation level, and transmission power to reach a given error probability bound. A further resource allocation approach based on the mean SNR for cooperative networks is proposed for cooperative networks which perform dynamic OFDMA subcarrier allocation [16]. Here, standard graph theory is used to find the best combination of source, relay, and OFDM subcarrier in order to maximize the capacity per subcarrier. Furthermore, heuristics to averageout the utilization among the nodes are proposed to prevent exhaustion of the relay’s batteries due to frequent usage. Further channel-related factors are spatial and autocorrelation of the channel. However, observing these factors is non-trivial and early studies show that, compared to the mean SNR, their effect on the performance is negligible [27, 28, 34]. Position/topology-based allocation: In reference [18] Lin et al. proposed a cooperation partner selection scheme based on geographical information. With known node locations, e.g., obtained via GPS, the distance can be considered for optimizing partner selection and/or cooperation level adjustment. The knowledge of topology information is also the basis of the centralized partner selection scheme introduced by Shi et al. [24]. Here the access point aims to allocate optimal partners to minimize the maximum outage probability for each nodes in the cell. Traffic-based allocation: Even the traffic type may be considered for adjusting cooperation parameters. For example, in reference [31] Xu et al. combine cooperative coding with code rate allocation according to multimedia traffic priority. In practical cooperative systems, the traffic class may be more relevant since it defines the optimization goal. For example, while with non-real time traffic, e.g., downloading a web page or file, the optimization goal is to maximize the data rate, with real-time traffic, e.g., Voice over IP (VoIP) telephony, latency and packet loss have to be minimized.

168

Stefan Valentin et al.

Figure 8.6. Hidden nodes at cooperation partners may decrease the user cooperative diversity gain through unwanted interference.

While one approach is the exact consideration of only one factor, e.g., the geographic positions of nodes, in some scenarios considering multiple factors for a single decision may be more feasible. For example, in the scenario “moving train” nodes may only cooperate with relatively fixed partners within the same train. As shown in Figure 8.3 this ensures high diversity gain which may be required to reach a base station outside of the train. Additionally, traffic information may be exploited. For example, a node performing VoIP telephony, may only select partners located in the train without real-time traffic constraints, e.g., downloading a web page.

8.3.3 Medium Access Control Cooperation requires additional functions at the MAC sublayer. Throughout the following discussion we point out practical implications related to the popular IEEE 802.11 standards family. Hidden terminals: The better a cooperation partner can decode the signal sent from the source, the more user cooperative diversity can be gained. Unfortunately, nodes in the proximity of the partner become new hidden nodes and, if not silenced by the MAC, may cause unwanted interference at the partner. Consider the example topology depicted in Figure 8.6. Let us assume collision avoidance based on RTS/CTS exchange. With the RTS packet the source indicates that it has data to send to the destination and silences all nodes in its proximity. Likewise, the CTS packet from the destination confirms the transmission request and silences all nodes in the proximity of the destination. The cooperation partner receives both RTS and CTS packets, but some nodes in the proximity of the cooperation partner might not be aware of the transmission. Consequently, these nodes may interfere at the partner, thus preventing it from cooperation. We conclude that the presence of a cooperation partner enlarges the communication range, for which a conventional RTS/CTS exchange does not suffice. Choosing an opportunistic strategy, we may ignore the problem and hope for the best. We must then be aware that the user cooperative diversity gain may decrease, which becomes more significant the more users participate. Instead, a cooperative MAC sublayer may extend the simple RTS/CTS exchange to a three-party exchange that also includes the partner. This requires additional

8 Implementing Cooperative Wireless Networks

169

Figure 8.7. Partner selection according to channel qualities.

MAC complexity and, therefore, overhead, but likely benefits the user cooperative diversity gain. Partner selection: When the channel quality between source and destination is superb, cooperation would waste time, frequency, and space that could otherwise be used for communication. Consequently, we need metrics to decide whether cooperation is needed for any particular transmission and we must implement the decision somewhere. Basically, three different approaches exist. Firstly, the source evaluates the current channel condition and actively requests a partner for cooperation (i.e., let the source decide). Secondly, the destination evaluates the current channel condition and actively requests a partner for cooperation (i.e., let the destination decide). Both are active approaches. In a third approach, the source has faith that neighboring nodes have some notion about the wireless channel and cooperate when necessary (i.e., let the neighbors decide). In this case, both source and destination remain passive. All three approaches have their advantages and drawbacks with respect to implementation and performance, which we will later point out in Section 8.3.3. Partner selection can depend on several factors such as inter-node and uplink channel qualities, traffic class, and data rate. To simplify the following discussion we consider channel states as the only selection criterion. Figure 8.7 shows a topology where a source can choose between three potential partners. It also shows the assumed channel states. The direct transmission may be error-prone due to a channel having a medium quality. Therefore, the source decides to cooperate. If it were to consider only the inter-node channel, the source would select n3 as its partner. Unfortunately, that partner cannot communicate with the destination. Instead, the source should select n2 since its inter-node channel is good and the uplink channel is as well. Node n1 also has good uplink but only medium inter-node channel. For sender-initiated cooperation, the sender requires knowledge of the internode channel state as well as the uplink channel state of every potential partner. This information requires two-hop information that can only be obtained by the sender through the exchange of additional control packets. The same holds for destination-initiated cooperation. Information about both channels is easier obtained by the potential partners themselves as this only requires onehop information. This makes partner-initiated cooperation cheaper to deploy in terms of overhead, but introduces a new problem. If several partners exist, how do they coordinate themselves such that only a well-defined number of them cooperate? Both sender-initiated and destination-initiated cooperation are

170

Stefan Valentin et al. centralized approaches whereas partner-initiated cooperation requires a distributed algorithm for deciding whether or not to cooperate. In reference [1], Bletsas et al. propose opportunistic relaying as a distributed partner selection scheme without requiring global topology information. Potential partners that overhear an RTS/CTS exchange start a timer inversely proportional to a channel estimate γ. This estimate is based on the received signal strength of both packets that indicate the quality of the inter-node channel γSi (the instantaneous SNR between source node and relay i) and the partner’s uplink channel γiD (the instantaneous SNR between relay i and destination node). For regenerative relaying, e.g., DaF, γ = min(γSi , γiD ) yields an appropriate estimate, whereas γ=

2γSi γiD γSi + γiD

is more suited for non-regenerative relaying, e.g., AaF. When the timer of a potential partner expires, the potential partner first senses the medium. If it cannot detect a signal, e.g., using clear channel assessment, it broadcasts a packet to announce its help. The timeout serves as a backoff in which the node with the earliest timeout becomes the cooperating partner. Since all nodes must sense the channel before making an announcement, only one node can win the contention assuming the announcement is of sufficient duration. If potential partners may be hidden from each other, source and destination must announce the winner of the timeout period through a flag packet, therefore causing additional signaling overhead.

Cooperative MAC Patterns We now go one step further and relate the identified issues to the design of cooperative MAC patterns. Such patterns describe an ordered sequence of frames, e.g., RTS, CTS, ACK, and interframe spacings, e.g., SIFS, DIFS, that reserve the medium for a specific time span to perform cooperation. Patterns depend on who initiates cooperation, which time span needs to be reserved for cooperation, and on the cooperative mode. Here, as a first step, we concentrate on who initiates cooperation. We propose patterns for sender-initiated, partner-initiated, and destination-initiated cooperation, and we identify advantages and disadvantages of these patterns. Sender-initiated cooperation: Figure 8.8 shows a cooperative pattern for senderinitiated cooperation. Cooperative relaying is initiated by a CRTS packet. The CRTS is an extended RTS that additionally includes the address of the selected partner. Its duration field must be set such that it reserves the medium for the forwarding phase of the partner as well. The pattern exactly defines when which node has to send a packet. The partner confirms the cooperation request with an ordinary CTS, which is followed by the confirmation of the destination. To distinguish both CTS packets, the partner must use its own address as the receiver address in the CTS packet. If the partner cannot cooperate for some reason, indicated by the missing interleaved CTS, the sender can use the second

8 Implementing Cooperative Wireless Networks

171

Figure 8.8. Sender-initiated cooperative MAC pattern.

Figure 8.9. Destination-initiated cooperative MAC pattern.

phase for a redundant transmission of its own data. This is possible because the second phase has already been reserved by the duration specified in the CRTS packet. Destination-initiated cooperation: Figure 8.9 shows a cooperative pattern for destination-initiated cooperation. Here, the data transmission starts immediately after an ordinary RTS/CTS exchange. Upon receiving the data, the destination sends a RTH packet in which it specifies the address of the partner that should forward redundancy in the second phase. The RTH packet can also be used as an acknowledgment in case the destination was already able to decode the data sent during the first transmission without error. The RTH packet might then inform the sender that it may send another packet of its own data during the second phase. As an alternative, the RTH could inform the partner that it may send its own data if it wants to. In contrast to sender-initiated cooperation, destination-initiated cooperation remains compatible to the IEEE 802.11 MAC since the RTS/CTS exchange sets up the NAV for the entire cooperative transmission. There is no need for legacy stations to understand the RTH packet, whereas with sender-initiated cooperation, the CRTS packet must be understood by all participating stations. Partner-initiated cooperation (proactive): Partner-initiated cooperation can either be proactive or reactive as defined in reference [2]. In proactive schemes the partner is selected before data transmission, whereas in reactive schemes the partner is selected after transmission. Figure 8.10 shows a pattern for a proactive scheme. Here, the partner sends a CTH packet after it has overheard the initial RTS/CTS sequence. When neither sender nor destination specify a partner, the

172

Stefan Valentin et al.

Figure 8.10. Partner-initiated cooperative MAC pattern (proactive).

Figure 8.11. Partner-initiated cooperative MAC pattern (reactive). potential partners must coordinate themselves such that only one partner offers its help. Whatever the partner selection scheme may be, it directly influences the MAC pattern. When applying the distributed partner selection scheme by Bletsas et al. [1], the CTH packet serves as the announcement of a potential partner. Although opportunistic relaying seems appealing at first, it suffers from a drawback when applied to IEEE-802.11-based systems. Since partners wait for a time span inversely proportional to their channel estimates, the NAV cannot be set accurately beforehand. A possible but inefficient solution to the problem is to use a worst-case waiting time when setting up the NAV during the RTS/CTS exchange. Partner-initiated cooperation (reactive): In contrast to proactive cooperation, partners can use the data frame as an implicit CTH (see Figure 8.11). This makes the reactive pattern more efficient in time, as the CTH packet and its associated SIFS can be omitted. If the second phase should be used for a successive data transmission by the sender when no partner replies, an additional waiting time is required prior to the second phase to detect the absence of the partner’s data frame. Furthermore, in reactive schemes the correct decoding of the sender’s data at the potential partners can be used as an additional partner selection criterion. All of the above patterns assume that there always exists a second phase even if no partner exists. This simplifies pattern design as only the NAV needs to be set up appropriately to reserve the medium for the entire cooperative transmission. Such an approach is feasible both for sender-initiated and destination-initiated cooperation, which are inherently centralized algorithms. Partner-initiated cooperation tends to

8 Implementing Cooperative Wireless Networks

173

Figure 8.12. Two SORBAS SDR-based prototyping units.

be distributed, hence fixed MAC patterns are rather inappropriate. Distributed approaches can be implemented easily when another contention phase is used for the partner.

8.4 Towards Feasibility – Implementing Cooperative Systems The research community has now reached a point where the theoretical background is understood well enough to go one step beyond simulation and implement cooperative relaying systems. Two practical contributions have already been made, which are briefly illustrated and lessons learned from them are discussed. Liu et al. proposed CoopMAC as an amendment to IEEE 802.11 WLANs [20]. Every node maintains a cooperation table containing potential partners as well as an estimate of the achievable data rate based on channel qualities by overhearing ongoing transmissions. When a node has data to send, it picks a potential partner from its list and addresses both the destination and the chosen partner in the RTS frame (CoopMAC is a sender-initiated cooperation protocol) along with the estimate data rate. The potential partner replies with a HTS frame if it can sustain the rate, thereby participating in cooperation. The data frame is then transmitted to the destination via a two-hop path established through the relay instead of direct transmission. Liu et al. implemented CoopMAC using off-the-shelf WLAN interface cards on a Linux platform [19]. This approach could not fully benefit from cooperation as off-the-shelf hardware does not allow modifications of all layers to the extent needed. The physical layer is on silicon allowing no modifications (e.g., for coding or receiver combining). Powerful and general prototypes can only be built using SDRs as these provide access to all layers in a programmable way. Figure 8.12 shows the commercially available SORBAS prototyping testbed as an example. Further details about this prototyping system can be found in reference [25].

174

Stefan Valentin et al.

Bletsas and Lippman used low-cost embedded SDRs to implement opportunistic relaying [2]. They considered minimization of the overall reception power as the primary challenge to prolong battery life of nodes. With regard to this challenge opportunistic relaying is appealing. Being a proactive partner selection scheme, the partner is determined before the actual data transmission begins. From the set of potential partners only one node needs to receive while the remaining ones may become idle. Unfortunately, many questions remain open. Opportunistic relaying is indeed appealing when compared to reactive schemes in the partner-initiated cooperation domain. It remains to be shown whether opportunistic relaying can also outperform sender-initiated or destination-initiated cooperation approaches. Bletsas and Lippman did not consider these approaches in their discussion of opportunistic relaying. Furthermore, their implementation is not concerned with standard integration as opposed to the implementation of CoopMAC. Before focusing on further extensions to the Bletsas and Lippman approach, a comparison of the implemented cooperative relaying approaches like CoopMAC and opportunistic relaying is needed. Which one does perform better in practice, considering that CoopMAC achieves backward compatibility to the IEEE 802.11 standard and opportunistic relaying does not? From a manufacturer point of view, a standard-compatible cooperative relaying implementation is more suitable than a protocol designed from scratch. The following steps need to be done now for an implementation of a cooperative wireless communication system: Resource allocation: Cooperative-aware resource allocation is crucial for mobile cooperative wireless networks where node and channel conditions may change frequently. A resource allocation scheme requires an optimization algorithm and functions to derive the optimization objective, e.g., from the inspected traffic, to measure observables, and to adjust controllables, e.g. choose a partner or a cooperation scheme. In cooperative networks, these functions may be placed at multiple layers and nodes. System design has to make sure that the information flow between these nodes and layers is efficiently organized to avoid spending the received gain for control overhead. Compared to non-cooperative systems in cooperative networks this design question is even more relevant since here many nodes are involved in a single resource allocation decision. Buffering: Nodes must be able to buffer several packets belonging to a transmission, namely the original packet and all relayed packets. Buffering must be done on the symbol-level and not on bit-level when maximum ratio combining is used. Cooperative decoding: Decoding at the receiver typically operates on the bitstream of a single received packet. With cooperation, several packets, signals, or symbols (at least two) will reach the destination, namely the original and all the relayed versions. Prior to decoding these versions must be suitably combined such that the decoder can produce a meaningful output. MAC protocols: The design of a cooperative MAC protocol is based on a cooperative MAC pattern. We have already elaborated several patterns – namely source-initiated, destination-initiated, and partner-initiated – and shown their advantages and disadvantages. A cooperative MAC may also implement more than one pattern and dynamically switch between them. A switch between patterns cannot occur while a pattern is active, but it can occur at idle times. When a particular pattern yields advantages for a particular scenario, then based on physical layer information that pattern should be chosen. Other stations can

8 Implementing Cooperative Wireless Networks

175

detect the pattern by comparing the structure of the received packets with the structure defined by the pattern. Therefore, no additional control packets are needed to decide upon the patterns used. Implementing patterns: Cooperative patterns increase the complexity of the communication system in contrast to conventional ones. Also, they are more susceptible to errors simply because there exist more possibilities to invoke errors (e.g., a packet forming a substantial part of the pattern is severely corrupted and, thus, lost). Therefore, the protocol automaton belonging to a particular cooperative pattern has a lot more states and must handle a lot more scenarios compared to its conventional pattern counterpart. This complexity has to be handled efficiently.

8.5 Conclusion In the previous sections we have introduced user cooperative relaying as a promising approach to increase transmission performance in wireless multi-user scenarios. We have provided a survey and classification of cooperative relaying schemes which allow nodes to act as a multiple antenna system by sharing their antennas and time slots. In addition to the cooperative relaying scheme as such, additional functionality is required: Choosing among several schemes and their optimal parameterization (e.g., partner set or cooperation level) depending on the concrete scenario, cooperative MAC protocols (e.g., reserving the wireless channel at all cooperating nodes), legacy integration of these protocols, or efficient implementation. The following open questions at multiple layers of the protocol stack have to be answered: Cooperation-aware resource allocation: In practical systems cooperative relaying may be combined with resource allocation. This requires new functions to observe scenario factors, define optimization objectives (e.g., by monitoring the traffic type), find an optimal resource allocation, and control the system parameters. Depending on the scenario, all these functions may have to be solved under strict timing constraints and further aspects, such as fairness and traffic or user-based prioritization may have to be considered. Cooperative Medium Access Control: Conventional MAC patterns must be enhanced to accommodate the cooperative transmissions of partners. We have outlined basic differences in pattern design, depending on whether cooperation is senderinitiated, destination-initiated, or partner-initiated. Standard integration and implementation: Formalisms help us to approach cooperative relaying using algorithmic and information-theoretic approaches. Using SDRs these algorithms can be transformed easily into prototypes. Implementations of cooperative relaying schemes do not necessarily have to involve the physical layer when conventional coding is used. With the appropriate MAC pattern, a simple cooperative SaF system can be implemented by modifying the IEEE 802.11 MAC layer only. Having a cooperative MAC, physical layer extensions can be done in a next step to incorporate different cooperative modes and coding schemes. In order to provide transparent usage of user cooperation schemes, the above protocols have to be integrated into future mesh, WMAN, WLAN, or cellular network standards. These standards or amendments should focus on the parameters

176

Stefan Valentin et al.

of the PHY and the cooperation scheme, and on MAC/Data Link Control protocols rather than defining details of solving optimization or cooperation problems. This ensures inter-node compatibility, while enabling the freedom for device manufacturers to choose the integrated optimization and cooperation algorithms.

References 1. A. Bletsas, A. Khisti, D. P. Reed, and A. Lippman. A simple cooperative diversity method based on network path selection. IEEE Journal on Selected Areas in Communications, 24(3):659–672, March 2006. 2. A. Bletsas and A. Lippman. Implementing cooperative diversity antenna arrays with commodity hardware. IEEE Communications Magazine, 44:33–40, December 2006. 3. Y. Chen, S. Kishore, and J. Li. Wireless diversity through network coding. In Proc. of Wireless Communications and Networking Conference (WCNC), 2006, volume 3, pages 1681–1686, April 2006. 4. T. M. Cover and A. A. El Gamal. Capacity theorems for the relay channel. IEEE Transactions on Information Theory, 25(5):572–584, September 1979. 5. F. H. P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! Springer, 2006. 6. R. Gallager. Communications and Cryptography: Two Sides of One Tapestry. in Engineering & Computer Science. Kluwer, 1994. 7. J. Hagenauer. Rate-compatible punctured convolutional codes (RCPC codes) and their applications. IEEE Transactions on Communications, 36(4):389–400, April 1988. 8. A. Høst-Madsen. Capacity bounds for cooperative diversity. IEEE Transactions on Information Theory, 52(4):1522–1544, April 2006. 9. T. E. Hunter and A. Nosratinia. Cooperation diversity through coding. In Proc. of IEEE International Symposium on Information Theory (ISIT), page 220, July 2002. 10. T. E. Hunter, S. Sanayei, and A. Nosratinia. Outage analysis of coded cooperation. IEEE Transactions on Information Theory, 52(2):375–391, February 2006. 11. M. Janani, A. Hedayat, T. E. Hunter, and A. Nosratinia. Coded cooperation in wireless communications: Space-time transmission and iterative decoding. IEEE Transactions on Signal Processing, 52(2):362–371, February 2004. 12. T. Korakis, S. Narayanan, A. Bagri, and S. Panwar. Implementing a cooperative MAC protocol for wireless LANs. In Proc. of IEEE International Conference on Communications (ICC), June 2006. 13. G. Kramer, M. Gastpar, and P. Gupta. Cooperative strategies and capacity theorems for relay networks. IEEE Transactions on Information Theory, 51(9):3037–3063, September 2005. 14. J. N. Laneman, G. W. Wornell, and D. N. C. Tse. An efficient protocol for realizing cooperative diversity in wireless networks. In Proc. of IEEE International Symposium on Information Theory (ISIT), page 294, June 2001. 15. J. N. Laneman, G. W. Wornell, and D. N. C. Tse. Cooperative diversity in wireless networks: Efficient protocols and outage behavior. IEEE Transactions on Information Theory, 50(12):3062–3080, December 2004.

8 Implementing Cooperative Wireless Networks

177

16. G. Li and H. Liu. Resource allocation for OFDMA relay networks with fairness constraints. IEEE Journal on Selected Areas in Communications, 24(11):2061– 2069, November 2006. 17. Y. Li, B. Vucetic, T.F. Wong, and M. Dohler. Distributed turbo coding with soft information relaying in multihop relay networks. IEEE Journal on Selected Areas in Communications, 24(11):2040–2050, November 2006. 18. Z. Lin, E. Erkip, and A. Stefanov. Cooperative regions and partner choice in coded cooperative systems. IEEE Transactions on Communications, 54(7):1323–1334, July 2006. 19. P. Liu, Z. Tao, Z. Lin, E. Erkip, and S. Panwar. Cooperative wireless communications: A cross-layer approach. IEEE Wireless Communications, 13:84–92, August 2006. 20. P. Liu, Z. Tao, and S. Panwar. A cooperative MAC protocol for wireless local area networks. In Proc. of IEEE International Conference on Communications (ICC), volume 5, pages 2962–2968, May 2005. 21. A. Nosratinia, T. E. Hunter, and A. Hedayat. Cooperative communication in wireless networks. IEEE Communications Magazine, 42(10):74–80, October 2004. 22. A. Sendonaris, E. Erkip, and B. Aazhang. Increasing uplink capacity via user cooperation diversity. In Proc. of IEEE International Symposium on Information Theory (ISIT), page 156, August 1998. 23. A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity. Part I. System description. IEEE Transactions on Communications, 51(11):1927–1938, November 2003. 24. J. Shi, G. Yu, Z. Zhang, Y. Chen, and P. Qiu. Partial channel state information based cooperative relaying and partner selection. In Proc. of IEEE Wireless Communications and Networking Conference (WCNC), March 2007. 25. Signalion GmbH. SORBAS 101: Signalion software radio based protyping system. Available at: http://www.signalion.com/, 2005. 26. V. Stankovic, A. Høst-Madsen, and Z. Xiong. Cooperative diversity for wireless ad hoc networks. IEEE Signal Processing Magazine, 23(5):37–49, September 2006. 27. S. Valentin and H. Karl. Analyzing the effect of asymmetric mobility and channel configurations on the outage performance of coded cooperative systems. In Proc. of the European Wireless Conference (EW), April 2007. 28. S. Valentin and H. Karl. Effect of user mobility in coded cooperative systems with joint partner and cooperation level selection. In Proc. of IEEE Wireless Communications and Networking Conference (WCNC), March 2007. 29. E. C. van der Meulen. Three-terminal communication channels. In Advances in Applied Probability, volume 3, pages 120–154, 1971. 30. D. H. Woldegebreal and H. Karl. Network-coding-based adaptive decode and forward cooperative transmission in a wireless network: outage analysis. In Proc. of the European Wireless conference (EW), April 2007. 31. X. Xu, D. Gunduz, Erkip, and Y. Wang. Layered cooperative source and channel coding. In Proc. of IEEE Conference on Communications, volume 2, pages 1200–1204, May 2005. 32. B. Zhao and M.C. Valenti. Distributed turbo coded diversity for the relay channel. IEEE Electronics Letters, 39(10):786–787, May 2003.

178

Stefan Valentin et al.

33. B. Zhao and M.C. Valenti. Some new adaptive protocols for the wireless relay channel. In Proc. of Allerton Conference on Communication, Control, and Computing, October 2003. 34. S. A. Zummo. Performance analysis of coded cooperation diversity in wireless networks. Wireless Communications and Mobile Computing, July 2006.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks Peter Rost and Gerhard Fettweis Technische Universit¨ at Dresden, Vodafone Chair Mobile Communications Systems, Dresden, Germany [rost|fettweis]@ifn.et.tu-dresden.de

Summary. Multiterminal relaying is a promising extension for conventional mobile communications systems as it is able to increase coverage and throughput of these systems. Due to practical constraints most relaying protocols employ half-duplex relay nodes. This implies a rate loss due to the necessity of assigning orthogonal channels for the source-relay and relay-destination communication. We will start with an overview of existing work on half-duplex relays and present protocols using alternatingly transmitting relay nodes. Only one relay is sending at any given time and at least two relay terminals are employed, such that a kind of large scale spatial duplexing based on a time or frequency division duplex system is implemented. We combine this idea with compress-and-forward and decode-and-forward approaches to present achievable rates for the discrete memoryless relay channel. Both approaches are then combined to a mixed strategy and applied to a network with more than one relay transmitting at a time. We conclude the chapter with an overview of specific protocols which were proposed and analyzed in the context of wireless fading channels and use alternatingly transmitting relays.

9.1 Introduction With the global success of wireless communications systems, such as cellular networks, the question arises how to utilize and design mobile communications networks more (cost) efficiently. So far, the operation of cellular networks is divided into a broadcast channel in the downlink, and a multiple-access channel in the uplink. Beside infrastructure based networks, ad hoc networks relying on IEEE 802.11 (WLAN), Bluetooth, and other standards draw more attention as they operate in a decentralized manner without central control. In [10] Gupta and Kumar derived upper and lower bounds on the transport capacity for different kinds of ad hoc networks. In their fundamental work they considered only point-to-point links, i.e., one source node is only communicating with exactly one destination node. Conversely, each transmission not originating from the source node is considered as interference at the destination node. Under a given power constraint it can be shown that the achievable rate on one point-to-point link is not improved by an increase in the number of nodes.

179 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 179–197. c 2007 Springer.

180

Peter Rost and Gerhard Fettweis

The growing importance of infrastructure based wireless networks as well as ad hoc networks in telecommunication poses the question how to exploit a network of wireless terminals to increase capacity and coverage, optimize power distribution, reduce usage of backhaul infrastructure and so forth. One answer to this question was presented by van der Meulen and his fundamental work on relay channels in [34, 35] where terminal cooperation is utilized to increase achievable data rates. In the simplest case of a three terminal network, a relay node supports the information transmission between a source and destination node. Cover and El Gamal presented fundamental results for this one relay case in [4]; in their work they proposed two basic coding techniques – decode-and-forward and compress-and-forward as well as a mixed strategy with partial decoding. They further introduced the degraded relay channel for which the decode-and-forward technique achieves capacity. Nevertheless, so far the actual capacity of the general relay channel is not known but upper-bounded by the cut-set bound [5, Ch. 14.10] (which is achievable for certain relay channels, e.g., the degraded relay channel and the general relay channel with feedback [4]). In recent publications the analysis of the three terminal case was extended to the multi-terminal case: Kramer et al. presented in [17] a comprehensive overview of different coding strategies for networks of relay nodes, e. g., a generalized decode-andforward and compress-and-forward approach (as well as a mixed strategy of both), where the former utilizes regular encoding (all codebooks used in the network have the same size) and sliding-window decoding approach (see also [2, 16, 40, 41]). Furthermore, they presented results for the discrete memoryless relay channel as well as different wireless models. A generalization of the decode-and-forward approach based on irregular encoding (codebooks might be of different size) and successive decoding strategy of [4, Theorem 1] is given by Gupta and Kumar in [11], using a multi-level relaying scenario. They further extended this general approach to a scenario with more than one active source-destination pair and derived the transport capacity of different wireless networks. Beside the previous two decode-and-forward strategies, a third approach based on regular encoding/backward decoding was proposed in [36] (see [17] for a comprehensive overview of all three strategies). As a ¨ ur et al. [43] who proposed a disfinal example we want to mention the work of Ozg¨ tributed MIMO cooperation scheme as well as the work of Gastpar and Vetterli [7] who analyzed a network coding approach under the restriction of only one active source-destination pair. Most of the previously mentioned work considers full-duplex terminals. However, this might be hard to implement in practice due to the coupling of the transmit signal into the receiver path in realistic transceiver frontends. This problem was examined by Laneman et al. [18] who imposed the orthogonality constraint, i.e., a relay node is not able to both transmit and receive in the same time-frequency bin. As an immediate consequence, the relay nodes must act either in a time or frequency division manner, i.e., at first the source transmits relevant information to the relay node. This will either repeat, reencode or amplify and retransmit this information in a second phase. Two-phase protocols as proposed in [18] must increase the spectral efficiency on the individual links to achieve the same end-to-end spectral efficiency as a direct transmission. This tradeoff is quantified by the multiplexing-diversity tradeoff [42]. Azarian et al. analyzed in [1] different relaying protocols concerning their multiplexing-diversity tradeoff and proposed a dynamic decode-and-forward protocol as well as a generalized nonorthogonal amplify-and-forward (originally pre-

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

181

sented by Nabar et al. in [20, 21] for the one-relay case). The basic principle of both protocols is that the relay only listens to a part of the source’s transmission and repeats this concurrently to the source in a second phase. This leads to a receive collision of the source and relay message but still achieves better results concerning the multiplexing-diversity tradeoff [1]. In [19] Laneman and Wornell extended the work of [18] to the multi-terminal case and proposed a two-phase protocol achieving better results concerning the multiplexing-diversity tradeoff than a repetition based protocol. Herhold et al. analyzed different relaying protocols in a multihop relay network in [13]. Finally, we want to mention the work of Kramer [15] who analyzed relay networks under the SLoT constraints, i.e., a relay is at any time in sleep, listen or talk mode. He analyzed different mode assignment strategies and shows achievable rates for a decode-and-forward as well as a partial-decode-and-forward approach. The purpose of this chapter is to present relaying protocols based on multiple relay terminals alternatingly transmitting such that only one relay node is transmitting at a given time. With this kind of protocols the source is able to transmit continuously. This allows us to avoid the multiplexing loss due to the half-duplex constraint. Unlike the proposals in [1, 20, 21], at any time at least one relay node is transmitting, so that the complete source information is received and retransmitted by the relays. Although this comes at the cost of lower diversity than the dynamic decode-and-forward or nonorthogonal amplify-and-forward protocols achieve, it is essential in scenarios with shadowed areas where the source-destination link is very weak and the relays must retransmit the complete source message. Furthermore, we show how each relay can assist the decoding process at each other relay to increase the reliability at the relay nodes, which again is of crucial importance in shadowing scenarios (in contrast to the multi-level approach in [11] where only relays of lower levels support relays of higher levels). Finally, we want to mention that the discussed protocols are spectrally efficient, i.e., although they do not provide the same diversity gain as the proposals in [19], they are appropriate for the high rate regime as they are able to support spectral efficiencies R > C/2 where C denotes the channel capacity per link in a symmetric scenario [1, 18]. After a description of the underlying relay network model in Section 9.2 we will present two protocols in Section 9.3 which generalize the compress-and-forward and decode-and-forward approach based on irregular encoding/successive decoding and show a multi-antenna reception and multi-antenna transmission behavior, respectively. Additionally, we will discuss a mixed technique combining both approaches. The focus of this chapter is to present the basic ideas, hence we only provide results for the discrete memoryless relay channel and discuss some protocols for wireless fading channels in Section 9.4.

9.2 Nomenclature and Relay Network Model In the following we will use bold lowercase letters (x) to denote vectors, non-italic uppercase letters (X) to denote random variables, italic uppercase letters (N ) to denote constant scalars and italic lowercase letters (n) to denote scalar indices. Futhermore, X is used to denote ordered sets, kX k to denote the cardinality of a set and [b; b + k] to denote the ordered set of numbers (b, b + 1, . . . , b + k) with [b; b + k] = ∅ for k < 0. We further define the index r over the set [1; N ] and define

182

Peter Rost and Gerhard Fettweis

addition and subtraction on the ring [1; N ], i.e., r + k := mod (r + k − 1, N ) + 1 and r − k := mod (r − k − 1, N ) + 1. Let Xk be a random variable parameterized using k then {Xk : c(k)} denotes the set of all Xk which satisfy the given constraint c(k). Furthermore, let π(R) be the set of all permutations of a set R and πj (r) the r-th element in the j-th permutation of R, i.e., πj ∈ π(R). We consider in this chapter a network of N + 2 nodes: the set of N relay nodes t ∈ R := [1; N ], the source node s = N + 1 and the destination node d = N + 2. The relay channel is defined over all possible channel inputs (x1 , · · · , xN , xs ) ∈ X1 × · · · XN × Xs and channel outputs (y1 , · · · , yN , yd ) ∈ Y1 × · · · YN × Yd with Xi and Yj denoting the input and output alphabets, respectively. We further use yR to refer to the vector of all yt with t ∈ R. Using this notation the discrete memoryless relay channel is defined by the time-invariant joint probability density function (pdf) pYR ,Yd |Xs ,XR (yR , yd |xs , xR ), i. e., the channel output depends only on the channel inputs in the same block but not on the block itself and previous channel inputs. If the arguments of a pdf are lower case letters of the random variables for which the pdf is defined we write p(x|y) instead of pX|Y (x|y). Due to reasons of readability we will use in the following the index r as a short-cut for the relay node t = πj (r) (the r-th element of some permutation πj ∈ π(R)), when this does not create any confusion, e.g., xr is used to abbreviate xπj (r) . An example network is shown in Figure 9.1 for N = 4 relay nodes.

9.3 Protocols for Half-Duplex Relay Nodes Based on the previously described network model this section will present two different protocols for half-duplex relay terminals. We consider only the case of a single relay transmitting in a certain block and a single source-destination pair. Nevertheless, there are straightforward as well as more involved extensions to the presented concepts possible which will not be discussed here. To select the currently sending relay we choose a certain permutation πj ∈ π (R) which defines the order in which the relays are transmitting. Let the source message W consist of B blocks wb , b ∈ [1; B], each of nR bits. The source transmits in block b the n-sequence xs (wb ) (we describe in the sequel how these sequences are chosen). Furthermore, relay πj (1) transmits the signal xπj (1) (m1 ). Assume that relay node r transmits in block b the signal xr (mb ) we select select r0 = r + k to transmit in block b0 = b + k the signal xr0 (mb0 ). Concurrently, all other N − 1 relay nodes R\{r0 } receive in block b0 the superposition of the source and the relay signal xs (wb0 ) and xr0 (mb0 ), respectively (see Figure 9.1 for an illustrative example). Since the last source block wB is received by N − 1 relay nodes which subsequently transmit useful information to the destination, it follows that the last block wB is decoded by the destination after block B + N − 1. Furthermore, due to the introduced order (by πj ) we can guarantee that each relay can receive N − 1 consecutive source and relay blocks, which improves the relay cooperation to increase the reliability.

9.3.1 A Compress-And-Forward Based Approach At first we present a protocol based on the compress-and-forward approach proposed in [4, Theorem 6]. The strategy of this approach relies on Wyner-Ziv coding (see

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

183

1 4 s

d 3 2 (a) In block b = 1, s and r = 1 transmit.

1 4 s

d 3 2 (b) In block b = 3, s and r = 3 transmit.

Figure 9.1. One example deployment for N = 4 relay nodes. Dashed circles denote nodes in receive mode and solid circles nodes in transmit mode. In block b the relay node r, r = mod (b − 1, N ) + 1, is in transmit and all other relays are in receive mode. This is shown here only for the blocks b = 1 and b = 3 as the remaining two scenarios are easy to derive.

[38, 39] and [8, 9] for multiple sources), i.e., source coding within a given distortion ˆ such that and side information at the decoder. A source X is described by a X the expected distortion between both is less than a predefined value D. The rate distortion theory now states that the rate to describe X needs to satisfy R(D) > ˆ If the decoder can use the side information Y it can be shown that RY (D) > I(X; X). ˆ − I(X; ˆ Y) suffices to describe the source X (since I(X; ˆ Y) ≥ 0 we have I(X; X) R(D) ≥ RY (D)). This idea is now generalized to a network of half-duplex relays as follows: let relay r be transmitting the n-sequence xr (mb )1 in block b with mb ∈ [1; 2nRr ]. According to the previous description it is able to observe the previous N − 1 consecutive channel outputs in blocks [b − 1; b − N + 1], i.e., yr (b − 1), · · · , yr (b − N + 1). We assume that relay r at first decodes the messages of the relays, i. e., xr−k (mb−k ), k ∈ [1; N − 1], mb−k ∈ [1; 2nRr−k ], n > 0, where the subscript r − k is only used for 1

We show in the proof of Theorem 1 how the described sequences are created.

184

Peter Rost and Gerhard Fettweis

Table 9.1. Outline of the compress-and-forward based coding scheme with B = 7 source blocks and N = 3 relay nodes (πj = {1, 2, 3}). The signals yˆr,k (ur,b−k |xr−k (mb−k )) determine the indices mb,k and are decoded at the destination using Wyner-Ziv coding (the table uses the shorthand notation yˆr,k (ur,b−k ) for reasons of brevity). b 1 2 3 4 5 6 7 8 9

s t=1 t=2 t=3 xs (w1 ) ∅, ∅ xs (w2 ) yˆ2,1 (u2,1 ), ∅ xs (w3 ) yˆ3,1 (u3,2 ), yˆ3,2 (u3,1 ) xs (w4 ) yˆ1,1 (u1,3 ), yˆ1,2 (u1,2 ) xs (w5 ) yˆ2,1 (u2,4 ), yˆ2,2 (u2,3 ) xs (w6 ) yˆ3,1 (u3,5 ), yˆ3,2 (u3,4 ) xs (w7 ) yˆ1,1 (u1,6 ), yˆ1,2 (u1,5 ) ∅ yˆ2,1 (u2,7 ), yˆ2,2 (u2,6 ) ∅ ∅, yˆ3,2 (u3,7 )

the sake of readability (the relay transmitting the signal corresponding to x· (mb−k ) is uniquely defined by the introduced order πj ∈ π(R)). Knowing the relay signals in blocks [b − 1; b − N + 1], relay r selects for each yr (b − k) a quantized verˆ sion yˆr,k (ur,b−k |xr−k (mb−k )), ur,b−k ∈ [1; 2nRr,k ], according to a distortion measure d(yr , yˆr ) and within a given distortion (note that the quantization depends on both the relay node and the index k). Using the random binning argument [32] we select ˆ r,k , i.e., different ur,b−k for each ur,b−k an index mb,k , mb,k ∈ [1; 2nRr,k ], Rr,k ≤ R might be assigned to the same mb,k . The vector of these N P − 1 indices mb,k now determines the index mb = (mb,1 , · · · , mb,N −1 ), i.e., Rr = k Rr,k . Finally, this index mb is used to determine the relay signal xr (mb ) in block b. In the proof of Theorem 1 we will give a more formal description of this coding scheme based on a random coding argument and typical sequence decoding. ˆ r,[1;N −1] the destination exploits the side To decode the described estimates y information provided by the own channel outputs yd (b − k) as well as the estimates decoded after blocks [b−1; b−N +2]. Finally, using the quantized versions yˆr−k+1,N −k and the own observation yd (b − N + 1) the destination decodes the source block wb−N +1 . An exemplary outline of this coding approach is given for B = 7 source blocks and N = 3 relay nodes in Table 9.1. Consider the source block xs (w4 ) in this table. Relay nodes t = 2 and t = 3 are transmitting x2 (m5 ) and x3 (m6 ) with m5 = (m5,1 , m5,2 ) and m6 = (m6,1 , m6,2 ). The indices m5,1 and m6,2 are determined by the estimates yˆ2,1 (u2,4 |x1 (m4 )) and yˆ3,2 (u3,4 |x1 (m4 )), respectively (where the indices u[2,3],4 need not be the same). The destination uses yd (4) to decode yˆ2,1 (u2,4 ) and decodes yˆ3,2 (u3,4 ) using yd (4) and yˆ2,1 (u2,4 ). Using both estimates and the own channel output the destination then decodes xs (w4 ) at the end of block b = 6. As outlined in [17] we achieve with this approach a multi-antenna reception behavior (which can be seen from the following Theorem). Furthermore, there is a tradeoff between the number of copies and the fidelity of the copies, i.e., whether it is preferable to send many copies each with less information or less copies each with more information.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

185

Theorem 1. The achievable rate of the previously described compress-and-forward approach in a network of half-duplex relay terminals is given by n o ˆ (r−k+1),(N −k) : k ∈ [1; N − 1] |Xr−N +1 R ≤ max min sup I Xs ; Yd , Y (9.1) πj

r

with the side conditions (r ∈ [1; N ] , l ∈ [1; N − 1]) X Rr,k = Rr < min (I(Xr ; Yd ), I(Xr ; Yr+l ))

(9.2)

k

and n o ˆ (r+k),k ; Yr+k |Xr , Yd , Y ˆ (r+k−i+1),(k−i+1) : i ∈ [2; k] , R(r+k),k > I Y

(9.3)

where (9.2) is the bound due to the inter-relay (I(Xr ; Yr+l )) and relay-destination communication (I(Xr ; Yd )) and (9.3) is implied by Wyner-Ziv source coding. The supremum is taken over p xs , xr , yˆr0 ,(r0 −r) : r0 ∈ [1; N ] \ {r} , yR\{πj (r)} , yd = Y (9.4) p(ˆ yr0 ,(r0 −r) |yr0 , xr ), p(xs )p(xr )p(yd , yR\{πj (r)} |xs , xr ) r 0 ∈[1;N ]\{r}

which is the joint pdf of the system depending on the current πj and the currently transmitting r. Furthermore, we need to take the minimum over all relays r ∈ [1; N ] and the maximum over all possible orders πj ∈ π(R). From the source channel separation theorem [5, Ch. 14.10] we know that the separation into a channel coding region (eq. (9.2)) and a source coding region (eq. (9.3)) does not generally lead to optimal results (see also [17, Remark 19]). The proof of Theorem 1 relies on strongly typical sequences which are defined in the following for completeness. Let p(x1 , x2 , · · · , xk ) be a fixed joint pdf of the discrete random variables X1 , · · · , Xk defined on X1 × X2 · · · Xk . Furthermore, Q let S ⊆ {X1 , · · · , Xk }, and Sn be n independent copies of S with p(S n = sn ) = n i=1 p(si ) where si is the i-th element of sn . At first we define the set of jointly -typical n-sequences. n n Definition 1. The set of jointly -typical n-sequences (xn 1 , x2 , · · · , xk ) (also called weakly typical) is defined by n n n n n n A(n) := (xn 1 , x2 , · · · , xk ) ∈ X1 × X2 · · · Xk : (9.5) o 1 − log p(sn ) − H(S) < , ∀S ⊆ {X1 , · · · , Xk } n n for a given > 0. Now let N (x1 , · · · , xk |xn 1 , · · · , xk ) be the number of occurrences n of the tuple (x1 , · · · , xk ) in the tuple of sequences (xn 1 , · · · , xk ).

Definition 2. The set of -strongly typical sequences (also called strongly typical) is defined by

186 A∗(n)

Peter Rost and Gerhard Fettweis n n n n n n := (xn 1 , x2 , · · · , xk ) ∈ X1 × X2 · · · Xk : 1 n n N (x1 , x2 , · · · , xk ) | (xn , x , · · · , x ) − p(x , x , · · · , x ) 1 2 k 1 2 k n < , kX1 × X2 × · · · × Xk k

(9.6)

o n n p(x1 , x2 , · · · , xk ) = 0 ⇒ N (x1 , x2 , · · · , xk ) | (xn 1 , x2 , · · · , xk ) = 0 for a given > 0. The following proof is based on the asymptotic equipartition property (AEP) (see [5, Theorem 14.2.1 and Lemma 13.6.2]) which can be defined for weak and strong typicality and holds in general for ergodic sources (see [14], [5, Ch. 15.7] and [31] which uses the terms entropy-typical and frequency-typical)2 . One can show using the AEP that an arbitrarily small probability of error for a sufficiently large block length n can be achieved if a typical sequence decoder is employed. We further use for our proof the Markov lemma (see [5, Lemma 14.8.1]) which requires the definition of strong typicality and we make usage of the random binning argument introduced by Slepian and Wolf in [32]. Proof. (of Theorem 1) We give here an outline of the proof for the achievability of the rate given in (9.1). The proof is intentionally created in the same way as [4, Theorem 6] to allow an easy understanding of the proposed protocol: at first we describe a random coding scheme used in our proof. Then we describe the encoding and decoding procedure to achieve the described rate.

Random Coding nR 1. The source creates i.i.d. n-length sequences xs (w) each drawn accordQn2 ing to p(x ) = p(x s si ) where xsi denotes the i-th element of xs and i=1 w ∈ 1, . . . , 2nR . nRr i.i.d. n-length sequences 2. Each relay r creates nRra codebook consisting of 2 Q xr (mr ), mr ∈ 1; 2 , each drawn according to p(xr ) = n i=1 p(xri ) where xri denotes the i-th element of xr and vector of the indices with mr asPthe −1 mr,1 , · · · , mr,(N −1) , mr,k ∈ 1; 2nRr,k , i. e., Rr = N k=1 Rr,k . 3. Create, for each xr−k (mr ) with r ∈ [1; N ], k ∈ [1; N − 1] , mr ∈ 1; 2nRr−k , ˆ

2nRr,k i.i.d. n-length sequences yˆr,k (u(k) |xr−k (mr )), drawn according to Q ˆ p(ˆ yr,k |xr−k (mr )) = n yr,ki |x(r−k)i (mr )), u(k) ∈ [1; 2Rr,k ], where yˆr,ki dei=1 p(ˆ ˆ r,k denotes the necessary rate to notes the i-th element of yˆr,k . Note that R describe the channel output yr (b − k) by a yˆr,k within a given distortion and Rr,k is the rate distortion function if side information is given at the destination (the destination channel outputs yd ). 4. We further introduce a random partitioning at each relay r with N −1 mappings such that each u(k) randomly mapped independently into one of 2nRr,k cells isnR Sr,k (mk ), mk ∈ 1; 2 r,k , r ∈ [1; N ], k ∈ [1; N − 1], according to a uniform distribution. 2

While weak typicality can be applied to both discrete and continuous random variables, the concept of strong typicality as defined here can only be applied to discrete random variables.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

187

Encoding Let us assume that the relay r, transmitting in block b, successfully decoded xr−k (mb−k ), and it created for each of the last N − 1 observations the tuple ∗(n) (yr (b − k), yˆr,k (ur,b−k |xr−k (mb−k )), xr−k (mb−k )) ∈ A , k ∈ [1; N −1]. With ur,b−k ∈ Sr,k (mb,k ), it transmits xr (mb ) with mb = mb,1 , . . . , mb,(N −1) , mb,k ∈ nR 1; 2 r,k . Concurrently, the source sends xs (wb ). We assume that the previous N − 1 steps were error free.

Decoding The decoding procedure at the end of block b is as follows (at the end of block b > N − 1 the source index wb−N +1 is decoded): 1. The destination at first decodes xr (mb ). This is done by searching for a uniquely typical xr (mb ) with yd (b), which is possible iff Rr < I(Xr ; Yd ) and n is sufficiently large (resulting from the channel coding theorem [30]). 2. In step 2, the destination creates the sets of those u ˜r,b−k , k ∈ [1; N − 1], such that the quantizations of relay r are jointly typical with the destination channel outputs yd (b − k), already decoded quantizations and the relay messages xr−k (mb−k ): ˜r,b−k : xr−k (mb−k ), yd (b − k), . . . Lk (yd (b − k)) := u . u(r−i+1),(b−k) |xr−k (mb−k )) : i ∈ [1; k] ∈ A∗(n) yˆ(r−i+1),(k−i+1) (˜ The assumption that all previous N − 1 decoding steps were error free ensures that previously decoded relay quantizations are correctly decoded and satisfy for k ≥ 2 xr−k (mb−k ), yd (b − k), . i ∈ [2; k] yˆ(r−i+1),(k−i+1) (u(r−i+1),(b−k) |xr−k (mb−k )) ∈ A∗(n) ur,b−k |xr−k (mb−k )) such that Afterwards it chooses for each k an estimate yˆr,k (˜ ˜r,b−k ∈ Sr,k (mb,k ) ∩ Lk (yd (b − k)) , ∃˜ ur,b−k : u which succeeds, i.e., u ˜r,b−k = ur,b−k , with arbitrarily low probability of error iff ˆ r,k ; Yd , Y ˆ (r−1),(k−1) , . . . , Y ˆ (r−k+1),1 |Xr−k + Rr,k ˆ r,k < I Y R n o ˆ r,k ; Yd , Y ˆ (r−i+1),(k−i+1) : i ∈ [2; k] |Xr−k + Rr,k
188

Peter Rost and Gerhard Fettweis

3. Using

yˆr,(N −1) (ur,b−N +1 |xr−N +1 (mb−N +1 )), . . . ,

yˆ(r−N +2),1 (u(r−N +2),(b−N +1) |xr−N +1 (mb−N +1 )) = yˆ(r−k+1),(N −k) (u(r−k+1),(b−N +1) |xr−N +1 (mb−N +1 ) : k ∈ [1; N − 1] ˜b−N +1 ) iff the destination now chooses one xs (w ∃w ˜b−N +1 : yˆ(r−k+1),(N −k) (u(r−k+1),(b−N +1) |xr−N +1 (mb−N +1 ) : k ∈ [1; N − 1] , . . . ˜b−N +1 ) ∈ A∗(n) . yd (b − N + 1), xr−N +1 (mb−N +1 ), xs (w We can state that w ˜b−N +1 = wb−N +1 with arbitrarily low probability of error iff n o ˆ (r−k+1),(N −k) : k ∈ [1; N − 1] |Xr−N +1 R < I Xs ; Yd , Y (9.8) and n is sufficiently large. 4. Furthermore, all other relays decode the relay message xr (mb ) iff Rr < I(Xr ; Yr+k )

(9.9)

for k ∈ [1; N − 1] and create for each k the following tuple yr+k (b), yˆ(r+k),k u(r+k),b |xr (mb ) , xr (mb ) ∈ A∗(n) which is possible iff ˆ (r+k),k > I(Y ˆ r+k,k ; Yr+k |Xr ) R

(9.10)

and n is sufficiently large. From the previous points one can see that B + N − 1 blocks are necessary to communicate B blocks, i.e., a rate loss (N −1)/B · R is implied which goes to 0 as B → ∞. We can further reformulate (9.7) without loss of generality to n o ˆ (r+k),k < I Y ˆ (r+k),k ; Yd , Y ˆ (r+k−i+1),(k−i+1) : i ∈ [2; k] |Xr + R(r+k),k R which implies using (9.10) that n o ˆ r+k,k ; Yr+k |Xr I Y

(9.11)

which is a result of the employed Wyner-Ziv source coding. Remark 1. A special case of Theorem 1 is the case of N = 2 relay nodes where the achievable rate is given by ˆ 1,1 |X2 , sup I Xs ; Yd , Y ˆ 2,1 |X1 (9.12) R ≤ min sup I Xs ; Yd , Y which would coincide with the rate given in [4, Theorem 6], if both relays are on the same position (and therefore building a full-duplex relay). In this case also the condition in (9.2) due to the inter-relay communication could be ignored.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

189

Remark 2. If relay r chooses the same distortion Dr = d (yr , yˆr,k ) for each k ∈ [1; N − 1], one can see from (9.3) (and (9.7)) that the necessary rate to communicate the k-th quantized observation is monotonically decreasing with k, i.e., Rr,1 ≥ Rr,2 ≥ · · · ≥ Rr,(N −1) .

9.3.2 A Decode-And-Forward Based Approach After the compress-and-forward based approach we present in this section a protocol applying the decode-and-forward strategy from [4, Theorem 1] which uses irregular encoding/successive decoding, block Markov superposition encoding and the random binning/partitioning argument introduced by Slepian and Wolf in [32]. The basic idea is that the relay node (in the three-terminal case) decodes the source message and supports the source node by transmitting a partition index associated to the source message which is used by the destination to decode the source message. We now apply this idea in a similar way as above in the compress-and-forward case. Each relay decodes the source messages (not necessarily all) and transmits the partition indices in the same way as above. A significant difference to the previous protocol is that only one relay is decoding at a time. Remember, in the previous protocol each relay was creating its estimate yˆr at each time instant. In this protocol only the relay transmitting in block b will decode after block b − 1. This offers the chance that each relay can gain on the information sent by the other relays to improve its own decoding. Furthermore, consider the case that relay r is not able to decode the source message xs (wb ) using only on the channel output yr (b). But, if the relay can exploit the information sent by the other relay nodes regarding this source message it might be able to decode xs (wb ). Therefore, not necessarily all relay nodes decode all source messages (usually only those nodes with sufficiently good source-relay channel conditions) but we require all relay nodes to decode all other relay transmissions. This shows again that the order in which the relays are transmitting is of crucial importance, therefore we will again use the set π(R) as defined in the last section. Theorem 2. The achievable rate of the decode-and-forward based approach is given by R < max min sup min(Rs,d , Rs,r ) πj

(9.13)

r

Rs,d = I(Xs ; Yd |Xr−N +1 ) +

N −1 X

R(r−k+1),(N −k)

(9.14)

k=1

Rs,r =

min

k∈[1;N −1],Rr,k >0

I(Xs ; Yr |Xr−k ) +

k−1 X

Rr−i,k−i

(9.15)

i=1

with the side condition X Rr,k = Rr < min (I(Xr ; Yd ), I(Xr ; Yr+k )) ,

(9.16)

k

for r ∈ [1; N ], k ∈ [1; N − 1] and the joint pdf p xs , xr , yR\{πj (r)} , yd = p(xr )p(xs |xr )p yR\{πj (r)} , yd |xs , xr

(9.17)

190

Peter Rost and Gerhard Fettweis

which depends r. The supremum in (9.13) is over all on the currently transmitting joint pdf p xs , xr , yR\{πj (r)} , yd . Furthermore, we need to take the minimum over all relays and the maximum over all possible orders πj ∈ π(R). The rate Rs,d denotes the limit on R induced by the source to destination channel (including the supporting relay nodes) and Rs,r denotes the limit on R to communicate the source message to the relay nodes. Furthermore, as also outlined in [17], we achieve with this approach a multi-antenna transmission behavior. For the following proof of the theorem we require weak typicality. Proof. (of Theorem 2) We give here an outline of the proof for achievability by providing a random coding scheme achieving the rate in (9.13). One can see that the proof is again intentionally created similar to [4, Theorem 1] to allow an easy understanding of the proposed protocol.

Random Coding 1. Define 2nR conditionally i.i.d. n-length sequences according nR xs (w|xr (s)) drawn Q to p(xs |xr (s)) = n , r ∈ [1; N ], s ∈ 1; 2nRr and i=1 p(xsi |xri (s)), w ∈ 1; 2 xsi and xri denoting the i-th element of xs and xr , respectively. The dependence between source and relay messages is used so that the source can assist the relay transmission, e.g., by coherent transmission. 2. Each relay creates a codebook consisting of 2nRr i.i.d. Q n-length sequences xr (sr ), sr ∈ 1; 2nRr , each drawn according to p(xr ) = n p(x ). Further let sr be i=1 ri P 1 the vector of indices sr,1 , · · · , sr,(N −1) , sr,k ∈ 1; 2nRr,k and Rr = N k=1 Rr,k . 3. Finally we introduce a random partitioning such that each x (w|x (s)) is rans r domly mapped into one of 2nRr,k cells Sr,k (sk ), sk ∈ 1; 2nRr,k , r ∈ [1; N ], k ∈ [1; N − 1]. We introduce the possibility that Rr,k = 0, i. e., the relay r transmitting in block b does not decode the source symbol xs (wb−k |xr (sb−k )).

Encoding Consider block b in which relay r is transmitting. Assume that relay r correctly decoded all those wb−k for which Rr,k > 0 and let wb−k ∈ Sr,k (sb,k ). It transmits in block b the n-sequence xr (sb ) with sb = sb,1 , . . . , sb,(N −1) and the source transmits xs (wb |xr (sb )), k ∈ [1; N − 1]. This is possible since the source node knows the mappings Sr,k and therefore can determine by itself the message sent by the relay. Again, we assume that the previous N − 1 steps were error free.

Decoding At the end of block b the following decoding procedure is applied (at the end of block b > N − 1 the source index wb−N +1 is decoded): 1. The destination decodes xr (sb ) sent by relay r in block b. This is done by searching for a uniquely typical xr (sb ) with yd (b), which is possible with arbitrarily low probability of error iff Rr < I(Xr ; Yd ) and n is sufficiently large.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

191

2. In the next step the destination creates a set of those w ˜b−N +1 which can be the correct source index: n ˜b−N +1 : xs (w ˜b−N +1 |xr−N +1 (sb−N +1 )), · · · L yd (b − N + 1) = w o. xr−N +1 (sb−N +1 ), yd (b − N + 1) ∈ A∗(n) Since the destination also knows the bin indices we can state that −1 N\ S(r−k+1),(N −k) (s(b−k+1),(N −k) ) wb−N +1 = L yd (b − N + 1) ∩ k=1

with arbitrarily low probability of error iff R < I(Xs ; Yd |Xr−N +1 ) +

N −1 X

R(r−k+1),(N −k)

(9.18)

k=1

and n is sufficiently large. 3. Let r0 = r + 1 be the next relay sending in block b0 = b + 1. At first r0 needs to decode xr0 −k (sb0 −k ), k ∈ [1; N − 1], which is possible iff Rr0 −k < I (Xr0 −k ; Yr0 ). The relay r0 further creates the sets n ˜b0 −k : xs (w ˜b0 −k |xr0 −k (sb0 −k )), · · · Lk yr0 (b0 − k) = w o , yr0 (b0 − k), xr0 −k (sb0 −k ) ∈ A∗(n) for k ∈ [1; N − 1]. Using these sets and the previously decoded indices the relay can now decode the source messages: \ k−1 wb0 −k = Lk yr0 (b0 − k) ∩ S(r0 −i),(k−i) (s(b0 −i),(k−i) ) i=1

which succeeds (for all k such that Rr0 ,k > 0) iff R < I(Xs ; Yr0 |Xr0 −k ) +

k−1 X

Rr0 −i,k−i

(9.19)

i=1

and n is sufficiently large. Remark 3. Once again consider the case of N = 2 relay nodes which gives the (intuitively) achievable rate R < min (sup min (I(Xs ; Yd |X2 ) + R1,1 , I(Xs ; Y1 |X2 )) , · · · sup min (I(Xs ; Yd |X1 ) + R2,1 , I(Xs ; Y2 |X1 )))

(9.20)

which coincides with [4, Theorem 1] if both relays are on the same position (again building a full-duplex relay). In this case also the additional side condition in (9.16) due to the inter-relay communication can be ignored. Remark 4. The analysis of this protocol assumes that each relay decodes all other relay transmissions to gain from this additional information. To make the protocol more flexible one could weaken this constraint so that not all relays must decode all relay transmissions (or only partially decode as in [17, Theorem 5]).

192

Peter Rost and Gerhard Fettweis

Remark 5. We allow that some relays need not to decode all source transmissions which makes the analysis of the achievable rate a little bit more complicated. If we tighten this constraint and require that all relays have to decode all source transmissions then the condition in (9.15) simplifies to R < min (I(Xs ; Y1 |XN ), I(Xs ; Y2 |X1 ), . . . , I(Xs ; YN |XN −1 ))

(9.21)

which is a rather tight condition in comparison to (9.15).

9.3.3 Mixed Strategies In [4, Theorem 7] a combination of the compress-and-forward and decode-andforward approach using partial decoding was presented. Furthermore, in [17] all relays were divided into two groups: one group participating in a decode-and-forward cooperation and one part using compress-and-forward. In a similar way we can combine the protocols presented in the last two sections. For the sake of brevity we discuss here only one possible mixed strategy although, there exist a variety of possibilities to combine both approaches. Consider the two sets RCF ⊆ R and RDF ⊆ R where RCF contains all relay nodes which use the compress-and-forward approach and RDF contains all nodes using decode-andforward. We further define RCF ∩ RDF = ∅. As before, the nodes t ∈ RCF offer additional gains due to a multi-antenna reception behavior and the relays t ∈ RDF due to a multi-antenna transmission behavior. Furthermore, the nodes t ∈ RDF can improve their own decoding by exploiting the quantized observations of other relay nodes which belong to RCF . This introduces more degrees of freedom to the design of the actual relaying protocol but makes it also more complicated. For instance, the quantization of the relay node outputs has to be done differently for the destination and the relay nodes t ∈ RDF (see also the multiple description source coding problem with successive refinement and unstructured side information [6, 12]). Additionally, we have to consider in this case a broadcasting problem as the nodes in RCF need to send different information to the destination and the relay nodes [3]. Table 9.2 shows a simple example for N = 4 relay nodes, B = 8 source blocks, RCF = {1, 3} and RDF = {2, 4}. Consider in this example the source block xs (w4 ): node t = 1 sends the bin index determined by the quantized version of y1 (4). Relay t = 2 can use this index to benefit from yˆ1,1 (u1,4 ) to decode xs (w4 ) and determine the index s4,1 according to the random paritioning. Finally, the destination can use yd (4), the quantized relay observation yˆ1,1 (u1,4 ) and the binning index s4,1 to decode xs (w4 ). One can see that this kind of mixed strategy is still too static in its structure, which makes it necessary to weaken the constraint that only one relay node is allowed to transmit at a time. Therefore, we take the power set P(R) and use the set PI (R) ⊂ P(R) of those Rp ∈ P(R), kRp k ≥ 2, which define the groups of relay nodes cooperating in the previously described way. If |PI (R)| > 1 we have the case that more than one group is concurrently cooperating using the described protocols, i. e., two or more relays are transmitting at the same time. Now take for instance two sets Rp00 , Rp0 ∈ PI (R), Rp0 6= Rp00 , and let Rp0 ∩ Rp00 6= ∅, i. e., we can further improve the flexibility of the protocol by letting a relay node cooperate with more than one group. Roughly speaking we define clusters of relay nodes (the sets Rp ∈ PI ) which need not be disjoint and each of these clusters cooperates in the

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

193

Table 9.2. Outline of the mixed coding scheme with B = 7 source blocks, πj = {1, 2, 3, 4}, RCF = {1, 3} and RDF = {2, 4}. In the mixed strategy we combine a compress-and-forward part which uses the quantized observations yˆr,k (ur,b−k |xr−k (sb−k )) (the table only shows the abbreviated notation yˆr,k (ur,b−k )) and a decode-and-forward part which decodes the source messages xs (wb ) determining sb,k . b s 1 xs (w1 ) 2 xs (w2 )

t=1 ∅, ∅

t=2

t=3

∅, ∅ yˆ3,1 (u3,2 ), yˆ3,2 (u3,1 )

3 xs (w3 ) 4 xs (w4 )

s4 = (s2,1 , s1,2 )

yˆ1,1 (u1,4 ), 5 xs (w5 ) yˆ1,2 (u1,3 ) 6 xs (w6 )

s6 = (s4,1 , s3,2 ) yˆ3,1 (u3,6 ), yˆ3,2 (u3,5 )

7 xs (w7 ) 8 xs (w8 ) 9

∅

10

∅

t=4

s8 = (s6,1 , s5,2 ) yˆ1,1 (u1,8 ), yˆ1,2 (u1,7 ) s10 = (s8,1 , s7,2 )

described way. The number, size and overlap of the clusters are design parameters depending on the actual scenario. For brevity we forbear from presenting a complex proof of the achievable rates and show instead in the next section some practical applications of the discussed protocols.

9.4 Application to Wireless Communications We conclude this chapter with a short discussion of some practical issues on the basis of published work about alternatingly transmitting half-duplex relays. The first description of such a protocol was given in [23] which combines the amplifyand-forward (AF) approach [18] and the idea of Delay Diversity Codes (DDC) [37]. Consider a four-terminal network with two relays, where only one relay is transmitting/listening at a certain time. Both relays receive a superposition of the source and the relay signal which is amplified by a factor βR and retransmitted by the respective relay node. One can imagine that this approach could lead to a high noise level as each relay is amplifying and retransmitting the signal of the other relay node and its own noise. To avoid such a high noise level, an interference cancellation at one of both relays was proposed. Under the knowledge of the respective channel conditions, one relay subtracts its last sent signal which was received, amplified and resent by the other relay node. In [22] the capacity of a similar approach based on amplify-and-forward and a network of half-duplex relay terminals was analyzed. Rankov and Wittneben proposed in [24, 25] protocols based on amplify-andforward and decode-and-forward which use alternatingly transmitting half-duplex

194

Peter Rost and Gerhard Fettweis

relays. In comparison to the previous one, they consider only scenarios without a direct link between source and destination. A scheme for Direct Sequence Code Division Multiple Access (DS-CDMA) was proposed by Ribeiro et al. in [26]. In their work all nodes, i.e., source and both relays, use different pseudo noise spreading sequences and therefore the relays and the destination can separate the different signals. Another scheme for a four-terminal network was proposed in [27, 28] which uses maximal ratio combining and successive interference cancellation at the destination. It is shown that this approach offers a performance advantage for low SNR values. However, the protocol is interference limited and only approaches first order diversity in the high SNR regime. All these proposals have in common that they are restricted to or at least only analyzed for a scenario with two relay nodes and without an explicit cooperation between both terminals. In [29], by contrast, an approach was presented which implements a Space-Time Trellis Code (STTC) [33] by a network of half-duplex relay nodes. The basic idea is that each relay receives the superposition of the source and relay signals according to a predefined STTC (or a DDC in the simplest case) and employs a maximum a-posteriori (MAP) decoder to decode the source messages. With an increasing number of relay nodes, the relays are able to observe more source/relay messages and therefore decode more reliably. Beside the additional diversity the protocol also gains in performance due to the usage of the STTC. In comparison to the previously mentioned work this approach is easily extendable to a network of relay nodes. Furthermore it implements an inter-relay cooperation and can therefore be seen as a special case of the presented decode-and-forward approach analyzed in Section 9.3.2. The presented protocols are preferable over the approaches mentioned in Section 9.1 in certain scenarios of wireless communications: •

•

•

They ensure a continuous source-destination transmission to make the relaying protocol more transparent, i.e., the source does not necessarily have to know whether the relays are supporting the transmission or not (for instance if a simple DDC is used). Furthermore, the complete source transmission is retransmitted by the relay nodes which might be necessary if the source-destination link suffers from severe channel conditions and does not allow a reliable communication. Furthermore, the protocols are easily extendable to networks and need not increase the rate on the individual links which avoids the typical rate loss in half-duplex relay protcols. The proposal in [29] partly implements the suggested inter-relay cooperation to improve the reliability of the relay nodes.

9.5 Summary In this chapter we presented different protocols for networks of relays which are subject to the orthogonality constraint, i.e., a relay node can not transmit and receive on the same time-frequency resource. Many protocols which were proposed so far use a strategy based on orthogonal channels, i.e., source and relay transmit in different time/frequency slots. This implies a rate loss which can be overcome if both source and relay use the same resources which is done for instance in the nonorthogonal amplify-and-forward protocol. Another approach is to use two or more relays

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

195

which alternatingly transmit. With this proposal we achieve a kind of space division duplexing (SDD) based on a time division duplex (TDD) or frequency division duplexing (FDD) system. In line with this idea we presented different scalable protocols for networks of half-duplex relay nodes, based on compress-and-forward and decode-and-forward. We further showed how to combine these approaches in a mixed strategy and how to extend them to a network with multiple concurrently transmitting relay nodes. However, it still remains open how the protocols discussed in Section 9.3 can be adapted to wireless fading channels. As Laneman et al. showed in [18] the straightforward application of the decode-and-forward approach can not achieve full diversity in a wireless system subject to channel fading. Furthermore, we did not consider the possibility of a feedback channel which makes the protocol more flexible and might increase the performance significantly. Finally, we want to mention that to the authors’ knowledge so far only little work is published on relaying in realistic systems, e.g., system level results about the capacity-coverage tradeoff, cost-benefits analyses or the amount of necessary signaling to implement relaying. Acknowledgement. The authors are gratefully indepted to Dr. Wolfgang Rave, Ernesto Zimmermann and Dr. Patrick Herhold for valuable discussions and their comments that helped enhance the presentation of this chapter. Part of this work has been performed in the framework of the IST project IST4-027756 WINNER II, which is partly funded by the European Union. The authors would like to acknowledge the contributions of their colleagues, although the views expressed are those of the authors and do not necessarily represent the project. This information reflects the consortiums view, the Community is not liable for any use that may be made of any of the information contained therein.

References 1. K. Azarian, H. El Gamal, and P. Schniter. On the achievable diversitymultiplexing tradeoff in half-duplex cooperative channels. IEEE Transactions on Information Theory, 51(12):4152–4172, December 2005. 2. A.B. Carleial. Multiple-Access Channels with Different Generalized Feedback Signals. IEEE Transactions on Information Theory, IT-28(6):841–850, November 1982. 3. T.M. Cover. Broadcast channels. IEEE Transactions on Information Theory, 18(1):2–14, January 1972. 4. T.M. Cover and A. El Gamal. Capacity Theorems for the Relay Channel. IEEE Transactions on Information Theory, 25(5):572–584, September 1979. 5. T.M. Cover and J.A. Thomas. Elements of Information Theory. John Wiley & Sons, Inc., 1991. 6. A. El Gamal and T. Cover. Achievable rates for multiple descriptions. IEEE Transactions on Information Theory, 28(6):851–857, November 1982. 7. M. Gastpar and M. Vetterli. On the Capacity of Large Gaussian Relay Networks. IEEE Transactions on Information Theory, 51(3):765–779, March 2005. 8. Michael Gastpar. On Wyner-Ziv Networks. In 37th Asilomar Conference on Signals, Systems, and Computers, volume 1, pages 855–859, Pacific Grove (CA), USA, November 2003.

196

Peter Rost and Gerhard Fettweis

9. Michael Gastpar. The Wyner-Ziv Problem With Multiple Sources. IEEE Transactions on Information Theory, 50(11):2762–2768, November 2004. 10. P. Gupta and P.R. Kumar. The Capacity of Wireless Networks. IEEE Transactions on Information Theory, 46(2):388–404, March 2000. 11. P. Gupta and P.R. Kumar. Towards an information theory of large networks: An achievable rate region. IEEE Transactions on Information Theory, 49(8):1877– 1894, August 2003. 12. C. Heegard and T. Berger. Rate distortion when side information may be absent. IEEE Transactions on Information Theory, 31(6):727–734, November 1985. 13. P. Herhold, E. Zimmermann, and G. Fettweis. Cooperative Multi-Hop Transmission in Wireless Networks. Journal on Computer Networks, 49(3):299–324, October 2005. 14. A.I. Khinchin. Mathematical Foundations of Information Theory. Dover Publications, Inc., New York, 1957. 15. G. Kramer. Models and Theory for Relay Channels with Receive Constraints. In 42nd Allerton Conference on Communication, Control and Computing, Monticello (IL), USA, September 2004. 16. G. Kramer, M. Gastpar, and P. Gupta. Capacity theorems for wireless relay channels. In 41st Allerton Conference on Communication, Control and Computing, Monticello (IL), USA, October 2003. 17. G. Kramer, M. Gastpar, and P. Gupta. Cooperative Strategies and Capacity Theorems for Relay Networks. IEEE Transactions on Information Theory, 51(9):3037–3063, September 2005. 18. J.N. Laneman, D.N.C. Tse, and G.W. Wornell. Cooperative diversity in wireless networks: Efficient protocols and outage behavior. IEEE Transactions on Information Theory, 50(12):3062–3080, December 2004. 19. J.N. Laneman and G.W. Wornell. Distributed space-time-coded protocols for exploiting cooperative diversity in wireless networks. IEEE Transactions on Information Theory, 49(10):2415–2425, October 2003. 20. R.U. Nabar, H. B¨ olcskei, and F.W. Kneubuehler. Fading Relay Channels: Performance Limits and Space-Time Signal Design. IEEE Journal on Selected Areas in Communications, 22(6):1099–1109, August 2004. 21. R.U. Nabar, F.W. Kneubuehler, and H. B¨ olcskei. Performance limits of amplifyand-forward based fading relay channels. In IEEE International Conference on Acoustics, Speech, and Signal Processing, pages 565–568, Montr`eal, Canada, May 2004. 22. T.J. Oechtering and H. Boche. Capacity of a Gaussian FIR Linear Relay Network. In International Conference on Wireless Networks, Communications and Mobile Computing, Maui (Hawaii), USA, June 2005. 23. T.J. Oechtering and A. Sezgin. A new Cooperative Transmission Scheme using the Space-Time Delay Code. In ITG Workshop on Smart Antennas, Munich, Germany, March 2004. 24. B. Rankov and A. Wittneben. Spectral Efficient Protocols for Non-Regenerative Half-Duplex Relaying. In 43rd Allerton Conference on Communication, Control and Computing, Monticello (IL), USA, October 2005. 25. B. Rankov and A. Wittneben. Spectral Efficient Protocols for Half-duplex Fading Relay Channels. IEEE Journal on Selected Areas in Communications, 2007. to appear.

9 Scalable Cooperation in Multi-Terminal Half-Duplex Relay Networks

197

26. A. Ribeiro, X. Cai, and G. B. Giannakis. Opportunistic Multipath for Bandwidth-Efficient Cooperative Networking. In IEEE International Conference on Acoustics, Speech and Signal Processing, Montr`eal, Canada, May 2004. 27. P. Rost and G. Fettweis. A Cooperative Relaying Scheme Without The Need For Modulation With Increased Spectral Efficiency. In IEEE Vehicular Technology Conference, Montr`eal, Canada, September 2006. 28. P. Rost and G. Fettweis. BER Performance Evaluation of a Continuous Transmission Cooperative Relaying Protocol. In IEEE Communications Theory Workshop, Dorado (PR), USA, May 2006. 29. P. Rost and G. Fettweis. A Distributed Space-Time Trellis Coding Approach for Multi-Terminal Relay Networks. In IEEE International Conference on Communications, Glasgow, Scotland, June 2007. 30. C.E. Shannon. Mathematical theory of communication. Bell System Technical Journal, 27:379–423 and 623–656, July and October 1948. 31. P.C. Shields. The Interactions Between Ergodic Theory and Information Theory. IEEE Transactions on Information Theory, 44(6):2079–2093, October 1998. 32. D. Slepian and J.K. Wolf. Noiseless Coding of Correlated Information Sources. IEEE Transactions on Information Theory, IT-19(4):471–480, July 1973. 33. V. Tarokh, N. Seshadri, and A.R. Calderbank. Space-Time Codes for High Data Rate Wireless Communications : Performance criterion and Code Construction. IEEE Transactions on Information Theory, 44(2):744–765, March 1998. 34. E.C. van der Meulen. Transmission of Information in a T-Terminal Discrete Memoryless Channel. PhD thesis, University of California, Berkeley (CA), USA, 1968. 35. E.C. van der Meulen. Three-Terminal Communication Channels. Advances in Applied Probability, 3(1):120–154, 1971. 36. F.M.J. Willems and E.C. van der Meulen. The Discrete Memoryless MultipleAccess Channel with Cribbing Encoders. IEEE Transactions on Information Theory, IT-31(3):313–327, May 1985. 37. A. Wittneben. Basestation modulation diversity for digital SIMULCAST. In IEEE Vehicular Technology Conference, pages 848–853, St. Louis (MO), USA, May 1991. 38. A.D. Wyner. On Source Coding with Side Information at the Decoder. IEEE Transactions on Information Theory, IT-21(3):294–300, May 1975. 39. A.D. Wyner and J. Ziv. The Rate-Distortion Function for Source Coding with Side Information at the Decoder. IEEE Transactions on Information Theory, IT-22(1):1–10, January 1976. 40. L.-L. Xie and P.R. Kumar. A Network Information Theory for Wireless Communication: Scaling Laws and Optimal Operation. IEEE Transactions on Information Theory, 50(5):748–767, May 2004. 41. L.-L. Xie and P.R. Kumar. An Achievable Rate for the Multiple-Level Relay Channel. IEEE Transactions on Information Theory, 51(4):1348–1358, April 2005. 42. L. Zheng and D.N.C. Tse. Diversity and Multiplexing: A Fundamental Tradeoff in Multiple-Antenna Channels. IEEE Transactions on Information Theory, 49(5):1073–1096, May 2003. ¨ ur, O. L´evˆeque, and D. Tse. How does the Information Capacity of Ad 43. A. Ozg¨ Hoc Networks scale? In 44th Allerton Conference on Communication, Control and Computing, St. Louis (MO), USA, September 2006.

10 Trigger Management and Mobile Node Cooperation Jukka M¨ akel¨ a, Kostas Pentikousis, Mikko Majanen, and Jyrki Huusko VTT Technical Research Centre of Finland [email protected]

Summary. This chapter addresses one of the challenges in cooperative networking, namely, mobility support in a heterogeneous ambient network environment. We motivate the need for efficient mechanisms for handling the large amount of network and channel state information required in assisting fast handovers and network and service adaptation strategies. Managing a variety of network and protocol events and triggering information is a challenging task even in a homogeneous networking environment when different mobility schemes (node, network, session) and application adaptation are considered and, not unexpectedly, the heterogeneity of access networks increases further the amount of such information. We present triggering management mechanisms which efficiently handle triggering information at node and network level, dealing with a greater variety of events originating from any component of the node’s protocol stack as well as mobility management entities within the network. We then discuss the benefits of arranging mobile nodes into specific mobile routing groups, and how such approaches can benefit from the availability of the triggering management mechanisms in an ambient network environment. Key words: ambient connectivity; mobility management; routing group; event triggering; handover decission-making

10.1 Introduction Today’s mobile devices are already capable of running demanding network applications and services, and the devices may also have multiple network interfaces for wireless and wired access networks and thus, they can provide a variety of connectivity options for users. Nevertheless, state-of-the-art mobile protocol stacks can only handle a small set of event notifications, typically related to radio access network (RAN) connectivity, user mobility, and load balancing. For example, signal strength deterioration generally leads to a base station handover in cellular voice; 2G/3G mobile phones typically opt for 3G connectivity when the user moves into a new area; and, sustained high data traffic loads may force the Universal Terrestrial Radio Access (UTRA) transport function to reallocate resources (and even perform a handover) in WCDMA 3G/UMTS networks [9].

199 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 199–211. c 2007 Springer.

200

Jukka M¨ akel¨ a et al.

In Section 10.2 we argue that we need a framework to handle a much larger set of notifications caused by events that originate not only from the lower layers of the protocol stack (physical, data link, and network), as in the examples, but from the upper layers (session, transport, and application) as well. Then, in Section 10.3, we present mechanisms that allow mobile devices to manage mobility events and the associated notifications, which we refer to as triggers in the remainder. These trigger management mechanisms lay the foundation upon which sophisticated handover operations can be performed, and establish an extensible framework where new sources of triggers can be included as necessary. For example, as we will see in Section 10.4, triggers can be very useful in forming routing groups in an ambient network setting. After introducing the routing group concept, for scenarios where a set of mobile nodes move in unison, we overview the gateway selection architecture in Section 10.4.3 and conclude the chapter summarizing the main points.

10.2 Mobility Triggers Different kinds of events may trigger mobility management actions: traditional radio link specific conditions, context-dependent, security-related, upper-layer requirements and other system-, application- or user-dependent events. To cater for all these events, general and coherent mechanisms are needed to enable mobility triggering and to identify related events on different protocol layers in a distributed system. Trigger sources and trigger information will have to be included in the mobility architecture. We concur with Eisl et al. [4] that the decision process (arbitration between triggers, policies), and in particular the relative roles and cooperation of the involved networks and devices regarding mobility triggering and rule setting, needs to be handled by a generic and cohesive framework. Mobility actions should be executed based on unambiguous decisions, even if there were several conflicting triggers. Some of the events may be seen as forcing triggers, while some might be suggesting hints, either predicting or triggering. Eisl et al. [4] explain that the difference between triggering and predicting is that the latter enables anticipation of a seamless handover. The aim is that based on an analysis of different triggers, such general functionality can be developed for mobility management like deciding about handovers.

10.2.1 TRG Producers and Consumers We defined a common triggering framework (TRG) [6] which receives “events” from the sources (TRG Producers), process them, and generates “triggers” which are then dispatched to interested parties, called TRG Consumers. In this framework, producers register with TRG before starting to send measurement and events. The registration can be seen as a form of contract between TRG and the producer. The latter affirms its commitment to report events it deems important for further propagation, while TRG guarantees that triggers based on these events will be delivered to interested consumers. Conventional event sources include, for example, different radio interfaces reporting events associated with radio access characteristics, such as, current or average network capacity load, signal-to-noise ratio (SNR), dropped frames ratio, received signal strength indication (RSSI), to name a few; the battery

10 Trigger Management and Mobile Node Cooperation

201

reporting its SOC; SecMon reporting security alerts; and the system CPU load. Mobility protocol state transitions and routing table updates can also be reported by designated producers along with changes in the radio access, system settings, and user preferences. Consumers include firstly the handover decision-making process, but also user applications, mobility protocols, and other functions interested in optimizing their performance in a mobile, multi-access network environment. Consumers have to contact TRG and declare their interest in a particular (set of) trigger(s). A mobility protocol can take advantage of triggers that inform it about not only the activation of a given link (“link up”), but also about crossing a threshold in the battery state of charge or the RSSI, or any combination of two or more events. For example, in a wireless sensor network, gateway nodes may decide to kick start the process to elect a different set of gateway nodes if their traffic load is too high and their battery state of charge is too low.

10.2.2 The Role of TRG TRG is mostly concerned with mobility-related events, and any other information that can assist handover decision-making, thus, the first category of TRG consumers includes the mobility management protocol(s) employed in the device. Because of that, TRG has to be very compact and perhaps trade some flexibility in order to deal with events expeditiously. It is important to highlight our design decisions regarding TRG. Triggers are meant to facilitate mobility management in ways that are not available today and improve application performance but should not be required for proper operation. The mantra is that anything that works today without triggers should continue doing so regardless if it has access to triggers in the future. TRG only provides the reliable information stream that can enable the mobility management protocol, for example the Host Identity Protocol (HIP) [7] or Mobile IP (MIP) [8] [3], to take handover decisions in an educated manner. While receiving the (raw) event information from the producers, TRG maintains an internal repository of triggers generated based on the incoming data and filtering criteria registered by each consumer. It is the job of TRG to take “raw” readings and deliver standardized notifications (in a single format), based on the consumerdictated criteria. For example, TRG may have three different consumers registered to receive triggers based on these measurements: a monitoring tool, a voice over IP (VoIP) client, and an email client, which are interested in (a) the readings as they occur, (b) an aggregate/mean value every so often, and (c) a single notification when a certain threshold is exceeded, respectively. All three consumers receive their respective triggers based on the frequency they specify, and they only need to express this to a single entity: TRG. TRG can easily deal also such measurements like UMTS cell load, CPU utilization, or end to end measured round trip times in a similar way factoring out common functionality. Without TRG, several consumers interested in the same or similar information need to poll system components and do the correlation on their own, replicating functionality over and over again. With the introduction of TRG, information flows become more straightforward, less convoluted and, as explained above, with added functionality factored out in a single reference point in the architecture of the mobile system, as illustrated in Figure 10.1.

202

Jukka M¨ akel¨ a et al.

4 ... VoIP Email O iiii LLL r8 i r i i r LLL r iiii LLL rrriiiiii r r i r ii / TRG ojUi MobileIP o 8 O fLLULUUU Battery r r LLL UUUUUU rrr UUUU LLL r r UUUU L rrr UU ... SecMon CPU MutliRadio 4 8 O h hhhqqqq h h h hhhh qqqq hhhh qq hhhh

MobileP2P fL

UMTS

WiFi

BT&IR

Figure 10.1. The role of TRG in future mobile systems: single point of reference.

Our approach lends itself well to developing a standard application programming interface (API) for receiving triggers and allows consumers to request triggers based on their own filtering criteria and temporal patterns, as is explained in the following Section, which introduces our architecture for trigger management.

10.3 A Trigger Management Architecture The bird’s eye-view of the TRG framework is shown in Figure 10.2 and comprises (a) sources (or producers) feeding relatively fast-changing information about events; (b) trigger consumers, which receive notifications in the form of standardized triggers about events they are interested in; and (c) the implementation of TRG, which includes data stores and internal logic. The TRG implementation processes the events received from the producers, generating triggers based on consumer-provided (filtering) rules and making sure that all system-wide policies are enforced.

Figure 10.2. Schematic of TRG.

10 Trigger Management and Mobile Node Cooperation

203

Designating entities as producers and consumers of triggers, as described above, is a central part of the design (see also Figure 10.3). Consumers must state their need to receive triggers and can choose to stop receiving them anytime. The same entity may be simultaneously acting as both a consumer and a producer. For example, it can receive all triggers originating from RAN events, but opt to receive only the upper-layer triggers associated with security policy violations. In the former case, the consumer takes advantage of the trigger grouping and classification functionality [6]; in the latter, it additionally requests trigger filtering. Consumers can, of course, use these triggers to generate their own and serve as a producer for other entities. We expect that TRG will be used to guide handover decision and execution. In particular, consumers can use triggers to derive whether the mobile device is moving within a single network or it is crossing different access technology boundaries, and whether the addressing scheme, trust and provider domains should be changed accordingly. The TRG architectural requirements address functional, performance, and security issues. As shown in Figure 10.3, the core TRG implementation is partitioned in three major components, namely, trigger event collection (Section 10.3.1), processing (Section 10.3.2), and storage (Section 10.3.3), described next. The figure also depicts examples for TRG event sources (access technology, HIP, MIP, and TCP) and TRG consumers (applications, TCP, MIP, and HIP). As mentioned earlier, the same component may act as a trigger source and consumer at the same time, such as the case of HIP in the figure. In short, events are collected from the corresponding sources and are handed over to the trigger processing engine which is responsible for time-stamping and reformatting triggers (if necessary), and assigning them to the appropriate group. Consumers-specific rules for filtering and the filtering itself are handled during trigger processing. Processed triggers typically have an expiration time, after which they are automatically removed from the active triggers repository. We support the application of different triggering policies (Section 10.3.4), defined as a set of classification, filtering, trust, and authorization criteria/rules. This allows our implementation to enforce a different policy at different times or when the node operates in different contexts.

10.3.1 Triggering Events Collection Triggering events collection is a function in TRG, which receives events from various sources in the network system via the trigger collection interface. A TRG implementation may contain several event collectors, which may be distributed and which may be responsible for collecting different types of events. The need for different event collectors arises from the fact the origin of an event source can be a hardware device, a system component implemented in kernel space, or an application implemented in user space. For example, each device driver could implement its own event collection functionality, which would be capable of handling triggering events produced by the specific device only. Moreover, sources can also be located in the network such as at active network elements or at the user’s home network. Finally, a particular TRG implementation can act as a consumer to another one located in a different node. Thus, orchestrating the collaboration of, perhaps, several collection entities is needed in order to efficiently gather a larger amount of events.

204

Jukka M¨ akel¨ a et al.

Figure 10.3. TRG Architectutal Components.

10.3.2 Trigger Processing We want to handle triggers using a common format and reformat any “legacy” ones into the chosen standardized format once they arrive via the triggering event collection. This allows us to tap into existing event sources, which are not yet compatible with our TRG and will be instrumental in migrating “legacy” systems to the new framework. New sources should implement the trigger event collection functionality and use the trigger collection interface in order register their triggers and to make them available to consumers. Consumers can subscribe by specifying a set of triggers (and, optionally, filtering rules) and are expected to unsubscribe when they do not wish to receive them any longer. For each consumer subscription, the TRG processing component makes sure that filters are formatted correctly, may supply default filters for certain consumers, and performs the actual filtering. Basic rules can also be used as building blocks for crafting more sophisticated rules.

10.3.3 Trigger Repository The trigger repository aims at meeting the stringent requirements placed by mobility management, but can be used to store non-mobility triggers. The basic primitives include adding, removing, updating, and disseminating triggers in a standardized format. Each stored trigger has an associated lifetime and is removed automatically once it expires.

10 Trigger Management and Mobile Node Cooperation

205

10.3.4 TRG Policies and Rules The availability, on the one hand, of a system-wide policy and, on the other, consumer-supplied filters lies at the center of our TRG design. These two are orthogonal, providing flexibility and adaptability. System policies ensure that only designated consumers can receive certain groups or types of triggers. For example, a node may operate under different policies regarding network attachment depending on whether the user is on a business or a leisure trip. Policies can also establish different trigger classification and groupings in different contexts and are typically stored in a separate repository, accessible to the TRG implementation. Filters allow a consumer to focus a trigger subscription. For example, a monitoring application may be interested in receiving all network utilization measurements, while a VoIP application may be interested in a receiving a trigger when utilization exceeds a threshold and the user is in a call. The VoIP application can even be an intermittent trigger consumer, subscribing and unsubscribing to receive triggers as needed.

10.4 Routing Group Cooperation In traditional networking solutions, mobility is handled on a per-node basis, that is, each node is handled individually. However, with the recent emergence of wireless personal area networks there is an increasing number of scenarios in which a number of nodes move together, such as, for example, a group of users onboard an intercity train. Under such circumstances, there is potential for several optimization strategies in arranging group communication. Since the mobile devices are all moving together, they will all be handing over at the same time. The amount of signaling required for a handover can be reduced by aggregating the mobility management procedures to a single node. Devices without inherent mobility support can also exploit the mobility capabilities of the other nodes in the moving network. Also the routing between the nodes can be enhanced locally by using advanced ad-hoc routing protocols. Network-wide routing can be optimized by using the fact that certain nodes are moving close together. The nodes in the moving group can also use shared applications. Aggregated mobility management and use of advanced ad-hoc routing protocols require some kind of agreement between the nodes, whereas other optimizations do not need any kind of agreements between nodes. Because of the different classes of optimization that are possible, Surtees et al. distinguish the following two concepts [10]: A (physical) cluster is a group of nodes that are physically near to each other, are likely to stay near to each other, and are able to communicate. A routing group (RG) may be formed by (a subset of) the nodes within the (physical) cluster. Formation of a routing group requires information exchange and agreement between the nodes involved. The distinction between a (physical) cluster and a RG is that the nodes in the cluster are not aware of group membership and as such cannot actively participate in RG optimisations, whilst the nodes in the RG are aware of group membership and can actively support functionality for routing and mobility optimizations.

206

Jukka M¨ akel¨ a et al.

In order to make use of RGs and the associated optimizations, it must be possible to form a RG. However, the formation and maintenance of RGs bring significant challenges. At first, the cluster should be identified with minimal overhead for the nodes in order the gains from clustering could outweigh the costs of formation. Secondly, the stability of the clusters should be able to be evaluated (e.g., by using mobility patterns) so that it can be determined whether it is worth to form a RG. The RG environment extends the basic peer-to-peer (P2P) communication environment to peer-to-multiple peers environment, where information consistency may be an issue, e.g., the information regarding RG membership decisions. RG may include several different radio access technologies and be multi-homed and have a radius greater than one hop (i.e., multi-hop); all these bring their own challenges to RG formation and maintenance. One of the main objectives of the triggering mechanism is to deal with triggers that could originate a handover process. In addition to this, the triggering mechanism can be used for providing triggers for routing group management. This function will also be related to hand off of active sessions (due to change in default router or route) pertaining to a node joining the routing group. Depending on the routing group algorithms, triggers could be used for a variety of functions. These triggers could be the same ones or combinations of those collected by the collection function of the triggering mechanism.

10.4.1 Routing Group Formation In order to transit from the cluster to a RG, a RG formation algorithm is used to identify which nodes in the cluster should become members of the RG. The formation algorithm can be modeled in several iterative steps, illustrated in Figure 10.4, and are as follows [10]: Triggering RG formation This first stage identifies that it may be worthwhile to try to form or join a RG based on internal or external triggers received by the node. The triggers that initiate routing group formation range from locally generated indications that a new interface is available with appropriate characteristics to support local routing optimizations, through to more complex triggers, for example, generated by the network only when the environment is appropriate for routing group formation to take place. Regarding routing group management, the first category of trigger is quite general, and may initiate attempts to form routing groups even though it may not be worthwhile to do so. Therefore, additional context information about the environment within which the endpoint is currently operating may be useful to augment this trigger. The latter type of trigger is more complex to generate, and requires more intelligence within the network to establish when the environment is suitable for routing group formation. This information could also be used to augment the former, more general trigger information above. Where trigger information is generated by the network, some means to transport this information to the endpoint is required as for any other sort of network generated trigger. Basic communications between the joining node and one (or more) of its neighboring nodes are also established in this phase.

10 Trigger Management and Mobile Node Cooperation

207

Basic exchange Information is exchanged to find out whether there is a stable group of nodes, such that a RG may be formed. Stability of the group of nodes is a key factor in determining whether routing group formation is feasible. Iterative exchange Depending on the algorithm and the criteria, further refinements of the RG may take place. This will include consideration of more complex policy such as trust agreements and user policies. Nodes with specific roles may be elected during this stage, e.g., gateways and cluster heads (the blue and red nodes in the figure, respectively), and at the end of this stage, mobility management may be delegated to gateway node(s). Maintenance This phase takes care of nodes leaving or joining the group; special actions are triggered, if a node in special role (such as gateway or cluster head) is leaving or lost.

Figure 10.4. RG formation phases.

The benefits of creating a RG and establishing routing and mobility optimizations include, for example, reduction in communication costs, mobility management overhead, battery consumption, and better quality of service. On the other hand, resources are used for creating and maintaining the RG. Whether it is useful to form a RG depends highly on the topological stability of the nodes in the original cluster. In the following sections, we present one RG formation protocol together with an architecture for gateway selection in more detail and discuss their performance.

10.4.2 Stability-Based Multi-Hop Clustering Protocol Associativity is defined as periods of spatial, temporal, connection and signal stability and it is usually estimated by sending periodical Hello messages. The concept of associativity was originally presented by C.K. Toh with the ABR routing protocol [12]. Later on, this concept has been used also in clustering, for example in ABC protocol [1]. The RG formation protocol presented in [11] assesses also nodes’ stabilities with an associativity-based method. However, in this protocol the stability is measured not only by using associativity ticks, but the stability estimates are enhanced with link quality information (using SNR or PER). The protocol has three phases: neighbor discovery, cluster head selection and maintenance.

208

Jukka M¨ akel¨ a et al.

Neighbor Discovery The RG formation is started with the identification of the physical cluster. The nodes’ proximity and movement patterns are estimated by monitoring the neighboring nodes and by assigning them a stability value. For this, all nodes transmit “Hello” messages every TH seconds (hello interval). Every time such a message is received within a certain time-frame, the stability value of the sending node is incremented. Correspondingly, if the message is not received in that time, stability value is decreased. Also, if the SNR of the received message is bad or message is totally lost (sequence numbers are used for detecting missing messages), the stability value is decreased. “Hello” messages are broadcasted but not forwarded, so they are received only by the first neighbors of the sending node. When the stability value of the node exceeds a certain threshold, the node is considered stable. The nodes continue sending “Hello” messages even after the RG is formed. Thus, the stability of thenodes is monitored continuously.

Cluster Head Election Each stable node calculates a cluster head suitability value for itself. The suitability value is a weighted sum of the number of neighbors, their stabilities, and the remaining battery level of the node itself. This value is broadcasted to all stable nodes in a role advertisement message. When a node receives such a message, it compares the suitability value to its own suitability value and if that is bigger, sends a new role advertisement message within a certain time window in order to be part of the election process. After the time window has closed the node with biggest suitability value broadcasts a role claim message to other nodes informing that it is now the cluster head. RG can be seen formed after the cluster head has been elected.

Maintenance Phase In the maintenance phase, new nodes may join the RG or former members leave. Both cases can be detected from the hello messages. Also, two RGs may merge if nodes belonging to different RGs form stable link between them. The decision of accepting new members or merging RGs can be made by the cluster head, or alternatively, all nodes can be authorized to make it in a distributed manner. Tenhunen et al. [11] explore two scenarios in order to evaluate the RG formation protocol. In the first scenario, the purpose was to validate the functionality of the protocol and to find suitable values for the essential protocol parameters, especially the “Hello” interval TH . The scenario was run in 50 randomly generated topologies, each of which had 75 static nodes distributed randomly in a 100 m×100 m area. The scenario verified that the protocol is able to identify physical clusters and form RGs out of nodes residing close to each other. A hello interval of 0.5 seconds was found to be a good compromise between adaptability and control message overhead: shorter intervals could enable more reliable stability estimates and hence faster adaptability to topology changes, but on the other hand, the number of sent control messages increased drastically. The second scenario evaluated how the protocol maintains the RG’s integrity while it passes through static nodes. The scenario consisted of a group of nodes

10 Trigger Management and Mobile Node Cooperation

209

randomly placed inside a rectangular 7 m × 7 m area, and one static node, through which the group moved with varying speeds. The desired behavior in this case is that the static node does not join the moving RG. For each RG size (2, 3, 5, 10, and 20 nodes) ten random topologies were generated. From the results it could be seen that the protocol needs different parameter values (especially the stability threshold value) in order to work properly in different user scenarios and environments.

10.4.3 An Overview of the Gateway Selection Architecture Gateway (GW) is a service providing access to outside of the RG. There are basically three approaches for discovering and selecting gateways: Proactive approach—the gateways broadcast advertisement messages to the whole routing group. Nodes select the GW for themselves based on the advertisements they receive. Reactive approach—the routing group nodes needing GW service broadcast request messages to the whole RG. Gateway nodes respond to these requests and the requesting nodes select the GW for themselves based on the responses they receive. Hybrid approach—the GW advertisement messages’ life-time can be limited, for example, to a few hops; nodes inside this range use proactive approach for GW selection, whereas nodes outside this range use reactive approach. The Gateway Selection Architecture (GSA) [2] was developed based on the assumption that some RG nodes are not capable or willing to do GW selection for themselves. GSA introduces special Gateway Selector (GWS) nodes that make the GW decision on behalf of other nodes. Usually, GWS is the same node as the cluster head. Thus, it has much more information about the RG than other RG nodes and it should be able to do more sophisticated GW selections for others. GSA follows a hybrid approach where gateways send their advertisements only to the GWS. Nodes needing GW service send a request to the GWS. GWS selects the most suitable GW for the requesting node based on the information it has received from the GW advertisements and service requirements from the GW request messages. In GSA, the advertisements and requests are unicasted to the GWS node; this is a major difference compared to proactive and reactive approaches, in which these messages are broadcasted. GSA is independent from the mobility protocol used. With little changes in signaling, the GSA can be taken into use for example with MIP and HIP protocols as explained in [5]. The performance of GSA is evaluated in [5]. In that paper, a commuter train scenario was used, in which a group of nodes (from 3 to 98 nodes) were moving together in a train. One of the nodes was set to provide GW services for other nodes in MIP-like manner. The RG formation protocol described in previous section was used for selecting GWS node (being the same as the cluster head). The signaling overhead for discovering and selecting the gateway (in this case, MIP’s FA) was studied by calculating the average number of sent and processed (i.e., sent, forwarded, received, and dropped) control messages per mobile node. GSA was compared both to proactive and reactive approaches, as well as to a case where every mobile node used MIP individually. Clearly, the last MIP case underperformed all the other approaches.

210

Jukka M¨ akel¨ a et al.

GSA sent little bit more messages than proactive and reactive approaches, but, since they were unicasted instead of broadcasted messages, GSA performed best in terms of processed messages in RGs with more than 15 nodes.

10.5 Conclusion The mushrooming of different wireless access networks such as IEEE 802.11 based “WiFi” solutions, IEEE 802.16 “WiMAX” and 3GPP systems has generated the demand for purely media independent handover procedures. The requirements rising also from the mobility management, service initiated connection controlling, service quality guarantees and seamless cooperation of different access networks have created demand for efficient event handling mechanism in network devices, which can utilize both access network specific events and events initiated by different network protocols, applications and services. In this chapter we presented a generic triggering management mechanism, which is able to cope with the aforementioned requirements and handle different type of triggering events for mobility management purposes in small devices. Even though we considered mainly the mobility management decision-making process, similar triggering mechanism approach can be utilized in controlling the variety of other network and service functionalities such as application controlling, and QoS provisioning of the network. We deepened our triggering management solution by examining also the routing group formation for personal networking and we presented the group formation process utilizing triggering information by using the stability-based multi-hop clustering protocol. By aggregating e.g., the mobility functionalities into only a limited amount of nodes and utilizing the triggering management in controlling the group formation, we can argue that whole system will benefit from proposed solutions in terms of bandwidth utilization and complexity. As a conclusion, the proposed triggering management solution fits well also for currently ongoing work on media independent handovers (IEEE 802.21), providing the means for both cross-layer information aided vertical and horizontal hand offs. Acknowledgement. The better part of this work has been carried out in the framework of the Ambient Networks project, which is partially funded by the Commission of the European Union. The enlightening and fruitful discussions with our colleagues in the Ambient Networks project were instrumental in furthering our work. The views expressed in this paper are solely those of the authors and do not necessarily represent the views of their employer, the Ambient Networks project, or the Commission of the European Union.

References 1. Y. Choi and D. Park. Associativity Based Clustering and Query Stride for Ondemand Routing Protocols in Ad Hoc Networks. In Journal of Communications and Networks (KICS), volume 4, March 2002.

10 Trigger Management and Mobile Node Cooperation

211

2. Michael Eyrich, Mikko Majanen, Eranga Perera, Ralf Toenjes, Roksana Boreli, and Tim Leinmueller. GSA: An Architecture for Optimising Gateway Selection in Dynamic Routing Groups. In The 16th Annual IEEE International Symposium on Personal Indoor and Mobile Radio Communications, Sept. 11–14, Berlin, Germany, September 2005. 3. T.R. Henderson. Host mobility for IP networks: a comparison. In IEEE Network, volume 17, pages 18–26, 2003. 4. Jochen Eisl et al. Mobility architecture & framework - d4.2 core report, March 2005. IST-2002-507134-AN/WP4/D4.2. 5. Mikko Majanen and Kostas Pentikousis. An Evaluation of the Ambient Networks Gateway Selection Architecture. In The Third International Conference on Wireless and Mobile Communications (ICWMC 2007), Guadeloupe, French Caribbean, March 2007. 6. Jukka M¨ akel¨ a and Kostas Pentikousis. Triggering Management Mechanism. In International Symposium on Wireless Pervasive Computing, San Juan, Puerto Rico, February 2007, February 2007. 7. R. Moskowitz and P. Nikander. Host Identity Protocol (HIP) Architecture. In Internet RFC 4423, May 2006. 8. C. E. Perkins. Mobile IP. In IEEE Communications Magazine, volume 40, pages 66–82, March 2002. 9. P. Prasad, W. Mohr, and W. Konh¨ auser. Third Generation Mobile Communication Systems. MA: Artech House Publishers, Boston, 2005. 10. Abigail Surtees, Ram´ on Ag¨ uero Calvo, Jari Tenhunen, Michele Rossi, and Daniel Hollos. Routing Group Formation in Ambient Networks. In 14th IST Mobile & Wireless Communications Summit, June 19–22, Dresden, Germany, June 2005. 11. Jari Tenhunen, Ville Typp¨ o, and Marko Jurvansuu. Stability-Based MultiHop Clustering Protocol. In The 16th Annual IEEE International Symposium on Personal Indoor and Mobile Radio Communications, Sept. 11–14, Berlin, Germany, September 2005. 12. C. K. Toh. Associativity-Based Routing Fro Ad-Hoc Mobile Networks. In Kluwer Wireless Personal Communications, volume 4, pages 103–139, March 1997.

11 Cooperative Mobile Positioning in 4G Wireless Networks Simone Frattasi1 and Marco Monti2 1

2

CTIF Denmark, Dpt. of Electronic Systems, APNet Group, Aalborg University [email protected] CTIF Italy, Dpt. of Electronic Engineering, University of Rome Tor Vergata [email protected]

Summary. Cooperative mobile positioning is raising up as a promising new branch of wireless location, in which several research directions are being explored. In this chapter, we apply the cooperative mobile positioning framework as an innovative solution for positioning determination in 4G wireless networks by introducing the Ad-Coop Positioning System (ACPS). The ACPS is supported by a hybrid cellular ad-hoc architecture, where the cellular network has a centralized control over the adhoc connections among pairs of mobiles. Specifically, peer-to-peer communications are exploited in a mesh fashion within cellular-established clusters for cooperationaided localization purposes (from that, the word ad-coop is derived). The numerical results exposed in the chapter will show that thanks to the spatial proximity and spatial diversity within a group of cooperative mobiles, our proposal has the potential to enhance the location estimation accuracy with respect to conventional hybrid positioning techniques in stand-alone cellular networks.

11.1 Introduction Geolocation, i.e., location estimation in terms of geographic coordinates of a mobile with respect to a specific coordinate system, has gained considerable attention over the past decade, especially since the FCC in 1996 passed a mandate requiring cellular providers in the USA to generate accurate location estimates for E-911 services [7] (see Table 11.11 ). Such a mandate has been extended also to the EU in 2003, where mobile positioning is considered even a more critical issue, due to the continually increasing mobile originated E-112 calls [20]. Nevertheless, the EU will most likely follow USA and Japan in requiring high positioning accuracy from 2010, when Galileo will be fully operational [14]. In the meanwhile, the research in the field of wireless location has been boosted as an important public safety feature, which can also add many other potential applications to future cellular systems [22]: 1

The FCC requirements are expressed in terms of CEP. For instance, CEP67 = 100 m means that at least 67% of the radial positioning errors are smaller than 100 m.

213 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 213–233. c 2007 Springer.

214

Simone Frattasi and Marco Monti

user routing and navigation, location-sensitive billing, information browsing, fraud protection, user / asset tracking, fleet management, intelligent transportation systems, mobile yellow pages, planning for networking, catching of information closer to the user, managing and enhancing wireless resource allocation, system design and management. Table 11.1. FCC requirements [8]. Specification Network-based Mobile-based CEP67 FCC-N 1: 0.1 km FCC-M 1: 0.05 km CEP95 FCC-N 2: 0.3 km FCC-M 2: 0.15 km

The basic concept behind any mobile positioning system is to measure some key parameters extracted from wireless signals either received at the mobile from some fixed reference points, such as BSs or satellites, or viceversa. According to the place in which the location estimation is performed, these systems are referred to as mobilebased or network-based (in this latter case, the measurements obtained at the BSs are relayed to a central site for further processing and data fusion). The underlying technologies used in mobile positioning systems are either satellite-based or terrestrial-based, which differ from each other in terms of accuracy, coverage, cost, power consumption, application environment (outdoor or indoor), and system impact. Satellite-based technologies are employed mainly for outdoor applications and are represented by the GPS. Even though the GPS is the most popular solution on the market, the introduction of handsets with built-in GPS receivers leads to an increased cost, size, battery consumption, and a long time for a full market penetration [22]. Furthermore, it is sometimes unfeasible in dense urban environments to obtain any sort of location information due to the impossibility of having a clear view of at least four satellites, or due to signal blocking and multipath. Terrestrialbased technologies have the same drawback of the GPS in multipath environments and in NLOS conditions, when no accurate environmental information is available. Hence, investigations have started in connection with 4G in order to define a solution – probably integrated with the GPS –, which would be able to provide location information with a high level of accuracy anywhere and anytime. Toward this end, in addition to novel air-interface technologies and collocated antenna technologies, some major modifications in the wireless network architecture itself are required. The most promising architectural upgrade relies on the use of a combination of the cellular network model with the P2P one, which is usually used only in a special class of wireless networks called ad-hoc networks. Whereas in conventional cellular networks mobile hosts operate in a purely peer-agnostic fashion, in ad-hoc networks, they act cooperatively as routers or relays for other hosts, where communications are enabled through multi-hopping without the need for a centralized BS. By using transmission powers that are just large enough to ensure network connectivity, the ad-hoc network model achieves several performance benefits over the cellular one, including better spatial reuse characteristics and lower energy consumption [12]. It is straightforward to realize that a hybrid network model, such as the cellular ad-hoc one, is the most natural type of environment in which cooperation not only between users or terminals, but also between networks can be established and best flourish. Cooperation is indeed a raising alchemic paradigm in wireless communications,

11 Cooperative Mobile Positioning in 4G Wireless Networks

215

which gives the designers the potentials to achieve enhancements in terms of data rate, coverage, and energy consumption [9]. These goals can be achieved either by exploiting exclusive cooperative stations (e.g., fixed or mobile relay stations / access points) or short-range communications among neighboring mobiles. To the best of our knowledge, there is no research dealing with the impact of cooperation on location estimation accuracy in such hybrid types of networks. Cooperative mobile positioning is raising up as a promising new branch of wireless location, in which several research directions are being explored (e.g., positioning, tracking, data fusion, clustering, etc.). Its identifying concept is the exploitation of likely reliable mobile-to-mobile measurements to increase the location estimation accuracy of a wireless system, which would have been provided otherwise only with usually unreliable fixed-to-mobile measurements. In this chapter, we apply the cooperative mobile positioning framework as an innovative solution for positioning determination in 4G wireless networks by introducing the Ad-Coop Positioning System (ACPS). The ACPS is supported by a hybrid cellular ad-hoc architecture, where the cellular network has a centralized control over the ad-hoc connections among pairs of mobiles. Specifically, P2P communications are exploited in a mesh fashion within cellular-established clusters for cooperation-aided localization purposes (from that, the word ad-coop is derived). Indeed, upon receiving a location information request from a certain MS, the home BS forms a cluster in the surroundings of that MS, where the latter can be elected as CH and its neighbours as CMs. While the available BSs perform time / time-difference and angle-based measurements for each BS-CM link, each CM performs range-based measurements for each CM-CM link. Finally, the position of each CM is obtained by a novel data fusion method, which appropriately combines the available long- and short-range location information. The numerical results exposed in Section 11.5 will show that thanks to the spatial proximity and spatial diversity within a group of cooperative mobiles, our proposal has the potential to enhance the location estimation accuracy with respect to conventional terrestrial positioning systems in stand-alone cellular networks. The rest of the chapter is organized as follows: Section 11.2 presents the related work on wireless location, with a detailed look into hybrid positioning techniques and NLOS error mitigation techniques; Section 11.3 introduces the ACPS with its system architecture and its data fusion method; Section 11.4 describes the models used to carry out the simulation results; Section 11.5 discusses the performance of the ACPS evaluated via computer simulations; and Section 11.6 illustrates the potential impact of cognition on the proposed framework. Finally, the concluding remarks are given in Section 11.7.

11.2 Related Work Wireless location technologies mainly involve the measurements of TOA, TDOA, AOA, or RSS of radio signals either received or transmitted by an MS. Specifically, it can be recognized that: •

The dominant wireless location technologies for cellular networks are based on TOA, TDOA, and AOA measurements. Whereas TOA-based schemes require a

216

•

Simone Frattasi and Marco Monti tight synchronization between the clocks of the transmitting BS and the receiving MS, TDOA-based schemes are free from this restriction, but their accuracy is highly dependent on the relative geometric location of the BSs [26]. Generally, both TOA- and TDOA-based schemes achieve better accuracy than AOA-based schemes, but they both require at least three BSs for 2D positioning. AOAbased schemes, on the other hand, can involve a minimum of two BSs, but they are highly range dependent: a small error in the angle measurement results in a large location error when the MS is far away from any BS involved. Also hybrid location techniques, which utilize combinations of time, time-difference and angle measurements, have been proposed for cellular networks [5]- [24]. They are especially useful in hearability-restricted conditions or low infrastructure environments, i.e., when power control is employed or the density of BSs is low; The dominant wireless location technologies for ad-hoc networks are based on TOA and RSS measurements. Although the latter are in general easily available, RSS-based schemes have been circumvented in cellular networks because of their dependency on the distance between the MS and the BSs involved. In ad-hoc networks, instead, since the distances are sensibly reduced and thus the likehood to have a fixed reference point close to the MS is substantial, they can perform as well as TOA-based schemes [19]. AOA-based schemes are limited by the need of deploying antenna arrays at the user terminal, which can negatively contribute to increase its size. Finally, besides the traditional multi-lateration techniques that involve TOA and RSS measurements, also cooperative location techniques have been proposed for ad-hoc networks. For example, in wireless ad-hoc sensor networks, TOA and RSS measurements are not only made between unknownand known-location sensors, but also between unknown- and unknown-location sensors. The additional information gained from the measurements between pairs of unknown-location sensors can enhance the accuracy and robustness of the localization system [19]. However, these location techniques are usually exploited in a multi-hop fashion, since each sensor is placed K ≥ 1 hops from a fixed reference point and can only communicate with its one-hop neighbours. This is because low-capability, energy-conserving devices such sensors do not include a power amplifier and thus lack the energy necessary to perform long-range communications to the remote BS.

To the best of our knowledge, most of the existing studies in the literature treat the location estimation separately for cellular and ad-hoc networks.

11.2.1 Hybrid Positioning Techniques In order to make an accurate location estimation, traditional time-based schemes in cellular networks combine several estimates obtained at different BSs. However, in situations where the MS is much closer to one BS (home site) than the others, their accuracy is significantly degraded because of the relatively low SNR of the received MS signal at the neighboring BSs. Such accuracy is further reduced due to some interference mitigation technique, such as power control, which requires the MS to decrease its transmitted power when it approaches a BS, causing what is known as the hearability problem [24]. In such conditions, hybrid techniques, which combine TOA, TDOA and AOA estimates can reduce the number of BSs to involve in the location estimation. The simplest hybrid method is exposed in [5], where a hybrid

11 Cooperative Mobile Positioning in 4G Wireless Networks

217

TOA / AOA positioning (HTAP) technique is used to combine the range and the direction of the MS only with respect to the home BS. However, the HTAP is mostly feasible in micro-cellular environments where the MS may be relatively close to the home BS, so that the effect of the angular error is reduced. Furthermore, it is only suited for 2D location determination, since just one BS is used for location purpose. Many other hybrid location techniques can be used, such as combining TDOA and AOA estimates. For example, the hybrid TDOA / AOA technique proposed in [3] achieves good accuracy in small error conditions, but it requires measurements at four or more BSs for 2D positioning.

11.2.2 NLOS Error Mitigation Techniques The accuracy of the above-mentioned techniques depends on the propagation conditions of the wireless channel. In case of LOS propagation between the MS and all the BSs available for location purpose, the main error source is Gaussian measurement noise, and a high location estimation accuracy can be achieved. Otherwise, since the direct path from the MS to a BS can be blocked by buildings and other obstacles, the range and angle measurements will include a NLOS error due to single or multiple reflections or diffractions. In order to protect location estimates from NLOS error corruption, NLOS error mitigation techniques have been investigated extensively in the literature [4]. Most techniques assume that NLOS corrupted measurements only consist of a small portion of the total measurements and thus can be treated as outliers. The algorithms based on this consideration only work well with a large size of samples and a small number of outliers, e.g., when there are a large number of BSs available and the majority is being LOS with the MS. However, in a practical cellular system, the number of available BSs is always limited and multiple NLOS BSs are likely to occur. As a consequence, these algorithms do not provide any realistic improvement in location estimation accuracy and, in fact, can perform worse than traditional ones [24]. Several approaches have been proposed to enhance the location estimation accuracy for hybrid positioning techniques when the majority of BSs is NLOS. In [6], the location estimate obtained by the HTAP is used as initial position guess (x0 , y0 ) for the MS in an unequal weighted least square (UW-LS) algorithm, where the weights of the objective function and the number of TOA measurements engaged in the location estimation are dynamically adjusted according to the distance between (x0 , y0 ) and the available BSs. The UW-LS meets the FCC requirement of 100 m location accuracy in 67% of the cases when seven BSs are available. In a more realistic scenario with only three BSs detectable from the MS, the HTAP standalone can suffice and outperform the UW-LS. This is because while many NLOS BSs (e.g., seven) increase the chance of mutual NLOS error cancelling, few NLOS BSs (e.g., three) increase the tendency to greatly bias the final location estimate [4]. Many other algorithms can be used to mitigate NLOS errors, such as the hybrid TOA / AOA algorithm (HTA) and the hybrid lines of position algorithm (HLOP) proposed in [24]. The first is based on a constrained NLLS minimization procedure, which explicitly attempts to reduce the effect of NLOS by using bounds on the range and angle errors inferred from the geometry. The second is based on a linear LS minimization procedure, which intrinsically mitigates the effect of NLOS errors.

218

Simone Frattasi and Marco Monti

However, both algorithms fall in the same trap as for [6] when all the available BSs are NLOS.

11.3 The Ad-Coop Positioning System 11.3.1 System Architecture Figure 11.1 shows the hybrid cellular ad-hoc system architecture under consideration for the ACPS. Upon receiving a location information request from a certain MS (CM1 ), the home BS (BS1 ) forms a cluster in the surroundings of that MS, where the latter can be elected as CH and its neighbours as CMs (CM2 and CM3 )2 . When the cluster has been formed, three types of measurements are carried out for location purpose: (1A) TOA measurements: each CM can measure the time of arrival of the pilot signals from the available BSs by cross-correlating each of them with an internal generated pilot signal. This type of measurement can only be used if each CM has a clock accurately synchronized with the available BSs; (1B) TDOA measurements: each CM can measure the time difference between the arrival of a pilot signal from the home BS and a neighbouring BS by cross-correlating them [15]. The pilot signal from each of the neighboring BSs can only be used if its SINR at the receiving CM is above a certain threshold; (2) AOA measurements: with an adaptive antenna array, each of the available BSs steers its antenna spot beam to track the dedicated backward-link pilot signal from each CM and thus obtains its arriving azimuth angle (with respect to a specified reference direction). If we consider a cellular system based on CDMA, in order to avoid the signal degradation due to the near-far effect, this type of measurement can only be performed by the home BS [3]; and (3) RSS measurements: each CM can measure the received power of the signals coming from neighbouring CMs. This may be done during normal data communications without presenting any additional bandwidth or energy requirements. Finally, the location determination of each CM can be either CH-based or CH-assisted / network-based (the choice mainly depends on the computational power at the CH). In the first case, while each available BS retrieves on the backward-link the AOA measurements for each BS-CM link, the CH relays back to the home BS all the TOA / TDOA and RSS measurements respectively obtained on each BS-CM link and CM-CM link, and the calculations are performed by a specific server in the network. In the second case, instead, while the home BS transmits all the AOA measurements to the CH via the forward-link, the CH retrieves all the TOA / TDOA and RSS measurements respectively obtained on each BS-CM link and CM-CM link, and the calculations are directly performed at the CH. In this chapter, we consider a hybrid UMTS / WLAN 802.11a system, in which we make the assumption that none of the CM is equipped with the GPS. Nevertheless, note that the ACPS can be easily extended to: (1) A GPS-aided system, which 2

The formation of the cluster and the election of the CH are issues do not directly tackled in this chapter. For completeness we can mention that a clustering method may take into consideration parameters like fixed-to-mobile (BS-MS) and mobileto-mobile (MS-MS) channel conditions, and spatial distribution, speed, remaining battery power, homogeneity / heterogeneity of the terminals.

11 Cooperative Mobile Positioning in 4G Wireless Networks

219

exploits the GPS-equipped CMs to enhance the location estimation accuracy over the cluster; and (2) An A-GPS system, which exploits the non GPS-equipped CMs to enhance the location estimation accuracy in areas where the GPS alone might not be sufficiently accurate, such as in dense urban environments, which actually represent the greatest interest of cellular network providers and service providers in general [18]. Note that in both cases the established cooperation brings advantages to all the CMs within the cluster (GPS-equipped and non). Finally, it is worth highlighting that the same framework can be applied also indoors, where the cellular and ad-hoc segments may be replaced by WLAN communications in infrastructure and ad-hoc mode, respectively. As a consequence of all the previous considerations, our proposal might represent a possible global solution to the mobile localization problem.

Figure 11.1. ACPS: system architecture.

11.3.2 Data Fusion Method The location coordinates of BSj and the estimated location coordinates of M Si / CMi in a Cartesian coordinate system are denoted respectively as: iT h (11.1) x[j] = x[j] y [j] and h ˆ (i) = x x ˆ(i)

yˆ(i)

iT

(11.2)

For simplicity of presentation, only the x and y coordinates are considered in the derivations (this corresponds to a case where BSj and M Si / CMi are located on a relatively flat plane). Although the z coordinate is ignored, the proposed data fusion method can be easily extended to the 3D case. Moreover, since we need to correlate the information obtained from all the BSs and the MSs / CMs, it is useful to define: " #T ˆ= x x ˆ(1)

yˆ(1) . . . x ˆ(n)

yˆ(n)

(11.3)

Figure 11.2 shows the proposed data fusion method, which is detailed described as follows: (1) For each MS / CM, a first location estimate is obtained from a conventional location algorithm found in the literature. As a reference, we have considered

220

Simone Frattasi and Marco Monti

Figure 11.2. ACPS: data fusion method.

the HTAP in [5] and the HLOP in [24], for available TOA and AOA measurements. Considering that we usually have at disposal more than a single TOA or AOA measurement, the quantities used in this calculation are the median values calculated over each set of measurements. Note that the median value is preferred to the mean value, due to its robustness against eventual biased measurements; (2A) If cooperation is off, in order to determine the final position estimate of each MS, each initial position guess obtained in Step (1) and each set of TOA and AOA estimates relative to each MS are used independently in an unconstrained WNLLS minimization procedure (see Section 11.3.2); and (2B) If cooperation is on, in order to determine the final position estimates of all CMs, all initial position guesses obtained in Step (1) and all sets of TOA and AOA estimates relative to all CMs are used simultaneously in an unconstrained WNLLS minimization procedure (see Section 11.3.2).

11 Cooperative Mobile Positioning in 4G Wireless Networks

221

Before presenting the proposed WNLLS solution, we introduce the following definitions of distance, which will be extensively used in the next subsection: • (i)[j] dˆk = (i)(j) = dˆk

rn oT n o ˆ (i) − x[j] ˆ (i) − x[j] x x rn o n o ˆ (i) − x ˆ (j) x

T

ˆ (i) − x ˆ (j) x

(11.4) (11.5)

(i)[j]

where dˆk is the estimated distance between BSj and M Si / CMi at iteration k of the minimization routine when considering available TOA measurements, (i)(j) and dˆk is the estimated distance between CMj and CMi at iteration k of the minimization routine when considering RSS measurements; • dˆ(i)[j] = (t(i)[j] − t0 )c dˆ(i)(j) = d0 10

(11.6)

P0 −P (i)(j) 10n

(11.7) (i)[j] ˆ is the estimated distance between BSj and M Si / CMi derived where d from t(i)[j] – the TOA of BSj ’s signal at M Si / CMi , c = 3 × 108 m/s is the speed of light, t0 is the time instant at which BSj begins to transmit, d(i)(j) is the estimated distance between CMj and CMi derived from the received power P (i)(j) at one of the peers by maximum likelihood estimation set that the transmitted power is Pt = 17 dBm, P0 is the received power in dBm at a short reference distance d0 , and n is the path-loss exponent (see Section IV.B.2.).

Weighted Non Linear Least Square The objective functions to be minimized in case of cooperation off and on are respectively3 : ) ) ( ( P 2 2 (i)[j] (i)[1] N (i) (i) (i) ˆ ˆ ˆ = j=1 γt I x Jt x Jα x + γα , i = 1, · · · , n (11.8) and ( ( ) ) 2 2 P PN Pn (i)[j] (i)[1] n (i) (i) ˆ ˆ ˆ = i=1 j=1 γt Jt x Jα x + + I x i=1 γα )2 ( Pn P (i)(j) ˆ (i) + n Jp x i=1 j=1 γp

(11.9)

where n is the total number of MSs / CMs, N is the total number of BSs, and the functions J (•) are defined as follows: ˆ (i) = dˆ(i)[j] − dˆ(i)[j] (11.10) Jt x k

3

The unconstrained minimization procedure employed to search for the minimum of the objective functions is the trust region method defined in [1].

222

Simone Frattasi and Marco Monti yˆ(i) − y [1] ˆ (i) = α(i)[1] − arctan k(i) Jα x x ˆk − x[1] ˆ (i) = dˆ(i)(j) − dˆ(i)(j) Jp x k

(11.11) (11.12) (i)

(i)

xk , yˆk ) where (x[1] , y [1] ) are the location coordinates of the home BS, being BS1 , (ˆ are the estimated location coordinates of M Si / CMi at iteration k of the minimiza(i)[j] tion routine, and α(i)[1] is the measured AOA for M Si / CMi 4 . The weights γt , (i)[1] (i)(j) γα and γp in Eq. (11.8) and Eq. (11.9) are appropriately selected in order to reflect the reliability of TOA, AOA, and RSS measurements: =

1 σt2(i)[j]

(11.13)

γα(i)[1] =

1 2 σα (i)[1]

(11.14)

γp(i)(j) =

1 σP2 (i)(j)

(11.15)

(i)[j]

γt

2 2 where σt2(i)[j] , σα (i)[1] and σP (i)(j) are directly derived from the available measurements. According to the choices made for the weights, it can be expected that in case of cooperation on a high location estimation accuracy will be achieved both in LOS and in NLOS conditions. This is because, when all the CMs are either in LOS or in NLOS towards the available BSs, the WNLLS extracts the set of location estimates, which maximizes the probability that the polygon created by the conjunction of the initial position guesses has in total the closest length combination of edges with respect to the distances defined by the available TOA and RSS measurements. Since CM-CM links are likely to have better channel quality than BS-CM links, they will be assigned a higher weight in the objective function and thus they will tend to geometrically constrain the final solution, diminishing its location error. Furthermore, it has to be noticed that due to the spatial diversity towards the available BSs experienced by proximate CMs within the cluster, there is the possibility that some BS-CM links are in LOS. As a consequence, the WNLLS will treat the corresponding CMs as anchors for the others, and the remarkable result is that the location estimation accuracy can be reduced and equalized over the cluster to the one of the CMs in LOS.

11.4 Simulation Models In this section, we list the models used to carry out the simulation results of Section 11.5. Specifically, the statistical models for TOA and AOA estimation errors are taken from [5], where the differentiation between LOS and NLOS is based on the results concerning the LOS probability of [25], from which is also extracted the channel model used for the estimation of RSS. Indeed, in [25] a statistical channel 4

Considering that we usually have at disposal more than a single TOA, AOA or RSS measurement, dˆ(i)[j] , α(i)[1] and dˆ(i)(j) are the median values calculated over each set of measurements.

11 Cooperative Mobile Positioning in 4G Wireless Networks

223

model is derived from a ray tracing simulation of the city of Bristol for either BSMS and MS-MS links in the frequencies of 2.1 GHz (UMTS) and 5.2 GHz (WLAN 802.11a), respectively.

11.4.1 Statistical Models for TOA and AOA Estimation Errors LOS The estimation errors are small and primarily due to equipment measurement errors: they have traditionally been assumed to be normally distributed with zero mean and small standard deviation, respectively cσT = 30 m and σΘ = 1 deg for TOAs and AOAs [5]5 . In particular, we assume that the standard deviation for TOA measurement errors associated with different BSs is identical.

NLOS (1) TOA: The estimation errors are large and primarily due to reflections or diffractions of the signal between a BS and a CM. As a consequence, the following exponential probability density function can be used to model the excess delay [11]: 1 τ P (τ ) = (11.16) exp − τrms τrms where τrms is the RMS delay spread, which is defined by: oε n τrms = η dˆ(i)[j] χ

(11.17)

where η is the median value of the RMS delay spread in µs at 1 km, dˆ(i)[j] is the distance between BSj and CMi in km, ε is an exponent with value between 0.5 and 1, and χ represents the log-normal shadow fading, so that X = 10 log χ is a Gaussian random variable with zero mean µχ = 0 and a standard deviation σχ that lies between 2-6 dB (see Table 11.2). Specifically, the autocorrelation of the shadow fading among the CMs is modeled by [23]: |∆x| ln2 cor

−d

ρ(∆x) = e

(11.18)

where ρ is the correlation coefficient, ∆x is the distance between the CH and a certain CM, and dcor = 20 m is the de-correlation distance. In practice, if χ1 is the log-normal component at position P1 , the one at position P2 , χ2 , which is ∆x away from P1 , is normally distributed with mean µ2 = ρ(∆x)χ1 dB and standard r h i2 deviation σ2 = σχ 1 − ρ(∆x) dB. (2) AOA: The estimation error on the angle measurement is considered to be Gaussian distributed with zero mean µΘ = 0 and standard deviation given by [5]: cτ σθ = (i)[j] (11.19) d where τ is the excess delay determined by Eq. (11.16). 5

In practice, the standard deviation of the measurement errors depends on the chip rate, the propagation environment and the home BS antenna parameters. It can be estimated based on the SNR and its typical values for various propagation conditions [3].

224

Simone Frattasi and Marco Monti Table 11.2. Parameters setting for different environmental types [11]. Channel type η [µs] (A) Urban Macrocells 0.4-1.0 (B) Urban Microcells 0.4 (C) Suburban Areas 0.3 (D) Rural Areas 0.1 (E) Mountainous Areas ≥ 0.5

ε 0.5 0.5 0.5 0.5 1.0

σχ [dB] 1.9-3.6 2.3 2.0-4.7 4.0-5.3 2.4-3.2

11.4.2 Statistical Channel Model for RSS Estimation LOS Probability Figure 11.3 shows the LOS probability as a function of the separation distance between the transmitter (Tx) and the receiver (Rx), for both BS-CM and CM-CM links. It can be observed that: (1) When the separation distance between Tx and Rx lies in the region of 10-20 m, which broadly corresponds to the value of the street’s width, the LOS probability is 1 (for this particular database and set of locations [25]); and (2) When the separation distance between Tx and Rx increases further, the LOS probability is higher for BS-CM links with respect to CM-CM links. This is because reducing the height of Tx from 15 m for BS-CM links to 1.5 m for CM-CM links, the signal propagation paths are affected not only by the surrounding buildings in the vicinity of Tx but also by the terrain. As a consequence, CM-CM links could also be NLOS even when no buildings lie in the propagation path.

Figure 11.3. LOS probability vs. separation distance between Tx and Rx.

11 Cooperative Mobile Positioning in 4G Wireless Networks

225

Path Loss The following slope / intercept statistical model is used for the mean outdoor path loss: L = L0 + 20log10 (f ) + 10nlog10 (d) (11.20) where d is the separation distance between Tx and Rx in m, f is the operating frequency in MHz, and L0 is the path loss in dB at a short reference distance d0 = 1 m. For CM-CM links, L is modeled in (1) LOS with L0 = −27.6 dB and n = 2 (free space path loss); and (2) NLOS with L0 = −51.22 dB and n = 5.82.

Shadowing The shadowing process Ls is characterized by a log-normal distribution, i.e., a normal distribution in dB, with a distance-dependent standard deviation given by: i h −(d−d0 ) σs = S 1 − e DS (11.21) where S is the maximum standard deviation in dB, and DS is the growth distance factor in m. For CM-CM links, σs is modeled in (1) LOS with S = 2 dB, DS = 36 m, and d0 = 0 m; and (2) NLOS with S = 23.4 dB, DS = 36 m, and d0 = 10 m.

Fast Fading The most popular models for fast fading in LOS and NLOS are respectively Rice and Rayleigh, where the ratio between the expected power of the dominant path ρ2 and the power of the Rayleigh components 2σ 2 is often expressed by the Ricean K-factor : K0 = ρ2 /2σ 2 . For CM-CM links, K0 is modeled in (1) LOS by: K0 = −nk d + Lk

(11.22)

with Lk = 23 dB and nk = 0.029; and (2) NLOS by: K0 = logN (µk , σk ) − 10

(11.23)

where logN (µk , σk ) is a log-normal distribution with mean µk = 2.43 dB and standard deviation σk = 0.45 dB.

11.5 Simulation Results Computer simulations have been performed in order to compare the location estimation accuracy of the ACPS with respect to conventional hybrid positioning techniques in stand-alone cellular networks. Simulated urban microcells of 1000 m of radius have been considered, where a hexagonal test cell is surrounded by 6 neighbouring cells (see Figure 11.4). Specifically, 1 ≤ N ≤ 7 BSs have been positioned in a 2D plane with location coordinates BS1 (0, 0) m, BS2 (1732, 1000) m, BS3 (1732, −1000) m, BS4 (0, 2000) m, BS5 (0, −2000) m, BS6 (−1732, 1000) m and BS7 (−1732, −1000) m. For simplicity, we have taken into account only one cluster

226

Simone Frattasi and Marco Monti

of 25 m of radius (all the CMs are in LOS with respect to each other), which embraces 1 ≤ n ≤ 7 CMs, where the CH is placed in the center and the other CMs are uniformly generated around it. Note that the speed of the CMs has been assumed to be very low, therefore during the number of iterations considered for the location estimation there is no relative movement between the cluster and the BSs, and within the cluster itself. The values of some of the parameters used in the simulations are summarized in Table 11.3. The performances of the ACPS are presented in the form of cumulative distribution functions of the cooperative root mean square error (CRMSE): CRM SE = E[RMSE]

(11.24)

T where RMSE = RM SE1 · · · RM SEn are the RMSEs for all the CMs, which are defined as: rn oT n o ˆ (i) − x(i) ˆ (i) − x(i) RM SEi = x x (11.25) r r (i)

where xr is the vector representing the real location coordinates of CMi 6 . From Eq. (11.24), we can observe that with the ACPS the accuracy experienced by the single MS has been replaced with the accuracy experienced by all the CMs.

Table 11.3. Simulation parameters. Parameters Values Cell radius 1000 m Cluster radius 25 m Number of BSs N =1−7 Number of CMs n=1−8 Measurements per BS I = 100 Measurements per CM i = 100 Channel type (see Table 11.2) (B)

11.5.1 Performance Dependency on the Number of CMs HTAP Figure 11.5(a) shows the comparison in terms of location estimation accuracy of the proposed data fusion method with and without cooperation for a variable number of CMs, when the CH is placed at (500,0) m and the initial guesses are calculated by the HTAP in [5]. It is observed that: (1) Regardless the number of CMs involved, the case with cooperation on always outperforms the case with cooperation off. In particular, for a set up with n = 8, the location estimation accuracy is increased in average of about 48% and for 95% of the cases of about 60% (see Table 11.4); (2) The CRMSE 6

In case of cooperation off, the CRMSE falls to be equal to the RMSE calculated by considering only one MS.

11 Cooperative Mobile Positioning in 4G Wireless Networks

227

Figure 11.4. The seven-cell and one-cluster system layout.

considerably drops when increasing the number of CMs. This is because the location estimates obtained in Step (1) of the data fusion method shall represent a polygonal configuration in the space, which respects the geometrical constrains imposed by the knowledge of the relative distances between the CMs. As a consequence, the more cooperative users join the cluster the more binds the selected solution has to respect and the more accurate the final location estimation is. For example, going from n = 2 to n = 8, the location estimation accuracy is increased in average of about 40% (see Table 11.4); and (4) With the specific simulation models taken into consideration, both cases always meet the FCC requirements.

HLOP Figure 11.5(b) shows the comparison in terms of location estimation accuracy of the proposed data fusion method with and without cooperation for a variable number of CMs, when the CH is placed at (500,0) m and the initial guesses are calculated by the HLOP in [24]. It is observed that: (1) Regardless the number of CMs involved, the case with cooperation on always outperforms the case with cooperation off. In particular, for a set up with n = 8, the location estimation accuracy is increased in

228

Simone Frattasi and Marco Monti

average of about 61% and for 95% of the cases of about 54% (see Table 11.4); (2) The CRMSE considerably drops when increasing the number of CMs. For example, going from n = 2 to n = 8, the location estimation accuracy is increased in average of about 49% (see Table 11.4); and (4) With the specific simulation models taken into consideration, both cases always meet the FCC requirements.

11.5.2 Performance Dependency on the Number of BSs Figure 11.6 shows the comparison in terms of location estimation accuracy of the proposed data fusion method with and without cooperation for a variable number of BSs, when the CH is placed at (500,0) m and the initial guesses are calculated by the HLOP. It is observed that: (1) The CRMSE considerably drops when increasing the number of BSs. For example, going from (N = 3, n = 8) to (N = 5, n = 8) the location estimation accuracy is increased in average of about 38% (see Table 11.4); and (3) The ACPS achieves an accuracy very much comparable to the GPS for the combination (N = 5, n = 8) [2].

Table 11.4. (C)RMSE statistics for a variable number of CMs and BSs.

HLOP HLOP

n=6 —,— —,— 46.51,35.65 49.30,116.1 —,— —,— 32.64,25.37 39.22,80.48 —,— —,— 18.77,12.83 21.30,43.87

n=8 Metr. —,— —,— 41.79,28.50 44.60,98.51 —,— —,— 29.16,23.37 33.74,77.56 —,— —,— 17.99,12.18 19.85,41.51

n=4 —,— —,— 50.53,37.22 52.35,125.0 —,— —,— 36.03,29.18 42.79,93.89 —,— —,— 20.98,14.30 23.70,48.92

(m,m)

n=2 —,— —,— 69.78,71.45 68.79,207.1 —,— —,— 57.67,46.04 66.49,144.2 —,— —,— 30.59,21.44 33.13,71.53

µ , σ 67% , 95%

n=1 80.66,91.85 OFF 76.07,244.4 1 —,— ON —,— 74.18,52.54 OFF 84.94,167.8 3 —,— ON —,— 38.01,24.36 OFF 42.91,83.48 5 —,— ON —,—

HTAP

Algo. Coop. N

11.6 Localization, Cooperation and Cognition 4G aims at providing higher data rates and a wide variety of applications, while serving as many users as possible. Unfortunately, limited available resources, such as power and frequency spectrum, are still the biggest constraints. Hence, in order to not slow down the economic and technological improvement of the wireless world, it is necessary to consider new solutions that exploit more efficiently the existing scarce resources. A recent way out of this problem is represented by the concept of cognitive radio, whose main scope is to achieve an efficient spectrum utilization by employing smart wireless devices [13]- [21]. As per Mitola’s cognition cycle, the properties of a CR are divided into two groups: (1) User-centric properties: information sharing and supervising, ability to understand and follow actions and choices of the users,

11 Cooperative Mobile Positioning in 4G Wireless Networks

(a) HTAP

(b) HLOP Figure 11.5. (C)RMSE vs. number of CMs.

229

230

Simone Frattasi and Marco Monti

(a) HLOP

(b) HLOP Figure 11.6. (C)RMSE vs. number of BSs.

11 Cooperative Mobile Positioning in 4G Wireless Networks

231

and learning over time to become more responsive and to anticipate the user needs; and (2) Technology-centric properties: spectrum monitoring, location-awareness, information and knowledge processing. Hence, it is clear that a CR needs to possess a certain amount of situation awareness, which is commonly defined as “the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning and the projection of their status in the near future” [17]. This implies to be able to access location information not only to predict what kind of service the user may want, but also to optimally manage the spectrum and limit the eventual interference. Indeed, geolocation is an essential information at the base of the learning process of a cognitive radio device and represents one method to access the “spectrum holes” [13]. In addition, a CR can benefit from an established cooperation among mobile nodes. There are in fact studies that have considered cooperation for reducing detection time and thus increasing the overall agility of a CR. It has been shown that a cooperation scheme under the amplify and forward protocol has lead to a reduction of the detection time for CRs up to 35% [10]- [16]. Within this framework, we can foreseen that the extension of the ACPS to support CRs would be of fundamental importance. On the one hand, each cognitive CM will be able to reliably sense the spectral environment over a wide bandwidth, detect the presence of active primary users and use the spectrum under a non-interfere policy. Note that a cognitive CM may appear to the legacy network as a primary user when performing BS-CM measurements and as a secondary user when performing CMCM measurements. On the other hand, the ACPS itself can be advantaged from the support of CRs for two main reasons: (1) A CR would be able to interface mobile nodes with different air interfaces, thanks to the SDR technology. This will obviously give a CR the possibility to increase the number of relative measures employable, which is then translated in the potential of increasing the location estimation accuracy of the ACPS; and (2) Higher the possibilities to interface a location-aware terminal, i.e., a terminal that can provide a trustable position information either because it is equipped with a GPS receiver or because it has already retrieved its location information in other ways. As a consequence, the data fusion method will treat the corresponding CMs as anchors for the others, and the remarkable result is that the location estimation accuracy can be reduced and equalized over the cluster to the one of those CMs.

11.7 Conclusions In this chapter, we have proposed an innovative solution for positioning determination in 4G wireless networks by introducing the Ad-Coop Positioning System (ACPS). The numerical results shown in the chapter have demonstrated that, regardless the number of cooperative users or base stations available, the ACPS enhances the location estimation accuracy with respect to conventional hybrid positioning techniques in stand-alone cellular networks. Hence, this work has demonstrated that the emerging paradigm of cooperation has a beneficial impact on wireless location’s performance. There are several areas in which this work could be expanded in the future: enhancement of the data fusion method, investigation of its tracking performance, etc. It is also our hope to evaluate this system in field tests.

232

Simone Frattasi and Marco Monti

References 1. Coleman TF and Li Y. An interior, trust region approach for nonlinear minimization subject to bounds. SIAM Journal on Optimization, 6:418–445, 1996. 2. Assistant Secretary of Defence for Command, Control, Communications, and Intelligence. Global positioning system – standard positioning service – performance standard. Technical report, Department of Defense, October 2001. 3. Cong L and Zhuang W. Hybrid tdoa/aoa mobile user location for wideband cdma cellular systems. IEEE Transactions on Wireless Communications, 1(3):439–447, July 2002. 4. Cong L and Zhuang W. Non-line-of-sight error mitigation in mobile location. Proceedings of the 23rd Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), March 2004. 5. Deng P and Fan PZ. An aoa assisted toa positioning system. Proceedings of the World Computer Congress/International Conference on Communication Technology (WCC/ICCT), 2:1501–1504, August 2000. 6. Deng P and Fan PZ. An efficient position-based dynamic location algorithm. Proceedings of the Autonomous Decentralized Systems Workshop, pages 36–39, September 2000. 7. FCC. Fcc acts to promote competition and public safety in enhanced wireless 911 services. Washington, DC: WT Rep. 99-27, September 1999. 8. FCC. Fcc wireless 911 requirements. Federal Communications Commission (FCC) Fact Sheet, 2001. 9. Fitzek FHP and Katz MD. Cooperation in Wireless Networks: Principles and Applications. Springer, 2006. 10. Ganesan G and Li YG. Agility improvement through cooperative diversity in cognitive radio. Proceedings of IEEE Global Telecommunications Conference (GLOBECOM ’05), 5:2505–2509, November 2005. 11. Greenstein LJ, Erceg V, Yeh YS, and Clark MV. A new path-gain/delay-spread propagation model for digital cellular channels. IEEE Transactions on Vehicular Technology, 46(2):477–485, May 1997. 12. Hsieh HY and Sivakumar R. On using peer-to-peer communication in cellular wireless data networks. IEEE Transactions on Mobile Computing, 3(1):57–72, January-February 2004. 13. Mitola J III. Cognitive radio for flexible mobile multimedia communications. Proceedings of IEEE International Workshop on Mobile Multimedia Communications (MoMuC’99), pages 3–10, November 1999. 14. Berg Insight. Gps and galileo in mobile handsets. Research Report, Berg Insight, November 2006. 15. Knapp CH and Carter GC. The generalized correlation method for estimation of time delay. IEEE Transactions on Acoustics, Speech and Signal Processing, 24(4):320–327, August 1976. 16. Laneman JN and Tse DNC. Cooperative diversity in wireless networks: Efficient protocols and outage behaviour. IEEE Transactions on Information Theory, 50:3062–3080, December 2004. 17. Endsley M. Design and evaluation for situation awareness enhancement. Proceedings of the 32nd Annual Meeting of the Human Factors Society, Riding the Wave of Innovation, pages 97–101, October 1988.

11 Cooperative Mobile Positioning in 4G Wireless Networks

233

18. McGuire M, Plataniotis KN, and Venetsanopoulos AN. Data fusion of power and time measurements for mobile terminal location. IEEE Transactions on Mobile Computing, 4(2):142–153, March-April 2005. 19. Patwari N, Ash JN, Kyperountas S, Hero AO, Moses RL, and Correal NS. Locating the nodes: Cooperative localization in wireless sensor networks. IEEE Signal Processing Magazine, 22(4):54–69, July 2005. 20. EU Institutions Press Release. Commission pushes for rapid deployment of location enhanced 112 emergency services. DN: IP/03/1122, July 2003. 21. Haykin S. Cognitive radio: Brain-empowered wireless communications. IEEE Journal on Selected Areas in Communications, 23(2):201–220, February 2005. 22. Sayed AH, Tarighat A, and Khajehnouri N. Network-based wireless location: Challenges faced in developing techniques for accurate wireless location information. IEEE Signal Processing Magazine, 22(4):24–40, July 2005. 23. Senarath G and et Al. Multi-hop relay system evaluation methodology (channel model and performance metric. IEEE 802.16’s Relay Task Group, October 2006. 24. Venkatraman S and Caffery J. Hybrid toa/aoa techniques for mobile location in non-line-of-sight environments. Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC), 1:274–278, March 2004. 25. Wang Z, Tameh EK, and Nix AR. Statistical peer-to-peer channel models for outdoor urban environments at 2ghz and 5ghz. Proceedings of the 60th IEEE Vehicular Technology Conference (VTC2004-Fall), September 2004. 26. Zhao Y. Mobile phone location determination and its impact on intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems, 1(1):55–64, March 2000.

12 Peer-to-Peer Information Retrieval Based on Fields of Interest Bertalan Forstner, Gergely Cs´ ucs, Imre Kel´enyi, and Hassan Charaf BUTE [bertalan.forstner|gergely.csucs|[email protected]| [email protected]]@aut.bme.hu

Summary. The task of efficient distributed information retrieval has been one of the most serious challenges in the history of information technology. With the spread of advanced mobile devices, the demand for an efficient file sharing protocol moved also into the mobile world. However, the mobile networks have special characteristics that should be taken into account when designing efficient resource sharing protocols for this area. The performance of the Peer-to-Peer protocols can be increased with the use of semantic information gathered from the shared files. However, as the optimal solution for an unstructured network requires a full knowledge of the files available in the network, certain kinds of heuristic methods should be designed to increase the probability of successful queries, without large protocol overhead and network traffic. In this chapter, we will present how the network topology can be quickly improved to increase hit rate with an appropriate protocol and algorithm using Bayesian process based on local decisions that infers from the fields of interest owned by the nodes.

12.1 Inspiration from Everyday Life Today the term Peer-to-Peer (P2P) network is not unknown for the computer users, as there are applications with different purposes that work in a fully distributed manner. We can exploit the advantages of the distributed architecture in different ways. File sharing, resource or computing power sharing, distributed workflow management, chat and voice calls through the distributed network without the presence of any server - just a few fields of application of the P2P technology. The strength of this concept may be in the similarity with the way in which we live: being in relationship with different people, communicating with them directly, performing common tasks in immediate manner, without an intermediate “center”. In the computer world, this center is called the server, and in some application areas, it has more disadvantages than advantages on the process it is intended to support. Servers are often performance bottlenecks, single point of failures, targets of attacks and abuses in the systems. That is why the P2P approach has been rediscovered again and again in the world of computer networks, as it enables direct communication between the participants.

235 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 235–249. c 2007 Springer.

236

Bertalan Forstner et al.

However, the thinking about the distributed manner often stops at the advantages, without deepening more in the concept to exploit the possibilities to cover the difficulties raised by the decentralized architecture. The most important of them is more than straightforward: with which other network elements should a computer be in direct relation, and how can it find these elements? The basic and easy answer to this question is any computer whom it can discover. However, this way of thinking in most cases is simply impermissibly wasteful with resources. This chapter is intended to be an eye-opener in this question to liberate the mind to think in the computer networks’ world as rational as in everyday life, and shows how easily all the discovered methods can be transplanted to the language of mathematic and computer technology. In the next few paragraphs an overview of the unstructured P2P networks will follow to give a basic understanding that is necessary to seize the concepts of our approach in the mobile Peer-to-Peer world. The first popular fully distributed P2P protocol was Gnutella, a file sharing network [8]. In a Gnutella network the participating computers (also called as nodes) initiate random connections (or links) to other nodes in the network. When connected, a node can send queries to its neighbors containing a search term. The neighbors will check their shared files whether any of them matches the search criteria, and reply with query hit if they think they have the requested file. The query initiator can then decide whether to download the file from its owner. Besides of analyzing the query, the neighbors forward it to their connected nodes in order to increase the probability of finding the requested file at any of the participating nodes. Of course, this message forwarding can generate huge network traffic because of the exponential growth of the number of queries, therefore every message contains a Time-to-Live (TTL) value that is decremented every time a node forwards the query. If the TTL reaches zero, the message is dropped. Usually this TTL value is initialized to around 7. The nodes reached by the query via this forwarding mechanism form a query propagation graph. The functioning of the protocol can be imagined as follows. The nodes are people, for example in a crowded conference hall. A man can talk only to the few people standing next to him as shouting can not be regarded as polite behavior in this company. He might want to know the answer to a question, for example, “Who was the first Gnutella client developed by?”. It is not sure that the neighbors are knowledgeable enough to answer this question, however, they can ask this question from the people standing next to them, and so on, until someone is found with the answer (“Justin Frankel and Tom Pepper” in our case). The basic, unstructured Gnutella file sharing protocol was found unscalable and inefficient, but later different extensions or brand new protocols turned up for different special purposes and with improved efficiency. The efforts dealing with this issue can be divided into two significantly different groups. The first group consists of the structured P2P networks (for example, [13], [14], [16], [19]). These protocols specify strict rules for the location of documents to be stored, or define which other peers a node can connect to. The idea is that finding a piece of information is much more easier if we exactly know the location of such kind of documents or we know the connection through which the query should be propagated in order to reach the appropriate node. Although these networks usually have good scalability properties, and their performance can be estimated quite accurately, they are becoming disadvantageous in networks with strong transient character: they can handle the

12 Advantages of Local Cooperation in Peer-to-Peer Networks

237

frequent changes in the network population with difficulties and at great resource expenses. The second approach examines unstructured networks such as the basic Gnutella protocol [8]. In that case there is no rule for the location of the documents to store, and the connections of the nodes are controlled by few simple rules. For that reason these systems have limited protocol overhead and can tolerate when nodes frequently enter to and leave from the network. With the spread of broadband wireless communication, the network applications move from desktop computers to intelligent mobile devices. There is an increasing need for mobile software that can locate and retrieve documents (texts, music or video files) that are in the fields of interest of the mobile user. Since the computing resources and the increased usability of the smartphones make these devices with high proliferation a suitable platform for representing different kinds of information, it becomes highly important to involve them into the distributed information retrieval world. We learnt this from the fact that our Gnutella-based P2P client for the Symbian mobile platform, the original Symella client software [17], was downloaded more than 10.000 times just on the first week of its appearance. There are a few circumstances that make mobile P2P information retrieval a bit different. The mobile communication costs are higher than that of wired communication, therefore it is more important to use effective P2P protocols in their case. With the spread of flat data communication prices and WLAN-enabled devices, the costs in money is negligible besides the costs in energy. The wireless traffic requires quite significant power resource, therefore the limited battery capacity does not enable unrestricted connection time. The connectivity of these devices is also limited because of network coverage. Therefore the used P2P protocol should also be suitable for an important specialty of the mobile environment, namely it should tolerate the strong transient character of these P2P clients. Because of this latter property the use of structured P2P networks is less efficient. The Peer-to-Peer approach enables to make information retrieval more efficient, using a model well-known from everyday life. In the real world, working relationship is established among the people with a labor of the same topic. For example, if one’s job is connected to the 19th century French literature, her associates will have the same field of interest and probably have experience, books (documents), that is, relevant information, on the topic. Back to our previous illustrative example, if some related information needs to be found in the conference hall, then probably nobody would start with asking random people, but the colleagues that are experts on that field. The Internet and Peer-to-Peer makes it possible to contact those people with whom we cannot enter into relations because of geographical or other barriers. In the basic peer-to-peer protocols the search for the documents starts with querying the randomly selected neighbors. However, there are some methods elaborated that change the connections of the nodes to other links that might ensure a higher hit rate. The first group of these algorithms tries to achieve better results based on run-time statistics ([15], [11], [3], [18]). The simplest way is to monitor the hit rates through each connection and abandon them when they do not deliver the expected number of results. Another way is to initiate direct connections to the effective nodes in few hops distance, however, that can enormously increase the load on popular nodes. The second group of the content-aware Peer-to-Peer algorithms uses some kind of metadata provided for the documents in the system. One of the early Gnutella

238

Bertalan Forstner et al.

extensions is the Query Routing Protocol (QRP). This utilizes the Ultrapeer scheme: the so-called “leaf” nodes with short connection time or narrow bandwidth connect to the long-living “ultrapeers” that are rich in resources. QRP suggests that leaves send the keywords of their available files in a hashed form to the connecting ultrapeer. This will decrease the query traffic as the ultrapeers can check in their query routing tables whether their leaf neighbors have files that match a keyword. The ultrapeers are connected to each other and to the standard Gnutella clients that do not implement this extension. This idea is improved in different ways in the semantic P2P information retrieval protocols or extensions. Some algorithms are used to retrieve keywords from the stored documents, and then query routing is made by this semantic information. However, most of these proposals do not utilize enough of the potential in the semantic information ([9], [10], etc). For example, they cannot generalize the collected semantic information. Semantic relationships in the concept hierarchy can be exploited from the available data. For example, if someone is searching for information about mammals, this query should be forwarded to a node with the keyword dog as it is hyponymic to mammal. We call a tree that describes this kind of “is a” relation of the words a taxonomy. There is a part of a taxonomy containing the relation of mammal and dog in Figure 12.1. There are several dictionaries available for the natural languages that contain semantic links between the words or concepts. One of the most popular dictionary in semantic information retrieval systems is WordNet [4], which is an online lexical reference system whose design is inspired by current psycholinguistic theories of human lexical memory. It contains synsets, that is, sets of concepts, along with their relations. Later on we will also use this taxonomy to categorize the documents into different topics.

Figure 12.1. A part of a taxonomy from WordNet.

It seems straightforward to transform the P2P network in such way that the nodes connect to those who have relevant documents for the user. However, there is an important aspect that cannot be neglected when developing semantic protocols. It is easy to see and is proven by mathematical models [7] that, because of the fully distributed property of the network, it requires quite a large amount of message

12 Advantages of Local Cooperation in Peer-to-Peer Networks

239

traffic to obtain and maintain local information about the documents stored at each node or their metadata, and it can be very inefficient in a transient unstructured network. It can also be complicated to obtain information about the fields of interest of the user. In the next section, we will show our solution for the mentioned issues.

12.2 Modeling Fields of Interests The basic idea of our approach follows directly from the metaphor in the previous section. Firstly, every P2P client should be capable of describing the fields of interests of its user, which will be approximated by the stored documents and the queries sent out. Such a field of interest is represented by a topic (a synset) in a taxonomy. Then they should approximate the number of distinct documents in each topic, which will be computed from the data obtained by the neighbors. This can help a node to find out the probability that a query for a document in some topic will be found through a connection to a given node. And from the other side, a node should be able to inform the connecting nodes about the approximate number of known documents and the expectable hit rate in a topic through its propagation tree. Obviously the gathered information reflects the knowledge of the neighboring nodes about the whole network, and therefore, it cannot be regarded fully accurate. All the nodes should sum up these data and time to time refine it with their observations. Based on that information, the nodes can dynamically replace their neighbors with new ones thorough which higher hit rates can be expected. In our approach four kinds of profiles are maintained by the nodes in order to find the most appropriate connections. Firstly, we introduce the semantic profile which characterizes the stored documents of a node. This profile can be a good prior until the client learns the fields of interest of the user from his dynamic behavior, because supposedly, the stored documents in some degree reflect these fields of interest. Secondly, the connection profile will be introduced. Connection profiles are maintained by each node for all of their connections. The reason is to be able to compare the expectable answer ratio of the different propagation paths reachable through different node connections. Then the reply profile is described, which is the profile that is sent by a node in its answers to help other nodes deciding whether they want to initiate a connection with it. The fourth taxonomy, the query profile, characterizes the queries sent out by the user.

12.2.1 The Semantic Profile As the effectiveness of the information retrieval system is planned to be increased with the help of semantic information, we needed a tool with which the nodes can categorize the stored documents. In absence of dynamic information this can be used as initial data to describe the fields of interests of the user. Moreover, as the nodes should send the maximum number of known documents in each topic, these values should also be stored. Therefore we introduced the semantic profile, which is a weighted taxonomy, where the nodes represent the different synsets describing the stored documents, along with the count of occurrence of the given synset. The stored documents contain enough information for the peers to compose their own semantic profile. There are several methods to acquire even quite complex

240

Bertalan Forstner et al.

ontology from documents ([1,10], or the IBM UIMA Framework), therefore the user does not need to manually give the keywords for each document, this method can be fully automated. When a new document is stored in a node, it updates its semantic profile by increasing a counter for each keyword’s synset and all of its hypernyms. As an example with the taxonomy part introduced in Figure 12.1, if a document has the word ‘cat’ as a related metadata (keyword), the system will store this information, because probably cats are in the fields of interest of the given user. Then the algorithm will search for all the hypernyms of ‘cat’ in the taxonomy, and also mark them as concepts in the fields of interest of the given user. In our specific example, these synsets will be ‘feline, felid’, ‘carnivore’, ‘placental, placental mammal’, ‘mammal’, ‘vertebrate, craniate’, ‘chordate’, ‘animal’, ‘living thing’, ‘object’ and finally ‘entity’; moving from the most specific concept to the generalized ones, where ‘entity’ is the most general noun in the taxonomy, that means, it is the root synset. When we add a new document with a keyword ‘dog’, the hypernym structure will be very similar, only on the 10th level will branch the two concept. Assuming that the stored documents reflect the fields of interest of the user, this means that the semantic profile of the user describes these fields of interest on a specific generalization level as accurate as the extracted concepts cover the content of the documents.

12.2.2 The Connection Profile As described in the introduction, the nodes with the advanced protocol select their neighbors in a way such that their in-topic queries will be answered with high probability. Therefore, they need the connection profiles of the neighboring nodes. It is also a weighted taxonomy where along with the concepts the probability of the expectable hit rate of a query for a document that can be described with the given concept s through the given connection C is stored. This probability will be denoted as PsC . However, as we will show it on two examples, using only the semantic profiles of the neighboring nodes in itself is not sufficient by all means to approximate the answer ratio, because the value of a connection also depends on the profile of the nodes accessible through the whole query propagation path. The situation in Figure 12.2 shows a part of a network of nodes (with small letters) and with their semantic profile. These profiles are restricted to two fields of interests (marked with capital letters) for the sake of simplicity. The number of stored documents in the topic is written next to the letters. Now consider node n with two fields of interests A and B. n should connect either to o or p because of the similarities of the profiles. Nevertheless, node o is a more valuable connection because of its neighbors: Although o and p have the same fields of interests or even the same documents, there are some more nodes with similar profile connected to node o. Another situation is shown in Figure 12.3, where node n has interest in two other topics (C and D). Nodes q and r propagate the same topics in their profile with the shown weights. Despite of the ratio of the topics in the profile of these nodes, n can expect higher hit rate through q in topic C, and more positive answers from r in topic D, because of the documents stored at their neighbors.

12 Advantages of Local Cooperation in Peer-to-Peer Networks

241

Figure 12.2. Part of a network.

Figure 12.3. Another part of a network. The whole propagation path determines the value of a connection. It should also be noticed that the previous statements are valid only if the stored documents are different for each node, which cannot be guaranteed or find out without larger amount of overhead messages. The conclusion of all these examples is that semantic protocols should consider the whole propagation path when calculating the value of a connection. This finding is not reflected in the semantic data-based unstructured protocols, therefore we elaborated an own solution. We can conclude the requirements for the connection strategy as follows: •

•

The goal is to characterize the probability that the node can reach a query hit in a given topic through a given connection. As we have learnt it from the mathematical models of semantic networks (for example [7]), the precise answer requires knowledge on the exact number of documents in the given topic. However, because of the transient property of the network, we do not even know the number of all documents in the network in a point in time, moreover, all efforts to collect such kind of data needs a rather big amount of messages. Also, because of the transient property, it cannot be assumed that gathering semantic data from all the nodes in the query propagation path can be cost

242

•

•

Bertalan Forstner et al. effective. A semantic description of the documents by all of the nodes in the query propagation path is quite a large amount of data. Keeping this information upto-date when nodes are connecting and leaving the network and are constantly changing their connections, seems quite an inefficient way. We can suppose that the nodes working with the semantic protocol use the same semantic document-analyzing algorithm, therefore, they extract the same keywords from a given document. And as we intend to use the advanced protocol with mobile devices, we suppose that computing resources at the clients are quite low. Therefore the protocol should not perform complex calculations.

We elaborated a solution that can be applied in these circumstances. It follows a Bayesian process to estimate the probability of finding a document in a specific topic through a given connection, using the semantic profile propagated by the newly connected node as prior information. With the Bayesian process we want to approximate the probability PsC that a direct connection C to the candidate node c can deliver positive answer for a query for a document in the topic s. (The candidate node will be denoted with small c and the connection to that node with capital C.) We can regard the different queries of a node independent, therefore the probability of getting exactly α successes and β negative answers out of α + β queries through a connection in a context s is given by the probability mass function of the binomial distribution. Using beta distribution is a good choice for representing the prior belief in the Bayesian estimation, because it is conjugate prior for binomial likelihood [2]. Returning to our problem, let PsC be a probability with beta distribution that represents the belief of a node that connection C gives positive answers in context s. Its parameters are α and β where α is the number of observations when connection C did give results for a query for a document in the topic s, and β is the number of observations when it did not. Initially, the prior is Beta(1, 1), the uniform distribution on [0, 1], which represents absence of information. However, as nodes might not spend much time in the network, we should use a more precise prior approximation, the reply profile, which is sent by the candidate node c as a weighted initial observation. The Reply profile for a node c is a taxonomy where, along with each synset s, the maximum number of known documents in that topic and the estimation of the answering ratio from node c are represented. To dispense with the exact mathematical background, we suggest that the connecting node regard the reply profile sent by the candidate node as an n-fold observation. The actual value of n can be set by the connecting node individually according to its past experience in the truthfulness of the propagated data or the expectable online time of the node. After each observation (that is, a query) the connecting node can update the hypothesis (the connection profile) in a given synset for each connection. We assume that the node with prior distribution Beta(α0 , β0 ) sends out queries for documents which are then found and downloaded. Suppose that the synset s exploited from the documents by the commonly agreed metadata-gathering algorithm characterizes the files. If α0 pieces of documents among all the downloaded files were found via connection C and β 0 times this connection gave no result, the profile for connection C should be updated according to the following equation: (PsC )0 = Beta(α0 + α0 , β0 + β 0 )

(12.1)

12 Advantages of Local Cooperation in Peer-to-Peer Networks

243

With the growth of the number of queries, the probability distribution function becomes close to a Dirac at PsC because of the properties of the Beta distribution. Although the probability distribution function for beta distribution is quite complex, the expected value of a beta random variable X can be calculated as easily as α . (12.2) α+β As there is no need to use many computational resources for calculating the expected value, this theory is very suitable to be used in mobile environment. Deeper information on the beta distribution and its properties can be found for example in [2]. E(X) =

12.2.3 The Reply Profile The aim of a reply profile is to represent the knowledge of a node about the number of documents in a topic along with their availability in the set of reachable documents as a probable answer ratio. The reply profile serves as prior knowledge for the connecting nodes that is later refined with the observations of these nodes. Therefore, it is more important for the profile’s probabilities to be quickly computable than to be absolutely precise. The reply profile contains the maximum number of known documents in each synset. This is calculated as the maximum number of documents known either by the node itself (from the semantic profile) or by its connected nodes. It can be seen that the maximum number of reached documents in the topic might be greater, because the stored documents are not necessarily overlapped, however, it is a good heuristics. The maximum known number of documents for the synsets is important because based on that data nodes can approximate their own answer ratio. The reply profile also stores the answering probability for each synset s. This is approximated similarly as the number of documents from the data in the connection profile and the number of owned documents in the given topic.

12.2.4 The Query Profile Our protocol extension transforms the connections of a node in a way such that the query propagation path contains nodes with similar interests as the given node. Therefore the fields of interests of the nodes should be represented in some way. As earlier mentioned, the semantic profile of a node can be regarded as a prior information for this reason, supposing that the stored documents more or less reflect the fields of interests of a user. However, the real interests can be exploited from the documents that are sent as query hits from other nodes. Therefore, each node will set up a structure in which these interests can be stored. This structure is the query profile, which is a weighted taxonomy where, along with the synsets, a probability is stored which value represents that a query issued by the given node is pointed for a document that can be described by the synset s. This probability will be denoted as Qs . Similarly to the connection profile, a probability value in the query profile for a synset s will be a hypothesis, which means that the node beliefs that a question sent out by the node will result in a reply of a document which can be described

244

Bertalan Forstner et al.

with the synset s. For the same considerations we will regard this belief as a beta distribution probability which is updated with the observations of the query hits. After each query hit the node exploits the keywords from the received documents, and updates the hypothesis of the given synset in its query profile. With the growth of the number of queries the probability distribution function becomes close to a Dirac at Qs . Again, the semantic profile can be regarded as an n-fold observation to be a prior for the query profile. The actual value of n can be set by the connecting node individually according to its past experience in the usefulness of these data, in general we set this value to the average number of queries per session. If there are no documents available in the store, the prior hypothesis should be Beta(1, 1) , the uniform distribution on [0, 1]. In Figure 12.4, we conclude the relationship and data flow between the different profiles. It shows how the default data for the Semantic profile and Query profile are generated from the document store. The query profile and the connection profiles for each connection are then maintained by the metadata acquired from the downloaded documents. The reply profile is calculated from the available data as described earlier.

12.3 Protocol Extension As we want the nodes to replace their connections in the way described in the previous section, we have to extend the unstructured protocol. As described in the beginning of this chapter, we will demonstrate our results based on the Gnutella protocol. We will enhance this basic protocol with a standard Gnutella Generic Extension Protocol (GGEP), therefore, it will be built on the basic Gnutella messages. As it is a semantic Peer-to-Peer protocol, we called it SemPeer . SemPeer extends the basic messages as follows. 1. When connecting, the participant nodes send their reply profiles to each other. 2. When a query reaches a hit at a non-neighbor node, the reached node attaches its reply profile to the answer (QueryHit) message. The operation of the SemPeer-enabled nodes differs in the following properties from the clients that support only the basic Gnutella protocol. 1. Each node constructs its semantic profile based on the stored documents 2. Each node maintains its query profile based on the resulting files of their queries. 3. Anytime a node receives a query hit through one or more of its connections, it modifies its connection profiles. If it does not receive any answer, it does not modify the profiles as there may be no such documents at all. 4. When a node receives a reply profile form another node as part of a QueryHit message, it decides whether to put that node in its connections list. This decision is based on an algorithm that compares the query profile of the given node and the reply profile of the new node. If the calculated similarity promises better results than any of the existing connected nodes, a node should consider initiating a new connection. The nodes might also consider other circumstances when deciding

12 Advantages of Local Cooperation in Peer-to-Peer Networks

245

Figure 12.4. Data flow between the different profiles.

to connect, such as bandwidth, uptime or reputation of the new node, or whether the connection fits the topology or the underlying physical network, etc.) 5. In case the new node accepts that connection, and the number of semantic links is less than the predefined maximum for the given topic, the new link can be established without any further considerations. If, however, the maximum number of semantic connections is already reached, then the new link must replace one of the existing connections, which has the lowest connection value (that is, the less similar one). In both situations the received semantic data should be stored for further comparisons. The algorithm that compares the profiles retrieves values which approximate the probability that a query issued by the given node can be answered by the nodes reached through the given connection. There are several algorithms to compare the similarities of taxonomies or concepts (for example [12]). We used such a method that compares the taxonomies only up to a given level in the concept hierarchy. This means that the nodes perform more and more refined comparisons from the generic

246

Bertalan Forstner et al.

level to the most specific ones when selecting their neighbors. With this solution, if there are enough nodes have been a selected field of interest in a given taxonomy level, they can be found very easily with the protocol. And until these nodes are found, the protocol provides other nodes with similar fields of interest in a more generic level. This approach has a further advantage. Namely, it is more probable to find nodes with a given and specific field of interest through the connections of the nodes that are similar in a more general level than through random connections. This quick network transformation is very advantageous in a transient P2P network.

12.4 Protocol Performance In order to evaluate the performance of the new protocol, we have adapted a simulating environment [6] for the extension. Having measurements with different typical parameters, we discovered in which cases can the application of the protocol result in improved performance. An extreme example is the case when the clients have only one, very concentrated field of interest, and they can store only a very limited set of files. This is the case with the smartphones with built-in or small capacity memory card, that barely can store a full music album. There were 16.000 nodes in the simulated network with 4-7 connections each. The nodes can store around 5 files to their musical taste. We differentiated 10 major music genres. We got surprisingly good results when a node sent out queries for music in its field of interest: after receiving replies for an average 5-15 queries, the protocol could construct such a layer over the standard random network that reached a sevenfold hit rate to the basic protocol (the query hit probability increased from 0.11 to as high as 0.69). Of course, if the topics of the queries or the stored documents are not so homogeneous, then the increase in the average hit rate is smaller, however, the advance in the performance can still be observed. We previously mentioned that we have constructed a mathematical model to calculate the maximum hit rate that can be theoretically reached with the semantic protocol extension. We used this model when designing the SemPeer protocol to approach the optimal value. In our article [5] we compared the hit rate predicted by the model with the simulated results in different contexts (nodes with multiple fields of interests, queries and stored documents that do not belong to the fields of interests of the owning node, etc.). We have found that when the files are shared in expectable number, then the designed protocol almost reaches the optimal hit rate predicted by the model. The strength of the protocol extension requires that the fields of interest of the users can be determined. In order to prove this we conducted a simple experiment with a Peer-to-Peer client software running on mobile phones. The software used a simple taxonomy that categorized every song downloaded via the P2P network based on the genre information in their ID3 tag. With the data provided by approximately 4000 active users, we found that almost everyone of them have two or three strongly marked genres (that is, fields of interest), and more than 70 percent of their downloads belong to one of these significant categories. Similar results could be observed when analyzing downloaded text documents for their keywords.

12 Advantages of Local Cooperation in Peer-to-Peer Networks

247

12.5 The Application of Our Results 12.5.1 Designing Symella The low computing performance and limited storage capacity of the mobil devices prevented to run fully functional P2P clients on them up to the last one or two years. With the spread of smartphones, this situation has changed. Our primary goal was a Gnutella client for handheld devices that adapts to the characteristics of the mobile environment, still capable to connect to clients running on any other platform. Moreover, we designed it modular to be easily extensible in order to implement and test the results of the latest researches and protocol extensions such as SemPeer. We selected the already mentioned Symbian OS as the mobile platform to develop for, as it is the most mature and most popular mobile operating system available of its kind. Our application was named Symella, after the connection of the words “Symbian” and “Gnutella”. There were numerous Gnutella clients available for desktop environment when designing Symella, however, these versions considered totally different aspects than those required by mobile applications, therefore, we should find new solutions for the most of the problems.

12.5.2 The Architecture of Symella In this subsection, we present the architecture of the application that fulfills the necessary requirements and is suitable for the Symbian OS (Figure 12.5). The engine of Symella is responsible for creating and managing the connections, to process the queries, and supervise the downloads. An intelligent host cache stores the properties of the Gnutella clients, and it constantly collects and classifies the addresses getting in from different sources in order to enable the Connection manager to use the best connections. The base of the classification is the time necessary to work up a connection. Moreover, reliable connections with high uptime are also stored. The Message Parser is responsible for processing and forwarding the messages.

Figure 12.5. The architecture of Symella.

248

Bertalan Forstner et al.

The best usage of the available bandwidth during download was a very emphasized aspect during application design, that is why we decided to support multithreaded download. That means that before starting to download, the Download fragmenter logically divide the file into multiple smaller parts, then these fragments are downloaded in parallel from different nodes. Another task of the downloading subsystem is to collect query hits and collect those that are pointing to the same file, that is, when a file is found at multiple nodes. It also communicates with the Local profile manager, which, in turn, uses the services of the Topic analyzer in order to manage the different profiles. Symbian constitutes not only the basis of the operating system, it is completed with a User Interface (UI) layer, which can be very different by device types. Programming the UI of a Nokia Communicator with wide display and full QWERTY keypad (Series 80 platform) is quite different from developing for the Sony Ericsson P900 series with touchscreen (UIQ platform). As we did not intend to come down to only one user interface, we prudently separated the platform-dependent and platform-independent parts. At the time of writing Symella supports the Series 80 and the S60 platforms. The latter one is very popular with the standard phones with normal-sized displays and standard ITU-9 keypads. The popularity of Symella is owed to the fact that this is the only, modern P2P client in the Symbian world. More than ten thousand active Symella users worldwide prove the demand for distributed mobile applications. From our aspect, however, is more valuable the clear and modular architecture that helps enhancing the P2P protocols to make distributed networks as prosperous as the real-life human relationships can be.

12.6 Conclusion In this chapter, we have shown the basic concept of the Peer-to-Peer networks, along with different approaches to increase the performance of its applications with the use of semantic information. Then we borrowed a solution from the everyday life, where people are in relationship with each other based on their fields of interest. We have shown a model, the four different profiles, which can help us to find similar nodes in the P2P network. And of course we have extended an existing protocol, the Gnutella, in order to be able to construct and use these profiles. Moreover, we have presented the architecture of a mobile Peer-to-Peer client, the Symella, which enables to conduct experiments with the new results.

References 1. H. Assadi. Construction of a regional ontology from text and its use within a documentary system. In Proc. of International Conference on Formal Ontology and Information Systems, (FOIS-98), Amsterdam, 1998. IOS Press. 2. James O. Berger. Statistical Decision Theory and Bayesian Analysis. Springer, second edition edition, 1985. 3. Hanhua Chen, Hai Jin, and Xiaomin Ning. Semantic peer-to-peer overlay for efficient content locating. In Proc. of MEGA’06, Harbin, China, Januar 2006.

12 Advantages of Local Cooperation in Peer-to-Peer Networks

249

4. Christiane Fellbaum. WordNet, An Electronic Lexical Database. MIT Press, 1998. ISBN 978-0-262-06197-1. 5. B. Forstner. An analytic model for peer-to-peer systems with semantic overlay network. In Proc. of AACS’06 Workshop, Budapest, Hungary, 2006. 6. B. Forstner, G. Cs´ ucs, K. Marossy, and H. Charaf. Evaluating performance of peer-to-peer protocols with an advanced simulator. In Proc. of Parallel And Distributed Computing And Networks PDCN2005, Innsbruck, Austria, February 2005. 7. Bertalan Forstner and Hassan Charaf. Modeling peer-to-peer networks with interest-based clusters. Transactions on Enformatika, Systems Sciences and Engineering, 8(2):38–43, Oktober 2005. 8. The gnutella protocol homepage. http://www.the-gdf.org. 9. Sam Joseph. P2p metadata search layers. In Proc. of Second International Workshop on Agents and Peer-to-Peer Computing (AP2PC 2003), 2003. 10. J.-U. Kietz, A. Maedche, and R. Volz. Semi-automatic ontology acquisition from a corporate intranet. In Proc. of Learning Language in Logic Workshop (LLL-2000), New Brunswick, N.J., 2000. IOS Press. 11. S. Shashidhar Merugu and E. Zegura. Adding structure to unstructured peer-topeer networks: the use of small-world graphs. Journal of Parallel and Distributed Computing, 65(2):142–153, Februar 2005. 12. Resnik P. Semantic similarity in taxonomy: An information-based measure and its application problems of ambiguity in natural language. Journal of Artificial Intelligence Research, 11:95–130, 1999. 13. Sylvia Ratnasamy, Paul Francis, Mark Handley, RichardKarp, and Scott Shenker. A scalable content-addressable network. In Proc. of SIGCOMM’2001, August 2001. 14. A. Rowstron and P. Druschel. Storage management and caching in past, a large-scale, persistent peer-to-peer storage utility. In Proc. of SOSP’01, 2001. 15. K. Sripanidkulchai, B. Maggs, and H. Zhang. Efficient content location using interest-based locality in peer-topeer systems. In Proc. of Infocom 2003, 2003. 16. Ion Stoica, Robert Morris, David Karger, Frans Kaashoek, and Hari Balakrishnan. Chord: A scalable peer-topeer lookup service for internet applications. In Proc. of SIGCOMM’2001, August 2001. 17. The symella application homepage. http://symella.aut.bme.hu. 18. B. Yang and H. Garcia-Molina. Efficient search in peer-to-peer networks. In Proceedings of the 22nd International Conference on Distributed Computing Systems (ICDCS), July 2002. 19. Ben Y. Zhao, John Kubiatowicz, and Anthony Joseph. Tapestry: An infrastructure for fault-tolerant wide-area location and routing. Technical Report UCB/CSD-01-1141, University of California at Berkeley, Computer Science Department, 2001.

Part III

Cognitive Networks

13 Introducing Cognitive Systems to the B3G Wireless World P. Demestichas, G. Dimitrakopoulos, K. Tsagkaris, and V. Stavroulaki, and A. Katidiotis University of Piraeus, Department of Digital Systems

Summary. The wireless world is rapidly evolving towards the Beyond the 3rd Generation (B3G) era, characterized by increased complexity, due to the heterogeneity of wireless infrastructure and the varying environment requirements. Such complex situations can be tackled through the exploitation of cognitive networking potentials. To this end, the chapter contains an overview of cognitive networks principles and then focuses on management functionality that is applicable (i) in cognitive network segments, (ii) in cognitive access points (APs) and (iii) in cognitive wireless terminals, presenting versatile solution approaches.

13.1 Introduction 13.1.1 The Wireless World Today Wireless communications attract significant research and development effort, reflected on the progress of work performed in international projects [3], as well as on the discussions in international fora [7]. This work results in a powerful, high-speed infrastructure that offers versatile solutions to the digital information society. In this context, the technological focus is on the cooperation and coexistence of legacy Radio Access Technology (RAT) standards with currently emerging ones. The current wireless landscape is characterized by a plethora of RATs, which can be roughly classified in two major families: •

•

Wireless wide area networking (WWAN) technologies, which include, among others, 2G/2.5G/3G mobile communications [5], the IEEE 802.16 suite [4], WiMAX [6] and broadcasting technologies [2]; Wireless short range networks (WShRNs), which include wireless local and personal area networks (WLANs/WPANs), as well as wireless sensor networks (WSNs) [4] [1] [8].

This situation is depicted on Figure 13.1. Regarding the backbone network architecture, legacy [5] or modern paradigms [16] can be followed. Moreover, the evolution of wireless access networks is frequently referred to as B3G (Beyond the 3rd Generation) systems [3] [7].

253 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 253–269. c 2007 Springer.

254

P. Demestichas et al.

Figure 13.1. Overview of the Wireless World in the B3G Era.

In the B3G era network operators (NOs) will have to address increased complexity, with respect to today. Complexity derives from two main sources: • •

The inevitable heterogeneity of the network and terminal infrastructure; The user requirements that associate the B3G era with advanced services/ applications, provided seamlessly and ubiquitously.

To meet these objectives, NOs have to deploy complex network topologies of heterogeneous nature. The different RATs will have to co-exist, and be complementarily (and efficiently) exploited. Each RAT has different capabilities, in terms of capacity, coverage, mobility support, cost, etc. Therefore, each RAT is best suited for handling certain situations. In this respect, a NO will have to rely on different RATs for raising the customer satisfaction, and achieving the required Quality of Service (QoS) levels, cost-effectively. QoS refers to performance (e.g., bit-rate, delay, etc.), availability (e.g., low blocking probability), reliability (e.g., low dropping or handover blocking probability), as well as security/safety (indicated also in [11]). An option for handling such complex situations is to design wireless B3G infrastructures by exploiting “cognitive networking” capabilities [9] [15]. In general, cognitive systems dispose the capability to retain knowledge from previous interactions with the environment and determine their behaviour according to this knowl-

13 Introducing Cognitive Systems to the B3G Wireless World

255

edge, as well as to other goals and policies, so as to adapt to external stimuli and optimize their performance. In the case of cognitive networks, this definition can be translated into the ability to dynamically select the network’s configuration, through management functionality that takes into account the context of operation (environment requirements and characteristics), goals and policies [12] (corresponding to principles), profiles (capabilities), and machine learning [13] [14] (for representing and managing knowledge and experience). Cognition as a concept may extend to numerous parts of a communication system, i.e., to network segments, access points and even terminals. The required management functionality may thus comprise numerous cooperating mechanisms, in the sense that different versions of it should be applicable, wherever cognition is introduced. This chapter, accordingly, emphasizes on how to manage networks and terminals that operate in accordance with the cognitive networking paradigm. Our main objectives are to provide: • •

The motivation for the development of cognitive networking technologies; The specification of novel, intelligent management functionality applicable in cognitive network segments (semi-distributed approach), cognitive access points (fully distributed approach) and cognitive wireless terminals.

The work of this chapter is based on background work done in [3]. The chapter focuses on extensions for migrating to the cognitive domain.

13.1.2 Motivation: Cognitive Networks and their Management Functionality The wireless world has been lately moving far beyond conventional 2G/3G systems, with the scope to offer seamless mobility to users, considering the ever-increasing user demands. Major facilitator of this trend, as aforementioned, is the concept of cognitive networks [10]. Cognitive networks contradict to legacy ones, since they are able to adapt their operation by (proactively or reactively) responding to external stimuli. This is achieved by disposing mechanisms that observe external conditions, retain valuable knowledge from interactions with the environment and plan their future actions accordingly. Their operation can be reflected on a feedback loop (see also [15]), like the one shown on Figure 13.2. Basic cognition operational principles envisage that the network continuously observes (monitors) the environment, looking

Figure 13.2. Cognitive Networks Operation Loop.

256

P. Demestichas et al.

for potential changes that can affect its operation. Observations form the basis for initiating machine-based reasoning to see if the reconfiguration process should be invoked. Once the decision is taken, the network acts accordingly. This loop is repeated inside a machine learning process, which leads to cognition. The loop is guided by a set of goals, which take the observations into account in planning actions. At this point of time, a reasonable question may arise: “Which is the optimum way to manage the diverse entities that form part of a cognitive network ”? The answer to this question might be complex. Specifically, the radio access networks have been classically designed and deployed to cover the traffic demand of the planned services in a static approach and by means of manual configuration of network elements, considering the busy hour traffic in each geographical zone. However, the continuously increasing demand also raised the need for the deployment of new technologies and networks which have to be optimally planned and managed by choosing among the options of finding new sites, co-locating sites or migrating to reconfigurable transceivers. Additionally, in the case of cognitive networks, novel functionality should efficiently plan and manage an ever-changing network, since it should adapt to external requirements that also change over time and space. Furthermore, since a cognitive network consists of numerous elements and terminals of highly heterogeneous natures, located in different places, a centralized management approach becomes prohibitively complex and inappropriate. Hence, distributed management approaches, relying on pertinent technologies, e.g., autonomic computing, are currently in the focus (e.g., see [9]). This approach can offer scalability, stability and modularity (which provides low complexity). In this respect, this chapter is claimed to provide scalable answers to the question of managing (supporting) cognition, by dividing the management process in 3 cooperative parts (tiers), as explained below and shown also on Figure 13.3. Figure 13.3 depicts the overall management architecture of a B3G infrastructure that operates in accordance with the cognitive networking paradigm. Its entities are organized in a hierarchical manner that consists of three cooperative tiers. The entity that controls a whole cognitive network segment is made up of mechanisms whose primary purpose is to coordinate with the backbone network, as well as the decisions of the management entities that are responsible for individual elements (APs). These entities correspond to the management functionality targeted at APs; let it be noted that a cognitive network segment comprises numerous elements, which might be APs, base stations or other reconfigurable elements. Finally, the figure depicts also that entities that are targeted to the management of cognitive terminals. The next sections provide details on the design of management functionality that corresponds to the above tiers. In particular, Section 13.2 deals with the semidistributed solution for managing a whole network segment (that operates in a cognitive radio context). Section 13.3 refers to a fully distributed approach, where management functionality is targeted at specific cognitive APs. Finally, Section 13.4 deals with cognitive wireless terminals. A summary and further research challenges conclude this chapter.

13 Introducing Cognitive Systems to the B3G Wireless World

257

Figure 13.3. Management Architecture for Cognitive Networks.

13.2 Management Functionality for Cognitive Network Segments This section discusses on an intelligent management functionality system for B3G network segments operating in accordance with the cognitive networking paradigm. This corresponds to the less distributed part of the overall management architecture, as previously described. The approach presented may be short-term oriented (when semi-distributed, as explained in the sequel), since it sounds rather attractive for NOs, due to its potentiality to permit a high level of network control. It still constitutes an approach that needs to be able to guarantee for an acceptable level of scalability, an aspect which is also of great importance.

13.2.1 Problem Description General Discussion The general definition of cognitive networks implies some very advanced capabilities, which spring from the necessity to encompass reconfiguration (change in the behaviour of the segment, reflected on parameters / infrastructure variations) features, enhanced by cognition capabilities. As part of the reconfiguration, at the PHY/MAC layers, there can be elements (hardware components, such as reconfigurable transceivers) that dynamically change the RATs they operate and the spectrum they use, in order to improve their QoS levels offered. On the other hand, such changes should be performed in the best possible way, i.e., the changes should be based on the applicable policies related to the context (combination of environment characteristics and specific event requirements) and on the selection of the most appropriate reconfiguration pattern. In general, this semi-distributed approach covers the motivation, requirements, functionality and engineering challenges for a distributed functionality, which yields a powerful and scalable means that leads to cognitive, wireless access, networks. Furthermore, it includes the development of novel mechanisms for automating the procedure of deciding for the optimum reconfiguration, thus facilitating and optimizing the planning and management mechanisms. Figure 13.4 provides the overall description of the management functionality proposed for managing a cognitive network segment. The proposed management mechanisms undertake decisions that affect the protocol stack in a cross-layer fashion. The next subsections provide information on the problem’s inputs (context, profiles and policies) and output (configuration of

258

P. Demestichas et al.

Figure 13.4. Management Functionality for Cognitive Network Segments.

behaviour), as well as (see Section 13.2.2) on the necessary cognition features that the functionality covers.

Context This part of the input is probably the most crucial, referring to the gist of the problem, due to its interactions with the environment, which constitute the primary reasons that urge a system to adapt to stimuli. Context information is monitored and discovered (sensed) for each element of the network segment and for its environment. Context information reveals the status of the elements, in the network segment, and of their environment (therefore performance, fault, etc. triggers will be covered). Context includes information on locations, time, traffic demand, mobility levels, interference conditions with managed elements, etc.

Profiles This part provides information on the candidate configurations of the elements of the segment, such as the set of transceivers of each element, the set of operating RATs, as well as the set of spectrum carriers. Moreover, this part also describes the profiles (e.g., preferences, requirements, constraints) of user classes, applications and terminals, etc.

Policies Policies designate rules and functionality that should be followed in context handling. Specifically, this part provides information on the NO policies with respect to reconfiguration strategies, i.e., NO preferences and priorities on goals to be achieved. These are related to the maximization of the QoS levels (performance, availability, reliability), and the minimization of cost factors (resource consumption). Furthermore, this part provides information on NO agreements with cooperative NOs. Policies and goals should guarantee for the maximum possible level of stability.

13 Introducing Cognitive Systems to the B3G Wireless World

259

Output The output provides actions that will determine the behavior of the network segments (configurations at all layers of the protocol stack of each element of the segment). At the physical and MAC (Medium Access Control) there will be the selection of RATs (Radio Access Technologies), spectrum, transmission power, as well as other parameters and algorithms depending on the network element and terminal capabilities. At the network level there will be actions related to element interconnection (legacy or emerging network topologies like mesh), routing and congestion control. At the application level, there will be actions related to the allocation of applications to QoS levels.

13.2.2 Cognitive Features The management functionality for cognitive network segments constitutes an essential step towards manners of planning and managing purely cognitive network segments. This is depicted on its input, output and the approached optimization techniques. Specifically, the proposed functionality is applicable at various levels of distribution, ranging from semi-distributed to fully-distributed schemes. The semidistributed schemes are a first step for introducing cognition in the B3G world, while they can also serve as benchmarks for the fully-distributed approaches, which can support and ease cognition, due to their lower level of dependence on lateral factors. Solution algorithms for this part of the overall management functionality for cognitive networks can be based on optimization techniques, in addition to machine learning functionality and artificial intelligence techniques. In fact, machine learning functionality and artificial intelligence algorithms attribute the management functionality with knowledge and experience (at least) in the areas described below and can qualify the functionality in the cognitive domain. Context information is obtained through interactions with the environment, which lead to reasoning and perception, through appropriate machine learning techniques. The network segment is thus able to gain knowledge from those interactions and be aware of the optimum behavior for various contexts. The constant updates in the information provided by the “profiles” part of the functionality leads to significant knowledge gains with respect to user behavior. This is essential in the quest for seamless provisioning of services and unparalleled quality, tailored to individual user needs. This implies that the process of serving users is facilitated and optimized through experience. Cognitive features lie also in the “policies” part of the management functionality. Specifically, the suitability and efficiency of different policies possesses great importance in handling versatile contextual situations. Consequently, learning the most optimum policy and the most appropriate goals to be achieved may become valuable for NOs in successfully (transparently, fast and securely) handling difficult conditions.

260

P. Demestichas et al.

13.3 Management Functionality for Cognitive Access Points This section deals with management functionality that operates in a fully-distributed fashion and it is applicable in the access points of a cognitive infrastructure (see also Figure 13.3). The general description of the problem that such functionality aims at solving remains immutable with respect to the management of complete network segments, i.e., based on context, profiles and policies information, it is targeted at deciding on the most appropriate behaviour of an access point of the infrastructure (of an individual element). The subsections below describe a potential solution to the problem, i.e., the Autonomic Management of Access Points (AMAP) functionality entity, as well as its features that enable and support cognition.

13.3.1 The Autonomic Management of Access Points (AMAP) Introduction and Functional Analysis The starting point for the functional analysis of the management functionality proposed is the introduction of the components of an AMAP functionality entity. Figure 13.5 depicts those components. The AMAPS functional architecture includes parts that provide: (i) profiles and agreements (PA); (ii) monitoring, discovery and context acquisition (MDCA); (iii) resource and service brokerage (RSB); (iv) configuration negotiation, selection and implementation (CNSI), based on policies, goals, profiles and context information. The following subsections describe the AMAP components in detail, whereas Section 13.3.2 is dedicated on the knowledge (cognition supportive) features of the components.

Profiles and Agreements The component should provide information on users, the capabilities of the equipment served by the access point, as well as the element (access point) capabilities. Managed access point. The access point part describes all potential configurations of each transceiver and thus the configuration of the access point. Equipment. This part of the component shows the capabilities of the equipment in the service area, in a manner analogous to that of the managed element. End-users. The information should specify applications, permissible QoS levels, the importance (utility), cost (allowable or tolerable), etc.

Monitoring, Discovery, Context Acquisition The component should provide the means for: (i) acquiring the context in which the managed access point operates; (ii) monitoring the efficiency with which each contextual situation is handled. (iii) discovering alternatives that can be used for context handling. Context acquisition. The context, in which the managed element operates, consists of the: (i) location; (ii) time epoch; (iii) traffic, mobility and interference conditions encountered, caused by the equipment/users in the service area.

13 Introducing Cognitive Systems to the B3G Wireless World

261

Figure 13.5. Functional Analysis of an AMAP entity.

Monitoring. This procedure collects basic data from the transceivers of the managed element. The basic data gathered from transceivers can be processed to produce aggregate data at various levels of abstractions. Discovery. This functionality of this part lies in the estimation of the performance of alternate configurations in a certain context. Therefore, this part of the component should have information on the achievable performance, e.g., interference levels, bitrate, coverage, associated with each alternate transceiver configuration.

Resource and Service Brokerage This component enables an AMAP entity to interact with other entities in its environment, so as to acquire data on the status of the “neighbouring” elements. Eventually (as addressed in the corresponding knowledge layer paragraph), the need for these interactions can be reduced, and each access point can develop knowledge on the capabilities of the near-by elements. However, the existence of such an entity can ensure smooth migration from current management models to the autonomics.

Configuration Negotiation, Selection and Implementation The role of this component is to consider the inputs (described in 2.1) and decide on the optimum behaviour of the managed access point. The AMAP entity takes

262

P. Demestichas et al.

into account policies, which specify rules (constraints) that should be respected, and functionality that should be followed for reaching the goals, which is designated by policies. In the light of the above, different contexts can be tackled, according to a certain strategy, through the implementation of the decisions taken.

13.3.2 Cognitive Features This section discusses on the cognitive features that the management functionality for cognitive access points encompasses. These features are divided in the different AMAP components, as described above. The “profiles and agreements” part of the functionality acts in a cognitive environment by developing knowledge on the user profile, related to the behaviour of a certain user, regarding applications, traffic and mobility, in a given time-epoch/userrole, location, and the user preferences for QoS levels in a given time-epoch/user-role, location, etc. Such learning techniques can increase the confidence on the accuracy of the profile information and in general they can be based on measurements and interactions with users/devices. The “monitoring, discovery and context acquisition” part can add robustness to the decisions procedures and also provide the means for reasoning, not only on what currently goes on, but also on what is likely to happen in the future. Therefore, learning mechanisms can yield the typical capabilities of alternate, candidate configurations, in certain contexts. Finally, context predictions are necessary for the proactive handling of situations. The “resource and service brokerage” part can develop knowledge regarding the capabilities of neighbouring elements, for facilitating seamless mobility and ubiquitous provision. Finally, the “configuration negotiation, selection and implementation” part of the functionality uses information on typical solutions used with respect to the selected behaviour of the access point, so as to increase the confidence on the most likely behaviour of policies, in handing certain contexts. This knowledge will enable faster and more reliable selection of configurations.

13.4 Management Functionality for Cognitive Wireless Terminals This section describes novel management functionality targeted at user equipment, i.e., it discusses on the part of the management system that enhances wireless terminals with cognitive networking capabilities (see Figure 13.3). The problem that this part aims at solving is again the one described in Section 13.2.1. In this respect, the next subsections present a potential solution approach, namely the “Cognitive Reconfigurable Equipment Management System” (C REMS) with its components, as well as its cognitive features.

13 Introducing Cognitive Systems to the B3G Wireless World

263

13.4.1 The Cognitive Reconfigurable Equipment Management System (C REMS) Introduction and Functional Analysis The management functionality that controls and supports cognitive wireless terminals (terminals often being referred to as “equipment”), namely the “Cognitive Reconfigurable Equipment Management System” (C REMS ), should enhance equipment with the capabilities described in the following. Managing user preferences and equipment capabilities, as well as network policies. This refers to the accurate description and representation of this information as well as configuring and updating the respective profiles and policies. Acquiring context information. This can be further analyzed into discovering the available access networks and monitoring the corresponding environment. Discovery mainly deals with periodically performing checks in order to determine whether a new and/or more appropriate alternative has become available. More specifically, it should be possible to identify possible entrance of a new RAT in the service area offering better opportunities (e.g., higher QoS provision, lower cost per QoS level and service). Monitoring refers to the accumulation of statistics from different RATs in order to assess their status and identifying whether a reconfiguration is necessary. Negotiating offers with the various available networks and selecting the most appropriate reconfiguration action for the equipment. The selection procedure may result into a switch from one network / RAT to another. This selection should be consistent with user preferences and equipment capabilities, and should also take into account specific service area region conditions and time zones of the day. Furthermore, the selection should not be restricted to technologies that are preinstalled in the element. On the contrary, the dynamic downloading, installation and validation of software components, required for the reconfiguration process and the support of a potentially additional RAT, should be provided. A Cognitive Reconfigurable Equipment Management System (C REMS), comprising the capabilities outlined in the previous, is structured of the following main components: “profiles and policies”, “monitoring and discovery” and “reconfiguration negotiation and selection” (see also Figure 13.6). The following subsections discuss on the functionality of the components of the proposed C REMS, as well as on the components’ cognitive features (13.4.2).

Profiles and Policies This component acquires, maintains and provides information on equipment capabilities user behavior, preferences, requirements and constraints, as well as network policies. The equipment profile, i.e., information related to the device capabilities, specifies a set of potential equipment configurations. More specifically, this set of potential configurations comprises the Radio Access Technologies (RATs) that the device is capable of operating and the associated spectrum and transmission power levels that can be used. Information related to the user profile specifies the set of applications that can be used by the terminal user and the sets of QoS levels associated with the use of an application. User related information also specifies the utility volume that is

264

P. Demestichas et al.

Figure 13.6. C REMS functionality. associated with use of an application at a certain quality level. The utility volume is a way of expressing the preference of some QoS levels, with respect to other permissible ones. In other words, high utility values for an application and QoS level indicate high interest of the user for the specific QoS level when using the application. Finally, the user profile may include information on the maximum price that the user is willing to pay in order to use certain applications at specific QoS levels. Network policies specify NO constraints that have to be taken into account in the handling of contextual situations. These constraints are additional to those specified by the user and equipment profiles. Regarding the user, policies can constrain the applications and QoS levels that are allowed to be used, with respect to those specified in the profiles. Regarding the device, policies can allow only a subset of the capability sets. Therefore, policies determine the RATs that are allowed for operation, the frequencies and transmission powers that can be used, when a permissible RAT is selected. Network policies may also specify algorithms and parameter values that should be used for context handling. Moreover, policies may optionally specify, for each contextual situation, a corresponding set of “best” configurations, which can be used as rules or mere suggestions. In the latter case, policies specify a mapping of each contextual situation to a set of configurations (rules or suggestions).

Monitoring and Discovery The role of the “monitoring and discovery” module is to acquire context information so as to perceive the current status of the equipment and its environment, mainly in the sense of identifying the available networks in a certain area and monitoring their status. In case any discrepancies are observed it may be rendered necessary to initiate

13 Introducing Cognitive Systems to the B3G Wireless World

265

re-selection of the most suitable configuration (radio access scheme). Monitoring is performed at regular time intervals. The monitoring procedure deals with the collection of information on the current connection and the applications used. Once this information has been collected, the values of the monitored parameters are compared with a set of predefined thresholds. In case violations are noted and if a reconfiguration is considered appropriate, the discovery process is initiated so as to identify potential alternate configurations, i.e., RATs offering better operating conditions. The discovery process can also be triggered at regular time intervals in order for the equipment to maintain an update view of its environment. The Monitoring and Discovery module derives, from the set of detected configurations, the subset of configurations that are in line with the constraints imposed by the user and equipment profiles. For example, if user preferences (in the corresponding profile) indicate a minimum of 256 Kbps for a video call service this will lead to a rejection of all the RATs that offer lower quality level. In case the alternate RAT(s) can be deployed, the Reconfiguration Negotiation and Selection module is triggered to re-evaluate the selected RATs.

Reconfiguration Negotiation and Selection The “reconfiguration negotiation and selection” component decides on the most appropriate configuration(s) for the managed device in terms of the obtained QoS levels, taking into account the current context, the user and equipment profiles, and the network policies. In principle, the above factors are embedded, either through relevant decision variables or through appropriate constraints, into an objective function that has to be optimized. As was introduced in the previous, network policies reflect constraints imposed by the NO in managing contextual situations and may specify, in certain cases, a mapping of each contextual situation to a set of configurations. Therefore, the Reconfiguration Negotiation and Selection component can handle the different contexts in the following manner. Firstly, if rules exist, then a configuration can be selected among the set of indicated ones, which are also feasible. Otherwise, if no rules can be followed, due to infeasibility, then a configuration can be selected, among the set of those deriving from the algorithms and those suggested by the policies. The existence of multiple “best” configurations per contextual situation has several advantages. First, it facilitates the achievement of appropriate QoS levels (higher reliability). Second, it enables the resolution of conflicts, i.e., the selection of a new configuration, in case the performance of the one in use is compromised. Third, it enables the equipment to participate in negotiation schemes.

13.4.2 Cognitive Features This section describes the basic cognitive features that the management functionality for cognitive wireless terminals encompasses, as well as the respective engineering challenges that need to be met, in order for these features to be seamlessly integrated with the basic C REMS characteristics. The cognitive features are divided in the different C REMS components, as previously described. Getting into detail on the “profiles and policies” component, its cognition related features lie in that it develops knowledge on the behavior, preferences, requirements and constraints of the user. The following issues are important:

266 •

•

P. Demestichas et al. The traffic and mobility behavior of the user, when a certain application is used, at a certain location and time. This information also reflects the applications preferred. The preferences for QoS levels of the user, when a certain application is used, at a certain location and time. User preferences are again expressed through the utility volume value that is associated with each application and QoS level.

The traffic behavior per application covers the frequency of application usage and the duration of the usage, whereas the mobility behavior covers aspects like the speed and direction. Traffic and mobility behavior have an inherent stochastic nature. Therefore, knowledge is required for increasing the confidence on the accuracy of the managed equipments’ understanding. Moreover, it is difficult to express the user preferences for QoS levels through utility values. The vast majority of users are not technology experts, so as to be in position to directly express preferences for QoS levels. Moreover, these preferences may evolve/change with time. Finally, for the time being only suggested approaches exist for the mapping of QoS levels to utility values. The cognitive part of the functionality of the component can rely on random variables, measured and reference values, interactions with the “monitoring and discovery” module and the user. Random variables can be associated with each traffic and mobility parameter (of each application), as well as the utility parameter (associated with each application and QoS level). Measurements come from the “monitoring and discovery” module, and provide the values of the traffic and mobility parameters, as well as the QoS of the application provision. A set of discrete, reference, candidate values can be associated with the utility parameter of each application and QoS level. The traffic and mobility measurements can be used for managing probabilistic relations. Each of these relations can be described as: “probability that a specific (traffic or mobility) parameter, will take a certain (measured or reference) value, in a given context (that includes the user, location, time, and the application used)”. A more complex approach is required for managing the probabilistic relations that refer to the user preferences (utility values). The “profiles and policies” module can acquire, from the “monitoring and discovery” part, the QoS levels, at which each application is provided. The user can be asked to express the satisfaction, according to a user-friendly, high-level, rating scale. The feedback can be used for enforcing the probabilities of those reference values that are near the measured value. The probabilistic relation is: “probability that the utility, will take a reference value, in a given end-user context (that covers the user, location, time/role, application used and the QoS level of the application provision)”. Regarding the cognitive functionality of the “monitoring and discovery” module, this requires capabilities for deriving: • •

Supplementary information on the current context. Predictions of future contexts.

The supplementary information on the current context adds robustness (stability, reliability) to the monitoring and discovery procedures. Robust monitoring answers the following essential question: “What are the most likely QoS levels, typically achieved by a given configuration, in a given context (traffic, mobility, interference conditions encountered by the equipment)”. Robust discovery answers the following essential question: “What are the most likely capa-

13 Introducing Cognitive Systems to the B3G Wireless World

267

bilities e.g., bit-rate and coverage, of a given alternate configuration (candidate for future use), in a certain context (e.g., location, time)”. Robust discovery is required as the configuration capabilities may not be constant in a dynamic, reconfigurable environment. Capabilities depend on the configurations used in the environment, by network elements, and the resulting interference conditions. Therefore, mechanisms are required for learning the typical capabilities of alternate, candidate configurations, in certain contexts (e.g., locations, time regions). Robust monitoring has a similar motivation. In general, the estimation of the QoS levels that are offered by a configuration is difficult, and probabilistic relations are required. Finally, context predictions are necessary for the proactive handling of situations, which is an essential feature in the context of cognitive systems. The cognitive part of the “monitoring and discovery” module can be based on sets of random variables, reference values and measurements. Regarding robust discovery, random variables can be associated with the alternate configurations and each configuration capability (e.g., bit-rate, coverage, etc.). It can be assumed that each capability (random variable) can take a value from a set of reference, discrete values. Measurements, reflecting the actual values of each alternate configuration’s capability, are conducted by the discovery process. Based on these data, probabilistic relations can be managed. Each of these relations can be described as: “probability that a configuration capability (random variable) will take a reference value, for a given alternate configuration, in a given context (e.g., location, time)”. In principle, with each measurement, there will be an increase in the probabilities of the reference values that are near the measured value, while those that are not near will be decreased. Regarding robust monitoring, random variables can be associated with the selected configuration, the contextual situation encountered (time-related information, applications used, interference conditions), and each QoS parameter. The monitoring functions provide measurements on all these aspects. Reference contexts can be considered, in order to reduce complexity, and avoid the maintenance of information for a large number of situations. Therefore, the “monitoring and discovery” component should acquire data on the QoS levels achieved by the configurations used, identify the nearest reference contextual situation (pattern matching), and update the appropriate probabilistic information. Each of these relations can be described as: “probability that a QoS parameter (random variable) will take a certain value (measured or reference), given the configuration and contextual condition”. Context predictions can be based on the monitoring of situations, in each time zone. Reference contexts can be used for reducing complexity, i.e., the number of situations considered in the probabilistic relations. In each time zone, the “monitoring and discovery” monitors the situation encountered, and should identify and record the occurrence of the corresponding “nearest” reference situation. In this respect, this module is also in position to record the transitions that occur, between reference situations, in successive time zones. The frequency of the transitions is a basic way for computing probabilities, and predicting future situations. Regarding the cognitive features of the “reconfiguration negotiation and selection” component, functionality is required for estimating the most likely behavior (objective function value it can achieve) of a specific configuration in a given context (traffic, mobility, interference conditions). The objective function achieved by a configuration relies on several factors of the contextual situation that are random, e.g., location, applications and QoS levels, interference conditions. Therefore, probabilistic models should be developed for increasing the confidence on the most likely

268

P. Demestichas et al.

behavior of configurations, in certain contexts. Knowledge is required for selecting configurations and policies faster and in a more reliable manner. The cognitive part of the Reconfiguration Negotiation and Selection component can rely on random variables, measured and reference values, interactions with the Monitoring and Discovery module. Random variables can map to the selected configuration, the context addressed and the objective function value. Measurements are provided by the Monitoring and Discovery module, and provide information fundamental for the computation of the objective function, and the identification of the contextual situations. Reference contextual situations can be used, in order to reduce complexity. The Reconfiguration Negotiation and Selection component computes the objective function values, identifies the closest reference context, and maintains probabilistic relations. Each relation can be described as the “probability that the objective function will take a certain value, given the configuration and contextual condition (traffic, mobility, interference conditions)”.

13.5 Conclusions Wireless communications are migrating towards the B3G era, which can efficiently be realized by means of cognitive networking technologies. Cognitive networks are able to reconfigure their own infrastructure, so as to proactively adapt to external stimuli. Cognitive networks are made up of management functionality and reconfigurable elements. Cognitive networks embrace a wide variety of research areas, because they affect all layers of the protocol stack. In this respect, This chapter has presented numerous fundamental ideas for the consolidation of cognitive systems. Beginning from the assumption that management functionality for cognitive networks extends to certain parts of the whole communication link, i.e., (i) the network segment, (ii) the APs of the segment and (iii) the terminals that operate within an AP, the different parts of a management functionality for cognitive networks have been presented and thoroughly analysed. Emphasis was placed on their cognitive features and how these (co)operate. As can be deduced from the analysis above, cognition is a concept that can facilitate the B3G vision. Cognitive features are able to enhance communication networks with advanced capabilities that will have a significant impact on the end user, in terms of the potentiality to seamlessly provide innovative services and applications, in a cost-efficient manner. An efficient management of cognitive networks and the terminals that operate in such context may sound complex, but it stands as a prerequisite for the penetration of B3G systems in the global market and for the consolidation of the B3G vision. In conclusion, the ideas presented in this chapter may be helpful regarding the further exploitation of cognitive networking principles, in the sense that cognitive networks represent a technological approach that will offer to users, operators, manufacturers, and service providers an extended portfolio of operational choices for accomplishing their business goals. Acknowledgement. This work was partially performed in project E2R II which has received research funding from the Community’s Sixth Framework programme. This

13 Introducing Cognitive Systems to the B3G Wireless World

269

chapter reflects only the authors’ views and the Community is not liable for any use that may be made of the information contained therein. The contributions of colleagues from E2R II consortium are hereby acknowledged.

References 1. Bluetooth. http://www.bluetooth.com, 2007. 2. Digital Video Broadcasting (DVB). http://www.dvb.org, 2007. 3. FP6/IST project E2R (End-to-End Reconfigurability). http://www.e2r2. motlabs.com, 2007. 4. Institute of Electrical and Electronics Engineers (IEEE) 802 standards. http: //www.ieee802.org, 2007. 5. Third (3rd ) Generation Partnership Project (3GPP)). http://www.3gpp.org, 2007. 6. WiMAX Forum. http://www.wimaxforum.org, 2007. 7. Wireless World Research Forum (WWRF)). http://www.wireless-worldresearch.org, 2007. 8. ZigBee Alliance. http://www.zigbee.org, 2007. 9. P. Demestichas, D. Boscovic, V. Stavroulaki, A. Lee, and J. Strassner. m@angel: autonomic management platform for seamless wireless cognitive connectivity to the mobile internet. IEEE Commun. Mag., 44:118–127, June 2006. 10. P. Demestichas, G. Dimitrakopoulos, J. Strassner, and D. Bourse. Introducing reconfigurability and cognitive networks concepts in the wireless world. Vehicular Technology Magazine, IEEE, 1(2):32–39, 2006. 11. W. Hasselbring and R. Reussner. Toward trustworthy software systems. Computer, 39(4):91–92, 2006. 12. J. Strassner, editor. Policy-based network management: solutions for the next generation. Morgan Kaufmann, August 2003. ISBN 1-5586-0859-1. 13. T. Mitchel, editor. Machine learning. ISBN 0-0704-2807-7. McGraw-Hill Science/Engineering/Math, March 1997. 14. R. E. Neapolitan, editor. Learning Bayesian Networks. ISBN 0-1301-2534-2. Prentice Hall, April 2003. 15. R. Thomas, L. DaSilva, and A. MacKenzie. Cognitive networks. In 1st IEEE Symposium on Dynamic Spectrum Access Networks 2005 (DySPAN 2005), pages 352–360, Baltimore, USA, November 2005. 16. Xudong Wang, Sunghyun Choi, and J.-P. Hubaux. Wireless mesh networking: theories, protocols, and systems. Wireless Communications, IEEE [see also IEEE Personal Communications], 13(2):8–9, 2006.

14 Architectures and Protocols for Next Generation Cognitive Networking B. S. Manoj1 , Ramesh R. Rao1 , and Michele Zorzi2 1

2

Department of Electrical and Computer Engineering, University of California San Diego, CA 92093 [bsmanoj|rrao]@ucsd.edu Department of Information Engineering, University of Padova, Italy, [email protected]

Summary. In the late 1990s the introduction of cognitive radios paved the way for physical layer capabilities in an entirely new dimension, namely cognition. Since then, significant amount of research has gone into software defined radios and cognitive radio systems. Recent innovations in physical layers for network interface cards and base station front-ends for cellular systems utilize the benefits of cognitive radio systems. While cognitive radio systems are well researched, less-studied cognitive networking appears to have significant potential to be part of next generation wireless systems. In this chapter, we present some benefits of cognitive wireless networks, different architectural paradigms for such systems, and a new architecture specifically designed for cognitive wireless networks. This chapter also presents the details of the CogNet AP, a cognitive network access point, as an early implementation of autonomous cognitive networking systems.

14.1 Introduction A cognitive radio [8,9] is different from traditional Software Defined Radios (SDRs). SDRs provide in software the radio frequency (RF) processing functions, for example, waveform synthesis, traditionally implemented in hardware, thereby making reconfiguration of the radio properties very easy. On the other hand, cognitive radios are intelligent radio devices which can learn their own capabilities, radio environment, user behavior, and the physical environment in order to execute complex adaptations and configure themselves to best suit the situation. Such devices can even alter their physical layer interfaces from one access technology to another by over-the-air downloading a new software-defined waveform. While software radios are fairly well understood, cognitive radios are actively under research. On the other hand, cognitive networking is in its early infancy. In most cases, except for MAC issues, there is relatively little focus on networking in general and the overall network system in particular.

271 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 271–284. c 2007 Springer.

272

B.S. Manoj et al.

14.2 Definition of Cognitive Networking Clarke et al. provided in [2] the first definition of cognitive capability for the Internet where a knowledge plane is assumed to have significant cognitive capability and to be able to build and maintain high level models of what the network is expected to do such that it can receive and execute high level instructions from the network administrators and report the result of such actions. The cognition capability expected from the cognitive plane is also expected to enable self correction and reconfiguration of the Internet if unexpected behaviors or failures happen. Such a system has the ability to translate the high level instructions to executable low level actions. The strategies suggested would involve the use of a unified approach which includes several system components with a global view of the events in the network. While the suggestions in [2] were useful to initiate potential benefits of using a cognitive paradigm for the Internet, no systems have been built or studied yet based on those concepts. An important work in this direction was done by Thomas et al. in [10]. They provided the following definition for cognitive network systems: “A cognitive network has a cognitive process that can perceive current network conditions, and then plan, decide and act on those conditions. The network can learn from these adaptations and use them to make future decisions, all while taking into account end-to-end goals.” This early definition has paved the way for more sophisticated definitions in [6] where the fully distributed nature of today’s computer networks is taken into account. While the above mentioned two definitions provided early aspects essential for cognitive networking, a more comprehensive definition is provided in [6] according to which a cognitive networking system has a distributed set of cognitive processes which collect spatio-temporally tagged network environmental information, including the network parameter behavior from every layer of the network, from every network element within a node, and from other network nodes in order to identify the right network parameters to be used for achieving the individual and end-toend network goals. The definition in [6] focused on an distributed approach which underlines the spatio-temporal tagging of information and collection, storage, and analysis of information for a larger networking perspective.

14.3 Architectures for Cognitive Networking We provide a classification of cognitive networking approaches in this section which is motivated by the need to identify the existing solutions within the general research areas in cognitive networking. Figure 14.1 shows such a classification. The first and most simplistic approach for cognitive networking is termed autonomous cognitive networking according to which the cognitive wireless networking devices observe and learn about their environment based on which appropriate actions are taken. While these devices have significant cognition capability, they do not communicate among themselves or with any central entity. There are several situations where autonomous behavior is essential, for example, where there are no centralized network resources. The second type of approach is more useful where there exist centralized resources such as central repositories. These two approaches will be described in subsections 14.3.1 and 14.3.2.

14 Architectures and Protocols for Next Generation Cognitive Networking

273

Figure 14.1. Classification of cognitive network architectures.

14.3.1 Autonomous Cognitive Networking In the autonomous cognitive networking architecture, nodes adapt by observing and learning from the environment. Similar to the traditional cognitive radio networks, a particular node observes its networking and radio environment in order to obtain cognitive information such as traffic periodicity, traffic pattern, and protocol parameters for every layer. This is illustrated in Figure 14.2 where each node follows the Observe, Orient, Decide, and Act (OODA) loop discussed in [1, 8–10] which acts as the classical state machine for cognitive systems. In the autonomous CogNet, the OODA loop is maintained independently by each of the CogNet nodes. The primary question in this kind of autonomous Cognitive Networking is how successful is the independent use of higher layer information for an autonomous CogNet node. An early prototype, CogNet AP, for studying the impact of the performance of such systems is proposed in [7] and briefly described in the following section.

Figure 14.2. The autonomous cognitive network approach.

Figure 14.3. The architecture of CogNet Access Point.

CogNet AP As the development of sophisticated software defined radios and cognitive radios are hindered by expensive subsystems, cognitive radio systems may take longer than

274

B.S. Manoj et al.

expected to be popular. On the other hand, the cognition activity applied at higher networking layers can be beneficial and therefore an inexpensive architecture could popularize the concept of cognitive networking. Cognitive Network Access Point (CogNet AP) is one such early prototype in this direction which has many applications and significant research potential. CogNet AP is built from off-the-shelf wireless hardware and open source software in order to demonstrate the benefits of cognitive networking. It gathers, processes, analyzes, and stores information available through its IEEE 802.11 standard based interfaces in order to build a cognitive local repository which holds spatio-temporally tagged network traffic information. The inexpensiveness of the components used for building CogNet AP shows the flexibility of building cognitive network elements compared to cognitive radio devices. The CogNet AP belongs to the category of autonomous CogNet nodes which have the capability of using higher layer traffic information for efficient management of the network resources. Figure 14.3 shows the architecture of CogNet AP. The CogNet AP has two network interfaces: (i) the service interface and (ii) the monitoring interface. The service interface is used for providing network services to the users or client nodes which are associated to the CogNet AP. The second interface is used for constantly monitoring the channels. Both these interfaces are built from commercial, off-the-shelf, and inexpensive IEEE 802.11 based WLAN Network Interface Cards. Here, the cognition plane constantly monitors the network and the radio environment and populates a local CogNet AP repository. The information is spatio-temporally tagged before being stored in the repository. The CogNet controller makes appropriate decisions for physical, MAC, network, and transport layers. The CogPlane within the CogNet AP receives coarse physical layer information and all the receivable MAC layer frames and higher layer packets through the monitoring interface. From the received packets, any information related to higher layer protocol packets is extracted. Since the CogNet AP is designed for 802.11 spectrum, it is necessary to monitor activities in all the 11 channels. Monitoring all the channels simultaneously is a resource-expensive activity, therefore, CogNet AP uses a channel sampling approach. The received frames are used to populate a number of data structures from which higher layer packet information is extracted. The CogNet AP builds statistical models based on its traffic observations. There are two levels of models built: MAC layer models and higher layer models. These models are tagged spatio-temporally in order to exploit the temporal behavior of the network activity in any given geospatial point. The temporal cycle is repeated and the traffic model parameters built for every corresponding hour are averaged. For example, in our system, a seven day 24 hour temporal cycle is used in which the week days from Monday to Sunday are represented as Day 1 to Day 7. The MAC layer parameters observed are the following: (i) mean inter-arrival time of MAC frames, (ii) mean inter-arrival time for different frametypes1 , (iii) mean length of frames, (iv) mean length of frames for different frametypes, (v) frame count, and (vi) count of frames for each frametype. In order to generate traffic models for all channels, the channel switching is done in a cyclic manner across the channel range (i.e., from channel 1 to channel 11 for 802.11b systems). 1

Frametype refers to the category of MAC layer frames defined as part of the IEEE 802.11 standard.

14 Architectures and Protocols for Next Generation Cognitive Networking

275

Similar to the MAC layer model, the network, transport, and application traffic models are also built from the packets received at the cognition plane. The main traffic parameters extracted for building traffic models are the following: (i) mean inter-arrival time for IP packets, (ii) mean inter-arrival time for different protocols, (iii) mean length of IP packets received, (iv) mean length of packets for different protocols, (v) packet count, and (vi) count of packets for different protocols. The measured parameters are averaged over multiple samples and temporally stamped for every hour of operation to build the hourly model as mentioned above. Thus CogNet AP generates cyclic hourly models for mapping the network activity in the 802.11 spectrum and builds a Network Activity Time Table (NATT) which can be used by the CogNet Controller to make decisions to improve the performance of devices. An example action is the service channel selection which decides the channel to be used for serving the users. Channel selection in 802.11 spectrum is important for network access points, especially WLAN APs, though conventional APs do not provide a mechanism for dynamically choosing the best possible channel. In conventional APs, the channel is configured manually by the user. The manual channel setting causes several problems for the residential users as well as the enterprise users of WLAN equipment. In most residential deployments, users do not modify the channels and instead work with the channel preset by the device manufacturers. Thus the use of the preset channel or the dependence on manual channel setting have led to certain channels being heavily used while the remaining ones are mostly unused. For example, since most equipment manufacturers preset their APs to channel 6 as the default channel, channel 6 is the most heavily used channel in most residential environments. In enterprise WLAN deployment applications, the manual channel selection has made the WLAN deployment a complicated and expensive process. For example, in an enterprise network, optimal channel allocation to the production network requires consideration of a lot of factors such as co-channel or adjacent channel interference. Therefore, CogNet AP with its cognitive abilities to sample channels for network activity and building models for every hour of the day could exploit the periodicity of traffic on every channel and dynamically use the best channel for every hour of the day. In order to choose the channel, the CogNet AP estimates the channel activity and cumulative channel activity of every channel. The activity of a given channel is defined as the mean number of frame transmissions occurred, averaged across all sample durations within a particular hour. Cumulative activity of a channel not only represents the activity of a particular channel, but also considers the activity in other overlapping channels. For example, the cumulative activity is estimated by the following equation:

CAi =

i+COF X

Ak

(14.1)

k=i−COF k∈CH

where CAi represents the cumulative activity in channel i, COF is the channel overlap factor which in 802.11b is three, CH is the set of channels in the system (set of channels 1 to 11 in 802.11b), Ak is the activity in a given channel k. Therefore, the cumulative activity is the sum of the activities in the overlapping neighboring channels. For the physical layer, the CogNet controller chooses the best possible channel based on the following equation.

276

B.S. Manoj et al.

OperatingChannel = arg min CAi i

(14.2)

Measurements taken from a CogNet AP based testbed in a residential apartment for about three weeks are provided here as an example. The objective of this measurement setup was to see the traffic pattern differences present in a real network environment. During the measurement period, the traffic parameters observed are averaged across the same time points. Figures 14.4 and 14.5 show the traffic pattern observed. The collected results were divided into days of a week. For example, the observed traffic averaged across three Sundays is shown in Figure 14.4. We noticed crowded activity in that particular locality on two different orthogonal channels (Channels 6 and 11) and very low activity on Channel 1. The day time traffic on Sundays showed to be almost normal whereas we noticed traffic surge in the evenings. This trend was similar across all the three orthogonal channels. We noticed no significant activity on the non-orthogonal channels (802.11b channels which are not in the set of 1, 6, and 11).

Figure 14.4. Observed traffic pattern for Sundays.

Figure 14.5. Observed traffic pattern for Mondays.

On Mondays, an entirely different pattern which was determined by the residents’ work and living pattern is noticed. For example, channel 11, which is the most heavily loaded channel on Sundays, experienced much less activity. In addition, during the hours from 9am till 8pm, the traffic seemed to be very low. After 9pm, once again the traffic on channel 11 started growing. This trend remained similar for other orthogonal channels as well. The throughput achieved by the end-user devices when communicating with the CogNet AP is also provided here. In this experiment, results from ten runs are provided and the throughput obtained from all these runs were averaged. The channel was set to one of the preset channels before transferring the files. The dynamic channel selection algorithm was turned off and files of size approximately 4MB were transferred from a client mobile computer to the CogNet AP. This measurement provided the throughput achieved when not using CogNet AP’s channel selection algorithm based on the traffic observations. During the second experiment with CogNet AP, the dynamic channel switching algorithm was turned on and the CogNet AP chose the best channel based on its traffic observations. The location of

14 Architectures and Protocols for Next Generation Cognitive Networking

277

CogNet AP as well as the mobile computer was not changed during both the above experiments. Though the data rate is set to auto-rate, the AP and the mobile computer were kept static at their locations in order to avoid any potential data rate changes during our experiments. A throughput improvement is noticed as shown in Figure 14.6. The average throughput obtained from the traditional AP is about 3.25 Mbps in the residential testbed (Figure 14.6) and a throughput improvement of 10-15% is observed.

Figure 14.6. Throughput result.

Figure 14.7. The distributed cognitive networking paradigm.

14.3.2 Distributed Cognitive Networking Distributed cognitive networking is another possible approach that is expected to better exploit cognitive capabilities. These cognitive devices can interact among themselves, with centralized data bases, and across a variety of heterogeneous systems in order to utilize the cognitive network information efficiently. For example, in the future it is very likely that networking entities of different kinds will co-exist and, therefore, any capability to work across heterogeneous cognitive networking forms may be a great advantage. Figure 14.7 shows the architecture of distributed cognitive networks where the cognitive networking system can work in a fully distributed fashion. The interaction with centralized or distributed data base storage which contains spatio-temporally tagged network information is a significant value addition of this approach. The following section presents an example of distributed cognitive networking architecture.

14.4 CogNet: Cognitive Complete Knowledge Network In this section, we present one of the recently proposed architectures [6] in cognitive networking called Cognitive Complete Knowledge Network (CogNet). In the CogNet architecture, a Cognitive Bus (CogBus) within a Cognitive Plane (CogPlane) is used for exchanging and passing the data and information necessary for the efficient functioning of the CogNet framework (see Figure 14.8). This architecture provides a unique solution that fills the gap in learning the spatio-temporal

278

B.S. Manoj et al.

aspects of the protocols’ performance at every layer. For example, the temporal and spatial periodicity of higher layer traffic is not utilized in any of the existing protocol solutions for medium access protocols [3, 5]. Thus, the primary advantage of CogNet is the network performance benefit that can be derived from the system wide cognitive capability. A Cognitive Executive Function (CEF) within the CogPlane provides an analytical approach for translating the observations to actable information within the spatio-temporal context. In addition, the CEF coordinates the use of CogBus, CogPlane, and the inter-nodal exchange of information.

Figure 14.8. The distributed CogNet Framework.

In comparison to the architecture proposed by Thomas et al. [10], which proposes a monolithic cognition layer, the CogNet architecture is fully distributed. For example, in CogNet a cognition module (also called a cognitive agent) is present at every layer and can gather information and control parameters related to that particular layer. In a way, the cognitive agent serves as a local sensor, controller, and actor of each particular layer. Two important aspects of this architecture are the following: (i) this architecture can still maintain the layered abstraction of the networking protocol stack which is one of the primary factors behind the successful evolution of today’s computer networks, and (ii) this architecture can simplify the complexity of cognitive processes which otherwise may become unmanageable. In addition, the semantic interpretation of network events, behavior of protocol parameters, and the actions taken at every layer can be more efficiently handled if each layer has a cognitive module of its own. The cognitive plane helps coordination of the cognitive modules and of the information and data exchange through an internal cognitive bus. Furthermore, the CogPlane also helps the communication with other CogNet enabled nodes. This is particularly important in our architecture because, in many situations, cognitive information can be better accumulated if there is a framework for communication between CogNet nodes.

14 Architectures and Protocols for Next Generation Cognitive Networking

279

The CogNet architecture uses a language for defining the end-user requirements or end-to-end goals. The difference between our approach and existing approaches lies in translating the end-user requirements and network observations into what needs to be executed at every layer. In our case, the CogPlane is responsible for translating the end-goals to the responsibilities or action items required for each layer. A joint layer optimization module within the CogPlane develops the interaction models across the layers. This module is very important, as an optimization that does not consider the potential negative impact can be counterproductive as discussed in detail in [4].

14.4.1 CogPlane and CogBus The primary elements that enable the cognitive information exchange in the CogNet architecture are the Cognitive Plane (CogPlane) and the CogNet Bus (CogBus). The design of CogPlane is very important in developing cognitive network architectures similar to CogNet. The cognitive modules at each layer of the networking protocol stack report their observations which will be collected by the CogPlane and stored in the local or remote repository. Upon the user applications’ request, the CEF within CogPlane executes optimization algorithms for joint optimization and scheduling of resources. These optimization algorithms will generate the proper parameters to be chosen at each of the network layers and the cognitive modules are responsible for reconfiguring the protocols at each layer. Thus the CogPlane provides an opportunity for dynamic resource allocation and management with the help of the past history of the user, the device, and the network. The joint resource optimization within the CogPlane is aimed at managing the resources and the scheduling framework across multiple layers in such a way that it can achieve a satisfactory user experience. Essentially, the CogPlane intends to do a fast service compositioning across the networking layers. In addition to the coordination among the cognitive modules within a given node, the CogPlane also helps coordination between cognitive modules across nodes. The CogPlane uses the inter-node cognitive information exchange module which runs protocols such as the Cognitive Information Exchange Protocol (CIEP) [6] to manage the inter-node communication across homogeneous or heterogeneous CogNet entities. In order to enable communication between the modules and the CogPlane and in order to achieve the cross layer cognition information exchange, a bus architecture providing a broadcast medium is used. This cross layer CogNet bus (CogBus) will provide an infrastructure for publishing or exchanging cognitive information across various layers. The CogBus is also used to override the intermediate layers between a set of source-destination layer pairs. For example, if necessary, the physical layer or MAC layer can now directly communicate with the application layer without passing through the intermediate layers. Such short-cut communication scenarios may be of significant benefit. The first challenge in designing a CogBus architecture is the requirement for a light weight design of the bus architecture. Since some of the layers in the protocol stack are implemented in the Operating System (OS) kernel, it is necessary to provide a simple and efficient design for CogBus. The second challenge in designing a CogBus is in the design of an information format for exchanging CogNet specific information across all layers, that must be simple as well as extensible. One example for such information format at the application layer is the eXtensible Markup Language (XML) which might need design level changes

280

B.S. Manoj et al.

to work in CogPlane. For example, XML is a text-based protocol which may be of high overhead, and therefore a binary-based light weight design for information exchange format may be preferable. CogNet uses the CIEP protocol [6] for exchanging information across the devices in the same networking eco-system or between devices and repositories. Implementation of CIEP depends on the nature of the network. For example, in CogNet, the CIEP exploits the benefits of the broadcast nature of the wireless channel in order to query the neighbor nodes. CogNet implementation of CIEP targeted the neighbor repositories though it is possible to utilize remote or centralized repositories as well. Using CIEP, a given CogNet node broadcasts a request to its neighbors asking for their past experiences including the network parameters and network performance they observed for a particular destination node or destination network for a specific temporal domain. The requesting CogNet node can broadcast a CIEP-Request packet containing the necessary information about source and destination and the requested parameters. Upon receiving a CIEP-Request packet, the neighbor CogNet node that maintains a local CogNet repository queries its storage and replies with the CIEP-Reply packet. The CIEP-Reply packet contains the information needed by the requesting CogNet node.

14.4.2 Case Study: CogTCP In this section, we present an example realization of CogNet with a simple transport layer solution which we call CogTCP. CogTCP is a TCP transport layer with a Cognitive Transport Module (CTM). The CTM has intra-layer, inter-layer, and inter-node cognitive capability. Intra-layer cognitive capability refers to the ability of the module to learn from various internal transport layer (e.g., TCP) functional modules (e.g., socket structures and transmission control blocks). In a situation where a busy Internet server accepts thousands of TCP connections every second from a large number of networks, every new TCP connection, today, has to undergo the same transport layer behavior, e.g., the slow start phase, congestion management phase, and transmission window behavior. In CogNet, mapping of the TCP behavior models, current and past, to the host addresses and network addresses in a spatio-temporal manner will help the wireless clients optimize the protocol parameters, thereby improving their performance. Example TCP parameters are: average congestion window size, slow start threshold, probability distribution of run time throughput, bandwidth delay product, smoothed round trip delay and its average value, and spatio-temporal distribution of throughput. The inter-layer learning capability refers to the ability of the CTM to interact with other layers through the cross layer cognition bus. The inter-node cognitive capability helps a node to obtain cognitive information from other nodes. In CogNet, the CTM helps inter-node (i.e., across the transport layers of different devices) information exchange across existing TCP connections through centralized nodes such as base stations/access points. Therefore, a new TCP connection in a wireless client can query the existing connections, clients, and/or the past history information in order to avoid slow start and unnecessary congestion-control related throughput degradation. Thus, the intra-layer, inter-layer, and inter-node cognitive information exchange in CTM is particularly useful in wireless systems where the network parameter behaviors are obtained from a repository. In such cases, the inter-layer interaction may also be useful in querying the external counterpart, proxy-servers, base stations, and/or

14 Architectures and Protocols for Next Generation Cognitive Networking

281

end-hosts, in order to obtain the necessary information for improving the host’s connection performance.

Simulation Results In this section, we present the basic simulation environment and the experimental setup used for a preliminary performance evaluation of the CogTCP solution. The simulation engine was built around the GlomoSim simulation tool and the routing protocol was Distance Vector along with the IEEE 802.11 DCF MAC layer protocol. The physical layer data rate used was 2 Mbps and the radio propagation model used was two-ray model. The network topology (Figure 14.9) chosen was a grid topology with grid dimension set to 300 meters and a transmission power of about 15.0 dBm which gives an approximate transmission range of 375 meters when simulated with two-ray propagation model. We used a 25 node network with an FTP server running on Node 24 and Nodes A and B running FTP clients. Initially, when Node A runs an FTP session, the CTM module within the transport layer keeps track of the parameters and updates them in the central repository which can be centralized or distributed, and is not shown explicitly in Figure 14.9. When Node B wants to open a TCP connection with the FTP server, it queries the central repository about the right protocol parameters observed by previous nodes (in this case Node A) within the spatio-temporal domain. Hence, Node B receives the stored values for the average congestion window and the slow start threshold from Node A, and uses them as its initial values of the congestion window and of the slow start threshold and begins its data transfer session with the FTP server. In order to estimate the advantages of CogTCP, we studied the behavior of congestion window and throughput achieved which are presented below.

Figure 14.9. The network topology used for simulation experiments.

Figures 14.10 and 14.11 show the congestion window evolution in time for both Node A and Node B. Figure 14.10 presents the congestion window variation for

282

B.S. Manoj et al.

Figure 14.10. Congestion window Vs time for 10KB file transfer.

Figure 14.11. Congestion window Vs time for 1MB file transfer.

short file transfer sessions. At time 20 s, Node A begins an FTP session with the FTP server for transferring a short file of length 10KB (chosen to illustrate a simple example). At time 20.8 s, Node A’s connection completes and it registers the connection parameters such as the current congestion window, average congestion window, and slow start threshold. At time 21 s, Node B initiates a new FTP connection with the FTP server and obtains the protocol parameters that are stored in the repository, i.e., the average congestion window (4196 Bytes) and the last value of the slow start threshold (16 KB) experienced by Node A. In this case, Node B uses the average congestion window from the repository as the initial congestion window, i.e., its initial congestion window is set to 4196 Bytes. The time evolution of the congestion window for the new connection originated by Node B, with TCP and CogTCP is shown in Figure 14.10. We noted that the average congestion window at the end of Node B’s connection improved from approximately 4000 Bytes to about 5376 Bytes when Node B used CogTCP. Therefore, Node B’s transport layer connection could benefit from the experience gained by Node A and made available through the CogNet repository. Figure 14.11 shows the congestion window as a function of time for long file transfer sessions where we noted improvement in the average congestion window and the transfer time. When Node B used CogTCP, the file transfer ended approximately at 115 seconds in comparison to normal TCP which took approximately about 130 seconds. The file transfer time for Node B’s session has been reduced by approximately 15 seconds when Node B used CogTCP. We ran a 100 seed simulation campaign for estimating throughput improvement in CogTCP due to the exploitation of the information available from repositories. We used short and long file transfer sessions for these experiments. For the short file transfer experiments, we set up FTP sessions with file size fixed at 10 KB. In these experiments we studied two versions of CogTCP, i.e., CogTCP-1 and CogTCP-2. CogTCP-1 uses the value of the average congestion window experienced by Node A as the initial congestion window whereas CogTCP-2 uses half that value as the initial congestion window. In both cases, Node B uses the same slow start threshold. Initially we attempted data transfer sessions without background traffic, some results are presented in Table 14.1. The throughput performance for short files shows throughput improvement for CogTCP-1 compared to traditional TCP which does not exploit cognitive information. CogTCP-2 does not show significantly different performance compared to CogTCP-1 as the initial congestion window chosen was smaller than that of CogTCP-1, and the connection was terminated sooner than

14 Architectures and Protocols for Next Generation Cognitive Networking

283

Table 14.1. Throughput performance with and without background traffic. Throughput without background traffic (kbps) TCP CogTCP-1 CogTCP-2 Short file transfer 127 139 139 sessions (10 KB) Long file transfer 62 66 68 sessions (1 MB) Throughput with background traffic (kbps) TCP CogTCP-1 CogTCP-2 Short file transfer 120 135 135 sessions (10 KB) Long file transfer 55 57 58 sessions (1 MB) Long file transfer 44 50 50 sessions (10 MB)

expected due to the short file size used. Table 14.1 also shows the throughput performance over traditional TCP when CogTCP was used for transferring longer files of length 1 MB. Here again, we noticed similar throughput improvement for CogTCP when compared to traditional TCP. In this case, CogTCP-2 does show slightly better throughput than CogTCP-1. When CogTCP-2 takes half the value of the average congestion window obtained from the repository as its initial congestion window, it takes a longer time to touch the ceiling of the congestion window. The congestion window ceiling occurs when either the congestion window meets the receiver advertised window or there is a congestion-related packet loss. Therefore, for long connections, using half the neighbor’s average congestion window works slightly better than using the average congestion window. This also shows that the choice of parameters may be sensitive not only to time, space, source destination pair, but also to the session parameters for a particular data transfer session. A more detailed study to reveal the right choice of parameters is left for future research. The throughput performance in the presence of background traffic generated by Constant Bit Rate (CBR) sources is also shown in Table 14.1. The background CBR connections were originated from random source nodes and all terminated at Node 24. Also, in this experiment, we used short file transfer sessions (10 KB) and long file transfer sessions of 1 MB and 10 MB for the FTP sessions from Nodes A and B to the FTP server. When we used 10 CBR sessions for creating the background traffic, we noticed that the average of throughput achieved by TCP, CogTCP-1, and CogTCP-2 is reduced slightly when compared to the throughput achieved in the absence of background traffic. However, the relative throughput gain for CogTCP-1 and CogTCP-2 compared to traditional TCP remained almost the same as that of the experiments without background traffic.

14.5 Summary Cognitive networking has significant potential to contribute to the development of next generation wireless networks, especially when one considers the system level

284

B.S. Manoj et al.

optimization that these cognitive networks can provide. While the early concepts provided by Clarke et al. and Thomas et al. are good initial thoughts, they do not provide a solution which can be integrated with existing layer-oriented network architectures. CogNet is an architecture that does provide the unique property of co-existence with today’s network architecture. CogNet also provides the important benefits that can be derived from spatio-temporal information. In this chapter, we presented an architectural classification for Cognitive Wireless Networks, the design and results from an autonomous cognitive access point, and a novel distributed cognitive network architecture along with some performance results for a specific case study. Although much work remains to be done, early results provide an encouraging view on the future of cognitive wireless networking.

References 1. J. Boyd. A Discourse on Winning and Losing: Patterns of Conflict. In Lecture Notes, US Department of Defense, Pentagon, 1986. 2. D. D. Clark, C. Partridge, J. C. Ramming, and J. T. Wroclawski. A Knowledge Plane for the Internet. In Proceedings of ACM SIGCOMM 2003, August 2003. 3. C. Doerr, M. Nuefeld, J. Fifield, T. Weingart, D. C. Sicker, and D. Grunwald. MultiMAC – An Adaptive MAC Framework for Dynamic Radio Networking. In Proceedings of IEEE DySPAN 2005, pages 548–555, November 2005. 4. V. Kawadia and P. R. Kumar. A Cautionary Perspective of Cross-Layer Design. In IEEE Wireless Communications Magazine, pages 3–11, February 2005. 5. P. Kyasanur and N. H. Vaidya. Protocol Design for Multi-hop Dynamic Spectrum Access Networks. In Proceedings of IEEE DySPAN 2005, pages 645–648, November 2005. 6. B. S. Manoj, Michele Zorzi, and Ramesh R. Rao. CogNet: A Cognitive Complete Knowledge Network System. In Technical Report, Department of Electrical and Computer Engineering, University of California San Diego, CA, USA, January 2006. 7. B. S. Manoj, Michele Zorzi, and Ramesh R. Rao. CogNet AP: An Access Point for Cognitive Networks. In Technical Report, Department of Electrical and Computer Engineering, University of California San Diego, CA, USA, September 2006. 8. J. Mitola. Cognitive Radio: An Integrated Agent Architecture for Software Defined Radio. In Ph.D Thesis, Royal Institute of Technology (KTH), Sweden, 2000. 9. J. Mitola and G. Q. Maguire. Cognitive Radio: Making Software Radios More Personal. In IEEE Personal Communications Magazine, pages 13–18, Snowbird, UT, August 1999. 10. R. W. Thomas, L. A. DaSilva, and A. B. MacKenzie. Cognitive Networks. In Proceedings of IEEE DySPAN 2005, pages 352–360, November 2005.

15 Scheduling in Cognitive Networks Scheduling Variable Rate Links – Centralized and Decentralized Approaches

Chandrasekharan Raman, Jasvinder Singh, and Roy D. Yates, and Narayan B. Mandayam WINLAB, Rutgers - The State University of NJ. Summary. In this chapter, we present an optimization framework for link level and flow level scheduling in cognitive radio networks. In the centralized scheduling framework, a spectrum server coordinates the transmissions of a group of links sharing a common spectrum. With knowledge of the link gains in the network, the spectrum server schedules the on/off periods of the links so as to satisfy constraints on link fairness. We then compare the throughput regions of centralized scheduling and a probabilistic random access scheme, wherein in each slot, a link is active with a fixed probability chosen independent of other interfering links. We observe that for the case of two interfering links, the probabilistic scheme does not suffer any loss in the rate region relative to the centralized scheme if the interference between the links is sufficiently low. We then present a distributed algorithm where each link independently updates its transmission probability based on its measured throughput to achieve any desired feasible rate vector in the throughput region of the probabilistic scheme and prove its convergence. Finally, we present an optimization framework for end-to-end flow level scheduling of flows in network with mutually interfering links.

15.1 Introduction The emergence of unlicensed spectrum has spawned an impressive variety of important technologies and innovative uses, ranging from scientific and industrial to domestic applications and systems. Since these systems must adapt to a wide variety of unpredictable conditions, the emerging technologies called “cognitive radio” [17] offer significant potential benefits in system capacity and service quality. In their simplest embodiments (which are by no means simple to implement) cognitive radios can recognize the available systems and adjust their frequencies, waveforms and protocols to access those systems efficiently [16]. Not surprisingly, it is upon these difficult “design” issues that most current research activities are focused, as illustrated by some chapters in this book. For instance, some physical layer design issues related to cognitive radios are discussed in [5] while some issues related to sensing are addressed in [34, 35]. While these basic capabilities represent a difficult and significant step forward, they fail to fully illuminate the effects of

285 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 285–305. c 2007 Springer.

286

Chandrasekharan Raman et al.

cognitive behavior. When there exist methods by which cognitive radios can independently discover local information, a variety of physical layer, system and network layer protocols can be applied to allow cooperation and coexistence. However, such levels of cooperation and interoperability may not be possible when multiple services and systems must coexist. In a heterogeneous environment, some users may look to obtain high data rates without regard to energy efficiency; other users may wish to transmit at a fixed rate but with high efficiency. In certain applications, it will be important to enforce fairness constraints. In general, the system performance will have a multidimensional characterization. These dimensions often represent conflicting performance measures. In this chapter, we take a step towards characterizing some of these performance measures. In the realm of cognitive radio networks, two distinct sets of issues emerge. First, for a given set of transmitter and receiver technologies and a specified set of performance constraints, one must resolve the multidimensional boundaries of system performance. As we shall see, this is a difficult problem, even if complete system state information is available to all network nodes. Second, the collection of intelligent adaptation policies of the individual nodes constitute a large distributed system for spectrum allocation. A given set of distributed information gathering and exchange mechanisms may greatly influence the performance of the system. This work tries to separate these issues. We examine the boundaries of system performance under the assumption that efficient open access to spectrum can be resolved by an impartial “spectrum server” [4] that can obtain information about the interference environment through measurements contributed by different terminals, and then offer suggestions for efficient coordination to interested service subscribers. In this work, the spectrum server is a centralized scheduler that uses full knowledge of the network configuration to specify the activity patterns of the individual links. Our aim is to provide upper-bounds on the performance of distributed adaptive scheduling methods based on the performance of the spectrum server. We then characterize the loss in performance of a simple distributed random access scheme. It is recognized that the nodes of a cognitive radio network can interact in a variety of (arbitrary) ways. To distill these interactions, we see that each radio follows a transmission policy (specified here for instance by the spectrum server) that results in signals that vary over time, frequency, and space. This variation may be the result of adaptation to measurements of channels or interference. The performance of a particular signaling strategy depends on each receiver’s ability to resolve signals in the presence of interfering transmissions. In this work, we do not assume any specific physical layer transmission technique, but abstract out the essential features of any technique into a rate matrix. For a system of interfering wireless transmissions, a mathematical model starts with a basis for signal space. Each user employs a combination of basis functions to transmit in some or all of the signal dimensions. This work assumes a relatively simple signaling structure: signals are time-slotted transmissions over a frequencyflat channel. The interference that a receiver faces depends on the subset of nodes transmitting in that time slot. The structure of the transmitted signals (spread or unspread, coded or not) and the receiver technology determines the ability of the node to separate a desired signal from the interference. Thus, the data rate that a link can obtain in a time slot depends on the subset of nodes that are actively transmitting in that slot.

15 Scheduling in Cognitive Networks

287

There are many ways in which a spectrum server can coordinate a network of cognitive radios. The work in [18] considers the spectrum server’s role in demand responsive pricing and competitive spectrum allocation. In [27], the spectrum server aids the users’ decision in forming stable coalitions among themselves so as to maximize the sum throughput in the network. In this work, the spectrum server specifies the schedule of transmission for mutually interfering users sharing a common spectrum. Scheduling transmissions in a wireless network has been studied in various contexts outside of cognitive radio. Though link scheduling is a difficult problem to solve [15], joint optimization of scheduling and routing in a cross-layer framework has been studied by lot of researchers in recent years. In [9], a joint scheduling and power control strategy is proposed to maximize network throughput and energy efficiency of the system. Another direction in this problem is addressed in [8], where the authors look at the cross-layer issues of routing, scheduling and power control. Other related works include [19, 23, 40]. In the centralized scheduling framework, we assume that we obtain a non-zero rate in the links for any non-zero signal-to-interference ratio (SIR). If the link gains are known to the spectrum server, it can schedule the transmissions among the links to maximize the system throughput. The optimization problem, subject to minimum rate constraints in the individual links, is posed as a linear program. It is shown that when there is no minimum rate constraint, a fixed set of links (called the dominant mode) that maximizes the sum rate is operated all the time. In order to offset the inherent unfairness in the above solution, we introduce a minimum rate constraint in each link. We observe that the max-min fair rate allocation can be obtained in one step by solving a linear program which maximizes the minimum common rate among the links. The optimization framework for proportional fair scheduling is posed as a non-linear program. In the distributed scheduling framework, we compare the throughput regions of centralized scheduling and a probabilistic random access scheme, wherein in each slot, a link is active with a fixed probability chosen independent of other interfering links. We observe that for the case of two interfering links, the probabilistic scheme does not suffer any loss in the rate region relative to the centralized scheme if the interference between the links is sufficiently low. For more than two interfering links, the characterization of throughput rate region for the probabilistic scheme becomes intractable and similar observations are not easily forthcoming. However, we give a distributed algorithm where each link independently updates its transmission probability based on its measured throughput to achieve any desired feasible rate vector in the throughput region of the probabilistic scheme and prove its convergence. Finally, we present the cross-layer optimization framework for end-to-end flow scheduling for a network supporting variable rates because of mutual interference.

15.2 System Model Before we explain the system model, we comment on the notation of this chapter. We use boldface lowercase characters for vectors and boldface P uppercase for matrices. If a is a vector, aT denotes its transpose and aT b = i ai bi represents the inner product of the vectors a and b. The vector of all zeros and all ones are represented by 0 and 1 respectively. Inequalities between vectors are component-wise inequalities.

288

Chandrasekharan Raman et al.

Figure 15.1. Graph of network showing the nodes and directed links.

We consider a wireless network with N nodes forming L logical links sharing a common spectrum. The network can be represented as a directed graph G(V, E), where the nodes in the network are represented by the set of vertices V of the graph and the links are represented by a set of directed edges E. Therefore the cardinalities |V| = N and |E| = L. A directed edge from a node m to node n implies that m wishes to communicate data to node n. We consider the scenario where the spectrum server coordinates the activity of the set of L links to share the spectrum efficiently. Define the set of transmission modes T = {0, 1, . . . , M − 1}, where M = 2L denotes the number of possible transmission modes. Then the mode activity vector ti of mode i is a binary vector, indicating the on-off activity of the links. If ti = [t1i , t2i , . . . , tLi ]T is a mode activity vector, then 1, if link l is active under transmission mode i, (15.1) tli = 0, otherwise. Note that there are M possible transmission modes including the mode with activity [0, 0, . . . , 0]T in which all links are off. Figure 15.1 shows a representative network and Figure 15.2 shows a particular transmission mode for the set of links. Let the transmitter power on a link l ∈ E be Pl . If Glk is the link gain from the transmitter of link k to the receiver of link l and σl2 is the noise power at the receiver of link l, the Signal-to-Interference Ratio (SIR) γli at the receiver of link l in transmission mode i is given by tli Gll Pl . 2 k∈E,k6=l tki Glk Pk + σl

γli = P

(15.2)

It is assumed that the link gain between a transmitter and receiver takes into account the path loss and attenuation due to shadow fading. The data rate in each link depends on the SIR in that link. We assume that the transmitter can vary its data rate, possibly through a combination of adaptive modulation and coding. In particular, for a given mode, the transmitter and receiver on a link employ the highest rate that permits reliable communication given the link SIR in that mode.

15 Scheduling in Cognitive Networks

289

Figure 15.2. Graph of network showing transmission mode corresponding to (1 0 1 0).

For purposes of this study, we assume that the transmission of other links are treated as Gaussian noise and that a transmission on link l is reliable in a given mode i with a data rate cli = log(1 + γli ). (15.3) This same interference model is also employed in [6, 11]. We emphasize here that we do not consider any minimum SIR threshold required at the receiver. A non-zero γli in each transmission mode i defines a non-zero rate on the link l. Let xi be the fraction of time that transmission mode i is active and rl be the average data rate of link l. The average data rate in link l is the time average of the data rates of all the transmission modes that include link l. Thus, X rl = cli xi , (15.4) i

or in vector form, r = Cx,

(15.5)

where C = [c1 c2 . . . cM ] is an L × M matrix with non-negative entries, such that its jth column cj = [c1j , c2j , . . . , cLj ]T contains the rate obtained by each link in mode j. The idea of transmission modes to model a set of transmitter-receiver pairs has been used in [7, 43] previously. In this work, we use such a model as basis specify schedules that maximize the system throughput, subject to minimum rate and fairness constraints. In practice, a scheduler will specify a sequence of transmission modes. Typically, this would be done by constructing a frame with N time slots and allocating Nj time slots to each mode j. The fraction of time that mode j is active will be xj = Nj /N . For sufficiently large N , the ratio Nj /N can be made arbitrarily close to any xj ∈ [0, 1]. In this case, the average rate r in (15.5) will represent the average link data rates over one frame. For our analytical model, we optimize these average rates per frame by specification of the time fractions in x, without explicitly specifying

290

Chandrasekharan Raman et al.

the precise slots assigned to each mode. We denote the set of all feasible schedule vectors by X = {x : 1T x = 1, x ≥ 0}. (15.6) We note that this model has a number of desirable characteristics. First, we observe that all aspects of transmitter and receiver technology are embedded in the rate matrix C. For example, if the links employed CDMA spreading, Equation (15.2) for the SIR on link l in mode j would be appropriately modified, as in [44] for example, to reflect the transmitter spreading sequences and receiver filter vectors used in that mode. Similarly, if we were to assume a particular practical coding and modulation scheme, we would modify Equation (15.3) for the expected number of bits that we would expect to decode at a specified SIR. If the nodes employ multiaccess techniques like interference cancellation or multiuser detection, the rates in the C matrix are calculated according to the multiaccess rates that can be achieved in each transmission mode. Thus the general model allows for consideration of a large class of physical layer interactions. We employ the specific choices in Equations (15.2) and (15.3) to demonstrate tradeoffs in between average rates and various fairness constraints. In addition, consider the average link rates obtained by an arbitrary dynamic spectrum access system. Each link employs a dynamic policy, based on measurements and perhaps some side information, to determine when to be active. At any given time, some subset of links will be active and the rates obtained on each link will be determined by the interference generated by those active links. In short, any dynamic spectrum access system yields a series of transmission modes. The rate obtained by each link l in each mode j will be given by clj . To speak of average rates for the links, the collection of link access policies must yield an ergodic transmission mode process such that we can define xj as the fraction of time the system is in mode j. In this case, the average link data rates will be given by (15.5). In short, any set of average rates obtained by a dynamic spectrum access system can also be obtained by a centralized scheduler that specifies the identical time fraction xj for each mode j. Thus the centralized scheduler allows us to separate what average link rates can be obtained from the issue of whether a dynamic system can achieve those rates.

15.3 Maximum Sum Rate Scheduling A centralized scheduling approach is better suited to optimize the overall utility of the network as it can have access to global information about the nodes of the network. A wide variety of objective functions have been proposed in the literature depending on the driving application, e.g., [13]. In this work, as a first step, we are interested in finding the schedule that maximizes the sum of the average data rates over all links l = 1, 2, . . . , L, subject to constraints on the minimum rates for each link. The optimization problem for finding the maximum sum rate schedule can be posed as the following linear program (LP): max subject to

1T r

(15.7a)

r = Cx,

(15.7b)

r ≥ rmin ,

(15.7c)

x ∈ X.

(15.7d)

15 Scheduling in Cognitive Networks

291

P where, the objective function 1T r = i ri is the sum of average rates of the individual links and the constraint (15.7c) represents the minimum rate requirement of the links. The variables in the LP (15.7) are r and x. Rewriting the LP in terms of a single variable x, we have copt (rmin ) = max subject to

1T Cx

(15.8a)

Cx ≥ rmin ,

(15.8b)

x ∈ X.

(15.8c)

In the special case when minimum rate constraint rmin = 0, the transmission mode corresponding to the highest sum rate is always operated. The maximum sum rate copt (0) is equal to the maximum column sum of the rate matrix C [33]. This is true because L L X M M L X X X X cli xi = xi cli ≤ max cli , (15.9) 1T Cx = l=1 i=1

i=1

l=1

i

l=1

P

where the inequality in (15.9) is true since i xi = 1. Equality holds in (15.9) when P x = xopt = [0 0 . . . 1 . . . 0 0]T where the position of 1 in xopt is ˆi = arg maxi L l=1 cli . We refer to ˆi as the dominant transmission mode. The dominant mode could consist of a single link or a collection of links depending on the topology of the network. However, the links that are not a part of the dominant mode are not operated at all. Since we are maximizing a global objective function like sum rate of the network, there is an inherent unfairness in the system. In order to offset the unfairness in the system, we introduce a non-zero minimum rate requirement in the individual links of the network. In such cases, the optimal schedule balances the use of the dominant mode against modes that provide nonzero rates to the otherwise disadvantaged links. This, however, comes at the cost of reduction in the sum rate of the network. The loss in sum rate due to the minimum rate constraint was characterized in [33]. There exists a trade-off between the sum rate and individual rates of the links, i.e., when we increase the minimum rate requirement in the links, the sum rate obtained decreases. This is intuitively satisfying since the dominant mode, which offers the highest sum rate, is always turned on whenever there is no minimum rate requirement on the links. When the minimum rate requirement is increased from zero, other transmission modes are forced to be scheduled for transmission in order to satisfy the minimum rate requirement of the links. Since the modes other than the dominant mode always offer a lesser sum rate than the dominant mode, the sum rate decreases monotonically with increase in required minimum rate. The minimum rate requirement for all links in the network can be increased by trading off sum rate until it is infeasible to support the rate requirement in all links.

15.4 Fair Scheduling In the previous section, we observed that maximizing the sum rate without minimum rate constraints leads to unfairness among the links. More fundamentally, the underlying shared wireless medium and the global objective function bring up the question of fairness. We investigate fair scheduling strategies in this section. We start with the conventional max-min fair objective.

292

Chandrasekharan Raman et al.

15.4.1 Max-Min Fairness Definition 1 A vector of rates r is said to be max-min fair if it is feasible and for each l ∈ E, rl cannot be increased while maintaining feasibility without decreasing rl0 for some link l0 for which rl0 ≤ rl . Formally, for any other feasible allocation ˜ r, with r˜l > rl , there must exist some l0 such that r˜l0 < rl0 ≤ rl . Max-min fairness is well studied in the context of flow control of elastic traffic in data networks [1]. Fairness in wireless networks have been studied in [20,28,41]. Our model differs from the data networks model, wherein there could be many sessions flowing through multiple links of finite capacity. There may be several bottleneck links and a feasible rate allocation is max-min fair if and only if all flows pass through at least one bottleneck link [1]. However, in our model, each link has a minimum rate requirement to satisfy and it is not clear as to what constitute the system bottlenecks. Even if we could answer this question, the next important question is: how many bottlenecks are there in the system? Given that the max-min fair set of rates exist, how do we compute them? We try to answer these questions in this section. In order to obtain the max-min fair schedule in our setting, we begin by formulating the LP that maximizes the minimum common rate in all the links. The LP is given by r∗ = max

rmin

(15.10a)

subject to

r = Cx,

(15.10b)

r ≥ rmin 1,

(15.10c)

x ∈ X.

(15.10d)

We now have the following theorem. Theorem 1. If the link gains Glj are all non-zero, then any solution to the LP (15.10) results in the unique rate vector r∗ = r∗ 1. The proof of Theorem 1 appears in [33]. In the context of multi-hop flows in wireless networks, a similar result, called the solidarity property, was derived by Radunovic et al. in [32]. Since the LP (15.10) which maximizes the common rate among the links, achieves equal rates for all links, the following corollary follows from the definition of max-min fairness. Corollary 1. The solution x∗ obtained by solving the LP (15.10) results in the maxmin fair rate allocation. Note that the corollary reveals that the max-min fair rates can be obtained in one step by solving the LP (15.10), in contrast to the usual iterative computation of maxmin fair rates of multiple flows through finite capacity links in wired networks [1] and multihop flows in wireless networks [41]. This difference is due to the fact that there is a single bottleneck – the shared wireless bandwidth – in the network. As a result, so long as any link perceives non-zero interference from all other links in the network, the maximum common rate among the links is the max-min fair rate.

15 Scheduling in Cognitive Networks

293

15.4.2 Proportional Fairness Proportional fairness is another fairness criteria which is popular in the context of scheduling of wireless links. It has been studied in the context of multiuser diversity [42] and downlink scheduling for HDR [37]. Definition 2 A feasible vector of rates r is proportional fair if for any other feasible vector r0 , the aggregate of proportional change is negative: X ri0 − ri ≤ 0. ri i

(15.11)

In [21], Kelly proposed proportional fairness in the context of rate control for elastic traffic. It was also shown that the proportionally fair vector is the one that maximizes the sum of logarithms of the utility functions. We follow a similar approach to obtain the proportional fair rates, we solve the following non-linear optimization problem with linear constraints: X max log rl (15.12a) l

subject to

r = Cx,

(15.12b)

x ∈ X.

(15.12c)

The objective function of the above non-linear optimization problem is increasing and strictly concave. The constraint set is linear and hence the problem is a convex optimization problem [3]. This implies that the problem has a unique global maximum over the constraint set. The solution for such problems can be found out by gradient search algorithms. Energy conservation in wireless systems is essential to enhance the lifetime of each individual node and the overall life of the network [12]. In certain applications, it may be required that the schedule conforms to stringent energy constraints in the network [10]. Efficient scheduling with constraints on the energy in each node of the network is studied in [46].

15.5 Distributed Dynamic Spectrum Access Policies The centralized scheduling strategies that have been presented thus far, require information about all the links to be available at the spectrum server. There are many reasons why such a centralized scheduler may be difficult to implement in a real world situation, including: 1. Complete information about all the links and their channel gains should be known to the spectrum server to solve the scheduling problem precisely. 2. If the number of the links increases, the size of the LP increases exponentially. 3. The exchange of information between the centralized spectrum server and the individual transmitters may not be an easy task.

294

Chandrasekharan Raman et al.

An alternative to approaching the spectrum management problem in a centralized way is to adopt device-centric spectrum management schemes [48]. A devicecentric scheme corresponds to each user taking independent actions based on local interference measurements. Such actions might include channel switching [29, 30], deciding whether to transmit or not during a time slot [47], changing transmission power levels, or modulation waveforms [6, 39]. Under the general framework of transmission modes introduced earlier, such schemes will result in a time sequence of network transmission modes. If the devices use ergodic spectrum access policies, then their aggregate behavior will induce a probability distribution on the network transmission modes, and the average rates achieved by the devices under those set of policies can then be defined over the induced probability distribution. Since lack of coordination between the devices can only reduce the set of achievable rates by the devices, a centralized spectrum management scheme for a given network model will act as a benchmark for a decentralized or device-centric scheme. Device-centric random access schemes, e.g., ALOHA have been widely used in practical multiple access systems. The CSMA/CA schemes used in the IEEE 802.11 networks are very popular, thanks to the ease of implementation and decentralized control of these random access techniques. Of late, research effort has been directed towards analyzing the performance of these random access schemes. In [20, 45], the authors propose distributed approaches for fair random access. The throughput characteristics of random access schemes have been studied in in [22, 26]. A recent work [14] characterizes the Pareto boundary of the network throughput region as the family of solutions optimizing a weighted proportional fairness objective, parametrized by weights chosen by the links. The authors also propose a distributed random access scheme to achieve a desired point within the Pareto optimal boundary. In this section we present a simple device-centric spectrum access policy for the system model from the previous section (i.e., links use constant power and turn on and off in each time slot). Instead of following a transmission schedule computed and prescribed by a centralized entity, each link is active with a fixed probability chosen independently of the other links in each slot. A given set of transmission probabilities chosen by the users results into a unique average rate vector achieved at the links. We first compare the achievable rate region for this device-centric spectrum access scheme with that of the centralized case, and then provide a distributed algorithm to compute the transmission probabilities corresponding to a vector of average rates desired at the links.

15.5.1 Rate Regions We define the rate region as the set of rate vectors that can be achieved by a dynamic spectrum access scheme. The rate regions for the centralized scheduling scheme and the device-centric spectrum access scheme are discussed below.

Centralized Scheme In this scheme, a schedule is specified by the fractions of time each transmission mode is active. As discussed in the previous section, the spectrum server can be used to compute the the optimum time fractions of activity, to maximize a certain utility function. Let xj be the fraction of time transmission mode j is active and rl

15 Scheduling in Cognitive Networks

295

be the average data rate of link l. The average data rate in link l is the time average of the data rates of all the transmission modes that include link l. Thus, X rl = clj xj , (15.13) j

and in vector form, r = CL x.

(15.14)

The rate region for the centralized scheduling scheme is given by RS L := {(r1 , . . . , rL ) : r = CL x, x ∈ X }.

(15.15)

L vertices which are given by Clearly, the region RS L is a polytope defined by its 2 the column vectors of CL .

Device-Centric Spectrum Access Scheme In this scheme, link l transmits with a probability pl chosen independent of the other links in the network. The rate region for the random access scheme is given by RP L := {(r1 , . . . , rL ) : r = CL x, x = f (p), 0 ≤ p ≤ 1}

(15.16)

L

where f : RL → R2 is given by   (1 − p1 )(1 − p2 ) . . . (1 − pL )  p1 (1 − p2 ) . . . (1 − pL )      .. f (p) =  . .     (1 − p1 )p2 . . . pL

(15.17)

p1 . . . pL S It is easy to see that RP L ⊆ RL . Also, since f (·) is a continuous mapping, the set {x : x = f (p), 0 ≤ p ≤ 1} must be a closed and continuous region and therefore RP L must also be closed and continuous. Our aim will be to characterize the Pareto P S boundary of RP L and identifying the conditions, if any, under which RL ≡ RL . We consider the simple case L = 2 to obtain insight into the shape of the rate regions.

15.5.2 Characterization of Rate Region for the Decentralized Scheme Let α and β be the normalized rates achieved the links when both the links are simultaneously active.1 Then the rates on two links are r1 = p1 (1 − p2 ) + αp1 p2 ,

(15.18)

r2 = (1 − p1 )p2 + βp1 p2 .

(15.19)

The above equations can be rewritten as 1

The normalization is done so that the links get unit rate when they transmit in isolation.

296

Chandrasekharan Raman et al.

P P S Figure 15.3. RS L and RL for the case α + β ≥ 1. RL ≡ RL and is given by the area enclosed by OABC. B represents (α, β).

r1 = p2 (p1 α + (1 − p1 ).0) + (1 − p2 )(p1 .1 + (1 − p1 ).0),

(15.20)

r2 = p2 (p1 β + (1 − p1 ).1) + (1 − p2 )(p1 .0 + (1 − p1 ).0).

(15.21)

In vector form,

r1 r2

α 0 + (1 − p1 ) = p2 p1 β 1 1 0 + (1 − p1 ) + (1 − p2 ) p1 0 0

(15.22)

The above representation of the rate vector, as a nested convex combination of the polytope vertices, is useful in visualizing the rate region RP 2 . We now consider two different cases.

Low Interference Case : α + β ≥ 1 Figure 15.3 shows RS 2 . Any point in the quadrilateral OABC can be achieved using centralized scheduling. Notice that the vertices of the polytope OABC are the columns of C2 . For a given probability vector p = [p1 p2 ]T , the rate vector r given by (15.22) is shown as point F in Figure 15.3. As p1 varies between 0 and 1, points D and E completely trace the line segments AB and OC respectively. As p2 varies between 0 and 1, the point F traverses the line segment ED completely. Hence, it S can be seen that by varying p, the achieved rate region RP L is the same as RL [38]. Analytically,   (r1 , r2 ) :   1 0 ≤ r1 ≤ α ⇒ 0 ≤ r2 ≤ α−(1−β)r , RP . (15.23) L = α   1) α ≤ r1 ≤ 1 ⇒ 0 ≤ r2 ≤ β(1−r . 1−α

15 Scheduling in Cognitive Networks

297

High Interference Case : α + β < 1 In this case, RS 2 is given by the triangle formed by points O, A and C in Figure 15.4. As in the previous case, point F in Figure 15.4 corresponds to the rate vector r achieved for a given p = [p1 p2 ]T . If p1 = 1, the line segment DE coincides with BC. As p1 varies from 1 to 0, DE moves from BC to an intermediate position HG to finally AO (for p1 = 0) tracing out the region RP L as the area enclosed by OAHIC. Note that the boundary AHIC of the region is convex and contains two linear components AH and IC. The presence of linear component AH can be geometrically understood by observing that as DE moves from HG to AO, endpoint D always lies on the linear segment AH. In order to intuitively understand the presence of IC, it helps to notice that as p1 varies from 1 to 0, J, the point of intersection of DE and BC initially moves from B towards C, goes up to a certain point I, and then moves back towards B. The analytical characterization of the above region is given by [38]   (r1 , r2 ) :       2   α−(1−β)r1   0 ≤ r1 ≤ α ⇒ 0 ≤ r , 2 ≤ 1−β α P √ 2 . (15.24) RL = 2 ( (1−β)r −1) 1 α  ,   1−α   1−β < r1 < 1 − β ⇒ 0 ≤ r2 ≤     1−β ≤r ≤1 ⇒ 1) 0 ≤ r2 ≤ β(1−r . 1 1−α

P S Figure 15.4. RS L and RL for the case: α + β < 1. RL is given by the area enclosed P by OAC and RL is given by the area enclosed by OAHIC. B = (α, β).

15.5.3 Distributed Algorithm In this section, we present a distributed algorithm to compute the probability vector p corresponding to a feasible point in the rate region RP L . Each link updates its

298

Chandrasekharan Raman et al.

probability of transmission based on the rate it achieves in the previous slot. We start by identifying a property of the function ri (p) that is the key for proving the convergence of our distributed algorithm. The rate ri (p) achieved by link i in the random access scheme can be written as ri (p) =

M X j=1

cij

L Y

X

= pi

[tlj pl + (1 − tlj )(1 − pl )]

(15.25)

l=1

cij

j:tij =1

Y

[tlj pl + (1 − tlj )(1 − pl )]

(15.26)

[tlj pl + (1 − tlj )(1 − pl )]

(15.27)

l6=i

Let us define X

gi (p−i ) =

j:tij =1

cij

Y l6=i

where p−i = [p1 , . . . , pi−1 , pi+1 , . . . , pL ]T

(15.28)

Then ri (p) can be written as ri (p) = pi gi (p−i )

(15.29)

The following lemma states a property of ri (p), which is straight forward to prove [38]. Lemma 1. gi (.) is a positive and strictly decreasing function of pj for all j 6= i. Therefore, ri (.) is a strictly increasing function of pi and a strictly decreasing function of pj for all j 6= i. Now for each link i, consider the following iterative update of pi (n) based on the current rate ri (n) and the desired rate rid . In practice the current rate ri (n) is measured by averaging the rates obtained over many slots. pi (n + 1) =

rid pi (n) ri (n)

(15.30)

Theorem 2. Given a feasible rate vector rd ∈ RP L , if all the links perform the above iteration independently starting with p(0) = 0, then their iterations converge to a fixed point (p∗ , r∗ ) such that r∗ = rd and p(n) ≤ 1 for all n. Proof. Using (15.29), we can rewrite (15.30) as pi (n + 1) =

rid gi (p−i (n))

(15.31)

Substituting p(0) = 0 in the iteration, we get p(1) = rd and therefore p(1) ≥ p(0). Using Lemma 1 with the above fact, it follows that p(2) ≥ p(1) and in general p(n + 1) ≥ p(n) for all n. Therefore, if p(n) is bounded from above by 1, as n increases, it must converge to a fixed point p∗ and the corresponding r∗ is then equal to rd . We now prove that if rd is feasible, then p(n) remains bounded below 1. Feasibility of rd means that there exists 0 ≤ pd ≤ 1 such that

15 Scheduling in Cognitive Networks pdi =

rid gi (pd−i (n))

299 (15.32)

By definition, pd ≥ p(0). Using (15.31) and (15.32), we can see that pd ≥ p(1) and in general pd ≥ p(n) for all n. Therefore p(n) must also remain bounded below 1. In case the users choose an infeasible rd , the above iteration will lead to a situation where some pi (n)’s exceed 1. To avoid such infeasible conditions, we can modify the iteration to the one given below. d ri pi (n + 1) = min (15.33) pi (n), 1 ri (n) The above iteration converges to the desired rate vector rd if it is feasible.

15.6 Cross Layer Scheduling of End-to-End Flows Building up on the model described in Section 15.2, in this section, we discuss crosslayer scheduling of end-to-end flows in a cognitive wireless network with mutually interfering links. End-to-end rate guarantees through link scheduling have been well studied for wired networks [31]. However, rate guarantees in wireless networks are difficult to handle because the shared wireless medium induces a large number of various scheduling constraints. The tutorial paper [25] summarizes some of the recent works on cross-layer optimization in wireless networks. In [21], the author describes a model in which there exist multiple flows through finite capacity links. The problem is to find a set of optimal rates and flows that maximize a utility function of the source rates subject to constant link capacity constraints. However, our objective is to schedule the links for transmission and find optimal rates and flows subject to rate constraints in the links, due to interference from other links. In [36], the authors propose a fair scheduling algorithm that guarantees end-to-end max-min fair rates. The scheduling constraints are such that active links at any slot must constitute a matching. Our work provides a schedule which maximizes the sum of utility function of the source rates of origin-destination (OD) pairs, and fair scheduling can be brought out as a special case. The scheduling constraints can be very general and includes schedules consisting of matchings. In [24], the authors describe a cross-layer rate control and scheduling scheme. The rate in a link is a generic function (called the rate-power function) of transmission power of the link. The rates are lower bounded by the sum of source flows in the links. In this work, we provide an explicit characterization for the rate of a link. The rates are variable and depend logarithmically on the Signal-to-Interference Ratio (SIR) of the individual links. Consider a wireless network with N nodes forming L links sharing a common spectrum. As described in Section 15.2, the network can be represented as a directed graph G. Let us assume that the network consists of K sessions. A session is specified by an origin-destination (OD) pair. A route r is a sequence of links forming a path in the graph G. We assume that there are R possible routes in the whole network. For a session k, the routes are specified by the L × R matrix Ak with entries ∈ {0, 1}, where

300

Chandrasekharan Raman et al. 1, if link l is a part of route r, (15.34) [Ak ]lr = 0, otherwise. Let fkj , k = 1, . . . , K, j = 1, . . . , R be the flow corresponding to the kth session in the jth route. If Rl denotes the set of routes passing through the link l, we can write the expression for the rate in link l in the session k as X fkj = aTkl fk , (15.35) rlk = j∈Rl T

where fk = [fk1 fk2 · · · fkR ] is the vector of flows for the kth session, k = 1, . . . , N , and aTkl is the lth row of Ak . Let rk = [r1k r2k · · · rLk ]T , we can then write the link rate vector equation, rk = Ak fk . (15.36) Thus the aggregate rates through links l = 1, 2, . . . , L are given by X Ak fk . (15.37) r= k

Each OD pair kP in our system gets a rate yk = 1T fk , k = 1, . . . , K. We are interested in maximizing k Uk (yk ), the sum of utility functions of the rates in each session. In a network of mutually L interfering links, all the M = 2L modes may not be valid for transmission. For instance, since the links share a common spectrum, the transmitter and receiver in a node operate in the same channel. Hence the node cannot transmit and receive simultaneously because of self-interference. We refer to this constraint as the duplexing constraint. The duplexing constraint implies that modes corresponding to the adjacent edges in the graph G are invalid, i.e., the modes should constitute a matching. We now specify the cross-layer optimization problem for maximizing the sum of utility functions of the rates in each session can be posed as the mathematical program: X max Uk (yk ) (15.38a) k

subject to

yk = 1T fk , k = 1, . . . , K,

(15.38b)

r = Cx, X Ak fk , r≥

(15.38c) (15.38d)

k

x ∈ X,

(15.38e)

fk ≥ 0, k = 1, . . . , K.

(15.38f)

The variables of the above optimization problem are x, fk , k = 1, . . . , K. If the utility function Uk (yk ) = yk , then (15.38) maximizes the sum of end-to-end flows of OD pairs. We then get a linear program which can be solved using standard techniques [2]. If Uk (yk ) = log(yk ), then (15.38) is a convex optimization problem, which solves for the proportional fair rates [21]. Note that (15.38c) and (15.38d) imply that the sum of the flows in each link is upper bounded by a quantity which depends on the schedule. However, in the model described in [21], each link has a finite capacity, which is a constant. The result of the optimization program (15.38) is a set of transmission modes along with the time fraction of operation of these modes and the flows in each route. Appropriate activity of the modes makes the transport of end-to-end flows possible.

15 Scheduling in Cognitive Networks

301

Figure 15.5. Network with 5 nodes in a line. Each link is of length d. Table 15.1. Routes taken by the flow for different node separation distances. Range of d

Route Transmission modes (Sequence of links) used 0≤d≤4 (1,5) {(1,5)} 48 (1,2,3,4,5) {(1,2)}, {(2,3)}, {(3,4)}, {(1,2),(4,5)}

15.7 Simulation Results We discuss a simple illustrative example for flow-level scheduling in this section. We consider a network of five nodes in a line as shown in Figure 15.5, each node separated by a distance d from the other node. We label the nodes 1 through 5. These nodes form the vertices V of a complete graph G. Thus, there are 5 C2 = 10 links in the network. Each node may be able to transmit to any other in the network in just one hop. Because of the duplexing constraint, the links that transmit in any slot constitute a matching in the complete graph G. Thus, the number is transmission modes in the network is equal to the number of non-trivial matchings in the G, i.e., 25. The effect of interference between the links are captured by the matrix C. The interference gain Glj between the transmitter of link j and the receiver of a link l is given by Glj = d−4 . The transmit powers are fixed for all transmissions. We consider a single session originating at node 1 and ending at node 5. Note that there are 8 paths in the network for this OD pair. The objective is to maximize the flow in the network for the given OD pair. Given the distance d between the nodes, we can calculate the SIR for links in every possible mode and then construct the matrix C for a fixed transmit power. By solving (15.38), we obtain the routes and the schedule for the modes required to obtain the optimal flows in these routes. Figure 15.6 shows the variation of sum rate of flows with the distance d between any two nodes in the network. For small values of d, the direct hop is the most optimal route. This also results in the highest sum rate since the mode with the single link (1, 5) can be used. When d increases, there is a four fold increase in the length of direct hop link. Hence the flow between the OD pair decreases rapidly due to the path loss. As d increases further, the single hop link is no more optimal and the flow takes more than one hop to reach the destination. In our example, when 4 < d ≤ 6, two hops are required to maximize the flow. Since the nodes are equally spaced apart, the first hop is at the node 3. The flow decreases with increasing d, but for a two hop case, the link distance increases twice as d. For values of d in the range 6 < d ≤ 8, the flow is maximized when it takes three hops to the destination.

302

Chandrasekharan Raman et al.

Figure 15.6. Variation of source rates with the distance between the nodes for a fixed transmit power. d = 10 corresponds to 20 dB received SNR.

For larger values of d, the flow is maximized when the each node transfers to the nearest neighbor and the flow takes 4 hops to the destination. The routes followed by the flows and the scheduled transmission modes are given in the Table 15.1.

15.8 Conclusion Though link and flow level scheduling has received enough attention in literature, new emerging radio technologies offer additional degrees of freedom to the scheduling problem. In this chapter, we introduced the spectrum server that calculates efficient and fair time schedules. Performance of these schedules provide upper bounds for distributed scheduling schemes. We then study the loss in performance of a memoryless decentralized scheduling policy and provide a distributed algorithm that achieves any feasible rate vector. Finally, we extend the link scheduling framework to end-to-end flows in a network of mutually interfering links that support variable rates.

References 1. D. Bertsekas and R. Gallager. Data Networks. Prentice-Hall, 1992.

15 Scheduling in Cognitive Networks

303

2. D. Bertsimas and J. Tsitsiklis. Introduction to Linear Optimization. Athena Scientific, 1997. 3. S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004. 4. M. Buddhikot, P. Kolodzy, S. Miller, K. Ryan, and J. Evans. DIMSUMNet: New directions in wireless networking using coordinated dynamic spectrum access. In IEEE WoWMoM, June 2005. 5. D. Cabric and R. W. Brodersen. Physical layer design issues unique to cognitive radio systems. In Proc. IEEE PIMRC 2005, 2005. 6. N. Clemens and C. Rose. Intelligent power allocation strategies in an unlicensed spectrum. In Proc. IEEE DySPAN, 2005. Baltimore, MD. 7. R. L. Cruz and A. V. Santhanam. Optimal link scheduling and power control in CDMA multihop wireless networks. In IEEE Globecom, pages 52–56, 2002. 8. R. L. Cruz and A. V. Santhanam. Optimal routing, link scheduling and power control in multi-hop wireless networks. IEEE Infocom, 2003. 9. T. ElBatt and A. Ephremides. Joint scheduling and power control for wireless ad hoc networks. Wireless Communications, IEEE Transactions on, 3(1):74–85, 2004. 10. A. Ephremides. Energy concerns in wireless networks. IEEE Wireless Communications, 9(4):48–59, Aug 2002. 11. R. Etkin, A. Parekh, and D. Tse. Spectrum sharing for unlicensed bands. In Proc. IEEE DySPAN, 2005. Baltimore, MD. 12. A. J. Goldsmith and S. B. Wicker. Design challenges for energy-constrained ad hoc wireless networks. IEEE Wireless Communications, 9(4):8–27, Aug 2002. 13. D. Goodman and N. Mandayam. Power control for wireless data. IEEE Personal Communications, 7:48–54, Apr 2000. 14. P. Gupta and A. Stolyar. Optimal throughput allocation in general randomaccess networks. In Proc. CISS, 2006. Princeton, NJ. 15. B. Hajek and G. Sasaki. Link scheduling in polynomial time. IEEE Trans. Info. Theory, 34(5):910–917, Sept 1988. 16. S. Haykin. Cognitive radio: brain-empowered wireless communications. IEEE Journal on Selected Areas in Communications, 23(2):201–220, Feb 2005. 17. J. Mitola III. Cognitive radio: an integrated agent architecture for software defined radio. PhD thesis, KTH Royal Institute of Technology, 2000. 18. O. Ileri, D. Samardzija, T. Sizer, and N. Mandayam. Demand responsive pricing and competitive spectrum allocation via a spectrum policy server. In Proc. IEEE DySPAN, Nov 2005. Baltimore, MD. 19. M. Johansson and L. Xiao. Cross-layer optimization of wireless networks using nonlinear column generation. IEEE Trans. Wireless Commun., 5(2):435–445, Feb 2006. 20. K. Kar, S. Sarkar, and L. Tassiulas. Achieving proportional fairness using local information in ALOHA networks. IEEE Trans. Auto. Control, 49(10):1858– 1862, Oct 2004. 21. F. Kelly. Charging and rate control for elastic traffic. Euro. Trans. Telecommun., 8:33–37, Jan/Feb 1997. 22. L. Kleinrock and F. Tobagi. Packet switching in radio channels: Part i: Carrier sense multiple access and their throughput delay characteristics. IEEE Trans. Communications, 23(12):1400–1412, Dec 1975.

304

Chandrasekharan Raman et al.

23. M. Kodialam and T. Nandagopal. Characterizing achievable rates in multi-hop wireless networks: the joint routing and scheduling problem. In ACM Mobicom, Sept 2003. San Diego, CA. 24. X. Lin and N. B. Shroff. The impact of imperfect scheduling on cross-layer rate control in multihop wireless networks. In Proc. IEEE Infocom, 2005. 25. X. Lin, N. B. Shroff, and R. Srikant. A tutorial on cross-layer optimization in wireless networks. IEEE Journal on Selected Areas in Communications, 24(8):1452–1463, Aug 2006. 26. J. Massey and P. Mathys. The collision channel without feedback. IEEE Trans. Info. Theory, 31(2):192–204, Mar 1985. 27. S. Mathur, L. Sankaranarayanan, and N. Mandayam. Coalitional games in receiver cooperation for spectrum sharing. In Proc. CISS, 2006. Princeton, NJ. 28. T. Nandagopal, T.-E. Kim, X. Gao, and V. Bhargavan. Achieving MAC layer fairness in wireless packet networks. In Proc. ACM Mobicom, pages 87–98, Aug 2000. 29. J. Neel and J. Reed. Performance of distributed dynamic frequency selection schemes for interference reducing networks. In Proc. MILCOM, 2006. Washington D.C. 30. N. Nie and C. Comaniciu. Adaptive channel allocation spectrum etiquette for cognitive radio networks. ACM MONET (Mobile Networks and Applications), special issue on Reconfigurable Radio Technologies in support of ubiquitous seamless computing, to appear 2006. 31. A. Parekh and R. Gallager. A generalized processor sharing approach to flow control - the single node case. IEEE/ACM Trans. Networking, 1(3):344–357, Jun 1993. 32. B. Radunovic and J. Y. L. Boudec. Rate performance objectives of multihop wireless networks. IEEE Trans. Mobile Computing, 3(4):334–349, Oct.-Dec 2004. 33. C. Raman, R. Yates, and N. Mandayam. Scheduling variable rate links via a spectrum server. In Proc. IEEE DySPAN, 2005. Baltimore, MD. 34. A. Sahai, N. Hoven, and R. Tandra. Some fundamental limits on cognitive radio. In Proc. of Allerton Conf. on Comm., Control and Computing, Oct 2004. 35. A. Sahai, R. Tandra, S. M. Mishra, and N. Hoven. Fundamental design trade-offs in cognitive radio systems. In Proc. TAPAS, Aug 2006. 36. S. Sarkar and L. Tassiulas. End-to-end bandwidth guarantees through fair local spectrum share in wireless ad hoc networks. IEEE Trans. Auto. Control, 50(9):1246–1259, Sept 2005. 37. S. Shakkottai and A. Stolyar. Scheduling for multiple flows sharing a timevarying channel: The exponential rule. American Mathematical Society Translations, 207, 2002. 38. J. Singh, C. Raman, R. Yates, and N. Mandayam. Random access for variable rate links. In Proc. MILCOM 2006, Oct 2006. Washington D.C. 39. V. Srivastava, J. Neel, A. MacKenzie, J. Hicks, L.A. DaSilva, J.H. Reed, and R. Gilles. Using game theory to analyze wireless ad hoc networks. IEEE Communications surveys and tutorials, 4th quarter 2005. 40. L. Tassiulas and A. Ephremides. Jointly optimal routing and scheduling in packet radio networks. IEEE Trans. Info. Theory, 38(1):165–168, Jan 1992. 41. L. Tassiulas and S. Sarkar. Maxmin fair scheduling in ad hoc wireless networks. IEEE Journal on Selected Areas in Communications, 23(1):163–173, Jan 2005.

15 Scheduling in Cognitive Networks

305

42. P. Viswanath, D. Tse, and R. Laroia. Opportunistic beamforming using dumb antennas. IEEE Trans. Info. Theory, 48(6):1277–1294, Jun 2002. 43. H. Viswanathan and S. Mukherjee. Throughput-range tradeoff of wireless mesh backhaul networks. IEEE Journal on Selected Areas in Communications, 24(3):593–602, Mar 2006. 44. A. J. Viterbi. CDMA: Principles of Spread Spectrum Communication. Addison Wesley, 1995. 45. X. Wang and K. Kar. Distributed approaches for proportional and max-min fairness in random access ad hoc networks. In Proc. CISS, 2006. Princeton, NJ. 46. R. Yates, C. Raman, and N. Mandayam. Fair and efficient scheduling of variable rate links via a spectrum server. In Proc. IEEE ICC, Jun 2006. Istanbul, Turkey. 47. Q. Zhao, L. Tong, A. Swami, and Y. Chen. Decentralized cognitive MAC for opportunistic spectrum access in ad hoc networks: A POMDP framework. IEEE Journal on Selected Areas in Communications, to appear 2007. 48. H. Zheng and L. Cao. Device-centric spectrum management. In Proc. IEEE DySPAN, 2005. Baltimore, MD.

16 Design of Terminals and Infrastructure Components for Cognitive Wireless Networks A Platform Perspective

Alexander Vießmann, Admir Burnic, Christoph Spiegel, Arjang Hessamian-Alinejad, Andreas Waadt, Guido H. Bruck, and Peter Jung Lehrstuhl f¨ ur KommunikationsTechnik Universit¨ at Duisburg-Essen, Bldg. BB, Oststraße 99, D-47057 Duisburg, Germany [email protected] Summary. Cooperation in wireless networks will facilitate a new dimension in the evolution of multimedia communications, setting out from the today’s situation with a multitude of communication standards and radio interfaces both in the licensed and the unlicensed domain. In order to pave the way towards cooperative networks, the deployment of cognitive wireless solutions, which will form the communication platforms, will be a key asset. In the future, we expect to see an increasingly flexible, ad-hoc utilization of the available spectrum in the unlicensed domain and a coexsitence of communication standards in the licensed bands. In this communication, the authors will illustrate a platform based approach towards cognitive wireless communications. Also, the authors will present three software defined radio concepts designed by the members of the Lehrstuhl f¨ ur KommunikationsTechnik, namely the HAWK (Highly Adaptable Wireless Kit), the FALCON (Flexible Access Logic for COmmunication Networks), and the MUSTANG (MUlti-STAndard single chip transceiver for the Next Generation), finally, giving an outlook on their wireless optical communication device, termed ARGOS. HAWK, FALCON, MUSTANG and ARGOS form the basis of the Lehrstuhl f¨ ur KommunikationsTechnik’s cognitive wireless platform, termed PROMETHEUS. The setup of PROMETHEUS and its functionality shall be illustrated together with selected measurement results.

16.1 Ubiquitous Wireless Multimedia Multimedia [6] is commonly defined as the computer based, integrated generation, manipulation, presentation, storage and communication of mutually independent pieces of information, which consist of at least one continuous, i.e., time dependent, and one discrete, i.e., time independent, medium. Figure 16.1 shows a snapshot of a possible multimedia compilation: The avatar of a human interacts on the upper left side of Figure 16.1 with a virtual reality (VR) application, manipulating a two-dimensional projection of a three-dimensional background formed by artistic patterns in which a futuristic view of a sports car design study is integrated. The

307 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 307–325. c 2007 Springer.

308

Alexander Vießmann et al.

outcome of these mainpulations is shown in the remainder of Figure 16.1. Current developments include hypermedia, i.e., a mark-up language based presentation of multimedia as well as virtual reality. A prominent recent approach has been “Second Life”, a networked VR application which has found an increasing acceptance in our information society.

Figure 16.1. A multimedia compilation.

The furthering and evolution of such VR applications will require cooperative wireless networks. Hence, the deployment of increasingly flexible cognitive wireless solutions, which will form the communication platforms, will be an important step. Reconfigurability of transceivers for wireless networks will become increasingly important in the forthcoming decades. Wireless networks will evolve their still limited set of services to a great variety of applications, and the today’s set of wireless terminal types will expand considerably. A single homogeneous network like UMTS (Universal Mobile Telecommunications System) [3] will not be able to provide the requested versatile services and applications alone. Only heterogeneous networks consisting of wired and wireless networks will form a catalyst for the evolution of such a diverse mobile world. Future wireless communication systems will hence hierarchically integrate a broad variety of wireless networks into a common structure, encompassing e.g. UMTS/WCDMA (Wideband Code Division Multiple Access) based cellular mobile systems, OFDM (Orthogonal Frequency Division Multiplexing) [2] based W-LANs (Wireless Local Area Networks), e.g., IEEE 802.11a/g, WiMAX and Mobile WiMAX (Worldwide Interoperability for Microwave Access, IEEE 802.16d/e), and inexpensive personal-area networks like Bluetooth. In addition, future proofness of networks has become a key asset in the support of both an economical and a comfortable operation of wireless terminals and the corresponding infrastructure components. To enable the future proof evolution of existing networks, services and applications, appropriate terminal and infrastructure implementations are desirable.

16 Design of Terminals and Infrastructure Components

309

The deployment of communication systems strongly depends on the availability of appropriate microelectronics. Therefore, the combined approach to communications and microelectronic system design is crucial. The implementation of wireless terminals and infrastructure components can be eased when using a platform concept. To derive a viable platform concept, it is necessary to identify the tasks which must be carried out by the infrastructure components, in particular by base stations, and by the wireless terminals. Importantly, the technical requirements such as the power classes of wireless terminals and infrastructure components and the different service classes w.r.t. the supported information rates should be clarified. Flexible and reliable hardware/software architectures, allowing the concurrent processing of different controlling tasks for wireless terminals will be required to meet the aforementioned goals. Particularly, the reconfiguration capabilities of the mentioned wireless terminals and infrastructure components, e.g., the balancing of the amount of hardware and software in their design, need to be excellent. In particular, the transceiver implmentation must be carefully handled, e.g., w.r.t. the required wireless analog front-ends, supporting e.g., radio as well as unguided optical waves. Then, the mixed signal domain, e.g., the filtering and signal digitization, and the digital transceiver part, i.e., the digital front-end, the detection and decoding schemes as well as higher layer communication protocols, including the adaptation of medium access control (MAC) engines, must be considered. Finally, the power consumption, the bill of materials (BOM), including the area of microelectronic components, the level of integration, and the form factors must be taken into account.

16.2 Reconfigurability and Cognitive Modes of Operation Reconfigurability is not a new technique [4]. Already during the 1980s, reconfigurable receivers were developed for radio intelligence in the short wave range. However, reconfigurability became familiar to many radio developers with the publication of e.g., the special issues on software radios of the IEEE Communication Magazine [1]. A transceiver can be considered as a software radio (SR), if its communication functions are realized as programs running on a suitable processor [4]. An ideal SR directly samples the antenna output which does not seem feasible with respect to e.g., power consumption and linearity as well as resolution requirements for ADCs (Analog-to-Digital Converters). A software defined radio (SDR), however, is a practical and realizable version of an SR [4]: The received signals are sampled after a suitable band selection filter, usually in the base band or a low intermediate frequency (IF) band. The digital transceivers need at least one digital processor which will be running the software. Therefore, it must be decided which kind of processor shall be used, e.g., a microcontroller (µC) or microprocessor (µP), a digital signal processor (DSP) or an application specific standard product (ASSP) which includes several processor cores. Furthermore, the mode of operation w.r.t. the updating of communication software must be considered [7]. In addition, the adaptation of robust protocols, starting with frequency scanning and simultaneous operation of several wireless communications systems and applications, has to be included in the design process: It is hence

310

Alexander Vießmann et al.

a reasonable consequence to found the aforementioned platform concepts on SDR (Software Defined Radio) and CR (Cognitive Radio). The way of reconfiguring the terminal, in particular the realization of a processor with master controller and a Petri net based approach, which allows concurrent mode of operation, high reliability and secure applications, seems to be a pragmatic and yet rather novel approach towards the practical implementation of a cognitive wireless device [7]. The approach towards reconfiguration introduced by the authors consists of a master controller which is responsible for a reliable reconfiguration. In addition, there has to be a unit, which can communicate with the network, a PHY and MAC engine. This PHY and MAC engine needs software modules with signal processing algorithms for the data processing path. The third part is a memory, which contains these software modules. The master controller starts a cognitive operation in order to obtain the best reconfiguration and software modules needed for the SDR. The reconfiguration then consists of linking software modules existing in the memory, and installing them into the SDR to use the software modules in the regular signal processing chain. The master controller works with Petri net based software architecture. It needs a scalable control program by using e.g., Petri net compilers. The implementation and validation of the master controller based concept on a PCB (Printed Circuit Board) level integration will be done by setting out from currently used demonstrator platforms. It will be implemented in a single chip processor after validation of the PCB level integration.

Figure 16.2. Master controller concept for reliable reconfiguration of CRs.

16 Design of Terminals and Infrastructure Components

311

Figure 16.3. Platform based design flow.

The key component is the concept of the master controller including the appropriate Petri net based controlling schemes [7]. Some anticipated concurrent tasks are shown in Figure 16.2, namely the linking operation, the cognitive operation and the regular operation. After power on reset, the master controller starts a setup operation. The parameters and software modules for the last used wireless communication standard are then loaded from the single chip processor’s on-chip memory into its PHY and MAC engine, which is then ready to communicate using the wireless communication standards. Then, the regular operation (RO) is started, which operates the communication connections of the device. During the setup operation, the cognitive operation (CO) is prepared. It operates at the same time as the RO, but it senses for available communication connections. If it has been successful in perceiving a new connection, it asks for user interaction, if the new radio interface operation is desired. The user interaction may be skipped, if the user has allowed this. In case of a reconfiguration to take place, the linking operation (LO) is started. It connects appropriate binary PHY and MAC modules in order to adapt to the desired new communication connections and it validates the new transceiver chain. At the same time, the regular operation & prepare (RO&P) prepares the system for reconfiguration, e.g., update, reconfiguration and handover, while operating the communication connections at the same time.

312

Alexander Vießmann et al.

The validation of the linked modules is necessary to avoid conflicts which may arise when the linked signal processing path is not able to work properly with the new wireless communications standard and to increase reliability. A complete malfunction of the device may be possible, if no such validation is made. In case of a negative validation, the RO and CO start again as before. No change will be done in the signal processing path. In case of a positive validation a new setup process starts, which reconfigures the SDR starting the RO and CO including the new signal processing path for the newly adapted wireless communications link [7]. The master controller is Petri net based to simplify the software design, which leads to a better service quality of the controller. The main part of the whole system is the PHY and MAC engine, which runs the software modules for the different wireless modes and standards. It contains a processor for the digital signal processing [7]. The software modules must be validated with such a PHY and MAC engine before they are released for use in devices by storing them into the memory. A first device to test the suggested method is an SDR demonstrator [7] which should be developed according to the platform based design flow addressed in the next Section 16.3.

16.3 Platform Based Design Flow In order to develop cognitive wireless terminals and infrastructure components, the requirements on the reconfigurability and cognitive modes of operation, discussed in Section 16.2, must be taken into account. This leads to several prerequisites: •

•

•

•

Future Proofness: The selected concept of the cognitive wireless terminals and infrastructure components must be future proof, which leads to the scalability, the modularity and the availability and market impact requirements discussed below. Scalability: The deployed HW (Hardware), the FW (Firmware) and the application SW (Software) should be easily scalable to meet new requirements, e.g., of new or evolved communication standards; also, the simultaneous operation of several communication systems, e.g., of a Bluetooth link and a UMTS/WCDMA link at the same time, should be supported. Modularity: The HW, the FW and the application SW should be developed and implemented in a modular fashion, so that single modules can be easily replaced by e.g. their refurbished or evolved versions or by completely new modules. Availability and Market Impact: The terminal components and the IDEs (Integrated Design Environments), which are required for the development and implementation and which are to be acquired from third parties, must be easily available at the time of purchase and during the subsequent months and years to facilitate a long life time of the cognitive wireless terminals and the infrastructure components; furthermore, the long term support of these products by the vendors must be guaranteed; in addition only such components with a strong market potential or a high market acceptance should be used.

A platform based approach faciliates to meet all aforementioned requirements. In this Section 16.3 the authors will illustrate the platform develolpment by the example of a demonstrator realization. A demonstrator requires the design of digital,

16 Design of Terminals and Infrastructure Components

313

mixed signal and analog circuitry, the development and realization of FW and application SW, the integration of all HW, FW and application SW components and the validation of its functionality. The digital, mixed signal and analog circuitry can be made available either in the form of micro- and nanoelectronics, i.e., ICs (Integrated Chips), usually designed by using an HDL (Hardware Description Language) like the very high speed integrated circuits HDL or verilog HDL, both abbreviated as VHDL, or a block design entry method, or in the form of PCBs, which are designed by block schematic entry tools. The FW, as well as the needed HW drivers, is commonly realized by using either C or assembly languages. Application SW is either provided by deploying the Java or the .NET platforms. Since demonstrator realizations must be carried out in a limited amount of time with limited resources, it has to be first decided which parts of the demonstrator should be developed by the team and which parts should be acquired from a third party. In order to facilitate a most efficient design flow, the authors have decided to rely on a mix of commercially available HW and inhouse developed HW tailored to the specific needs of the demonstrators. The following tasks are carried out inhouse: • • • •

The design of digital, mixed signal and analog BB (Baseband), of IF (Intermediate Frequency) PCBs and of optical front-end PCBs. The development and realization of all FW and application SW. The integration of all HW, FW and application SW components. The validation of the demonstrator functionality.

The realization of an SDR and CR demonstrator does not necessarily require the synthesis of integrated chips based on HDLs. Therefore, the authors do not consider such an implementation inhouse, except for the programming of CPLDs (Complex Programmable Logic Devices). All integrated circuits, including the corresponding IDEs, are acquired from third parties. Also, the manufacturing of PCBs has been outsourced. The corresponding design flow used by the authors is sketched in Figure 16.3. From a top-down perspective, the design flow which will be discussed in what follows comprises • • • • •

the the the the the

sign-off level, system integration level, software level, board level and chip level.

Firstly, the selection of the programmable processors and processor cores, i.e., the DSPs, the µCs and the µPs is done. As already mentioned, the authors do not carry out the design of processors or ASSPs. Instead, commercial processors and processor cores are selected w.r.t. their importance in the wireless market. Therefore, only third party products are taken into account. A further selection criterion is the availability of flexible and powerful IDEs for the programming of the selected processors and processor cores. In order to be able to quickly include these processors in own designs, the authors usually acquire pre-commercial or commercial EVMs (Evaluation Modules) and PCBs which host the selected processors, also providing digital connectors. The selected commercial EVMs, commercial PCBs, including RF (Radio Frequency) PCBs, and commercial IDEs are taken from third parties. On the chip level shown in Figure 16.3, the authors carry out the programming of CPLDs. Commercially available programming tools are deployed for this task.

314

Alexander Vießmann et al.

Setting out from the aforementioned commercial products, special glue logic is commonly needed. On the board level of Figure 16.3, the design of special glue logic PCBs, including input-output (I/O) for the connection of analog and digital data sources and sinks as well as mixed signal circuitry for the interfacing of radio frequency and optical front-ends is accomplished. Furthermore, optical front-end PCBs are designed and implemented inhouse. Only commercial tools for the PCB design are deployed. The manufacturing of the specially required PCBs is carried out by the electronics workshop of the Universit¨ at Duisburg-Essen. On the software level depicted in Figure 16.3, the authors develop and implement both the FW required for the operation of the selected processors and processor cores as well as of the application SW, including the necessary HW drivers. Again, commercial FW/SW developemnt tools are used.

Figure 16.4. Block diagram of the WiMAX 802.16 communication system. The FW development requires the full understanding of the digital communication system which shall be integrated in the demonstrator. Figure 16.4 shows the functional block diagram of the WiMAX 802.16 communication link as an example. In the demonstrator setup, the data source and the data sink are computer terminals. These computer terminals run the application SW. Both computer terminals are connected to the transceivers via USB (Universal Serial Bus) connections. At the transmitter, the data randomizer, the channel coding including the puncturing and the interleavig, the symbol mapping and the IFFT (Inverse Fast Fourier Transform) based OFDM modulation with the cyclic prefix extension, are realized in DSP FW modules. The DAC (Digital-to-Analog Converter) and the subsequent functionalities, e.g., the filtering and the mixing, are done by HW modules. At the receiver, the analog and mixed signal reception is also accomplished by appropriate

16 Design of Terminals and Infrastructure Components

315

HW components and their impairments, i.e., phase noise, frequency errors, I-Q imbalance effects, the perturbation by AWGN (Additive White Gaussian Noise) and the sample rate errors occuring in the ADC, are reflected in Figure 16.4. All other modules are implemented in DSP FW. After coarse and fine synchronization of time and frequency, the RSSI (Radio Signal Strength Indication) followed by the FFT (Fast Fourier Transform) based OFDM demodulation is carried out. Then, intermodulation distortions and frequency errors are removed and the channel estimation is accomplished. The equalization, which is based on matched filtering, computes L-values (LLRs, Log-Likelihood Ratios). The equalizer output is fed into the symbol de-mapper. After de-interleaving, channel decoding and de-randomization, the detected data are delivered to the data sink.

Figure 16.5. Simulation results obtained for the WiMAX IEEE802.16d communication standard in the case of a single tap flat Rayleigh fading channel model, QPSK modulated data symbols and a carrier frequency of 3.5 GHz.

An important first step in the implementation of transceivers is the performance calibration, based on theoretical performance evaluations and on the generation of simulation results. A typical benchmark is usually achieved when well-known channel models are deployed, e.g., a single tap flat Rayleigh fading channel model. Figure 16.5 shows simulation results obtained for the WiMAX IEEE802.16d communication standard in the case of the aforementioned single tap flat Rayleigh fading channel model, QPSK modulated data symbols and a carrier frequency of 3.5 GHz. Both the bit error ratio (BER) and the block error ratio (BLER) results are shown versus the

316

Alexander Vießmann et al.

average signal-to-noise ratio (SNR) 10 log10 (Eb /N0 ) at the output of the equalizer, i.e., before the channel decoding is carried out. The BLER reflects the number of OFDM symbols which contain at least one erroneous bit versus the number of transmitted OFDM symbols. It is obvious that the BLER is greater than the BER at a given SNR, typically by a factor of 17. This reflects the fact, that bit errors occur in bursts which is a well-known feature of time varying channels. In the case of a mobile speed of 3 km/h, there is only a neglibible variation of the channel during an OFDM symbol. Therefore, the simulated BER curve meets the theoretical performance bound of [5] s ! 1 Eb /N0 Pe = (16.1) 1− 2 1 + Eb /N0 perfectly. However, in the case of a velocity of 30 km/h, which means a Doppler frequency of almost 100 Hz, the time variation of the channel is considerable during an OFDM symbol which results in a measurable error floor. Assuming that a desirable QoS (Quality of Service) is achieved at a BLER equal to 10−2 , 10 log10 (Eb /N0 ) of ≈26.4 dB is required in the case of a mobile speed of 3 km/h whereas an SNR degradation of about 0.6 dB must be expected in the case of a velocity of 30 km/h.

Figure 16.6. PROMETHEUS platform. On the system integration level shown in Figure 16.3, all HW, FW and SW components are integrated in a terminal setup. Finally, on the sign-off level, cf. Figure 16.3, the test and validation of the demonstrator is carried out.

16.4 PROMETHEUS Platform 16.4.1 General Concept Figure 16.6 shows the current plaform concept used by the Lehrstuhl f¨ ur KommunikationsTechnik, termed PROMETHEUS. The PROMETHEUS consists of an MMI (Man Machine Interface), realized either by a computer terminal, see also Figure 16.4, or a touch screen, an ARM1026EJ-S based master controller including the storage of the communication systems to be supported and four transceiver engines, namely the HAWK, the FALCON, the MUSTANG and the ARGOS transceivers. These transceiver engines will be briefly discussed in the following Section 16.4.2.

16 Design of Terminals and Infrastructure Components

317

16.4.2 Transceiver Engines The ARGOS transceiver provides infrared communication and ranging capabilities for short range scenarios, see Figure 16.6. It has been implemented using proprietary communication techniques. The ARGOS transceiver incorporates a Texas Instruments TMS320C6416 DSK (DSP Starter Kit) with a single TMS320C6416 DSP per transceiver. Furthermore, it contains several daughtercards including an infrared wireless optical front-end. The MUSTANG has been designed as a multi-processor platform using up to 20 TMS320C6416 DSPs, simultaneously, deploying a commercial multi-processor operating system. It uses a commercial platform deploying the Texas Instruments interface module (TIM) connectivity standard. However, in order to facilitate a reliable real-time operation, the authors had to modify the original multi-processor operating system, appropriately. Furthermore, the digital MUSTANG platform has been interfaced to the mixed signal and analog circuitry available inhouse. Whereas the ARGOS, the HAWK and the FALCON transceivers are targetting low-cost and mid-range segments with a strong focus on the terminal side, the MUSTANG transceiver has been tailored to infrastructure components. It realizes the functionalities of the aforementioned transmission schemes and extends these by the deployment of smart antenna techniques like beamforming and joint predistortion, including novel and not yet published irregular sampling based versions. The FALCON uses regular sampling and has been tailored to single carrier schemes deploying CDMA (Code Division Multiple Access) and multicarrier schemes like OFDM. Hence, CDMA based communication systems like UMTS/WCDMA and UMTS/TD-CDMA as well as OFDM based communication systems like WiMAX and IEEE 802.11a/g are supported by the FALCON. In Figure 16.7, a photo of the FALCON transceiver is shown. It connects to a computer terminal via USB, therefore, the MMI is not depicted in Figure 16.7. the FALCON transceiver consists of an antenna, an RF board provided by Atmel Duisburg which is mounted vertically, several analog and mixed signal daughtercards, developed by the authors, and a Texas Instruments TMS320C6416 DSK with a single TMS320C6416 DSP. The HAWK transceiver relies on irregular sampling based detection and has been tailored to single carrier schemes deploying FDMA (Frequency Division Multiple Access) and TDMA (Time Division Multiple Access), such as Bluetooth, DECT (Digital Enhanced Cordless Telecommunications) and GSM (Global System for Mobile Communciations). The HAWK transceiver therefore supports binary and HOM (Higher Order Modulation) schemes, including FM (Frequency Modulation) schemes like MSK (Minimum Shift Keying), GMSK (Gaussian Minimum Shift Keying) and GFSK (Gaussian Frequency Shift Keying), as well as linear modulation schemes such as QPSK (Quadrature Phase Shift Keying) and 8PSK (Octernary Phase Shift Keying) and their corresponding shifted versions deployed in Bluetooth, GSM and D-AMPS, as well as 16QAM (16-ary Quadrature Amplitude Modulation). A photo of the HAWK transceiver is depicted in Figure 16.8. It consists of a touch screen display, an antenna, an RF board provided by Atmel Duisburg, which is mounted directly below the touch screen display, several analog and mixed signal daughtercards and a Texas Instruments TMS320C6713 DSK with a single TMS320C6713 DSP.

318

Alexander Vießmann et al.

Figure 16.7. FALCON transceiver.

16.4.3 HAWK Transceiver The HW setup of the HAWK transceiver shall be further discussed as a representative example in this Section 16.4.3. Figure 16.9 shows a cut of the Texas Instruments TMS320C6713 DSK used in the HAWK transceiver. The DSP can be interfaced to daughtercards by the external peripheral interface connector, located on the left hand side, and the external memory interface connector, located on the right hand side. Figure 16.10 shows two views of the interface daughtercard developed by the authors. The interface daughtercard is connected to the Texas Instruments TMS320C6713 DSK by the aforementioned external peripheral interface connector

16 Design of Terminals and Infrastructure Components

319

Figure 16.8. HAWK transceiver.

and the external memory interface connector. The interface daughtercard contains a programming interface, an interface to the USB module, a serial test port, an interface to the man machine interface module, a TX/RX serial interface for the return channel, a TX and an RX interface to the radio frequency front-end module, a TX serial SMIQ (Rohde&Schwarz Vector Signal Generator) interface, an RX ESPI (Rohde&Schwarz Test Receiver) interface, a TX baseband interface, a programmable PLL (Phase Locked Loop), a TX and an RX CPLD, a TX and an RX FIFO (First-In First-Out) memory as well as a 14 bit DAC. The programming interface is used to modify the FW of the transceiver, which can e.g., be accomplished by the aforementioned ARM CPU. The USB module is used to communicate with the computer based data sink and data source. The serial test port faciliates debugging of the HW and the FW in the running transceiver. The man machine interface module is required to support the connection of touch screen displays and keyboards. The TX/RX serial interface for the return channel allows to emulate the wireless return channel for testing and evaluation purposes of handshake based communications,

320

Alexander Vießmann et al.

Figure 16.9. Cut of the Texas Instruments TMS320C6713 DSK.

such as hybrid ARQ (Automatic Repeat Request) protocols and their effect on the transmission quality. The TX and RX interfaces to the radio frequency front-end module support the integration of these RF modules. The TX serial SMIQ interface, the RX ESPI interface, and the TX baseband interface are used for transmission quality measurements. The programmable PLL facilitates the adaptation to varying clock requirements. The TX and RX CPLDs allow the programming and the reconfiguration of the mixed signal and the analog front-ends. The TX and an RX FIFO memories decouple the digital DSP system from the real-time mixed signal and analog front-end. The DAC converts the digital TX signals into analog baseband TX signals. In Figure 16.11, the modules which are connected to the interface daughtercard are shown. These modules are the man machine interface module, a USB module and the analog front-end modules. The man machine interface module integrates an SD (Secure Digital) card interface, hosting an SD card, and a serial interface module facilitating the connection to a touch screen dsiplay or, alternatively, the connection to a keyboard and a conventional display. The USB module is required for the connection to a computer based data sink and data source. The analog front-end modules comprise e.g., a comparator, the Atmel EVM TX and RX interfaces and

16 Design of Terminals and Infrastructure Components

321

Figure 16.10. Interface daughtercard connected to the TMS320C6713 DSK.

322

Alexander Vießmann et al.

Figure 16.11. Modules connected to the interface daughtercard and the Texas Instruments TMS320C6713 DSK. the Atmel EVM configuration interface. The comparator realizes the functionality of a 1 bit ADC, which is required for the irregular sampling based receiver, being a zero-crossing detector.

16.5 Future Proofness of the PROMETHEUS Platform In this Section 16.5 we will consider how the platform based design flow described in Section 16.3 has been reflected in the PROMETHEUS platform. Firstly, we will consider the chip level, cf. Figure 16.3. All transceivers use the same DSP family, namely the TMS320C6xxx DSPs. Furthermore, only ARM based CPUs, in the case of the PROMETHEUS platform the ARM1026EJ-S CPU, are integrated. Also, all transceivers use the same PLL, FIFO memory, CPLD, and DAC families. Next, we will discuss the board level of Figure 16.3. In the HAWK, the FALCON and the ARGOS transceivers, the same DSK families are used with identical connectors to daughtercards. Only the MUSTANG transceiver requires a more elaborate digital platform owing to the need of multi-core processing. The RF front-ends belong to the same families, provided by Atmel Duisburg, which can be programmed and reconfigured by the same interface and the same programming commands. The daughtercards are modular and, hence, interchangeable. On the software level, all transceivers use the same design tool and design libraries. Furthermore, all FW has been programmed in C using the Texas Instruments Code Composer Studio (CCS). The FW has been implemented in a modular

16 Design of Terminals and Infrastructure Components

323

fashion, facilitating the refurbishment and replacement of single FW modules, which also simplifies the porting of FW modules from one transceiver to another. Also, all transceivers use the same application SW which integrates inhouse and third party SW modules written in Java, visual C++ and C#. The system integration level finds a modular HW/FW/SW concept with the same or similar connectivities. The development and evolution of HW, FW and application SW is done by using the versioning system CVS, including a detailed template based module description. This approach facilitates an innovative and reliable design procedure meeting high quality standards when working in both real and virtual teams. Furthermore, the sign-off level uses the same measurement and evaluation procedures. Typically, the authors measure the eye diagrams of transmit and receive signals, the spectrum and the fractional out-of-band power of the transmit and receive signals, the EVM (Error Vector Magnitude) of the transmit and the receive signals and the constellation diagrams, the IMD (Intermodulation Distortion) and the ENOB (Effective Number of Bits) and, finally, the QoS, e.g., in terms of uncoded and coded BER and BLER curves versus the receive SNR and receive power, which allow the evaluation of the receiver sensitivity.

Figure 16.12. FALCON receiver sensitivity in the case of the UMTS/WCDMA 384 kbit/s service.

324

Alexander Vießmann et al.

As an example, the measurement results of the FALCON receiver sensitivity in the case of the UMTS/WCDMA 384 kbit/s service shall be discussed. The measurement was carried out in the laboratory of the authors, hence, the scenario can be regarded as a pico cellular environment. Since the deployed RF modules were tailor-made for IEEE 802.11b/g W-LAN applications and not for UMTS, the carrier frequency was set to 2.484 GHz. In order to generate representative transmission signals, a Rohde&Schwarz SMIQ was used instead of a second FALCON transceiver. The FALCON transceiver under test was acting as the receiver. Since both the Rohde&Schwarz SMIQ and the FALCON transceiver were located on a laboratory table, a low mobility case was realized, the mobile speed was hence 0 km/h. Figure 16.12 shows the measured error ratios versus the measured receive power 10 log10 (PRX /1mW). First, the uncoded BER at the output of the RAKE receiver was measured. Also, the effect of the PSCH (Primary Synchronization Channel) on the UMTS/WCDMA performance can be seen from Figure 16.12 when comparing the BER of the regular WCDMA signal with the BER, which is measured in only those signal parts not influenced by the PSCH. During the measurement, the PSCH and the regular traffic channels had equal transmit powers. It can be seen from Figure 16.12 that the effect of the PSCH on the performance is rather small in the case of an uncoded BER greater than 10−3 . In the case of lower values of the uncoded BER, the performance degrades considerably. Additionally, the coded BER and the coded BLER at the output of the TurboCode decoder were measured. The QoS targets were chosen to be 10−3 in the case of the coded BER and 10−2 in the case of the coded BLER. The latter corresponds to a throughput of 99% of the maximum transmission payload. It follows from Figure 16.12 that a coded BER of 10−3 is achieved at 10 log10 (PRX /1mW) of ≈−89.7 dBm which is equivalent to a receive power of about 1 pW. The coded BLER of 10−2 requires 10 log10 (PRX /1mW) of ≈ − 92.0 dBm which corresponds to a receive power of about 0.63 pW. Both receiver sensitivity values can be regarded as superb, taking into account that the RF modules were tailor-made for IEEE 802.11b/g W-LAN applications and not for UMTS. In summary, the prerequisite of scalability has been met in the PROMETHEUS platform and it could be demonstrated that transceivers supporting various wireless communication systems can be realized with the chosen approach. Furthermore, according to the above discussion, the PROMETHEUS platform is modular. All components used in the PROMETHEUS platform have been available and have a strong market penetration. Therefore, the authors believe that their PROMETHEUS platform fulfills the future proofness prerequisite discussed in Section 16.3.

16.6 Conclusions In this communication, the authors proposed a platform based design paradigm for the realization of cognitive terminals and infrastructure components. It has been illustrated that such an approach lends itself to fulfill several important prerequisites such as scalability, modularity, availability and market impact, leading to a favourable future proof design. The authors presented a new platform, termed PROMETHEUS, which they developed following the aforementioned design paradigm, showing that the design leads to transceivers with high quality and excellent performance.

16 Design of Terminals and Infrastructure Components

325

Acknowledgement. The authors gratefully acknowledge the support of the electronics workshop of the Universit¨ at Duisburg-Essen. Furthermore, the authors are indepted to their cooperation partners and sponsors, in particular, to Atmel Duisburg, SAMSUNG Electronics, Infineon Technologies, SK Telecom, Rohde&Schwarz, LeCroy, Texas Instruments, EPCOS, Analog Devices, and the Deutsche Forschungsgemeinschaft (DFG), for their continuing support. Furthermore, the authors would like to thank Dr. Thomas E. Faber and Dr. Tobias Scholand for their invaluable contributions to the research and development which enabled the realization of the PROMETHEUS platform.

References 1. Special Issue on Software Radio. IEEE Communications Magazine, 33, Issue 5 1995. 2. J. Bingham. Multicarrier Modulation for Data Transmission: An Idea whose Time has come. In IEEE Communications Magazine, vol. 28, pp. 5-14 1990. 3. Harri Holma and Antti Toskala. WCDMA for UMTS, Second Edition. ISBN 978-0470844670. John Wiley and Sons, October 2002. 4. Friedrich K. Jondral. Software-Defined Radio-Basics and Evolution to Cognitive Radio. In EURASIP Journal on Wireless Communications and Networking, vol. 3, pp. 275-283 2005. 5. J. G. Proakis. Digital Communications, Fourth Edition. ISBN 0-07-118183-0. McGraw–Hill, 2004. 6. Ralf Steinmetz and Klara Nahrstedt. Multimedia Systems. ISBN 978-3540408673. Springer, January 2004. 7. A. Vießmann, R. Franke, G. H. Bruck, and P. Jung. Petri net based controller concept for congnitive radios in wireless access networks. In Journal of Communications (JCM), invited contribution, to appear.

17 Fundamental Limits of Cognitive Radio Networks Survey of Recent Results on Information Theoretic Limits

Natasha Devroye and Vahid Tarokh Harvard University [ndevroye|vahid]@seas.harvard.edu Summary. Cognitive radios have the potential to greatly improve spectral efficiency in wireless networks. In this chapter we explore the fundamental limits of communication in channels employing cognitive radios. We take an informationtheoretic approach, making use of information theory’s wide variety of tools and ability to characterize communication limits independent of actual implementations. The chapter first surveys information theoretic results on a simple wireless channel where a primary link and a secondary (or cognitive) link share the same spectrum: the cognitive radio channel. Our survey includes recent capacity and achievable rate region calculations for the case when the channel is known to all transmitters and receivers. We compare the rates achieved through new non-orthogonal schemes where the cognitive and primary user simultaneously use the channel to more traditional spectral gap filling solutions. In the second part of the chapter we outline new results on the limits of cognitive radio channels where the fading coefficients are known to different degrees at the nodes.

17.1 Introduction The advent of cognitive radios, or software defined radios able to adapt to the sensed environment, along with recent FCC initiatives allowing for secondary spectrum access promise more flexible, and potentially more efficient spectrum access. There are many questions and aspects to be tackled before before cognitive radios can seamlessly and opportunistically employ spectrum licensed to primary user(s). Of both theoretical and practical importance is the question: what are the fundamental limits of communication in the presence of one or more cognitive radios? Information theory provides an ideal framework for analyzing this question. The capacity and rate regions achieved in a network with cognitive radios provide fundamental, unquestionable limits of the possible communication. Such results provide benchmarks for the communication field, where researchers may gauge the efficiency of any practical cognitive radio system. In this chapter, we outline some of the recent theoretical advances pertaining to the limits of cognitive radio systems, first assuming all channel gains are known to all involved nodes, and then extending to the case

327 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 327–351. c 2007 Springer.

328

Natasha Devroye and Vahid Tarokh

when fading coefficients are known to different extents at the involved wireless devices. In both cases we emphasize how the cognitive radios alter classical information theoretic communication scenarios, and what is gained by their introduction. The FCC’s recent Secondary Markets Initiative (SMI, [17]) was sparked by empirical measurements showing that most of the time certain licensed frequency bands remain unused, and the natural desire to remedy this and increase spectral efficiency. Currently, spectrum is either unlicensed, creating a spectral free-for-all (as, for example in the 2.4 GHz band), or is licensed to certain primary users (such as for example cellular providers). The goal of the SMI is to remove unnecessary regulatory barriers to new secondary market oriented policies. Of the multiple possible types of secondary leasing [16], in this chapter we will focus on dynamic spectrum leasing. There, licensed users (which we will use interchangeably with the term primary users) ultimately hold the right to the spectrum. However, the primary license holder may wish to re-distribute or share his spectrum with other devices not necessarily in his own network. The motivation, and fascinating game theoretical and economic models for doing so is beyond the scope of this chapter, and we refer the interested reader to [22, 39, 42]. In dynamic spectrum leasing, the non-licensed secondary devices would opportunistically (dynamically) employ the spectrum according to the primary licensee’s regulations. Three main types of opportunistic employment of the primary spectrum are possible (although by no means exhaustive). 1. Interference-controlled: the primary license holder could stipulate for example maximal permissible secondary user interference levels, in effect guaranteeing the primary users certain transmission rates. This could allow primary and secondary users to transmit in the same bands, that is, in the same timefrequency-space-code blocks. The concept of interference temperature [16] has been introduced with goal of avoiding the compromise of the primary users’ spectrum: secondary devices should control their emissions such that the aggregate interference at the primary users is below a certain level (or interference temperature). 2. Interference-avoiding: a subset of the interference-controlled regime is that in which the primary licensee only allows secondary users to use its spectrum on the condition that its user suffer no interference whatsoever. Such systems are much more restrictive than interference controlled systems, but are common in cognitive radio literature [22, 25, 43, 49]. The secondary user could adhere to this strict requirement by filling in spectral holes. That is, a secondary user would transmit only in the absence of primary users. 3. Interference-free: when cognitive devices exist in a network but have no information of their own to transmit, they could potentially act as relays, and collaborate with the primary users [38]. Rather than cause interference to the primary link, they boost it. Neglecting any other possibly active cognitive clusters [15], this system is interference-free. In this work, in order to obtain fundamental limits of communication in the presence of cognitive radios, we turn our attention mostly to the more general interferencecontrolled model. For both the interference-controlled as well as the interference-avoiding models, the secondary user must be able to perform some fundamental tasks. Specifically, we require a device which is able to sense the communication opportunities, and then

17 Fundamental Limits of Cognitive Radio Networks

329

take actions based on the sensed information. In this chapter, such actions will include transmitting (or refraining from transmitting) and adapting their modulation and/or coding strategies so as to “better” employ the sensed spectral environment. For the most part, “better” will mean larger rates. Cognitive radios, a term coined by Mitola [36] are special types of software-defined radios (SDR), and are natural candidates for secondary users. SDRs are wireless communication devices with the ability to transmit and receive using a variety of protocols and modulation schemes (enabled by reconfigurable software rather than hardware). Cognitive radios are SDRs which can furthermore become “cognitive”, and, as dictated by the software, adapt their behavior to the wireless surroundings without user intervention. Such radios can make decisions based on the availability of nearby collaborative nodes, or on the regulations dictated by their current location and/or spectral conditions. In this chapter, secondary users will be assumed to have abilities similar to cognitive radios. We use the terms secondary user and cognitive user interchangeably.

17.1.1 Chapter Outline The theoretical limits of cognitive radio systems are still being explored. As such, most models considered thus far are fairly simplified. In this chapter, we focus on wireless networks in which only a single primary user-primary receiver pair is present. First, in Section 17.2 we will explore achievable rate and capacity regions in which a single cognitive, or secondary, user is present. We define the cognitive radio channel (also known in the information theory community as the interference channel with degraded message sets) and review results on the achievable and capacity regions for both the discrete channel and Gaussian noise channel. For the first part, the channel parameters are assumed to be known to all involved nodes. In the second part of this chapter, we survey recent results on information theoretic limits of cognitive channels in which the fading coefficients are known to different degrees at the four nodes.

17.2 Fundamental Limits of Cognitive Radio Channels: Perfect CSI Consider first the simplest scenario in which a primary user and a secondary user co-exist in the same wireless channel, as shown in Figure 17.1. The primary (sender, receiver) pair is denoted by (S1 , R1 ), while the cognitive (sender, receiver) pair is denoted by (S2 , R2 ). Each sender wishes to transmit its own independent information to its receiver. We wish to determine the fundamental limits of the communication possible, as a function of both the wireless channel, and the transmitter and receiver capabilities (such as power constraints, and the cognitive nature of S2 ). We assume perfect channel state information (CSI), that is, all parameters characterizing the channel are assumed to be known perfectly to the senders and receivers. Before turning to definitions of variables and rates in this channel, we look qualitatively at different possible communication scenarios in this simple channel. The amount of data that can be transmitted from the senders S1 , S2 to the receivers R1 , R2 naturally depends on the amount of cooperation possible between transmitters and receivers. In this chapter, we consider different possibilities for

330

Natasha Devroye and Vahid Tarokh

Figure 17.1. A simple channel in which the primary transmitter S1 wishes to transmit a message to the primary receiver R1 and the secondary (or cognitive) transmitter S2 wishes to transmit a message to its receiver R2 . We explore the rates R1 and R2 that are achievable in this channel. transmitter cooperation1 , and so three categories of transmitter cooperation may be as shown in Figure 17.2, and can be summarized as:

Figure 17.2. (a) Competitive behavior, the interference channel. The transmitters may not cooperate. (b) Cognitive behavior, the cognitive radio channel. Asymmetric transmitter cooperation. (c) Cooperative behavior, the two antenna broadcast channel. The transmitters, but not the receivers, may fully and symmetrically cooperate.

1. Competitive behavior: The two transmitters transmit independent messages. There is no cooperation in sending the messages, and thus the two users compete for the channel. This is the same channel as the 2 sender, 2 receiver interference channel. This channel, introduced in [1,45] was consequently studied by, among many others, [5, 6, 21, 34, 40, 41]. Although the capacity region of this channel is known in a few cases, its capacity in its most general setting remains an open problem. 1

Transmitter cooperation seems to be of greater theoretical interest [26] than receiver cooperation. The case for cognitive receivers is also interesting and could form a nice topic for future research.

17 Fundamental Limits of Cognitive Radio Networks

331

2. Cognitive behavior: In this channel, the one way double arrow from S1 to S2 indicates asymmetric cooperation between the transmitters. This asymmetric cooperation is a result of S2 knowing S1 ’s message, but not vice-versa. We idealize the concept of message knowledge: whenever S2 is able to hear and decode the message of S1 , we assume it has full a-priori, or non-causal knowledge. We use the term cognitive behavior to emphasize the need for S2 to be a “smart” device capable of altering its transmission strategy according to the message of the primary user. We can motivate considering asymmetric side information in practice in three ways: • Cognitive networks: depending on the device capabilities, as well as the geometry and channel gains between the various nodes, certain cognitive nodes may be able to hear and/or obtain the messages to be transmitted by other nodes. These messages would need to be obtained in real time, and could exploit the geometric gains between cooperating transmitters relative to receivers in, for example, a 2 phase protocol [14]. • Automatic Repeat reQuest (ARQ) system: a cognitive transmitter, under suitable channel conditions (if it has a better channel to the primary transmitting node than the primary receiver), could decode the primary user’s transmitted message during an initial transmission attempt. In the event that the primary receiver was not able to correctly decode the message, and it must be re-transmitted, the cognitive user would already have the tobe-transmitted message, or asymmetric side information, at no extra cost in terms of overhead in obtaining the message. • Heterogeneous sensors: consider a network of wireless sensors in which a sensor S2 has a better sensing capability than another sensor S1 and thus is able to sense two events, while S1 is only able to sense one. Thus, when they wish to transmit, they must do so under an asymmetric side-information assumption: sensor S2 has two messages, and the other has just one. This scenario is considered in [48]. 3. Cooperative behavior: The two transmitters know each others’ messages (two way double arrows) and can thus fully and symmetrically cooperate in their transmission. The channel pictured in Figure 17.2 (c) may be thought of as a two antenna sender, two single antenna receivers broadcast channel [46]. The capacity region of the general broadcast channel is still unknown, save for certain cases [8, 9, 18, 32]. Of course, many achievable regions have been developed; the largest to date was computed in [31]. In Gaussian noise, much more can be said about the broadcast channel. The capacity region of a broadcast channel with single antennas coincides with the region of a degraded broadcast channel and is a classical result, outlined in Section 14.6.3 of [9]. In contrast, the capacity region of the Gaussian MIMO broadcast channel was recently found in the most nontrivial result of [46]. The gain of dirty-paper coding over time division multiple access (TDMA, the users time share the channel) when broadcasting information from a single base station to multiple users is explored in [27]. There, the authors find the sum-rate of broadcasting using the optimal dirty-paper coding strategy is at most min(number transmit antennas, number receivers) that of TDMA. When S2 is a cognitive radio and S1 is not, both the competitive and cognitive behaviors seem plausible. Cooperative behavior would require the primary and secondary transmitters to cooperate, a situation which is unlikely in the event that

332

Natasha Devroye and Vahid Tarokh

the primary user is non-cognitive in nature. In the context of secondary markets, it is also more reasonable to assume that primary users should be able to continue transmitting in the same way whether secondary users are present or not. This rules out the cooperative behavior. We will however make use of cooperative behavior as an outer bound, and interesting comparison, to what can be hoped to be achieved by channels employing cognitive radios. The cooperative behavior corresponds to the classical broadcast channel, while the competitive behavior reduces to the classical interference channel. We thus turn our attention to the much less studied behavior which spans and in a sense interpolates between the symmetric cooperative and competitive behaviors. We call this behavior asymmetric cognitive behavior. In this section we will consider one example of cognitive behavior: a two sender, two receiver (with two independent messages) interference channel with asymmetric, non-causal message knowledge at one of the transmitters, as shown in Figure 17.2(b). Certain asymmetric (in transmitter cooperation) channels have been considered in the literature: for example in [44], the capacity region of a multiple access channel with asymmetric cooperation between the two transmitters is computed. The authors in [24] consider a channel which could involve asymmetric transmitter cooperation, and explore the conditions under which the capacity of this channel coincides with the capacity of the channel in which both messages are decoded at both receivers. In [12, 13] the authors introduced the cognitive radio channel, which captures the most basic form of asymmetric transmitter cooperation for the interference channel. This same model was subsequently studied in [28], where the capacity region for the Gaussian cognitive radio channel under weak interference is found, as well as [48], with an analogous result for the discrete memoryless case. The fundamental way a cognitive transmitter differs from an interference channel is the presence of partial transmitter side-information. This is, intuitively speaking, where all gains are expected to come from, as it allows the cognitive transmitter to mitigate, or cancel interference from the primary user. As argued before, a cognitive radio, before it encodes and transmits its information, possesses both the primary user’s message (we assume it also knows this primary user’s encoding of the message) as well as its own message. Thus, if the cognitive radio were to simultaneously transmit with the primary user, it would know what interference (the primary user signal) its receiver would suffer. Note that in order to fully know the interference at the cognitive receiver, the cognitive transmitter also needs to know the channel between S1 and R2 .2 All nodes possess the conditional distribution p(y1 , y2 |x1 , x2 ) which fully describes the wireless channel. The channel capacity of a channel with input X, non-causal transmitter side information S, output Y , is given by the wellknown formula obtained by Gel’fand and Pinsker [19] as C = max [I(U ; Y ) − I(U ; S)] ,

(17.1)

p(u,x|s)

where U is an auxiliary random variable chosen to make the channel U → Y appear causal. We refer to the coding technique used in [19] as Gel’fand-Pinkser coding. By applying Gel’fand-Pinsker’s result to the Gaussian noise and interference case, Costa [7] achieves the capacity of an interference-free channel by careful selection 2

In all results for the single user scenario, we assume all nodes S1 , S2 , R1 , and R2 know each others’ channels (full CSI assumption).

17 Fundamental Limits of Cognitive Radio Networks

333

of X, U . That is, when the input X to the channel is Gaussian, and the auxiliary variable U is of the form U = X + αS for some parameter α whose optimal value is equal to the ratio of the signal power to the signal plus noise power, the interference S is completely mitigated, and the capacity is shown to be equal to that of an interference-free channel! Dirty paper coding is the term first used by Costa [7] to describe the aforementioned technique which completely mitigates a-priori known interference over an input power constrained additive white Gaussian noise channel. Many authors use the dirty-paper coding, or Gel’fand-Pinsker coding technique in channels where non-causal side information is present at the transmitters. The power of this technique was recently demonstrated in the capacity region calculation of the MIMO Gaussian broadcast channel [46], where the achievable dirty-paper coding region is shown to be capacity-achieving. We now proceed to find achievable rate, and in some channel scenarios, capacity, regions of cognitive radio channels. In order to build intuition, we start off by paraphrasing the work of [28], in which the capacity region of a channel of the form of Figure 17.2(b) with Gaussian noise is found for weak interference. The Gaussian region is particularly intuitive as we can evaluate the region and demonstrate the rate regions obtained graphically. This allows us to compare cognitive regions with the classical competitive and cooperative regions. We then extend results to the more abstratc discrete memoryless channel case, where we highlight the results of [14] and [48].

17.2.1 Gaussian Noise We define a 2 × 2 cognitive radio channel CCOG , also known as an interference channel with degraded message sets (IC-DMS) in [28, 48] as in Figure 17.3, to be two point-to-point channels S1 → R1 and S2 → R2 in which the sender S2 is given, in a non-causal manner (i.e., by a genie), the encoded message X1 which the sender S1 will transmit. Thus, S1 is the primary user, and S2 is a secondary, cognitive user. Let X1 and X2 be the random variable inputs to the channel, and let Y1 and Y2 be the random variable outputs of the channel. The conditional probabilities of the channel CCOG are fully described by P (y1 |x1 , x2 ) and P (y2 |x1 , x2 ). We first consider the case when these conditional distributions relate the input and output as in Eqns. (17.3), shown in Figure 17.3(a)3 X1 Y1 1 a21 N1 = + (17.2) Y2 a12 1 X2 N2 N1 ∼ N (0, σ12 ),

N2 ∼ N (0, σ22 )

(17.3)

An (n, K1 , K2 , ) code for the cognitive radio channel consists of K1 codewords n n n xn 1 (i) ∈ X1 for S1 , and K1 · K2 codewords x2 (i, j) ∈ X2 for S2 , i ∈ {1, 2, . . . , K1 }, j ∈ {1, 2, . . . , K2 }, which together form the codebook, revealed to both senders and 3

notice that we have assumed the channel between (S1 , R1 ), as well as (S2 , R2 ) are all unit. This can be assumed WLOG by multiplying the entire receive chain at R1 by any (non-infinite) 1/a211 , and the receive chain at R2 by 1/a222 without altering the achievable and/or capacity results.

334

Natasha Devroye and Vahid Tarokh

receivers such that the average error probabilities under some decoding scheme are less than . A rate pair (R1 , R2 ) is said to be achievable for the cognitive radio channel if there exists a sequence of (n, 2dnR1 e , 2dnR2 e , n ) codes such that n → 0 as n → ∞. An achievable region is a closed subset of the positive quadrant of R2 of achievable rate pairs. The capacity region is the closure of the set of all achievable rate pairs (R1 , R2 ). We refer the interested reader to [9,14,28] for further details and subtleties regarding these definitions. We will drop the block length index n when contextually clear.

Figure 17.3. (a) The Gaussian cognitive radio channel. (b) The discrete memoryless cognitive radio channel.

We first consider the results of [28] on the Gaussian interference channel with degraded message sets (IC-DMS or equivalently the Gaussian cognitive radio channel). The authors of this work are particularly interested in determining the maximal rate at which the secondary cognitive user may transmit such that the primary user’s rate remains unchanged (that is, the primary user’s rate continues to be the same as if there were no interference). However, the authors not only obtain this single point in the capacity region, but rather the entire for weak interference (a21 < 1). They require the primary receiver to employ a single-user decoder, which would be the case if no cognitive user were present. In essence, these two conditions, which they term co-existence conditions, require the cognitive user to remain transparent to the primary user. Of particular interest is that in the proof of the capacity region, the co-existence conditions are relaxed (allowing for joint codebook design between primary and secondary users), and the authors show that the capacity achieving coding/decoding scheme in fact satisfy these co-existence conditions, that is, that the primary user decoder behaves as a single user decoder. The main results of [28] stated in their Theorems 3.1 and 4.1 are summarized in the following single theorem. Here the primary user is expected power limited to P1 , the secondary user is expected power limited to P2 , and the noises at the two receivers are Gaussian of zero mean and variance N1 and N2 respectively. Theorem 1. The capacity region of the IC-DMS defined in (17.2) is given by the union, over all α ∈ [0, 1], of the rate regions √ √ P1 +a21 αP2 )2 0 ≤ R1 ≤ 12 log2 1 + ( 1+a 2 (1−α)P 0 ≤ R2 ≤

1 2

21

log2 (1 + (1 − α)P2 )

2

17 Fundamental Limits of Cognitive Radio Networks

335

In particular, the maximal rate R2 (or capacity) at which a cognitive user may transmit such that the primary user’s rate R1 remains as in the interference-free regime (R1 = 12 log2 (1 + P1 /N )) is given by R2 =

1 log2 (1 + (1 − a∗ )P2 ) 2

as long as a21 < 1,and a∗ is 1  √ p 2 P1 1 + a221 P2 (1 + P1 ) − 1 ∗   √ . a = a21 P2 (1 + P1 )

Both these results are obtained using a Gaussian encoder at both the primary and cognitive transmitters. For more precise definitions of achievability in this channel, we refer to [28]. We paraphrase their achievability results here. The primary user generates its 2nR1 codewords, X1n (block length n), by drawing the coordinates i.i.d. according to N (0, P1 ), where we recall P1 is the expected noise power constraint. Then, since the cognitive radio knows the message the primary user, it can form the primary user’s encoding X1n , and performs superposition coding as: r αP2 n n n ˆ X 2 = X2 + X1 , P1 ˆ 2n encodes one of the 2nR2 messages, and is genwhere α ∈ [0, 1]. The codeword X erated by performing Costa precoding [7] (dirty-paper coding). Costa showed that ˆ 2n statistito optimize the rate achieved by this dirty-paper coding, one selects X n cally independently from X1 , and thus i.i.d. Gaussian. Encoding is done using a standard information theoretic binning technique, which treats the message X1n as non-causally known interference. In order to satisfy the average power constraint ˆ 2n must be N (0, (1 − α)P2 ). The parameter α of P2 on the components of X2n , X allows for a trade-off at the cognitive transmitter between aiding the primary transmitter (α close to 1) or transmitting, using a dirty paper coding technique, its own message (α close to 0). A converse, resulting in the capacity region of the cognitive radio channel under weak interference, is given in [28] and is based on the conditional entropy power inequality, and results from [46]. We illustrate this region for three different values of the channel parameters a12 , a21 , and compare it to the Gaussian MIMO broadcast channel region (in which the two transmitters may cooperate), the achievable rate region for the interference channel region obtained in [21] (the largest known to date for the Gaussian noise case) where the two transmitters must operate independently, and the time-sharing region where the two transmitters take turns using the channel, thus creating two independent channels and avoiding any interference. The capacity region for the Gaussian MIMO broadcast channel with two single antenna receivers and one transmitter with two antennas subject to per antenna power constraints of P1 and P2 respectively, is given by Eqn. (17.4), which may be obtained from the general formulation in [4,46].

336

Natasha Devroye and Vahid Tarokh Convex hull {(R1 , R2 ) : R1 ≤ R2 ≤ R2 ≤

H1 (B1 +B2 )H1T +Q1 T H1 BT2 H1 +Q 1 H2 B2 H2 +Q2 1 log 2 2 Q2 1 2

1 2

log2

log2 1 +

P2 Q2

S

R1 ≤

1 2

R2 ≤

1 2

R2 ≤

1 2

H1 B1 H1t +Q1 Q1 H2 (B1 +B2 )H2T +Q2 log2 T +Q H B H 2 1 2 2 P2 log2 1 + Q 2

log2

for any 2x2 matrices B1 , B2 such that B2 0 B1 0, P1 c B1 + B2 c P2 c2 ≤ P 1 P 2 } (17.4) Here X 0 denotes that the matrix X is positive semi-definite, and we defined H1 = [1 a21 ] and H2 = [a12 1]. The achievable rate region of [21] used in these figures (as the “interference channel” achievable region) assumes the same Gaussian input distribution as in [14] and is omitted for brevity. The time sharing region for the interference channel (which requires coordination between the two transmitters, but only at the time sharing, and not coding level) is given by the region of Eqn. (17.5).     [ P1 1 P2 1 4 Time-share region = , (1 − α) log2 1 + ) (α log2 1 +  2 N1 2 N2  0≤α≤1

(17.5) From the Figure 17.4,17.5 and 17.6 we see that both users – not only the incumbent S2 which has the extra message knowledge – benefit from behaving in a cognitive, rather than simple time-sharing manner. This is as expected, as the decreasing α boosts R2 rates, while increasing α (of Theorem 1) boosts R1 rates, and so gracefully combining the two will yield benefits to both users. Time-sharing is in essence what would be theoretically achievable in spectral-gap filling models for cognitive radio channels. That is, under the assumption that an incumbent cognitive were to perfectly sense the gaps in the spectrum, and fill them by transmitting at the capacity of the point-to-point channel between (S2 , R2 ), the best rate region one can hope to achieve is the time-sharing rate region. Where we operate on the boundary of this region depends entirely on what percentage of the time the primary user is on the channel (or the α parameter of (17.5)). The rates achieved by the cognitive user thus depend on the primary user’s channel utilization. It is important to note that in the region and coding scheme described in Theorem 1 that the cognitive user’s choice of the power-sharing parameter α in essence determines where on the boundary of the cognitive IC-DMS region we operate. Thus, if such a scheme were to be implemented, one would need to ensure that cognitive radios would not select an α that would be too detrimental to the rate of the primary user. Nonetheless, theoretically, the presence of the incumbent cognitive radio S2 can be beneficial to S1 , and could provide incentives for the introduction of such schemes.

17 Fundamental Limits of Cognitive Radio Networks

337

Achievable rate regions at SNR 10, a21=0.8, a12=0.2 2 MIMO broadcast channel Cognitive channel Interference channel Time−sharing

1.8 1.6 1.4

R2

1.2 1 0.8 0.6 0.4 0.2 0

0

0.5

1

1.5

2

2.5

3

R1 Figure 17.4. Capacity region of the Gaussian 2 × 1 MIMO two receiver broadcast channel (outer), cognitive channel (middle), achievable region of the interference channel (second smallest) and time-sharing (innermost) region for Gaussian noise powers N1 = N2 = 1, power constraints P1 = P2 = 10 at the two transmitters, and channel parameters a12 = 0.8, a21 = 0.2.

17.2.2 Discrete Memoryless Channel In the previous subsection, we outlined the capacity region for the Gaussian cognitive radio channel (or IC-DMS) in the weak interference (a21 < 1) regime. Gaussian noise channels have in general been more well studied and understood than general discrete channel models (where channel inputs and outputs take on a set of discrete values) and lend themselves well to intuition and numerical calculation. In this section we outline the current state-of-the-art for discrete cognitive channels, highlighting results from both [48] and [12, 14]. We first look at the achievable rate region defined in [14] and extended in [12]. The channel model is shown in Figure 17.3(b) where the inputs X1 , X2 to the channel and the outputs Y1 , Y2 of the channel are now discrete, and the conditional distribution p(y1 , y2 |x1 , x2 ) which characterizes the channel is discrete rather than Gaussian. Since this channel resembles an interference channel with non-causal transmit side information at one of the encoders, the authors of [14] derive an achievable rate region by combining the best to date known achievable rate region for the interference channel, that of [21], with the results for non-causal transmit side information, Gel’fand-Pinsker coding [19]. We briefly outline the achievable rate region of [12,14]. In [21], an achievable region for the interference channel is found by first considering a modified problem and then establishing a correspondence between the achievable rates of the modified and the original channel models. We proceed in the m same fashion. The channel CCOG , defined as in Figure 17.7 introduces many new auxiliary random variables, whose purposes can be made intuitively clear by relating

338

Natasha Devroye and Vahid Tarokh

Achievable rate regions at SNR 10, a21=a12=0.55 2.5 MIMO broadcast channel Cognitive channel Interference channel Time−sharing

2

R2

1.5

1

0.5

0

0

0.5

1

1.5

2

2.5

R1 Figure 17.5. Capacity region of the Gaussian 2 × 1 MIMO two receiver broadcast channel (outer), cognitive channel (middle), achievable region of the interference channel (second smallest) and time-sharing (innermost) region for Gaussian noise powers N1 = N2 = 1, power constraints P1 = P2 = 10 at the two transmitters, and channel parameters a12 = 0.55, a21 = 0.55.

them to auxiliary random variables in previously studied channels. They are defined and described in Table 17.1. Standard definitions of achievable rates and regions are employed [9, 13] and omitted for brevity. Then an achievable region for the discrete cognitive radio channel is given by: 4

Theorem 2. Let Z =(Y1 ,Y2 ,X1 ,X2 ,V11 ,V12 ,V21 ,V22 ,W ), be as shown in Figure 17.7. Let P be the set of distributions on Z that can be decomposed into the form P (w) × [P (m11 |w)P (m12 |w)P (x1 |m11 , m12 , w)] × [P (a11 |m11 , w)P (a12 |m12 , w)] × [P (m21 |v11 , v12 , w)P (m22 |v11 , v12 , w)] × [P (x2 |m21 , m22 , a11 , a12 , w)] P (y1 |x1 , x2 )P (y2 |x1 , x2 ),

(17.6) 4

where P (y1 |x1 , x2 ) and P (y2 |x1 , x2 ) are fixed by the channel. Let T1 = {11, 12, 21} 4

and T2 = {12, 21, 22}. For any Z ∈ P, let S(Z) be the set of all rate tuples (R11 , R12 , R21 , R22 ) of non-negative real numbers such that there exist non-negative reals L11 , L12 , L21 , L22 satisfying:

17 Fundamental Limits of Cognitive Radio Networks

339

Achievable rate regions at SNR 10, a21=0.2, a12=0.8 3 MIMO broadcast channel Cognitive channel Interference channel Time−sharing

2.5

R2

2

1.5

1

0.5

0

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

R1 Figure 17.6. Capacity region of the Gaussian 2 × 1 MIMO two receiver broadcast channel (outer), cognitive channel (middle), achievable region of the interference channel (second smallest) and time-sharing (innermost) region for Gaussian noise powers N1 = N2 = 1, power constraints P1 = P2 = 10 at the two transmitters, and channel parameters a12 = 0.2, a21 = 0.8.

S1 M

11

M

12

A11 A12

M

V

V

11

V

12

21

21

M V

22

22

R1

X1

Y1

X2

Y2

S2

R2

Figure 17.7. The modified cognitive radio channel with auxiliary random variables M11 , M12 and M21 , M22 , inputs X1 and X2 , and outputs Y1 and Y2 . The auxiliary random variable A11 , A12 associated with S2 , aids in the transmission of M11 and M12 respectively. The vectors V11 , V12 , V21 and V22 denote the effective random variables encoding the transmission of the private and public messages.

340

Natasha Devroye and Vahid Tarokh Table 17.1. Description of random variables and rates in Theorem 2. (Random) variable names M11 , M22

V11

(Random) variable descriptions Private information from S1 → R1 and S2 → R2 resp. M12 , M21 Public information from S1 → (R1 , R2 ) and S2 → (R1 , R2 ) resp. R11 , R22 Rate between S1 → R1 and S2 → R2 resp. R12 , R21 Rate between S1 → (R1 , R2 ) and S2 → (R1 , R2 ) resp. A11 , A12 Variables at S2 that aid in transmitting M11 , M12 resp. = (M11 , A11 ), V12 = (M12 , A12 ) Vector helping transmit the private/public (resp.) information of S1 V21 = M21 , V22 = M22 Public and private message of S2 . Also the auxiliary random variables for Gel’fand-Pinsker coding W Time-sharing random variable, independent of messages

! \

X

T ⊂{11,12}

t∈T

≤ I(X1 ; MT |MT )

Rt

R11 = L11 R12 = L12

\

X

T ⊂T1

t1 ∈T

\

X

T ⊂T2

t2 ∈T

(17.7) (17.8) (17.9)

R21 ≤ L21 − I(V21 ; V11 , V12 )

(17.10)

R22 ≤ L22 − I(V22 ; V11 , V12 ) !

(17.11)

Lt1

≤ I(Y1 , VT ; VT |W ) + f (VT |W )

(17.12)

≤ I(Y2 , VT ; VT |W ) + f (VT |W ),

(17.13)

! Lt2

where f (vT ) denotes the divergence between the joint distribution of the random variables VT in (17.6) and their product distribution (where all components are independent). T denotes the complement of the subset T with respect to T1 in (17.12), with respect to T2 in (17.13), and VT denotes the vector of Vi such that i ∈ T . Let S be the closure of ∪Z∈P S(Z). Then any pair (R11 + R12 , R21 + R22 ) for which (R11 , R12 , R21 , R22 ) ∈ S is achievable for CCOG . Proof outline: The main intuition is as follows: the equations in (17.7) ensure that when S2 is presented with X1 by the genie, the auxiliary variables M11 and M12 can be recovered. Eqs. (17.12) and (17.13) correspond to the equations for two overlapping MAC channels seen between the effective random variables VT1 → R1 , and VT2 → R2 . Eqs. (17.10) and (17.11) are necessary for the Gel’fand-Pinsker [19] coding scheme to work (I(V21 ; V11 , V12 ) and I(V22 ; V11 , V12 ) are the penalties for using non-causal side information). The f (VT ) terms correspond to the highly unlikely events of certain variables being correctly decoded despite others being in

17 Fundamental Limits of Cognitive Radio Networks

341

error. Intuitively, the sender S2 could aid in transmitting the message of S1 (the A11 , A12 random variables) or it could dirty paper code against the interference it will see (the M21 , M22 variables). The theorem smoothly interpolates between these two options. Details may be found in [12, 14]. The work [48] also considers the discrete memoryless IC-DMS (or discrete cognitive radio channel), and looks at the Gaussian IC-DMS as a special case. The authors in this work are motivated by a sensor network in which one sensor has better sensing capabilities than another. The one with the better channel is thus able to detect two sensed events, while another is only able to detect one. This problem then reduces to the interference channel with degraded message sets (where the message of one user is a subset of the other user’s message). The authors define three types of weak interference (as opposed to the very strong and strong interference typically seen in the interference channel literature [5]), an achievable rate region, outer bounds, and conditions under which these outer bounds are tight. They then look at a Gaussian noise example in which their region is tight, and for which the result is as described in the capacity region of [28]. We summarize some of their main results in the single following theorem. It provides an inner and an outer bound on the discrete IC-DMS, which turns out to be the capacity region for the types of interference specified. Theorem 3. Inner bound: Let Rin be the set of all rate pairs (R1 , R2 ) (same as in the cognitive radio channel) such that R1 ≤ I(V, X1 ; Y1 ) R2 ≤ I(U ; Y2 ) − I(U ; V, X1 ) for the probability distribution p(x1 , x2 , u, v, y1 , y2 ) that factors as p(v, x1 )p(u|v, x1 )p(x2 |u)p(y1 , y2 |x1 , x2 ). Then Rin is an achievable rate region for the IC-DMS where transmitter S2 knows both messages and transmitter S1 only knows one. Outer bound: Define Ro to be the set of all rate pairs (R1 , R2 ) such that R1 ≤ I(V, X1 ; Y1 ) R2 ≤ I(X1 ; Y2 |X1 ) R1 + R2 ≤ I(V, X1 ; Y1 ) + I(X2 ; Y2 |V, X1 ), for the probability distribution p(x1 , x2 , v, y1 , y2 ) that factors as p(v, x1 )p(x2 |v)p(y1 , y2 |x1 , x2 ). Then Ro is an outer bound for the capacity of the IC-DMS. Capacity conditions: If there exists a probability transition matrix q1 (y2 |x2 , y1 ) such that X p(y2 |x1 , x2 ) = p(y1 |x1 , x2 )q1 (y2 |x2 , y1 ), y1

or if there exists a probability transition matrix q2 (y1 |x1 , y2 ) such that X p(y2 |x1 , x2 )q2 (y1 |x1 , y2 ), p(y1 |x1 , x2 ) = y2

342

Natasha Devroye and Vahid Tarokh

then the set of all rate pairs (R1 , R2 ) such that R1 ≤ I(V, X1 ; Y1 )

(17.14)

R2 ≤ I(X2 ; Y2 |V, X1 )

(17.15)

for the probability distribution p(x1 , x2 , y1 , y2 ) that factors as p(v, x1 )p(x2 |v)p(y1 , y2 |x1 , x2 ), is the capacity region of the IC-DMS. Since the channel of [48] is the same as the cognitive radio channel [12, 14], direct comparisons between their resepective bounds may be made. Whereas the outer bounds are equivalent, due to the fact that the inner bounds for the discrete memoryless channel involve non-trivial unions over all distributions of a certain form, it is unclear a priori which region is more general. However, the authors demonstrate that all Gaussian weak interference channels satisfy the capacity conditions of the theorem, and thus the region of (17.14)-(17.15) is the capacity region. This capacity region in the Gaussian noise case is shown to be explicitly equal to that of [28], and, numerically, to that of of [12], specialized to the Gaussian noise case.

17.2.3 Further Results Numerous other works have considered fundamental limits of systems involving cognitive radios. The work [24] considers again the cognitive radio channel, referred to as the interference channel with unidirectional cooperation. There, one set of conditions for which the capacity region of the channel coincides with that of the channel in which both messages are required at both receivers is derived. Notice that in the cognitive radio channel this added condition, of being able to decode both messages at both receivers, is not assumed. This is related to the work [23] on the compound multiple access channel with common information, in which the capacity region for another set of strong interference-type conditions is computed. Notice that whereas [48] considers weak interference conditions, [24] considers strong interference conditions. Their results on the cognitive radio channel capacity read as follows: Theorem 4. For an interference channel with unidirectional cooperation satisfying I(X2 ; Y2 |X1 ) ≤ I(X2 ; Y1 |X1 ) I(X1 , X2 ; Y1 ) ≤ I(X1 , X2 ; Y2 ) for all joint distributions on X1 and X2 , the capacity region C is given by [ R2 ≤ I(X2 ; Y2 |X1 ) , C= (R1 , R2 ) : R1 + R2 ≤ I(X1 , X2 ; Y1 ) where the union is over joint distributions p(x1 , x2 , y1 , y2 ). The case of causal (rather than non-causal) message knowledge is considered in [14], where various protocols for which S2 causally obtains the message of S1 are presented. In [12] the simple 2 × 2 cognitive radio channel is extended to the

17 Fundamental Limits of Cognitive Radio Networks

343

multiple access scenario. That is, the authors consider a channel in which a cognitive (q → 1) multiple access channel has non-causal knowledge of the messages of the primary (p → 1) multiple access channel. In a similar vein, the work [33] considers a multiple access channel with channel state known non-causally at one one of the encoders. Inner and outer bounds are obtained for this channel, with the interesting conclusion that when channel state information is present at only one of the encoders, full interference mitigation (achieved in Costa’s dirty-paper coding) is not possible. This is in sharp contrast to the complete state (or interference) mitigation that is possible in both dirty paper coding for a point-to-point channel [7] as well as generalized dirty paper coding in a multiple access channel in which the state is known at both encoders [30]. This channel differs of course from that of [12] as only a single multiple access channel is considered. Although up until now we have emphasized the gains to be made when a primary user and a secondary cognitive user simultaneously transmit, a simpler strategy by which cognitive radios may improve spectral efficiency is by sensing and filling in spectral gaps. The achievable rate region for this spectral gap filling region, when a single frequency band is shared over time, is given by the timesharing regions of Figure 17.4, 17.5, 17.6. The work in [25] and [43] addresses issues involved in the opportunistic sensing of and communication over such spectral holes. Notice that the time-sharing regions correspond to temporal holes rather than frequency holes. In [25, 43] capacity inner and outer bounds for a cognitive transmitter-receiver pair acting as secondary users in a network of primary users are derived. According to the authors, the capacity is limited by the distributed and dynamic nature of the spectral activity which these cognitive radios wish to exploit. The authors use the term distributed to denote the different views of local spectral activity at the cognitive transmitter and receiver. In addition to the spectrum availability being locationdependent, it will also vary with time, depending on the data that must be sent at different moments. The authors use the term dynamic to indicate the temporal variation of the spectral activity of the primary users. We refer the interested reader to [25, 43] for further details. Cognitive radios could also serve as relays in a network. In this section, we have highlighted results for when a cognitive radio has its own information to transmit. However, when no such information is at hand, it could act as a relay and aid a primary user in transmitting its message. For this type of communication, information theoretic limits can be found in the literature pertaining to the classical relay channel. The relay channel, which in its simplest and most classical form is a three-terminal channel with one source, one relay (without its own information to transmit) and one destination, were introduced by van der Meulen [44], and various variations of the problem were later studied by others [2], [10]. The current state of the art is well summarized in [35]. Three major issues are ignored in the classical relay channel framework: the half-duplex constraint of most practical wireless systems, the compound nature, and the non-degraded nature of most wireless channels. Some of these issues are addressed in the collaborative communications framework of [38]. We have so far looked at information theoretic limits of channels with cognitive radios under the assumption that all nodes has full CSI. In this scenario, cognitive radios were beneficial in terms of overall achievable rates due to their ability to mitigate the primary user’s interference (to the cognitive receiver), or alternatively to strengthen it. The ability to mitigate interference is, qualitatively, dependent on

344

Natasha Devroye and Vahid Tarokh

having full CSI at the transmitters. In the next section, we explore the benefits of cognitive radios in fading channel, under varying assumptions of the CSI.

17.3 Fundamental Limits of Cognitive Radio Channels: Imperfect CSI and Fading Channels The information-theoretic study of fading channels is by no means a new field. The traditional notion of capacity is extended to include concepts such as the ergodic capacity, the distribution of capacity (leading to “capacity-versus-outage” results), delay-limited capacity, and compound channels, to mention a few popular measures. The large number of different perspectives on fading channels are a consequence not only of the different underlying fading models assumed, but also the different amounts, and types (noisy, perfect, distribution only) of channel state information (CSI) available to the transmitter(s) and/or receiver(s). For example, when CSI, modeled as a time-varying random process, is known perfectly to all transmitters and receivers, but varies over time in an ergodic fashion, then the transmitter may continually adjust its rate to the current channel conditions. The ergodic capacity (region) is thus defined and obtained, roughly speaking, by taking the expectation with respect to the channel parameter distribution, of the capacity (region). On the other hand, if the transmitter has no CSI, and one demands error-free transmission at all times, then the achievable rates are significantly reduced, roughly to the rates sustainable by the “worst” channel conditions. However, if, for a certain rate R one allows for channel outages with probability (during which reliable communication at the rate R is impossible) then we can define the outage capacity C as the maximal rate which can be achieved such that the probability of outage is less than . Because of the breadth and wealth of information on this subject matter, we refer interested readers to [3] for detailed references on information-theoretic and communications aspects of fading channels. Here, we review only some of the very recent results on cognitive radio channels when the channel coefficients are fading and partially unknown, or possibly unknown, to all or some of the nodes. These include the compound Gel’fand-Pinkser channel, carbon copying onto dirty paper (an analogy related to Costa’s dirty paper coding), as well as Gel’fand-Pinkser coding with uncertainty in the phase of the non-causally known interference.

17.3.1 The Compound Gel’fand-Pinsker Channel Most of the rate regions considered thus far have employed non-causal side information at the cognitive transmitter to mitigate primary user interference at the secondary receiver. The gains exploited by the cognitive transmitter extended directly from the results of Gel’fand-Pinsker [19] on channels with non-causal side information at the transmitter. These results assume all channel parameters are perfectly known a-priori by the transmitters and receivers. Some more realistic models of fading channels assume CSI, or fading coefficients (h is the input-output relation Y = hX + N ) are available to the receiver but not the transmitter. The work [37] generalizes the results of Gel’fand-Pinsker to the case where the channel is parameterized continuously by β (belonging to a compact set C) unknown to the transmit-

17 Fundamental Limits of Cognitive Radio Networks

345

ter, but assumed known at the receiver. For communication over a channel without side-information, such channels are often called compound channels4 [11, 47]. For the traditional discrete compound channel with input X, output Y and conditional probability mass function (PMF) PYβ |X where β denotes the unknown parameter at the transmitter, it is well known that the capacity, [47], is given by C = sup inf I β (X; Y ), PX

β

(17.16)

4

where I β (X; Y ) = I(PX , PYβ |X ) denotes the mutual information between X and Y given the realization of the channel parameter is β. The main result of [37] in the finite alphabet case for compound channels with side-information at the transmitter is given in Theorem 5. Theorem 5. The capacity C of the discrete memoryless compound channel with side information at the transmitter is bounded by CL ≤ C ≤ CU where h i β inf I (U ; Y |W ) − I(U ; S|W ) , CL = sup (17.17) PU |X,S,W ,PX|S,W ,PW

β∈C

CU =

sup β

{PU |X,S,W }β ,PX|S,W ,PW

h i β inf I (U ; Y |W ) − I(U ; S|W ) ,

β∈C

(17.18)

and the suprema are over all finite alphabet auxiliary random variables U and finite alphabet time-sharing random variables W and {PUβ|X,S,W }β denotes any family of distributions (i.e., in CU a distribution PU |X,S,W is chosen for each β before the minimization over β is performed). Since the joint distribution PX,S,W in CU does not depend on β, the upper bound CU is in general tighter than if a genie had revealed β to the transmitter. When C is a singleton (the degenerate case), the bounds reduce to the well known Gel’fandPinsker result. The authors of [37] then apply this result to the cognitive radio scenario, and consider the problem of encoding a message V with knowledge of a Gaussian interfering signal S of power Q. The encoder output X is also power constrained to P = Q and the signal received at the decoder is Y = β1 X + β2 S + Z where Z is independent Gaussian noise and the compound parameter is β := (β1 , β2 ). Similar to Costa’s scheme, they suggest U = X + αS, where α is now chosen as a function of the second order statistics of β1 and β2 as α=

(|µ1

µ∗1 µ2 SN R , + σ12 )SN R + 1

|2

(17.19)

where µi and σi2 are the mean and variance of βi . This choice for Ricean fading channels, where β1 and β2 have K-factors K1 and K2 respectively satisfies: 1. If K1 , K2 → ∞, then the scheme is identical to Costa’s with α = P/(P + N ) and the interference is perfectly mitigated. 4

In [47], the channel is said to be compound provided that β is unknown to either the transmitter, the receiver or both. Here, we are interested in the scenario where β is known to the receiver but not the transmitter.

346

Natasha Devroye and Vahid Tarokh

2. If either K1 → 0 or K2 → 0, the scheme treats the interferer as noise. 3. The performance does not depend on the phase difference between µ1 and µ2 as this choice of α rotates the mean channels so that their phases are aligned. The authors numerically found that the achievable rates for given outage probabilities are good over a wide range of K-factors, by comparing the rates to the outage capacity of an interference-free scenario and a scenario where the interference is treated as noise.

17.3.2 Carbon Copying onto Dirty Paper In a similar vein, the work [29] focuses on a generalization of the Gel’fand-Pinkser problem in which a single transmitter wishes to transmit a common message to multiple receivers (which they call the multicast channel). They consider memoryless channels in Gaussian and noiseless binary special cases. Each receiver experiences different interference which is non-causally known at the transmitter. This problem is equivalent to that of dirty-paper coding (between a single transmitter and a single receiver) where only imperfect knowledge of the state (interference) is available to the transmitter. The collection of possible (random) interferences may be thought of as the interferences at different receivers, and the problem reduces to finding the maximal rate at which we can communicate the single message to all users simultaneously.

Interference

S1

Z1

+

+

Y1

+

+

Y2

S2

Z2

Noise

X

Interference

Noise

Figure 17.8. The two user Gaussian multicast channel model of [29] with additive interference. The encoder knows the two independent interfering sequences S1 , S2 and wishes to transmit a single, common message to each of the two receivers. The channel is subject to additive white Gaussian noise Z1 , Z2 . Of particular interest to the cognitive radio channel with unknown fading is the Gaussian case considered in [29]. They restrict their attention to sending a common message to two users (see Figure 17.8) with different interferences, which they assume to be i.i.d. Gaussian. This corresponds to a single input-single output channel where the interfering sequence can take on one of two possible values. Although this would not be the case in a situation in which the interfering sequence S is known but the channel fading coefficient h is not (Gaussian noise input-output relation

17 Fundamental Limits of Cognitive Radio Networks

347

Y = X + hS + N ), the results and techniques are of great interest nonetheless. Their main results are upper and lower bounds on the channel capacity in this two user case, paraphrased in Theorem 6. Theorem 6. The Gaussian multicast channel with Gaussian noise of power N = 1, input power constraint P , independent interference sequences of zero mean and power Q is bounded according to R− ≤ C ≤ R+ , where5

and

 √  1 log2 (1 + P ) + 1 log2 P +Q+1+2 P Q Q≥4 4 Q 4 √ R++= 1 P +Q+1+2 P Q 1+P 1  log Q<4 2 Q/4+1 + 4 log2 4 Q/4+1  1 P    2 log2 1 + Q/2+1 R− 1 log P +Q/2+1 + 2 2 Q   1 log2 (1 + P ) 4

Q/2 < 1 1 2

log2

Q 2

1 ≤ Q/2 ≤ P + 1 Q/2 ≥ P + 1

The upper bound is obtained by considering a single-interference Gaussian channel, and arguing why the achievable rate for the two-interference channel of interest cannot be higher. The lower bound is an explicit expression of the maximization of PS 1 1 R− = max log2 1 + + (1 + PV ) , {(PS ,PV ):PS ≥0,PV ≥0,PS +PV ≤P } 2 PV + Q/2 + 1 4 which they show to be achievable through a combination of superposition coding, dirty paper coding, time-sharing, and representing the two interferences S1 , S2 as S1n = S n + V n S2n = S n − V n

S n = (S1n + S2N )/2 V n = (S1n − S2n )/2.

The main idea is to split the message into two parts: the first is dirty-paper coded to mitigate the “common interference” S. The second message is then, in a timesharing manner, first sent to the first receiver (again by dirty-paper coding against the remaining first interference) and then to the second receiver (also by dirty-paper coding against the remaining second interference). The superposition of these two parts is transmitted. The authors then consider robust dirty-paper coding. That is, the channel model is the usual Y = X + S + Z (Z is Gaussian noise), and the interference sequence S 6 is either β1 S0 or β2 S0 , where we interpret S0 as the interference known to the encoder, while the fading coefficient β1 , β2 is not. The lower bound on the channel capacity is then given by the expression

Cβ ≥

5

6

max

(Ps ,Pv ):Ps ≥0,Pv ≥0,Ps +Pv ≤P

PS 1 1 + log2 (1+Pv ) . log2 1+ 2 1 + (β1 − β2 )2 Q/4 + Pv 4

The upper bound can be further tightened by considering min(R+ , log2 (1 + P )), where the second expression corresponds to the multicasting rate when the interference is absent. We have dropped all n superscripts for clarity.

348

Natasha Devroye and Vahid Tarokh

Both the work [29] as well as [37] provide upper and lower bounds to dirtypaper coding when the interference signal is imperfectly known. Although these results do not give explicit rate regions for the cognitive radio channel with unknown coefficients, they provide valuable insights and techniques which could be of use to future results in this direction.

17.3.3 Gel’fand-Pinkser Coding with Unknown Phase Continuing in the line of work of [29, 37], the authors in [20] consider the channel Y = X + (|h|ejθ )S + Z

(17.20)

where the interference S and the fading amplitude |h| = 1 are known to the transmitter, while the phase of the fading coefficient, θ is unknown. They are interested in determining whether the decrease in rate due to the unknown phase is substantial enough to warrant trying to learn the phase (i.e., through a training scheme, taking into account the required overhead). They derive upper and lower bounds on the rates achievable under these conditions, as well as an upper bound on the rate for a given outage probability. An upper bound on the channel capacity with phase uncertainty is obtained by considering the same channel where the phase uncertainty is reduced to either 0 or π, resulting in Theorem 7. Theorem 7. The channel described by (17.20) with additive white Gaussian noise Z ∼ N (0, 1), average power constraint P , interference S ∼ N (0, Q) known at the transmitter, has achievable rates R bounded by 1 (P + Q + 1)2 R ≤ log2 . 2 4Q This bound is shown to be the same as ignoring interference at high signal to interference ratio (SIR), and to be loose at low SIR. The authors then derive achievable rates at low SIR in a method similar to [29]. By sectoring the interference uncertainty circle into a few sectors, and then time-sharing between dirty-paper coding the ‘central’ interference vector in each sector (treating the residual interference as noise), they are able to demonstrate that if the phase uncertainty is δφ ∈ [−π/2, π/2] (difference between the actual phase θ and the phase of the central angle used to dirty paper code against) then an achievable rate is given as in Theorem 8. Theorem 8. For a phase uncertainty of δφ, an achievable rate is given by P R = sup log2 , (δφ) α∈[0,1] where (δφ) = (1 − β(δφ))2 P + (α2 + β(δφ)2 − 2αβ(δφ) cos(φ))Q + β(δφ)2 and 4

β(θ) = .

P + α cos(θ)Q P +Q+1

17 Fundamental Limits of Cognitive Radio Networks

349

To note is that this bound is only useful at low SIR, and it improves upon the scheme which treats interference as noise. The authors go on to propose a phase estimation scheme with help from primary transmitter pilot tones, which allows them avoid wasting energy and time by transmitting in useless sectors.

17.4 Conclusion Recent results on information theoretic limits of wireless channels involving cognitive radios were outlined. The capacity region of the interference channel with degraded message sets, also known as the cognitive radio channel, has been found in the weak interference regime. This region demonstrates that behaving in a cognitive fashion, where primary and secondary messages are transmitted in a non-orthogonal fashion, is beneficial, in terms of achievable rates, for both the primary and secondary cognitive users. In particular, when the primary user’s signal is known at the cognitive transmitter, it is possible to achieve better rates than those offered by spectral gap filling solutions. However, these results depend on both the non-causal knowledge as well as having perfect CSI at the transmitters. When this is not the case, as shown in the second part of this chapter, interference mitigation techniques may suffer in terms of rate. In channels employing cognitive radios there is therefore a tradeoff between learning the channel and interference in order to mitigate it, and avoiding interference altogether by spectral gap filling.

References 1. R. Ahlswede. Multi-way communcation channels. In ProcISIT, September 1973. 2. M.R. Aref. Information flow in relay networks. Technical report, Stanford University, 1980. 3. E. Biglieri, J. Proakis, and S. Shamai (Shitz). Fading channels: Informationtheoretic and communication aspects. IEEE Trans. Inf. Theory, 44(6):2619– 2692, October 1998. 4. G. Caire and S. Shamai. On the achievable throughput of a multi-antenna gaussian broadcast channel. IEEE Trans. Inf. Theory, 49(7):1691–1705, July 2003. 5. A.B. Carleial. Interference channels. IEEE Trans. Inf. Theory, IT-24(1):60–70, January 1978. 6. M. Costa. On the gaussian interference channel. IEEE Trans. Inf. Theory, 31(5):607–615, September 1985. 7. M.H.M. Costa. Writing on dirty paper. IEEE Trans. Inf. Theory, IT-29:439– 441, May 1983. 8. T. Cover. Comments on broadcast channels. IEEE Trans. Inf. Theory, 44(6):2524–2530, September 1998. 9. T. Cover and J.A. Thomas. Elements of Information Theory. New York: John Wiley & Sons, 1991. 10. T. M. Cover and A. E. Gamal. Capacity theorems for the relay channel. IEEE Trans. Inf. Theory, 25(5):572–584, September 1979. 11. I. Csisz´ ar and J. K¨ orner. Information Theory: Coding Theorems for Discrete Memoryless Systems. Academic Press, New York, 1981.

350

Natasha Devroye and Vahid Tarokh

12. N. Devroye, P. Mitran, and V. Tarokh. Achievable rates in cognitive networks. In 2005 IEEE International Symposium on Information Theory, September 2005. 13. N. Devroye, P. Mitran, and V. Tarokh. Achievable rates in cognitive radio channels. In 39th Annual Conf. on Information Sciences and Systems (CISS), March 2005. 14. N. Devroye, P. Mitran, and V. Tarokh. Achievable rates in cognitive radio channels. IEEE Trans. Inf. Theory, 52(5):1813–1827, May 2006. 15. N. Devroye, P. Mitran, and V. Tarokh. Cognitive decomposition of wireless networks. In Proceedings of CROWNCOM, March 2006. 16. FCC. 17. FCC. Secondary markets initiative. 18. A. E. Gamal. A capacity of a class of broadcast channels. IEEE Trans. Inf. Theory, 25(2):166–169, March 1979. 19. S.I. Gel’fand and M.S. Pinsker. Coding for channels with random parameters. Probl. Contr. and Inf. Theory, 9(1):19–31, 1980. 20. P. Grover and A. Sahai. What is needed to exploit knowledge of primary transmissions? http://arxiv.org/abs/cs.IT/0702071. 21. T.S. Han and K. Kobayashi. A new achievable rate region for the interference channel. IEEE Trans. Inf. Theory, IT-27(1):49–60, 1981. 22. S. Haykin. Cognitive radio: brain-empowered wireless communications. IEEE J. Select. Areas Commun., 23(2):201–220, February 2005. 23. I.Maric, R.D. Yates, and G. Kramer. The strong interference channel with common information. In Proc. of Allerton Conference on Communications, Control and Computing, September 2005. 24. I.Maric, R.D. Yates, and G. Kramer. The strong interference channel with unidirectional cooperation. In Information Theory and Applications ITA Inaugural Workshop, February 2006. 25. S. Jafar and S. Srinivasa. Capacity limits of cognitive radio with distributed dynamic spectral activity. In Proc. of ICC, June 2006. 26. S.A. Jafar. Capacity with causal and non-causal side information – a unified view. IEEE Trans. Inf. Theory, 52(12):5468–5474, December 2006. 27. N. Jindal and A. Goldsmith. Dirty-paper coding versus tdma for mimo broadcast channels. IEEE Trans. Inf. Theory, 51(5):1783–1794, May 2005. 28. A. Jovicic and P. Viswanath. Cognitive radio: An information-theoretic perspective. Submitted to IEEE Trans. Inf. Theory, 2006. 29. A. Khisti, U.Erezand A.Lapidoth, and G. Wornell. Carbon copying into dirty paper. To appear in IEEE Trans. Inf. Theory, 2007. Available at http://arxiv.org/abs/cs.IT/0511095. 30. Y.H. Kim, A. Sutivong, and S. Sigurl’onsson. Multiple user writing on dirty paper. In Proc. of ISIT, June 2004. 31. K.Marton. A coding theorem for the discrete memoryless broadcast channel. IEEE Trans. Inf. Theory, 25:306–311, May 1979. 32. J. Korner and K. Marton. General broadcast channels with degraded message sets. IEEE Trans. Inf. Theory, 23(1):60–64, January 1979. 33. S. Kotagiri and J.N. Laneman. Multiple access channels with state information known at some encoders. Submitted to IEEE Trans. Inf. Theory, 2007. 34. G. Kramer. Outer bounds on the capacity of Gaussian interference channels. IEEE Trans. Inf. Theory, 50(3), March 2004. 35. G. Kramer, M. Gastpar, and P. Gupta. Cooperative strategies and capacity theorems for relay networks. IEEE Trans. Inf. Theory, 51(9), September 2005.

17 Fundamental Limits of Cognitive Radio Networks

351

36. J. Mitola. Cognitive Radio. PhD thesis, Royal Institute of Technology (KTH), 2000. 37. P. Mitran, N. Devroye, and V. Tarokh. On compound channels with sideinformation at the transmitter. IEEE Trans. Inf. Theory, 52(4):1745–1755, April 2006. 38. P. Mitran, H. Ochiai, and V. Tarokh. Space-time diversity enhancements using collaborative communication. IEEE Trans. Inf. Theory, 51(6):2041–2057, June 2005. 39. J. Neel, R.M. Buehrer, B.H. Reed, and R.P. Gilles. Game theoretic analysis of a network of cognitive radios. In The 2002 45th Midwest Symposium on Circuits and Systems, August 2002. 40. H. Sato. The capacity of Gaussian interference channel under strong interference. IEEE Trans. Inf. Theory, IT-27(6), November 1981. 41. H. Sato. Two user communication channels. IEEE Trans. Inf. Theory, IT-23, November 1985. 42. S.J. Silverman. Games theory and software defined radios. In MILCOM, October 2006. 43. S. Srinivasa, S.A. Jafar, and N. Jindal. On the capacity of the cognitive tracking channel. In Proc. of ISIT, July 2006. 44. E. C. van der Meulen. Three-terminal communication channels. Adv. Appl. Prob., 3:120–154, 1971. 45. E.C. van der Meulen. Transmission of information in a T-terminal discrete memoryless channel. Technical report, University of California, Berkeley, 1968. 46. H. Weingarten, Y. Steinberg, and S. Shamai. The capacity region of the Gaussian MIMO broadcast channel. IEEE Trans. Inf. Theory, 52(9):3936–3964, September 2006. 47. J. Wolfowitz. Coding Theorems of Information Theory. Springer-Verlag, New York, 1978. 48. W. Wu, S. Vishwanath, and A. Arapostathis. On the capacity of the interference channel with degraded message sets. Submitted to IEEE Trans. Inf. Theory, June 2006. 49. T. Yucek and H. Arslan. Spectrum characterization for opportunistic cognitive radio systems. In MILCOM, October 2006.

18 Spectrum Awareness: Techniques and Challenges for Active Spectrum Sensing Spectrally Efficient Communication is Emerging

Marko H¨ oyhty¨ a, Atso Hekkala, Marcos Katz, and Aarne M¨ ammel¨ a VTT [Marko.Hyhty|Atso.Hekkala|Marcos.Katz|Aarne.Mmmel]@vtt.fi Summary. Radio spectrum is not efficiently used, mainly due to the prevailing rigid frequency allocation policy. Only some bands of the spectrum - such as those bands used by cellular base stations - are heavily used. Many bands are not used at all or are used only part of the time. Radios using cognitive radio technology are aware of their frequency environment. They can improve spectral efficiency by sensing the environment and then filling the discovered gaps of unused licensed spectrum with their own transmissions. In this review, different spectrum awareness techniques for a cognitive radio system are classified and discussed. Advantages and challenges for each technique are presented. Awareness techniques are classified into passive and active techniques and definitions for these terms are suggested. The focus in this review is on active techniques. In addition, classifications based on the response time and topology are discussed. Primary challenges for active spectrum sensing are introduced and the need for cooperation between users is considered. Finally, emerging spectrum awareness techniques are outlined.

18.1 Introduction In general, frequency bands of the wireless communication spectrum are not efficiently used, mainly due to the prevailing rigid frequency allocation policy. Present communication systems use an approach originally formulated in the early 1920’s in the United States, largely as a consequence of the communication failures associated with the sinking of the Titanic in 1912 [48]. In that approach, different frequency bands are assigned to different users and service providers, and licenses are required to operate within those bands. From the technical point of view, this approach makes the design and implementation of a communication system easier as different systems operate in dedicated yet non-overlapping bands. In addition, spectrum licensing offers an effective way to guarantee adequate quality of service (QoS) for license-holders. However, exclusivity also leads to inefficient use of spectrum. The Federal Communications Commission (FCC) reported in [48] that “while some bands are heavily used - such as those bands used by cellular base stations many other bands are not in use or are used only part of the time”. Thus, there is a clear need for secondary systems that can dynamically exploit the available bands

353 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 353–372. c 2007 Springer.

354

Marko H¨ oyhty¨ a et al.

with a suitable transmitter power without interfering with the primary users (PUs) who have higher priority or legacy rights. Spectral efficiency plays an increasingly important role as future wireless communication systems will accommodate more and more users and high performance rich content (e.g., broadband) services. Cognitive radio (CR) technologies have been proposed for lower priority secondary systems aiming at improving spectral efficiency by sensing the environment and then filling the discovered gaps of unused licensed spectrum with their own transmissions [21], [27]. In addition, CR techniques can be used within a licensed network to improve the spectral efficiency. The aim of the cognitive radio is very practical and concentrates on the efficient use of natural resources, which include frequency, time, and transmitted energy. Spectrum sensing is a crucial task in a cognitive radio system. The transmissions of licensed users have to be reliably detected and spectrum sensing is thus the first step towards adaptive transmission in unused spectral bands without causing interference to PUs. The secondary system has to be spectrum aware in order to exploit the available spectrum efficiently. The secondary system can obtain the current spectrum use pattern either actively or passively as explained in the following sections. Spectrum use pattern is a pattern that shows which frequencies are occupied and which frequencies are available for use in a band of interest at a particular geographic location and a particular time. In literature terms, spectrum usage pattern, electromagnetic environment or radio environment [23], [27], and information about spectral environment [7], [8] have the same meaning. Horne classifies spectrum awareness into the two forms: (1) an opportunistic one, where the secondary system recognizes the spectrum use pattern individually by cooperative or non-cooperative sensing and (2) the sharing information approach where spectrum knowledge is distributed through beacons or control channels, or by sharing databases of existing users [23]. Terms active awareness and passive awareness are proposed in [17] to describe these methods. Cabric presents three different active spectrum sensing methods, namely energy detection, matched filter detection, and cyclostationary feature detection [8]. A very readable and thorough survey on cognitive radio networks is provided in [1]. The survey includes information about different functions for cognitive radios that operate in a network. In addition to the spectrum sensing techniques presented in [8], the interference temperature concept is added to the spectrum sensing classification in the survey. We propose a detailed and up-to-date classification of spectrum awareness techniques for CR in this overview. Both active and passive awareness methods are considered in the following sections.

18.2 A Classification of Spectrum Awareness Spectrum awareness can be classified into passive and active awareness as presented in Figure 18.1. In passive awareness, the spectrum use pattern is received outside from one’s own secondary communication system. Secondary users (SUs) can obtain spectrum resources by negotiating with primary users [52]. The spectrum use pattern can be obtained also from a server [42] or database [3]. In addition, in a policy based approach, the primary system use is defined a priori [34]. In the other approach, active awareness, secondary users actively sense the surrounding radio environment and adapt their transmission based on the measurements. This can be done in a

18 Spectrum Awareness: Techniques and Challenges

355

noncooperative manner, where nodes make their decisions independently based on the observations about the spectrum environment, or cooperatively, where local measurements are combined before decisions about spectrum use are made. In the active method, the primary users do not need to know anything about secondary users. CR systems may employ either or both forms of awareness, thus the discussed approaches should not be viewed as mutually exclusive.

Figure 18.1. Classification of spectrum awareness.

18.2.1 Passive Awareness Secondary systems can obtain information about the spectrum use pattern from the primary system during the operation. In addition, it is possible that the spectrum use is defined a priori which resembles the current licensed spectrum use. In systems based on negotiated spectrum use the primary system explicitly distributes information to the secondary users about the allocated frequencies and the available spectrum opportunities. For example, the base station of existing primary system, such as television, broadcasts its spectrum use pattern to the CR terminals (SUs). Negotiation can include technical (transmit power, location, frequency, modulation, etc.), financial (price, payment options, etc.), and service quality (signal-to-noise ratio, interference protection, etc.) parameters [52]. In practice, the parameters may depend on the characteristics of services offered by the primary and the secondary users. Beacon approach: A primary receiver can send beacons that advertise the availability of licensed spectrum for secondary usage. A beacon can either grant permission to access the spectrum or deny access [24], [33]. The beaconing approach eliminates the need for detection and prediction of spectrum opportunities by cognitive radios. However, a single beaconing approach using either a grant or a denial beacon may still suffer from reliability problems. The beacon signal may not reach secondary users due to shadowing, or might be misinterpreted by them. A dual beaconing approach is proposed in [33] to improve the reliability. In the approach, both grant and denial beacons are used. Cognitive radios utilize the spectrum only if a grant beacon is detected and no denial beacon is detected. If a grant beacon and a denial beacon from different primary receivers or base stations are detected at the same time, a CR refrains from spectrum utilization. Furthermore, beacon broadcasting can be used within a network to inform neighbours about available communication channels [58].

356

Marko H¨ oyhty¨ a et al.

Secondary markets: Guaranteed quality of service requirements can be met for both primary and secondary users only if primary users promise not to interfere. This is most likely only true for a fee. Secondary markets were proposed by the FCC to encourage licensees to make all or portions of their assigned frequencies available to secondary use [13]. However, the initial secondary market proposals were very rigid ones. It was expected that the license-holder notifies the FCC 21 days before a secondary user begins operation. They were not thinking of leases lasting only minutes or seconds. Indeed, the FCC defined a lease to be short term if it is under 360 days. Real-time secondary markets allow license-holders lease rights to secondary users to use the spectrum for the duration of license in a real-time manner [40]. Secondary users can make requests as needed and obtain the spectrum opportunities from licensees. Policy based: In the policy based approach, the radio regulation authority identifies a licensed band of the radio spectrum where use is low or the band is used with a deterministic pattern [34]. This band is then made available for secondary use. The authority assigns a set of policies that provide rules and constraints concerning how to use this available band. Policies can answer, for example, to questions such as [35]: “Am I allowed to use xyz MHz? How much power could I emit on frequency xyz? Could I double my bandwidth?” The set of policies are published in a machine-understandable form for download from servers of the radio regulation authority [34], [11]. Secondary devices repeatedly (e.g., once a day) seek for updates of policies that are relevant for their regulatory domain and update their information bases. After updating information, secondary users adapt their transmission parameters, like frequency and power, to meet policies. The fact that the use of a primary system is defined as a priori could lead to a rather static secondary use, without optimally exploiting spectrum holes (i.e., temporarily unused frequency band of primary user). Database: A primary system or the radio regulation authority can maintain a table or database of frequency resources in its server and both primary and secondary users can update this table. The database includes location information and an estimate of the interference range of the secondary user and it is most likely available over the Internet since it is pervasive, flexible, and low cost [3]. Frequencies used by the licensed system can be seen and checked from this table. When a secondary user needs to transmit, it checks the table, chooses an available band and reserves it to its use. Other secondary users can then see that this particular band is occupied by a secondary user and can choose other resources for their use. When a primary or secondary user stops transmitting, the associated band is released from the table and made available to other users. Secondary users have to check this table periodically to avoid interfering with the primary system. The primary system can start using the band reserved by a secondary user whenever there are no free bands left. If free bands exist, the primary system uses them instead of forcing a secondary user to stop using a band. In this way, the spectrum can be used very efficiently. However, this approach may require an infrastructure to operate, such as a separate network for retrieving database information. The approach could be quite rigid and thus, not so suitable for fast, dynamic and highly efficient spectrum use. Server: A Spectrum Server can be used to enable coexistence of radios in a shared environment in a centralized fashion [42]. The centralized spectrum server obtains information about neighbourhood and interference through local measurements from different terminals and then offer suggestions to the efficient spectrum

18 Spectrum Awareness: Techniques and Challenges

357

use. The spectrum information can be gathered from several separate secondary networks. Authors in [4] define the term coordinated dynamic spectrum access networks to describe situations wherein the access to the spectrum in a region is controlled and coordinated by a centralized entity called a Spectrum Broker. This can be seen as a compromise between static allocations and the opportunistic spectrum use. Service providers and users of the networks do not a priori own any spectrum; instead they obtain time bound rights from a regional spectrum broker to a part of the spectrum and configure it to offer the network service. The requirement for this approach is that part of the spectrum is allocated by regulating authorities such as the FCC for controlled dynamic access. In a certain region, a spectrum broker owns that allocated spectrum and leases it to the requesters. All these passive approaches are good in the sense that they can ensure interference-free communication to the primary system since the spectrum use is defined as a priori. The secondary system uses only frequencies accepted by the primary system or regulation authority. However, passive awareness increases the amount of needed control information in the system. Considerable signalling could be needed to disseminate frequency information. Furthermore, currently only active sensing techniques can be used to obtain spectrum awareness information since existing licensed systems and regulation authorities do not support passive awareness principles. They can still be very useful in the future. It is important to note that passive awareness approaches can be combined with opportunistic spectrum use. For example, policies can set some restrictions for the use of licensed spectrum. By exploiting active sensing and taking these constraints into account, a secondary user can use a band in a dynamic way.

18.2.2 Active Awareness In a physical carrier sensing, the nodes in the wireless network wait until the total received power from ongoing transmissions is below a certain threshold before they start to transmit [16]. Similarly, an active awareness type of cognitive radio terminal senses the surrounding radio environment before it starts to communicate. However, in a CR system a transmitting node also needs to sense if a primary user starts to use the same frequency band. Other significant difference is that the sensing bandwidth in cognitive radio systems can be very wide. Furthermore, a cognitive radio may also sense frequency bands of many different licensed systems simultaneously to identify the spectrum holes. Reliable sensing of spectrum is perhaps the most important attribute of CR as all the decisions on frequency allocation are based on this observation. Spectrum sensing is the very first step towards adaptive transmission in wide bandwidths without causing interference to primary users [5]. Spectrum sensing has two important roles [6]. First, it ensures interference-free communication for primary users. Second, it identifies spectrum opportunities for capacity increase of cognitive networks. The detection time should be very short in order to avoid wasting the opportunities. The key challenge of spectrum sensing is the detection of weak signals in a noisy environment with a very small probability of miss detection. The same problem is present in radar systems [46]. In the following sections, we give an overview of active spectrum sensing methods, highlighting the underlying principles, advantages and challenges.

358

Marko H¨ oyhty¨ a et al.

18.2.3 Response Time and Topology Spectrum awareness techniques can be divided based on the response time into slow and rapid systems, as shown in Figure 18.2. The response time is the time needed to estimate the prevailing spectrum use pattern. The information on the current spectrum use can be obtained in a very short period with some techniques whereas in some cases, it may take hours or even days. Requirements are different within different systems. Operation in TV bands is possible with slow methods but rapid methods are needed to efficiently exploit spectrum holes in bursty data transmission. Rapid methods take only fractions of a second to estimate the current spectrum use pattern while slow methods estimate the spectrum in periods which range from seconds to hours. Naturally, the latter cannot be utilized to estimate short temporal spectrum holes. Many passive approaches are slow. Policy and database approaches in which information update may take an hour or even a day are typical examples of slow response time techniques [3], [34]. Also, the secondary markets approach proposed by the FCC belongs definitely to this category [13]. Most active methods can obtain spectrum use pattern very rapidly. The spectrum sensing chip presented in [38] can achieve a scanning speed 18 GHz/s with 100 kHz resolution, requiring for instance only 27.8 ms to estimate the spectrum use pattern in a 200 MHz bandwidth. Beacon approaches also acquires spectral information rapidly [24]. If we need to exploit very short available periods in the loaded licensed spectrum, for example, to transmit an emergency message, rapid methods are required.

Figure 18.2. Classification of spectrum awareness based on response time and topology.

To show an example about a situation in which rapid methods are needed, we can think about data packet transmission. Data packet transmission on the source level in a network is commonly modelled by an ON/OFF source model [57] (also known as packet train model [28]). The ON-period is the period in which the burst of data packets are sent and the OFF-period is the period in which no packets are sent. This model is basically controlled by two major sets of parameters: 1) distribution of the ON/OFF periods and 2) distribution of packet arrivals within an ON-period. The typical mean value for OFF-period (ON/OFF source model with Pareto distribution) obtained in [57] is 10.5 s for a real-time traffic in an Ethernet local area network. To exploit spectrum holes like this one, the detection have to be completed within a short time. Both time and frequency resolutions of spectrum awareness methods are important. However, there is a tradeoff between these two resolutions. Good frequency resolution typically implies a coarse time resolution and vice versa. Frequency resolution should match the characteristic of one’s own equipment. It makes no sense to

18 Spectrum Awareness: Techniques and Challenges

359

make a spectrum sensor that has a very good frequency resolution if the CR transmitter cannot exploit that. If the transmitter resolution over its specified operating range is for instance 20 kHz, sensing with a 1 kHz resolution sensor is clearly a wasting of resources. Thus, spectrum sensing frequency resolution should be exactly as accurate as the frequency resolution of the CR transmitter. Time resolution of the CR defines the primary network within which the CR can operate. In addition to the response time, the time resolution includes time for spectrum allocation between users and also the time needed to reconfigure the transmitter. Thus, the overall time must be taken into account when thinking of which kind of spectrum awareness technique will be used in a specific system. Moreover, as presented in Figure 18.2, spectrum awareness can be classified according to the topology. Distributed solutions are good for the situations where an infrastructure supporting spectrum sensing cannot be implemented [7], [58]. Each node is responsible for being locally spectrum aware. Active sensing methods can operate in a distributed fashion as well [7]. The beacon approach presented in [24] can be distributed and secondary users may locally exploit available resources obtained via the beacon signals. The advantage of distributed approaches is the flexibility, which helps especially in spectrum sharing with other secondary systems [33]. In centralized awareness, the local information obtained via measurements or beacons can be sent to the centralized entity that controls the spectrum use in the network. It combines the local information and constructs a spectrum allocation pattern [55]. Database, policy, and spectrum server approaches also belong to the centralized solutions [3], [4], [34], [42]. The advantage of centralized approaches is that the spectrum use is tightly controlled and thus chaotic and unpredictable situations can be avoided. The drawback is that they increase the amount of needed control information in the system.

18.3 Spectrum Sensing Techniques To be able to sense very weak signals, cognitive radios must have significantly better sensitivity than conventional radios [8]. Requirements for radio frequency (RF) frontend and analog-to-digital converter (ADC) are very demanding. The requirements can be so tough that advanced techniques like beamforming are needed to make them feasible [6]. After reliable reception and sampling of a wideband signal, digital signal processing techniques are utilized to further increase radio sensitivity. Most of the recent work on spectrum sensing focuses on primary transmitter detection, which is based on local observations of secondary users. The goal of the spectrum sensing is to decide between the two hypotheses, namely [20] n(t) H0 x(t) = (18.1) hs(t) + n(t) H1 where x (t) is the complex signal received by the cognitive radio, s(t) is the transmitted signal of the primary user, n(t) is the additive white Gaussian noise (AWGN) and h is the complex gain of an ideal channel. The delay has not been taken into account. If the channel is not ideal, h and s(t) are convolved instead of multiplied.

360

Marko H¨ oyhty¨ a et al.

The null hypothesis H0 states that no licensed user is present in the observed spectrum band. The alternative hypothesis H1 indicates that some primary user signal exists. Generally, spectrum sensing techniques can be classified into two main types, primary transmitter detection and interference temperature concept as shown in Figure 18.3. Either a CR tries to detect the PU transmitter or characterize the level of harmful interference at the PU receiver. In addition, sensing techniques can be autonomously implemented (e.g., sensing carried out at a given node) or in a cooperative manner (e.g., implying sensing at multiple nodes).

Figure 18.3. Basic techniques for spectrum sensing.

18.3.1 Matched Filter Detection When a secondary user has a priori knowledge of the primary user signal at both physical and medium access control (MAC) layers, such as the pulse shape, modulation type, and the packet format, the optimal signal detection method is a matched filter, since it maximizes the output signal-to-noise ratio [8], [41]. The output of a matched filter have to be compared to a threshold to decide whether the PU signal is present or not. More discussion about the choosing of the threshold is provided in the next section. The main advantage is that the matched filter needs less time to achieve high processing gain due to coherent detection [43]. For demodulation, cognitive radio has to perform timing and carrier synchronization, even channel equalization [8]. When the SU has no accurate information on the PU signal, the matched filter performs poorly. However, most licensed systems include pilots, preambles, synchronization words or spreading codes to aid the coherent detection. A significant drawback in the use of matched filter is that it would require a dedicated sensing receiver for every PU signal type. Digital television (DTV) band is an example where matched filter detection can be performed [9]. A matched filter based cognitive radio can detect 63-sample pseudo-random noise sequences referred to as PN63 sequences in DTV signal.

18.3.2 Energy Detection If the PU signal is a priori unknown to the secondary receiver, the optimal detector is an energy detector, also known as a radiometer [43]. Energy detection is the classical method for detecting unknown signals. It measures the energy of the received

18 Spectrum Awareness: Techniques and Challenges

361

waveform over an observation time window [53]. First, the input signal is filtered with a bandpass filter (BPF) to select the bandwidth of interest. The filtered signal is squared and integrated over the observation interval. Finally, the output of the integrator is compared to a threshold to decide whether the primary user is present or not. When the spectral environment is analyzed in the digital domain, fast Fourier transform (FFT) based methods are usually used in order to obtain frequency response. FFT also generates the resolution in frequency domain. The periodogram method is a method to estimate power spectrum and it is based on the FFT [49]. The periodogram is a poor spectrum sensing method because of the large variance and bias of the estimate. Modified periodogram techniques have been developed to cope with such drawbacks [49]. A practical energy detection method for cognitive radio is Welch’s periodogram [56] which is depicted in Figure 18.4. In [21], also computationally feasible Thomson’s multitaper method [50] is proposed as a good candidate for spectrum sensing task. This method exploits multiple orthonormal windows for spectral estimation [21], [50].

Figure 18.4. Implementation of an energy detector using the Welch’s periodogram. The implementation simplicity of the energy detector is perhaps its key advantage. However, its performance is highly susceptible to noise level uncertainty. Noise level uncertainty refers to a situation where the noise variance is only approximately known. Considerable noise uncertainty may result from thermal noise and nonlinearities of components, and from inaccuracy of the noise estimator (due to limited time of averaging). In addition, the channel may contain nonstationary noise and an extraneous signal [47]. The noise uncertainty causes problems especially in the case of a simple energy detector because it is difficult to set the threshold properly without the knowledge of the accurate noise level. A pilot tone from the primary transmitter can be used to alleviate this problem [43]. In addition, an energy detector cannot differentiate between modulated signals, noise, and interference. Thus, it cannot benefit from adaptive signal processing for cancelling the interferer and it is also prone to false detection triggered by unintended signals. The performance of an energy detector is clearly degraded in fading environments and secondary users may need to cooperate in order to detect the presence of primary users [12], [20]. One challenge is setting the right threshold for detection. The problem is illustrated in Figure 18.5 where the probability density functions of the received signal with and without primary signal are shown. When the probability of missed detections is very low, the probability of false alarms increases, resulting in low spectrum utilization. On the other hand, a low probability of false alarms would result in high missed detection probability, which increases the probability of interfering with the primary users. This trade-off has to be carefully considered. In most radar detectors, the threshold is set in order to achieve a constant level of a false alarm [46]. The threshold level is raised and lowered during detection to

362

Marko H¨ oyhty¨ a et al.

maintain a constant probability of a false alarm. This approach is known as constant false alarm rate (CFAR) detection.

Figure 18.5. Trade-off between missed detections and false alarms.

A simple energy detector works poorly in frequency hopping spread spectrum signals. The channelized radiometer is a multi-channel receiver that has several energy detectors that integrate energy in many frequency bands simultaneously [36]. It is especially useful for detecting frequency hopping spread spectrum signals. It is possible to use FFT to approximate an ideal channelized radiometer [32]. An analysis of the effects of frequency sweeping on a channelized radiometer is presented in [31]. It is assumed that the signal to be detected uses slow frequency hopping and that sweeping is faster than hop dwell time. In a practical signal detection system, the instantaneous bandwidth may be limited. In frequency sweeping, the centre frequency is changed as a function of time to cover a wider bandwidth. Numerical examples in [31] demonstrate that if the number of hops observed per decision is small, sweeping is needed to attain the desired performance. In a fading channel, the best performance is obtained using fast sweeping techniques. The drawback of the channelized radiometer approach compared to a simple energy detector is the increased complexity.

18.3.3 Feature Detection The idea of the cyclostationary feature detection is to exploit the built-in periodicity of the modulated signal [8], [45]. Cyclostationary signals exhibit correlation between widely separated spectral components due to spectral redundancy caused by periodicity. The features are detected by analyzing the spectral correlation function (SCF) [18]. The possible features that can be extracted include a sine-wave carrier, symbol rate, and modulation type [6]. The main advantage of the feature detection is that it discriminates the noise energy from the modulated signal energy. This is because noise has no spectral correlation while modulated signals are cyclostationary with spectral correlation due to embedded redundancy of signal periodicities.

18 Spectrum Awareness: Techniques and Challenges

363

A block diagram of a cyclostationary feature detector is presented in Figure 18.6. The parameter α is the cycle frequency, representing the frequency separation of the correlated spectral components [19].

Figure 18.6. Implementation of a cyclostationary feature detector. Cyclostationary feature detector is more robust to noise uncertainty than an energy detector. Furthermore, feature detectors operate with lower SNR values than energy detectors since the latter do not exploit the information embedded in the received signal whereas feature detectors do. On the other hand, feature detectors are more complex to implement than their counterpart energy detectors [6]. Moreover, cyclostationary detectors require also longer observation times than energy detectors. Therefore, spectral holes of short duration cannot be exploited so efficiently if compared to sensing methods requiring shorter observation times. However, as highlighted in [15], feature detectors can achieve a huge processing gain over an energy detector. A feature detector can be capable of receiving signals more than 30 dB below the noise floor. The hidden node problem that might result when missing the presence of a PU signal becomes much less likely than with energy detectors. In [59] a feature detection method is proposed to continuously observe the signal of the PU in a cognitive radio system. In the demonstration in [59], a Global System for Mobile Communications (GSM) network is the PU and an orthogonal frequencydivision multiplexing (OFDM) based wireless local area network (WLAN) is the SU. Since the cyclic features of the WLAN and GSM signals are different enough, the channel can be continuously monitored, i.e., the SU can transmit or receive data and sense the spectrum simultaneously. In the conventional sensing approach, the reaction speed to the PU coming to the band is poorer since no detection is possible between two sensing periods. In addition, the throughput of the SU network is increased because no silent period is needed.

18.3.4 Interference Temperature Concept Interference is conventionally regulated in a transmitter-centric way. The idea is that interference can be controlled at the transmitter through the radiated power, location of individual transmitters and the out-of-band emissions [1]. However, interference actually takes place at the receivers as in the case depicted in Figure 18.7. The PU and SU terminals are separated by a physical obstacle opaque to radio signals. Two such terminals are said to be hidden from each other [51]. One good example of hidden terminal problem is also a digital TV which lies at the cell edge; the power of received signal can be barely above the sensitivity of the receiver [30]. If the CR is not capable of detecting TV signal, it can start to use the spectrum and interfere with the signal the digital TV is trying to decode. This problem can be avoided if the sensitivity of the CR receiver outperforms that of the primary user receiver by a large margin [8], [30]. A model for measuring interference referred to as interference temperature has been introduced by the FCC [14]. As opposed to current transmitter-centric

364

Marko H¨ oyhty¨ a et al.

Figure 18.7. Interference caused by shadowing.

approach, this model attempts to regulate interference at the receivers, where it is distinctly harmful. The interference temperature TI is a measure of the interference power within a given bandwidth [10]. It is specified in Kelvins and defined as TI = PI /kB,

(18.2)

where PI is the interference power in Watts over bandwidth B measured in Hertz, and k is Bolzmann’s constant (1.38·10−23 J/K). The concept of interference temperature attempts to characterize interference and noise with a simple number, taking just a single measurement by the CR. To be more exact, the interference temperature equals the constant noise temperature plus the interference term caused by interference environment [10]. The interference temperature limit TL characterizes the maximum amount of tolerable interference for a given frequency band in a particular location where the receiver can operate satisfactorily. Thus, TL allows regulating the received power rather than the transmitted power [10]. CR terminals operating in licensed frequency bands have to measure the current interference temperature and adjust their transmission in a way that they avoid raising the interference temperature over the limit. Thus, real-time interactions between transmitter and receiver in an adaptive manner are needed [21]. The following must hold at all licensed receivers: TI + PS /kB ≤ TL ,

(18.3)

The symbol PS is the received power of a transmitted secondary signal at the licensed receiver. Thus, for each transmission the interference temperature is measured and suitable transmitted power and bandwidth B are computed to meet QoS requirements without violating TL . Note that the attenuation is assumed to be known and that the interference temperature limit multiplied by the Bolzmann’s constant yields the corresponding upper limit on permissible power spectral density in a frequency band of interest [21]. That density is measured in Joules per second or, equivalently, Watts per Hertz.

18 Spectrum Awareness: Techniques and Challenges

365

Due to interrelations between interference temperature and bandwidth, measuring of the interference temperature may be necessarily done with an iterative algorithm like hill climbing or fixed point iteration [10]. The fundamental problem in the interference temperature model is that a cognitive radio user can only be aware of its precise location with the help of positioning system. Since primary receivers are usually passive devices, a secondary user cannot be aware of their precise locations [1]. In addition, if the effects of cognitive radio transmissions cannot be measured on all possible receivers, interference temperature measurement may not be feasible.

18.4 Spectrum Sensing Challenges Restricted sensing ability: Cognitive radios are in a way blind and they cannot perceive other radios efficiently. They have only a basic “sense of hearing”. This makes cognitive operation very demanding. Imagine a blind person at a road crossing point trying to conclude whether the road is free or not based only on his hearing. A CR can perceive its multidimensional surrounding radio environment with a single sense. Several challenges for the spectrum sensing exist that need to be investigated. Many open questions are related to the sensing ability and performance in wide bandwidths, interference temperature measurement, spectrum sensing in multi-user environments, and cooperative detection techniques. The main requirement for detection is a reliable, accurate, and fast detection of primary users. Advanced techniques are needed to sense very wide bandwidths rapidly and reliably. Wideband sensing: In [44] authors propose to use a limited target spectrum for spectrum sensing instead of very wide band detection. By using limited sensing bandwidth, sampling at or above Nyquist rate is possible even with current technology. The computational burden can be restricted to a reasonable level and the rather high cost analog front-end (including a wideband antenna, wideband amplifiers and mixers) required for a very wide spectrum scan can be avoided. Furthermore, it can be prevented that a single type of cognitive radio occupies the majority of spectrum opportunities. Authors propose that regulatory agencies should allocate spectrum bands for different types of cognitive radios depending on the intended range and the throughput requirements. Interference measurement: The interference temperature concept is also a promising approach for dynamic spectrum use. As long as cognitive radios do not exceed the interference temperature limit by their transmissions, they can use the band. Two primary challenges in interference temperature concept are [29]: “(1) the determination of the background interference environment as a function of spatial location and frequency; and (2) the in situ measurement of the interference temperature to determine optimal radio transmission parameters”. The latter refers to the fundamental problem that cognitive radios cannot be aware of the precise locations of primary receivers and they cannot measure the effects of their transmissions on all possible receivers. One possible approach for in situ monitoring is to use measurements from many fixed and mobile sites and integrate the data to create an overall power flux density map across a large area [29]. However, the spatial sampling is insufficient to accurately map multipath fading. In [45], a separate sensor network based sensing architecture is proposed to effectively address the interference temperature model.

366

Marko H¨ oyhty¨ a et al.

This model offers diversity to cope with fading. However, the density of sensor network should be quite high even to accurately map the shadowing. In addition, continuous low power communication is possible by separating the sensing and operational functions. This approach, however, requires a fixed supporting infrastructure for sensing, therefore limiting the possibilities to use the cognitive radio system everywhere. In [24] a beacon-based technique was proposed to alleviate the in situ problem. Primary receivers send beacons to inform SUs about frequencies they use. Beacons are transmitted always at the same power level. Thus, a SU can choose its operating frequency based on the lowest beacon power. In addition, assuming the channel to be reciprocal (i.e., fading is same in both directions), the SU can estimate the effect of its transmission quite accurately in the location of PU receiver and calculate its transmission power to meet interference temperature constraint. However, this approach requires changes to be done in the primary system. Spectrum sensing in multi-user environment: The environment in which cognitive radios operate consists usually of multiple secondary users and primary users [1]. In addition, the cognitive radio networks can be co-located with other secondary networks competing for the same spectral resources. Secondary users can interfere with each other in spectrum sensing which makes it more difficult to detect primary users reliably. Secondary signals can (1) be detected as primary signals or (2) mask primary signals and thus interfere with the PU detection. In such a multiuser environment, cooperation is needed to exploit spatial diversity. On the other hand, cooperation creates challenges as the spectrum sensing information needs to be distributed. Delays in cooperation have to be very short and the signalling overhead caused by sensing information distribution must be kept low. Otherwise there would be very few temporal resources that can be used for cognitive radio transmission. To overcome these problems, fast physical layer signalling, a boosting protocol, was proposed in [54] for centralized spectrum pooling systems. However, there are several open research challenges in multi-user network operation. To mention a few: What kind of cooperation is really needed to efficiently exploit the spatial diversity? In addition to the spectrum holes, what information should be distributed (location, transmitted power and frequency of different users)? How to cooperate with other secondary networks? Do primary networks need to be involved in the cooperation?

18.5 To Cooperate or Not to Cooperate? If all the SUs make their own decisions individually, spectrum sensing may not work reliably enough due to possible hidden terminals [51]. In order to improve the performance of the spectrum sensing, several authors have proposed cooperation among SUs [8], [20], [22]. The performance improvement of the cooperative spectrum sensing results from the exploitation of spatial diversity [37]. A PU might not be seen by an SU due to the instantaneous channel condition, but a group of SUs will avoid this problem as the statistics of their associated channels are independent. However, cooperation increases the amount of needed control information in a system. In a centralized cooperation, the locally sensed spectrum information of nodes will be sent to a common control channel, combined in the base station or in the conscious node (CNode) that plays a role of spectrum coordination in the network [2], and then broadcasted to the CR terminals in the network [55]. The role of the CNode differs from conventional access point in an ad hoc network because it has cognitive

18 Spectrum Awareness: Techniques and Challenges

367

capability and the link between nodes is basically established by peer-to-peer (P2P) communications. Cooperation can also be operating in a distributed fashion without centralized control. The fundamental question here is whether to cooperate or not. It is often useful to think analogies. Natural laws can sometimes be applied in different fields. There is a certain analogy between operation of a cognitive radio and the road traffic. A driver that comes to a crossing searches for a suitable gap in the traffic flow, drives his/her car to the gap, and accelerates to the speed of the other traffic. This flexible and shared use of the road increases clearly the efficiency of road use compared to the situation when the lane is occupied by a particular car type (e.g., taxi and bus lanes). This is especially true in a road in which traffic is sparse. A driver can make his/her decision independently and get instantly into the road at a suitable point and time. Instead, in a densely populated area where the traffic load is large, traffic lights are needed to ensure flexible road use and a fair deal between drivers. A comparable situation takes place in a cognitive radio network. In the place of traffic lights, the spectrum use in the network is controlled by a base station or a CNode. It gives permissions to different cognitive radio users to exploit different spectrum holes during their operation. In non-cooperative situations the nodes make their decisions independently based on the observations on the spectrum environment. This reduces the amount of needed control information in the network. However, the efficiency of the spectrum use may become lower and fairness between secondary users may not be kept. Cooperation is needed in a densely populated network whereas nodes can operate autonomously in a network that carries only sparse traffic.

18.6 Emerging Techniques Since energy detection techniques are simple and fast spectrum sensing method, they are widely adopted for spectrum sensing task in pure research as well as in the current standardization work. An energy detector is chosen by the the Institute of Electrical and Electronics Engineers (IEEE) 802.11y working group that will develop a standard for shared 802.11 operation with other users in the 3650-3700 MHz band. Though energy detection schemes are assumed in many studies, more advanced methods are needed to realize cognitive radio systems that can both offer interference-free communication to the licensed users and use the available spectrum opportunities in an efficient way. In addition to the fast sensing speed, the accuracy of spectrum sensing is a critical requirement. From the viewpoint of commercialization, low power consumption and simple implementation are desired [39]. TV bands are thought to be very promising spectrum resources for secondary use. A standard for utilizing cognitive radio technology in TV bands is being developed by the IEEE 802.22 working group. In [25], [39], a cognitive radio system with a dual-stage spectrum sensing is proposed. This approach combines coarse and fine sensing to meet the sensing speed and accuracy requirements of a cognitive radio system. First, a wavelet transform-based Multi-Resolution Spectrum Sensing (MRSS) energy detection method takes a snapshot of the current spectrum use pattern over the whole band of interest and identifies the occupancy of each spectrum segment. Second, a more sensitive time-domain feature detection method scrutinizes the candidate spectrum segments found free at the MRSS stage. MRSS processing

368

Marko H¨ oyhty¨ a et al.

is performed in the analog domain making possible low-power consumption and real-time operation. The IEEE has initiated the 1900 Standard series for Next Generation radio and spectrum management to address standards for dynamic spectrum use [26]. The 1900.2 working group is especially developing recommended practices for interference and coexistence analysis. The group will address the technical challenges in measuring and predicting interference [29]. Interference temperature concept is an interesting option for dynamic spectrum use. If the challenges in interference temperature measurement can be properly solved, the concept will provide a useful technique for dynamic spectrum management in the future. A beacon approach proposed in [24] appears to be quite a practical solution. Cooperative techniques will undoubtly emerge in the future. They are needed to utilize more efficiently available spectral resources by secondary systems. As mentioned, cooperation between secondary systems in active spectrum sensing helps to cope with hidden terminal problem. The probability of missed detection is significantly lower with cooperative detection than with individual sensing. In addition, a more balanced use of the spectrum in the spatial domain will be possible with cooperation. Most likely, cooperation between secondary and primary system will also be exploited in future communication systems. It helps the secondary system to avoid chaotic situations since the passively received spectrum use pattern shows the spectrum opportunities in advance. It can be thought as the yellow light in traffic light system. Instead of reactive operation it helps the driver (secondary user) to be proactive. Cooperation principles in spectrum sensing and sharing will improve the performance of the system. Moreover, without any cooperation, it is impossible to assure interference-free communication to the primary users.

18.7 Conclusions Spectral efficiency is playing an increasingly important role as future wireless communication systems will accommodate more and more users and high performance (e.g., broadband) services. There is a clear need for spectrum-aware systems that can dynamically exploit the available bands with suitable transmitter power without interfering the present users who have higher priority or legacy rights. In this review, different spectrum awareness techniques under current research are classified and discussed. We classified spectrum awareness into passive and active awareness and further divided these main classes into several subclasses. Techniques and challenges for each subclass were investigated and identified. In addition, classifications based on the response time and topology were provided. The main emphasis was in active spectrum sensing. Many challenges in which more research is needed were considered and emerging techniques in spectrum awareness were introduced. The role of cooperative techniques in spectrum sensing and sharing was also discussed. From the discussions provided it is clear that spectrum-aware communication will play a key role in emerging wireless and mobile communication systems. Acknowledgement. The authors would like to thank Mr. Mubaraq Mishra from Berkeley Wireless Research Center for his valuable comments.

18 Spectrum Awareness: Techniques and Challenges

369

References 1. I.F. Akyildiz, W.-Y. Lee, M.C. Vuran, and S. Mohanty. NeXt generation/dynamic spectrum access/cognitive radio wireless networks: A survey. Elsevier Computer Networks, 50:2127–2159, September 2006. 2. M.-G. Di Benedetto, G. Giancola, and M. D. Di Benedetto. Introducing consciousness in UWB networks by hybrid modelling of admission control. Springer Journal on Mobile Networks and Applications, 11:521–534, August 2006. 3. T.X. Brown. An analysis of unlicensed device operation in licensed broadcast service bands. In DySPAN 2005, pages 11–29, November 2005. 4. M.M. Buddhikot, P. Kolodzy, S. Miller, K. Ryan, and J. Evans. DIMSUMNet: New directions in wireless networking using coordinated dynamic spectrum access. In WoWMoM 2005, pages 78–85, June 2005. 5. D. Cabric and R.W. Brodersen. Physical layer design issues unique to cognitive radio systems. In PIMRC 2005, pages 759–763, September 2005. 6. D. Cabric et al. Novel radio architectures for UWB, 60 GHz, and cognitive wireless systems. EURASIP Journal on Wireless Communications and Networking, 2006:1–18, January 2006. 7. D. Cabric, S. M. Mishra, D. Willkomm, R. Brodersen, and A. Wolisz. A Cognitive radio approach for usage of virtual unlicensed spectrum. In 14th IST Mobile and Wireless Communication Summit, June 2005. 8. D. Cabric, S.M. Mishra, and R.W. Brodersen. Implementation issues in spectrum sensing for cognitive radios. In Asilomar Conference, pages 772–776, November 2004. 9. S.-Y. Chang. Analysis of proposed sensing schemes for IEEE802.22 WRAN. IEEE 802.22-06/0032r1 document, March 2006. 10. T.C. Clancy. Dynamic Spectrum Access in Cognitive Radio Networks. PhD thesis, University of Maryland, 2006. [Online]. http://www.cs.umd.edu/ %7Ejkatz/THESES/clancy.pdf. 11. DARPA XG. DARPA neXt Generation (XG) Program. [Online] http://www. darpa.mil/sto/smallunitops/xg.html. 12. F.F. Digham, M.-S. Alouini, and M.K. Simon. On the energy detection of unknown signals over fading channels. In ICC 2003, pages 3575–3579, May 2003. 13. Federal Communications Commission (FCC). Principles for promoting the efficient use of spectrum by encouraging the development of secondary markets. Report ET Docket no. 00-401, December 2000. 14. Federal Communications Commission (FCC). Notice of inquiry and proposed rulemaking. Report ET Docket no. 03-289, November 2003. 15. Federal Communications Commission (FCC). Notice of proposed rulemaking and order. Report ET Docket no. 03-322, December 2003. 16. J.A. Fuemmeler, N.H.Vaidya, and V.V. Veeravalli. Selecting transmit powers and carrier sense thresholds for CSMA protocols. Technical report, University of Illinois at Urbana-Champaign, October 2004. 17. T. Fujii and Y. Suzuki. Ad-hoc cognitive radio - Development to frequency sharing systems by using multi-hop network. In DySPAN 2005, pages 589–592, November 2005. 18. W.A. Gardner. Signal interception: A unifying theoretical framework for feature detection. IEEE Transactions on Communications, 36:897–906, August 1988.

370

Marko H¨ oyhty¨ a et al.

19. W.A. Gardner and C.M. Spooner. Signal interception: Performance advantages of cyclic-feature detectors. IEEE Transactions on Communications, 40:149–159, January 1992. 20. A. Ghasemi and E.S. Sousa. Collaborative spectrum sensing for opportunistic access in fading environments. In DySPAN 2005, pages 131–136, November 2005. 21. S. Haykin. Cognitive radio: Brain-empowered wireless communications. IEEE Journal on Selected Areas in Communications, 25:201–220, February 2005. 22. J. Hillenbrand, T. Weiss, and F. Jondral. Calculation of detection and false alarm probabilities in spectrum pooling systems. IEEE Communications Letters, 9:349–351, April 2005. 23. W.D. Horne. Adaptive spectrum access: Using the full spectrum space. In TPRC 2003, October 2003. 24. A.P. Hulbert. Spectrum sharing through beacons. In PIMRC 2005, pages 989– 993, September 2005. 25. Y. Hur et al. A wideband analog multi-resolution spectrum sensing (MRSS) technique for cognitive radio (CR) systems. In ISCAS 2006, pages 4090–4093, May 2006. 26. IEEE P1900. IEEE standard series on next generation radio and spectrum management. [Online] http://grouper.ieee.org/groups/emc/emc/ieee emcs sdcom/P1900-X Stds/ieee emcs - p1900-x main.htm. 27. J. Mitola III and G.Q. Maguire Jr. Cognitive radio: Making software radios more personal. IEEE Personal Communications, 6:13–18, August 1999. 28. R. Jain and S.A. Gouthier. Packet trains - Measurements and a new model for computer network traffic. IEEE Journal on Selected Areas in Communications, SAC-4:986–995, September 1986. 29. P.J. Kolodzy. Interference temperature: a metric for dynamic spectrum utilization. International Journal of Network Management, 16:103–113, March/April 2006. 30. W. Krenik and A. Batra. Cognitive radio techniques for wide area networks. In ACM IEEE Design Automation Conference 2005, pages 409–412, June 2005. 31. J. Lehtom¨ aki. Analysis of Energy Based Signal Detection. PhD thesis, University of Oulu, 2005. [Online]. http://herkules.oulu.fi/isbn9514279255/ isbn9514279255.pdf. 32. J. Lehtom¨ aki, M. Juntti, and H. Saarnisaari. Detection of frequency hopping signals with a sweeping channelized radiometer. In Asilomar Conference, pages 2178–2182, November 2004. 33. S. Mangold, A. Jarosch, and C. Monney. Operator assisted cognitive radio and dynamic spectrum assignment with dual beacons - Detailed evaluation. In COMSWARE 2006, pages 1–6, January 2006. 34. S. Mangold, Z. Zhong, K. Challapalli, and C.-T. Chou. Spectrum agile radios: Radio resource measurements for opportunistic spectrum usage. In GLOBECOM 2004, pages 3467–3471, November/December 2004. 35. P. Marshall. Policy languages, and specifically, How can we standardize the basis for extansible cognitive radio tools? SDR Forum Workshop on Cognitive radio, April 2004. 36. L.E. Miller, J.S. Lee, and D.J. Torrieri. Frequency-hopping signal detection using partial band coverage. IEEE Transactions on Aerospace and Electronics systems, 29:540–553, April 1993.

18 Spectrum Awareness: Techniques and Challenges

371

37. S.M. Mishra, A. Sahai, and R.W. Brodersen. Cooperative sensing among cognitive radios. In ICC 2006, June 2006. 38. R. Newgard, C. Conway, T. Journot, and J. Jiacinto. Broadband spectral sensing for dynamic spectrum allocation. In ISART 2006, pages 137–140, March 2006. 39. J. Park et al. Implementation issues of a wideband multi-resolution spectrum sensing (MRSS) technique for cognitive radio (CR) systems. In CrownCom 2006, June 2006. 40. J.M. Peha and S. Panichpabiboon. Real-time secondary markets for spectrum. In TPRC 2003, October 2003. 41. J.G. Proakis. Digital Communications. McGraw-Hill, New York, 4th, edition, 2001. 42. C. Raman, R.D. Yates, and N.B. Mandayam. Scheduling variable rate links via a spectrum server. In DySPAN 2005, pages 110–118, November 2005. 43. A. Sahai, N. Hoven, and R. Tandra. Some fundamental limits on cognitive radio. In Allerton Conference, pages 131–136, September-October 2005. 44. M.E. Sahin and H. Arslan. System design for cognitive radio communications. In CrownCom 2006, June 2006. 45. N. Sai Shankar, C. Cordeiro, and K. Challapalli. Spectrum agile radios: Utilization and sensing architectures. In DySPAN 2005, pages 160–169, November 2005. 46. M.I. Skolnik. Introduction to Radar Systems. McGraw-Hill, New York, 2002. 47. A. Sonnenschein and P.M. Fishman. Radiometric detection of spread-spectrum signals in noise of uncertain power. IEEE Transactions on Aerospace and Electronic Systems, 28:654–660, July 1992. 48. SPTF. FCC Spectrum Policy Task Force report. Report ET Docket no. 02-135, November 2002. 49. P. Stoica and R.L. Moses. Introduction to Spectral Analysis. Prentice Hall, New Jersey, 1997. 50. D.J. Thomson. Spectrum estimation and harmonic analysis. Proc. IEEE, 20:1055–1096, September 1982. 51. F.A. Tobagi and L. Kleinrock. Packet switching in radio channels: Part IIThe hidden terminal problem in carrier sense multiple-access and the busy-tone solution. IEEE Transactions on Communications, com-23:1417–1433, December 1975. 52. A. Tonmukayakul and M.B.H. Weiss. Secondary use of radio spectrum: A feasibility analysis. In TPRC 2004, October 2004. 53. H. Urkowitz. Energy detection of unknown deterministic signals. Proceedings of IEEE, 55:523–531, April 1967. 54. T. Weiss, J. Hillenbrand, A. Krohn, and F.K. Jondral. Efficient signalling of spectral resources in spectrum pooling systems. In 10th Symposium on Communications and Vehicular Technology in the Benelux, November 2003. 55. T. Weiss and F.K. Jondral. Spectrum pooling: An innovative strategy for the enhancement of spectrum efficiency. IEEE Communications Magazine, 42:S8– S14, March 2004. 56. P.D. Welch. The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Transactions on Audio and Electroacoustics, au-23:70–73, June 1967. 57. W. Willinger, M.S. Taqqu, R. Sherman, and D.V. Wilson. Self-similarity through high-variability: Statistical analysis of Ethernet LAN traffic at the source level. IEEE/ACM Transactions on Networking, 5:71–86, February 1997.

372

Marko H¨ oyhty¨ a et al.

58. J. Zhao, H. Zheng, and G.-H. Yang. Distributed coordination in dynamic spectrum allocation networks. In DySPAN 2005, pages 259–268, November 2005. ¨ 59. M. Oner and F.K. Jondral. Cyclostationary-based methods for the extraction of the channel allocation information in a spectrum pooling system. In IEEE Radio and Wireless Conference, pages 279–282, September 2004.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks Danijela Cabric and Robert Brodersen University of California, Berkeley [danijela|rb]@eecs.berkeley.edu

Summary. Recently, cognitive radios have been proposed as a possible solution to improve spectrum utilization via opportunistic spectrum sharing. Cognitive radios are considered lower priority or secondary users of spectrum allocated to a primary user. Their fundamental requirement is to avoid interference to potential primary users in their vicinity. Spectrum sensing has been identified as a key enabling functionality to ensure that cognitive radios would not interfere with primary users, by reliably detecting primary user signals. In addition, reliable sensing creates spectrum opportunities for capacity increase of cognitive networks. One of the key challenges in spectrum sensing is the robust detection of primary signals in highly negative signal-to-noise regimes (SNR). In this chapter, we present system design approaches to meet these challenges with techniques across physical and network layers. The design space is diverse as it involves various types of primary user signals, traffic patterns, and interference requirements. We start from the in-depth analysis of signal processing approaches and identify the regimes where these techniques are applicable. The analytical study is supported by experimental data that shows fundamental limits and practical gains possible to exploit using single radio sensing at the physical layer. In cases where sensing requirements cannot be met with a single radio sensor, network cooperation can be used to improve both robustness and sensing time. Using theoretical analysis we derive upped bounds on achievable gains and provide experimental results to show practical limitations in typical wireless environments. The goal of the chapter is to present a unified theoretical and practical system design view of spectrum sensing functionality together with guidelines for its implementation in a specific primary user band.

19.1 Spectrum Sensing for Cognitive Radio Networks 19.1.1 Requirements and Challenges New paradigm in spectrum sharing contrasts the conventional mechanism of accessing the spectrum based on spectrum allocation and licenses [5] [6] [7]. Since cognitive radios are secondary users of unoccupied spectrum they do not have a priori right to any frequency band. Their communication is strictly conditional on the

373 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 373–394. c 2007 Springer.

374

Danijela Cabric and Robert Brodersen

reliable detection of primary user transmissions in their vicinity. There are certainly a number of approaches that can be used to check the presence of primary user signals, like databases or beacons, but the only autonomous and flexible approach is based on measurements of the actual occupancy at a given location and time. Even in database and beacon approaches, spectrum sensing could add robustness and responsiveness to changes in the environment because it provides a real-time feedback. Therefore, we argue that spectrum sensing should be considered as an essential part of any cognitive radios system. Now, given that every cognitive radio should have some sensing capability, the question is what is the architecture and implementation of this radio component. Different spectrum utilization regimes might require different spectrum sensing architecture designs. The most flexible architecture scalable to regimes with higher spectrum utilization should allow a cognitive radio to simultaneously sense the spectral environment over a wide frequency band [2]. Within a wide band of interest there could be different types of primary user systems with various ranges of operation and traffic patterns. Based on these characteristics and actual sensing measurements cognitive radios can apply suitable spectrum sharing strategies. For example, packet based primary systems bands might be suitable for time sharing, while broadcast system offer frequency reuse spectrum sharing at non-overlapping geographic locations. More sophisticated approaches could use spatial dimension in addition to time and/or location. Even though there are a large variety of primary user systems, cognitive radio’s knowledge of their characteristics and requirements for interference protection can be abstracted by a few generally applicable parameters. Three critical requirements for sensing radio are the detection time, the probability of detection and the minimum detectable signal level. The required detection time and probability of detection are set by primary user tolerances to QoS degradation. While these are two conflicting requirements, a cognitive radio system goal is to minimize the detection time in order increase the time available to use the channel. The minimum detectable signal level is determined by a primary user receiver sensitivity and channel propagation environment between primary and cognitive systems [4]. For example, TV broadcast signals are much easer to detect than GPS signals, since the TV receivers’ sensitivity is tens of dBs worse than GPS receiver. Furthermore, a higher sensitivity in spectrum sensing could possibly allow cognitive radios to transmit at higher power.

19.1.2 System Design Options In general case, the key challenge of spectrum sensing is the detection of weak signals in noise with a very small probability of miss detection. As a result, a received signal is very likely to be below noise floor where classical detection and demodulation methods, used in positive SNR regimes, are not adequate. The new problem that cognitive radios face is the reliable detection in negative SNR regimes. On the other hand, a relaxed requirement is that signal does not need to be demodulated but only detected thus there is no need for synchronization, decoding or protocol compatibility. This setting opens the room for new signal processing techniques that provide sufficient processing gain to improve radio sensitivity to the levels where detection becomes reliable. Furthermore, it can allow a primary user’s identification based on knowledge of its signal characteristics.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

375

From the system design perspective, multiple radios need to simultaneously perform spectrum sensing and jointly decide on available spectrum opportunities through cooperation [3]. The system can exploit inherent diversity in individual measurements and improve detection by relying on users with favorable channel conditions. However, the network topology determines the required number of users and their optimal spatial distribution. The gain from user cooperation involves the trade-off between the overhead of the measurement exchange through a common control channel. Both latency and capacity of the control channel need to be minimized and scalable to large networks.

19.2 Signal Processing Techniques for Spectrum Sensing 19.2.1 Simple General Approach - Energy Detector The most basic approach for detecting signals in the presence of noise is based on energy measurement. It is particularly simple as it uses a non-coherent processing, thus applies to any signal type. It requires minimum information about the signal, including only signal bandwidth and carrier frequency. In communications and signal processing literature, energy detection is studied as a hypothesis testing problem and performance is measured by a resulting pair of detection and false alarm probabilities (Pd , Pf a ). While the high SN R regimes have been very well understood, it is unclear if the same analysis and performance apply in the highly negative SN R regime. Let us revisit the theoretical analysis first. The detection is the test of the following two hypotheses: H0 : y[n] = w[n] n = 1, ..., N

(19.1)

H1 : y[n] = s[n] + w[n] n = 1, ..., N

(19.2)

where N is the observation interval. Both signal s[n] and noise w[n] samples are modeled as independent Gaussian 2 random variables with zero mean and variance σx2 and σw , respectively. A decision statistic for energy detector is: ε(y) =

N X

y[n]2

(19.3)

n=1

The detection is performed by a threshold test on the measured energy. There are number of ways that a threshold can be set. In the spectrum sensing situations, γ

−N

threshold γ is set to meet the fixed Pf a = Q( σ√2 2N ). Note that the threshold depends only on the receiver noise. If the number of samples used in sensing is not limited, an energy detector can meet any desired Pd and Pf a simultaneously. The minimum required number of 2 samples is a function of the signal to noise ratio SN R = σx2 /σw : N = 2[(Q−1 (Pf a ) − Q−1 (Pd ))SN R−1 − Q−1 (Pd )]2

(19.4)

In the low SN R << 1 regime, number of samples required for the detection, that meets specified Pd and Pf a , scales as O(1/SN R2 ).

376

Danijela Cabric and Robert Brodersen

Unless the primary user system continuously broadcasts signal, the time available for sensing is limited by a duty cycle in primary user’s transmission. In that case, a cognitive radio can only use a fixed number of samples N to detect primary user signal. Probability of detection depends on chosen probability of false alarm Pf a , sensing interval N and signal to noise ratio SN R as: ! r 1 N SN R −1 Pd = Q Q (Pf a ) − (19.5) 1 + SN R 2 1 + SN R In terms of implementation, there are a number of choices for energy detection based sensors. The main design goal is to optimally filter the signal bandwidth and minimize the contribution of the out-of-band noise and interfering signals. Analog implementations require analog pre-filter with fixed bandwidth which becomes quite inflexible for simultaneous sensing of narrowband and wideband signals. Digital implementations offer more flexibility by using a FFT based spectral estimates. This architecture inherently supports various bandwidth types and allows sensing of multiple signals simultaneously. The size of the FFT is the critical parameter because larger FFT size improves the bandwidth resolution and detection of narrowband signals but at the same time increases the sensing time. Figure 19.1 presents the architecture for wideband energy detector spectrum sensor.

Figure 19.1. Digital implementation of wideband energy detector.

Energy detectors have simple implementation, but can not provide robust performance in highly negative SN R regimes. Our experimental study show that the theoretically predicted performance holds for SN Rs above −20 dB. However, below −20 dB SN R the detection becomes progressively harder and at −23 dB signal cannot be detected regardless of the sensing time duration (Figure 19.2). This deviation from theoretical prediction comes from the fact that noise variance in actual systems is not precisely known and could vary over time. This becomes particularly important when the signal strength is below the estimation error in the noise variance since. In that case, the detection threshold is set too high and weak signals could never be detected. Quantitatively, if there is a residual xdB noise estimation error, then the detection is impossible below SN Rwall = 10 · log10 [10(x/10) − 1]dB [8]. In addition to noise estimation error, there could be other interference sources in adjacent bands that due to the spectral leakage would add additional noise and further increase the minimum detectable signal.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

377

Figure 19.2. Performance of energy detector in negative SNR regime.

19.2.2 Exploiting Deterministic Signals - Coherent Processing Pilot Detection While energy detection is fairly general approach, it neglects the presence of deterministic signals that primary users embed in their transmission in order to perform synchronization and acquisition, for example. A special case of a pilot signal, frequently present in primary user systems, is a sinewave tone used for receiver synchronization. The power of the pilot tone is typically 1% to 10% of the total transmitted power. The benefit of pilot signals, if they are perfectly known to cognitive radio sensor, is that it can be processed coherently. Coherent processing achieves the best possible robustness with respect to noise. The optimal detector is the matched filter that projects the received signal in the direction of the pilot: T =

N X

y[n]xp [n]∗

(19.6)

n=1

Theoretical analysis shows that the required number of samples is a function of 2 the pilot SN Rp = σp2 /σw : N = [Q−1 (Pf a ) − Q−1 (Pd )]2 SN Rp−1

(19.7)

Since matched filter uses the optimal processing, it could achieve detection with the minimum possible number of samples. Therefore, the scaling law of N 1/SN Rp gives a lower bound on the sensing time performance for any possible sensing detector type. However, if pilot is weak with respect to the energy of the signal then the

378

Danijela Cabric and Robert Brodersen

required number of samples can be higher than for the energy detector. A matched filter is typically implemented in the digital domain, and its realization is illustrated in Figure 19.3. In order to obtain gains predicted by theoretical results, a sensor requires perfect synchronization with primary user pilot. In all practical receivers with noisy oscillators and circuitry, it is almost impossible to achieve perfect synchronization even in positive SN R. Now, the question is how much degradation is caused by the residual frequency offset. Let consider the sinewave pilot: xp [n] = Aej(w0 n+ϕ) . Suppose there is a frequency offset between the primary transmitter and sensing receiver: x fp [n] = xp [n]ejwn (19.8) Then, the decision statistic becomes: T˜ =

N X

y[n]xp [n]∗ e−jwn ∼ A∗ e−jϕ

n=1

N X

e−jwn

(19.9)

n=1

Figure 19.3. Implementation of coherent pilot detector. If the sensing interval N becomes comparable or larger than the period of the frequency offset (2π/w), then coherent processing gain decreases and eventually becomes equal to zero. Therefore, in the presence of frequency offsets the pilot detection has limits on sensing time and detectable signal levels (see Figure 19.4). This analysis was confirmed by an experimental study performed on the sinewave pilot at 2.4 GHz. Although the minimum detectable signal levels are far below the energy detector sensor, the limits are set by fairly tight frequency offset requirements (10 Hz at 2.4 GHz is equivalent to 0.04 ppm). In the next section we discuss approaches that can benefit from coherent processing of pilots but avoid tight synchronization requirements.

Mixed Approaches Consider the detector that performs two stage processing. First, it correlates the received signal with a known sequence of a period of time M , and then non-coherently averages these coherent pieces K times. The decision statistic now becomes: T =

K X M X [ y[m]f xp [kM + m]∗ ]2 k=1 m=1

(19.10)

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

379

Figure 19.4. Performance of pilot detector in negative SNR.

The total detection time is N = KM . Now probability of detection depends on the choices for M and K. It can be shown that: s Q−1 (Pf a ) K Pd = Q( p − (19.11) M SN Rp ) 2(1 + 2M SN Rp ) 1 + 2M SN Rp Under assumption that K is sufficiently large we can identify 3 limiting regimes based on the length of coherent processing M . Regime 1 if M SN Rp << 1 then r K −1 Pd = Q(Q (Pf a ) − (19.12) M SN Rp ) 2 which scales with K same as in energy detector case. The impact of M is in the increase of the SN R by logM in dB. In addition, it effectively moves the SN Rwall by logM dB. Regime 2 if M SN Rp >> 1 then r Q−1 (Pf a ) KM SN Rp Pd = Q( p − ) (19.13) 4 2M SN Rp In this regime detector behaves similar to the optimal coherent detector. Regime 3 if M SN Rp ≈ 1 the detector is in a transient regime, between coherent and non-coherent processing. The scaling of the sensing time with respect to M is presented in Figure 19.5. The special case where this approach is particularly convenient is partially coherent sinewave pilot sensing. The optimum matched filter can be realized using the

380

Danijela Cabric and Robert Brodersen

Figure 19.5. Performance of mixed approach detectors in negative SNR.

FFT with the length equal N. However, in practice it is expensive to implement large size FFTs and all efficient implementations have length that is power of 2. It is rather desirable to have fixed and short FFT processor that would perform partially coherent processing and then use the non-coherent averaging. The experimental study results presented in Figure 19.6 show the improvement obtained by partially coherent processing using longer FFT [9].

Packet Detection There is a large class of primary user systems that transmits signal in discontinuous time intervals using packets. Based on the proportion and regularity between idle and busy periods a cognitive radio system can establish the time domain spectrum sharing with such primary user systems. The main limitation in the sensing functionality is the time available for sensing, which is set by the packet length. While the energy in the packet is constrained, there is a deterministic part of the packet that might facilitate its coherent detection. It is almost general case that every packet has a fixed preambles sequence at the beginning in order to facilitate packet acquisition and synchronization. Preamble data is typically a pseudo noise sequence which is known to have good autocorrelation properties. Therefore, if a cognitive radio sensor knows the packet structure then it can perform coherent processing and improve the detection. Clearly, there are several approaches to perform the detection. One approach is to only perform energy detection, second would be to coherently process a preamble, while the third approach would jointly detect preamble and then energy in the packet. An optimal approach would take all three detectors decisions into account.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

381

Figure 19.6. Performance of mixed approach detectors in negative SNR.

In the combined detection approach, joint probability of detection is a function of individual detection per each processing stage. Furthermore, there is an optimal trade off between the probabilities of detection and false alarm in different stages of the processing that maximizes total probability of detection. Figure 19.7 shows that energy detector is a good candidate in SN Rs above −10 dB and short packet lengths (∼1000 symbols). Preamble detectors are superior in lower SN Rs but their performance strongly depends on the preamble size. Combined detector provides additional performance gain needed for robust detection.

19.2.3 Detecting Signal Features - Cyclostationary Processing Previous discussion shows that coherent processing always provides robust detection and if there is enough energy in the deterministic signal portion the sensing time is minimized. However, truly coherent signal processing comes at the price of synchronization. A natural question arises: is there any other information that cognitive radio sensor can exploit from primary signals to exploit some coherent processing? Some intuition can be gain by understanding what information is thrown away by a non-coherent energy detector. In essence, the energy detector threats noise and signal in the same way, as wide-sense stationary white processes. However, modulated signals are in general coupled with sine wave carriers, pulse trains, repeating spreading, hoping sequences, or cyclic prefixes which result in built-in periodicity. Even though the data is a wide-sense stationary random process, these modulated signals are characterized as cyclostationary, since their statistics, mean and autocorrelation, exhibit periodicity. This periodicity is introduced intentionally in the signal format so that a receiver can exploit it for: parameter estimation such as carrier phase, pulse

382

Danijela Cabric and Robert Brodersen

Figure 19.7. Comparison of different approaches for packet detection in negative SNR.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

383

timing, or direction of arrival [9]. This information can then be used for detection of a random signal with a particular modulation type in a background of noise and other modulated signals. Common analysis of wide-sense stationary random signals is based on autocorrelation function and power spectral density. On the other hand, cyclostationary signals exhibit correlation between widely separated spectral components due to spectral redundancy caused by periodicity. By analogy with the definition of conventional autocorrelation, one can define spectral correlation function (SCF):

Sxα (f ) = lim

Z

∆t/2

lim

T →∞ ∆t→∞

−∆t/2

1 XT (t, f + α/2)XT∗ (t, f − α/2)dt T

(19.14)

where finite time Fourier transform is given by: Z

t+t/2

XT (t, f ) =

x(u)e−j2πf u du

(19.15)

t−T /2

Spectral correlation function is also termed as cyclic spectrum. Unlike power spectrum density, which is real-valued one dimensional transform, the spectral correlation function is two dimensional transform, in general complex-valued and the parameter α is called cycle frequency. Power spectral density is a special case of a spectral correlation function for α = 0. The distinctive character of spectral redundancy makes signal selectivity possible. Signal analysis in cyclic spectrum domain preserves phase and frequency information related to timing parameters in modulated signals. As a result, overlapping features in the power spectrum density are non-overlapping feature in the cyclic spectrum. Different types of modulated signals (such as BPSK, QPSK, SQPSK) that have identical power spectral density functions can have highly distinct spectral correlation functions. Furthermore, stationary noise and interference exhibit no spectral correlation. The sufficient statistics used for the detection are obtained through non-linear squaring operation, therefore cyclostationary detectors also fall in the category of non-coherent detectors in terms of sensing time requirements. Given N samples divided in blocks of TF F T samples, spectral correlation function is estimated as: Sxα (f ) =

N 1 X XTF F T (n, f + α/2)XT∗F F T (n, f − α/2) N T n=0

(19.16)

where XTF F T (n, f ) is the NF F T around sample n. In case of high SN R, both power spectral density and spectral correlation function allow reliable detection. Furthermore, spectral correlation function allows modulation recognition based on the pattern of cycle frequencies. On the other hand in low SN R energy detector is highly susceptible to noise variations while distinct features of QPSK modulation in a spectral correlation function is preserved and widely separated from noise correlation. The performance of the feature detector also depends on how much energy a feature contains. Different modulation schemes have different features and energy associated with them. Here, we will take a common example of a class of amplitude modulated signals and analyze the feature associated with the symbol rate represented as:

384

Danijela Cabric and Robert Brodersen

x(t) =

∞ X

an q(t − nTs )

(19.17)

n=−∞

where an is data sequence, q(t) is a pulse shaping filter and Ts is a symbol rate. The cyclic spectrum of x(t) is given by: Sxα (f ) =

1 Q(f + α/2)Q∗ (f − α/2) 2Ts

(19.18)

for α = 1/Ts . Q(f ) is a pulse shaping filter in Fourier domain. In case of most commonly used square root raised cosine filter, the feature spans over: 1−β 1+β
(19.20)

Sxα = Sxα e−j2παt0

(19.21)

Since cyclic spectrum is a complex function, a time offset has different effect on it than on a power spectrum density, which is a real function. Now assume that there is a residual sampling clock offset ∆ from the ideal sampling clock T0 , i.e. T −T0 = ∆ where T is the sampling clock interval. Then, in the process of estimation of spectral correlation function every estimate obtained by correlation of FFT bins will have a phase offset. Lets assume that N FFTs are used to estimate Sexα (f ).

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks N 1 X α Sx (i, f ) Sexα (f ) = N i=1

385

(19.22)

where each FFT serves to estimate Sxα (f ) =

1 XTF F T (n, f + α/2)XT∗F F T (n, f − α/2) TF F T

(19.23)

Now, in the presence of clock offset estimations, each FFT block will have a time offset with respect to one another: Sxα (f ) = Sxα (i, f )e−j2παti

(19.24)

where ti = iNF F T δ and Sxα (i, f ) is the estimate with perfect sampling. In the process of averaging, these estimates are non-coherently added and resulting cyclic spectrum has the attenuated features: N X

sin(2πα∆NF F T N/2 −j2πα∆NF F T (N +1)/2 e sin(2πα∆NF F T /2 i=1 (19.25) The feature at α is smeared if ∆ ≈ 1/αN NF F T . Figure 19.8 illustrates SCF with perfect sampling and with sampling offset. Sampling time offset ∆ can be expressed in terms of sampling clock frequency offset δ as: α fα Sf x (f ) = Sx (f )

e−j2παti = Sxα (f )

δ (19.26) α(α − δ) Similar to the coherent pilot detector case the number of samples that can be used for sensing is limited to ∆=

α (19.27) δ The cancellation of the features due to sampling clock offset can be prevented by performing partially coherent feature processing. Based on the maximum expected sampling offset δmax , number of coherent averages M1 is chosen to be smaller than α/δmax NF F T and then two stage processing is performed as in the mixed pilot detector approach: N NF F T <

Sxα (f )0 =

1 TF F T

M1 M2 X X | X(k + mM2 , f + α/2)X ∗ (k + mM2 , f − α/2)|

(19.28)

m=1 k=1

where the total number of averages is N = M1 M2 . Feature detectors are always implemented in digital domain. Direct algorithms first compute the spectral components of the data through FFT, and then perform the spectral correlation directly on the spectral components. The computational complexity of a spectral correlation function estimator is easily estimated. For a stream of NF F T samples it requires a computation of NF F T point FFT, which requires NF F T ∗ logNF F T multiplications, and NF2 F T multiplication for cross multiplications. Note that this algorithm is extremely parallel so that the computation of

386

Danijela Cabric and Robert Brodersen

Figure 19.8. Spectral correlation function of 4 MHz QPSK signal with perfect sampling (top) and with 100Hz sampling offset (bottom).

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

387

the spectral correlation function can be organized across frequency or across cycle plane independently. Figure 19.9 presents the cyclostationary feature detector implementation that is robust to sampling offsets. The first stage averages SCF in the complex domain so that features are coherently added and the noise is cancelled. The second stage averages magnitude of the output of the first stage, and therefore it changes the processing from coherent to non-coherent. The output of the detector is obtained through integration of the energy in the SCF that lies in (α, f ) space where signal of interest has theoretically predicted features.

Figure 19.9. Implementation of a cyclostationary feature detector robust to sampling clock offsets.

Figure 19.10 presents the experimental results that compare feature detectors with energy detector in the presence of sampling clock offsets. Under stationary white noise, feature detectors (even with perfect sampling) have a performance loss with respect to energy detector. This is due to the fact that energy contained in the feature is related to the pulse shaping filter roll-off, as described by Eq. 19.16. In case of β = 0.5 the loss is approximately 6 dB. On the other hand, a sampling clock offset of 100 Hz at 64 MHz makes the detection of signals below −15 dB SNR impossible. However, once the proposed 2 stage averaging is deployed the detector achieves the desired probability of detection and false alarm at the penalty of increase detection time. The number of averages in the first stage is chosen from the Eq. 19.24. For SN Rs below −15 dB, the number of averages in the second stage has to be increased with respect to perfect sampling feature detection. The proposed scheme performs comparable even for sampling offsets of 1 KHz. It has been shown that energy detector is highly susceptible for the noise variance uncertainty that is contributed by temperature variations and out-of-band interference [7]. To test the robustness of feature detectors, we experimented with the adjacent channel interference coming from the commercial 802.11g WLAN with a continuous traffic generated by video camera data transfer between two laptops. Figure 19.11 shows the performance of both energy and feature detectors in the

388

Danijela Cabric and Robert Brodersen

Figure 19.10. Performance of cyclostationary feature detectors in negative SNR.

presence of adjacent band interference. Due to spectral leakage of the FFT, energy detector suffers from the large variation in the noise-plus-interference level. This variation progressively degrades the energy detector performance and at −18 dB SN R detection becomes impossible. On the other hand, feature detector robustly detects the weak signals and outperforms the energy detector. Note that there is a slight degradation in performance of feature detector as well due to leakage of the interference signal in SCF domain.

19.3 Network Level Techniques 19.3.1 Exploiting Diversity - Cooperative Sensing Up to this point we have considered spectrum sensing performed by a single radio in noise dominated channels. In fading channels, however, single radio sensing requirements are set by the worst case channel conditions introduced by multipath, shadowing and local interference. These conditions could easily result in SNR regimes where the detection would not be possible or could require unacceptably long time for the detection. However, there is inherent channel variability of signal strength at various locations therefore not all cognitive radios will experience the worst channel conditions. The simple way to exploit this channel diversity is to allow multiple radios to share their individual sensing measurements so that ones that have good channel conditions can provide reliable sensing for the whole network. The gains obtained through cooperation could be quantify through improvement of overall probability of the detection or through a decrease in required sensing time.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

389

Figure 19.11. Features of desired QPSK signal and the adjacent 802.11g signal (top) and detector performance in non-stationary noise due to adjacent band interference (bottom).

390

Danijela Cabric and Robert Brodersen

Under independent fading conditions often assumed for multipath, if radios are more than λ/2 apart, cooperation can be studied as a diversity gain in multiple antenna channels. Simple probabilistic analysis shows that if n radios combine independent measurements, then probability of detection of the system QD monotonically increases as QD = 1 − (1 − Pd )n . Unfortunately, the probability of false alarm for the system QF also monotonically increases as QF = 1 − (1 − Pf a )n . Another benefit of cooperation is that probability that every radio experience a deep fade decreases. If the combining rule is “OR”, meaning that it is sufficient that only one radio detects a signal, then the operating SN R is effectively the SN R of the best user. All radios that operate in SN Rs bellow the best users SN R will not be able to detect. As a result, target detection sensitivity can now be achieved by having less sensitive individual sensors. From the signal processing analysis, we have seen higher SN R corresponds to shorter sensing time. In case of energy detector, the reduction in the sensing time can be significant, as for 10 dB SN R improvement nominal sensing time can be decrease 100 times. We tested these hypotheses in typical office indoor environment channels. Each sensing radio used a simple energy detector with short sensing time (N=1000 samples) to detect both sinewave pilots and modulated signals. Figure 19.12 shows that with only few radios the detection achieves high reliability. Cooperation gains are higher for wideband signal as they can exploit frequency diversity in addition to user diversity. Narrowband sinewave sensing is dominantly affected by multipath fading.

Figure 19.12. Cooperative sensing gains for sinewave pilot signal and wideband QPSK signal.

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

391

19.3.2 Limitations in Cooperative Sensing The collaboration gain is maximized only if the radios exhibit independent fading channels. However, fading could be caused by shadowing that exhibits high correlation if two radios are blocked by the same obstacle. Commonly, a shadowing correlation is described by the coefficient ρ and modeled as an exponential function of distance: ρ = e−ad . Measurements of the shadowing in indoor environments show that the correlation coefficient is independent of wavelength over a frequency octave, but it is dependent on the topography. It was estimated that 90% correlation distance is typically 1 m, 50% is around 2 m, and slowly decays to 30% over 8 m. Therefore, in the limited area, increasing the number of radios introduces the correlation, which in effect limits the collaborative gain. Experimental results for indoor environments show that separation between the radios should be proportional to the size of the largest shadowing objects [1]. In particular, for typical office environments it is in the order of several meters (Figure 19.13).

Figure 19.13. Cooperation gain as a function of distance (2 radios).

In addition to radios separation, the important design parameter is the detection threshold used by each cooperating radio (see Figure 19.14). Two different types of threshold rules in the local decision process can be used: 1) a predetermined (fixed) threshold set by the centralized processor or 2) an independently estimated threshold based on the local noise and interference measurements. In the case of stationary environments with all radios being identical, these two rules would result in the same system performance. However, due to the presence of ambient interference caused by primary or cognitive radios in the vicinity, and local noise, temperature, and circuit variability, each radio sees different aggregate noise and interference.

392

Danijela Cabric and Robert Brodersen

This observation suggests that a fixed threshold might be suboptimal, and that in practical situations the estimated threshold would provide robustness and better gains. Experimental study shows we analyze that periodic estimation of noise and interference is critical. Suboptimal fixed thresholds introduce a gap of up to 25% in probability of detection. On the other hand, the estimation of threshold in each radio requires additional sensing time and more complex sensing receiver.

Figure 19.14. Cooperation gain as a function of threshold.

19.4 System Design Guidelines for Spectrum Sensing In the previous sections, we explored the fundamental trade-offs in the sensing problem involving sensing time and sensing sensitivity, i.e., minimum SN R where the signal can be reliably detected. We found that the primary user signal characteristics like active signal time available for signal processing, percentage of deterministic signal energy, and redundancy of the modulation present generalized parameters for optimal selection of spectrum sensing approach for the required SN R. The presented framework provided a classification of sensing schemes using coherent, non-coherent or mixed signal processing and their scaling of sensing time with respect to SN R. Taking practical receiver design constraints, we showed the limits on achievable SN Rs and maximum sensing time. From these findings, we propose the architecture of the physical and network layer spectrum sensing functionality (Figure 19.15). Since energy detector can quickly and reliably identify presence of strong primary user signals in the fairly high SN R regimes, it is suitable for selection of candidate bands where more elaborate processing should be done. Unfortunately,

19 Robust Spectrum Sensing Techniques for Cognitive Radio Networks

393

Figure 19.15. Cooperation gain as a function of threshold.

energy detector can not be used for reliable sensing in highly negative SN R regimes due to SN Rwall caused by estimation errors or variations in noise and interference. In addition, when the time available for sensing is limited, energy detector sensitivity level could be insufficient for the primary user requirement. For faster detection or higher sensitivity, energy detection should be combined with more sensitive coherent or mixed approaches. Exploiting coherent processing gain for deterministic pilots or preambles is beneficial if they contain at least a few percent of signal energy. However, achieving high sensitivities require longer processing time and then even small frequency offsets diminish the coherent processing gain. Mixed approaches provide a robust but yet simple way to exploit partial coherent processing gain and overcome the synchronization requirements. As a side effect, mixed approaches result in an increase of the required sensing time. In the absence of deterministic signals, sensing receiver can exploit modulation inefficiency and detect signal features. Although in principle non-coherent methods, feature detectors allow signal recognition and provide robustness to non-stationary noise and interference channel conditions. The sensing times of feature detectors are comparable to energy detectors sensing time within a constant factor. Their main advantage is that there is no limit in the achievable sensitivity, i.e., no SN Rwall . Their weakness is the sensitivity to sampling clock offsets, but it could be overcome by simple mixed processing approach. In order to guarantee high probability of detection and meet primary user requirements with minimized sensing time, cognitive radio networks should exploit user diversity using cooperation. Cooperation requires means of sharing sensing information among cognitive radios. For example, a common control channel can coordinate sensing of the group of cognitive radios. Using a control channel, a cognitive radio network should implement a designated sensing protocol in parallel with the regular medium access protocol. This is not only necessary but it could be rather critical in the design of cooperative sensing functionality. An additional information exchange among radios introduces latency which must be incorporated into overall system time. In the worst case, a sensing protocol overhead scales linearly with the number of radios and number of sensed frequency bands. Protocol latency is also proportional to the throughput of the control channel. Since control channel requires a physical communication channel it cannot be expected to have

394

Danijela Cabric and Robert Brodersen

a large throughput. In addition, simple sensing techniques such as energy detection require cannot distinguish transmission from primary systems from other cognitive radios. Therefore, when sensing takes place all cognitive radios in the area should resume transmission. This imposes additional silent time within and coordination with medium access protocol. The underlying signal processing on the physical layer determines what kind of information radios can share through cooperation. In the simplest case, when all radios use energy detection, reported information could be either hard decision or soft decision. In the case of hard decision reporting, the overall traffic is minimized as radios send only one bit per sensed frequency band. Sharing soft decision enables optimum detection, and allows identification of radios with poor channel and noise conditions. Furthermore, soft decisions allow estimation of the correlation between radios which can greatly minimize number of participants. In case coherent or feature detection is used, soft decisions provide information what radios perform robust and reliable detection. Furthermore, these techniques do not require quite periods. Further research is needed for development of flexible and scalable cooperation protocols that can optimally exploit physical layer spectrum sensing capabilities.

References 1. Cabric D., Tkachenko A., and Brodersen R. W. Spectrum Sensing Measurements of Pilot, Energy, and Collaborative Detection. In MILCOM, Washington, October 2006. 2. Cabric D., Mishra S. M., Willkomm D., Brodersen R. W., and A. Wolisz. A Cognitive Radio Approach for Usage of Virtual Unlicensed Spectrum. In IST Mobile Summit 2005, Germany, June 2005. 3. Cabric D., Mishra S.M., and Brodersen R. W. Implementation Issues in Spectrum Sensing for Cognitive Radios. In Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, November 2004. 4. Cabric D. and Brodersen R. W. Physical Layer Design Issues Unique to Cognitive Radio Systems. In PIMRC, Germany, September 2005. 5. FCC. Spectrum Policy Task Force Report. In ET-Docket No. 02-155, November 2002. 6. FCC. Notice of Proposed Rule Making. In ET-Docket No. 03-322, December 2003. 7. FCC. Notice of Proposed Rule Making. In ET-Docket No. 04-113, May 2004. 8. Tandra R. and Sahai A. Fundamental Limits on Detection in Low SNR. In WirelessComm Symposium on Signal Processing, Maui, June 2005. 9. Gardner W.A. Signal interception: performance advantages of cyclic-feature detectors. In Signal interception: performance advantages of cyclic-feature detectors, Maui, January 1992.

Part IV

Marrying Cooperation and Cognition in Wireless Networks

20 Cognitive Resource Manager A Cross-Layer Architecture for Implementing Cognitive Radio Networks

Marina Petrova and Petri M¨ ah¨ onen RWTH Aachen University, Department of Wireless Networks, Germany [mpe|pma]@mobnets.rwth-aachen.de Summary. Resource allocation and cross-layer design in the case of wireless networks have been mostly focused on well-defined and fixed algorithmic approaches with varying degrees of adaptivity. However, the complexity of the systems is increasing very rapidly and especially due to emergence of different cognitive radio approaches, more flexible and adaptive solutions for system optimization and parameter tuning are required. We argue that the principle of cognitive radio design can be used overall at the systems level. In this chapter, we introduce a new Cognitive Resource Manager (CRM) framework that is currently under the early development and research phase. This paper is focusing on describing the framework and architectural structure of the CRM. We further suggest that using machine learning based optimization in the flexible architectural context is a very promising approach. Several challenges including decision making under imperfect knowledge or missing information as well as coping with imprecise utility function definitions are also discussed. Finally, we address the issues of managing different time-scales and scheduling of the processes in the CRM implementation.

20.1 Introduction The trend of providing various technical solutions for making the wireless embedded system self-configurable, context aware and optimized with respect to, e.g., throughput, delay, spectrum efficiency, energy, QoS and scalability has become very popular in the last few years. Recently in the dynamic spectrum management domain the principle of spectrum agility, a.k.a. cognitive radio, has been introduced as new enabler for optimization. The original cognitive radio approach was introduced by Mitola in [35], where the main idea is to make the radios smart and able to learn, decide and act based on the environmental perceptions. Following the Mitola’s groundbreaking work, we apply the cognition principle in the whole system in order to enable cross-layer optimization and adaptive resource allocation. By cognition we mean ability of the system to use context information and machine learning techniques in the process of adaptation, auto-configuration and optimization. Being aware of the practical complexity issues and difficulty on taking in account widely different event time-scales of varying layers as pointed out

397 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 397–422. c 2007 Springer.

398

Marina Petrova and Petri M¨ ah¨ onen

by Kawadia and Kumar [29], we have chosen to build a full optimization software framework for cognitive radios, namely Cognitive Resource Manager (CRM). This is in small part similar approach to classical Radio Resource Manager (RRM) but the scope of our CRM is broader and most importantly the system is machine learning based. The cognitive cycle of our framework is in part inspired by the behavioural model of the cognitive radio and it is shown in Figure 20.1.

Figure 20.1. Behavioural model of the cognitive radio adopted as a reference model for the CRM implementation (in part modified from [35]).

One of the main motivations for our CRM work has been to progress the cognitive radio work towards all layers approach. With this work we want to establish a solid base for building a Cognitive Wireless Network (CWN), since a large part of the previous cognitive radio work has been still concentrating to the physical layer (see recent interesting work in the field reported in [26]). The clever crosslayering requires the view of the whole network stack, but we argue like others (e.g., [15, 31, 35]) that the machine learning should be an integral part of any autoconfigurable and adaptive system that tries to manage extreme complexity. Depending on the implementation complexity we foresee three modes of operation of the CRM framework. It can run on the terminal (laptop or a mobile phone) in a local manner, taking care of the local parameter and protocol optimization or in a centralized mode (e.g., running in an access point) where the cognitive cycle is executed only in the central point, which has control over several clients. The ultimate goal is to deploy the CRM in a distributed manner where the cooperation between the terminals will be enabled.

20 Cognitive Resource Manager

399

The ISM-band technologies and especially the IEEE 802.11 is our first target to test the framework. This is motivated by two facts. First, the current state-of-the art wireless technologies for short-range data communications are mostly operating in the unlicensed ISM-bands. The lack of cooperation between the various technologies as well as self-management and self-configuration will obviously lead to drastic degradation of the basic QoS due to interference and congestion in dense environments. Additionally, transparent exchange of information between different entities in the system, such as protocol stacks, operating systems and applications, is almost non-existent. This hinders the possibilities for smarter optimization. Hence, the possible gain achieved by using CRM in 802.11 domain is very high. Second, the IEEE 802.11 is very open for experimentation, and we have a quite free access to MAC-layer and PHY-layer parameters. The CRM concept is, as far as we are aware of, quite novel. There is a similar approach introduced by Virginia Tech., known as Cognitive Engine (CE) [44]. Both approaches have been introduced independently at the same time-frame. There are also some clear differences between the two approaches, ours being currently targeting more towards higher OSI-layers and cooperative networking based. Both groups are using, at least in part, also gnuRadio platform [1] for experimentation and prototyping. Similar type of work has been also started by Nolan et al. at Trinity College Dublin in Ireland [36]. In this chapter, we use the opportunity to also report and describe at the general level the on-going research work on implementing and testing CRM concept with Cognitive Radios. This is done to provide more concrete examples and also to show that in part we already have more than “just” ideas. These examples exhibit some of the early results and learned lessons from the prototype implementation work. We believe that early prototyping is a very necessary part of the research in this domain, in order to be able to not only test performance, but also formulate research problems more concretely.

20.2 CRM Framework In this section we introduce the concept of the Cognitive Resource Manager (CRM) as a main enabler of the future cognitive wireless networks. We see it as a vital architectural component which coupled with several other architectural blocks such as well-defined interfaces and a toolbox of advanced optimization techniques creates the so called CRM framework shown in Figure 20.21 . The key motivation for designing the CRM framework is to facilitate the system auto-configuration and optimization using learning and reasoning techniques inspired by the machine learning and artificial intelligence (AI) community. In fact the CRM framework should reflect an alternative way of performing cross-layer optimization compared to the solutions available so far. Achieving cross-layer optimization has been a research problem for long time now. Finding optimal end-to-end communication throughput and fairness in the 1

It is important to note that the Figure 20.1 shows only the functional components of the CRM framework and the possible interfaces between these components. The data flow and the decision and modeling processes are not presented in this figure.

400

Marina Petrova and Petri M¨ ah¨ onen

network, reducing latency and achieving better utilization of the available resources have been classical optimization targets so far. Instead of optimizing the operation of the separate layers individually cross-layer optimization is trying to find the best parameter settings using an exchange of information between the adjacent layers of the protocol stack. Strictly speaking, this concept is breaking the traditional layered approach of the OSI protocol stack where each layer is supposed to operate independently. However, so far, enabling interaction between the layers and still keeping the separation between them has been a good compromise. Still, because of the layered nature of the protocol stack there is no common way to exchange information directly between layers which are not adjacent. For example it would be extremely beneficial if the TCP had the ability to directly check the quality of the wireless link from the PHY before making a decision to slow down the transmission over the wireless link presuming that collisions have occurred. If the channel quality is bad it could mean that the packets are lost because of the erroneous channel and not because of a packet collisions. Understanding the wireless link quality better will help TCP not to cause extra degradation of throughput performance [46]. Enabling a free flow of information through the protocol stack requires well-defined interfaces and eventually a new system architecture. The CRM framework is built so that it allows the information from the different layers and additionally from the environment (measurements and observations) to be available for optimization purposes wherever it is needed.

Figure 20.2. Functional blocks and interfaces in the CRM framework.

In the following we will walk through the main components of the framework and highlight their functionalities. The CRM itself is a software entity that based on the information it gets from the application layer, the underlaying data link and

20 Cognitive Resource Manager

401

networking layers as well as from the operating system, can perform cross-layer optimization and efficient resource management using adequate advanced reasoning methods. As such it could, for example, optimally manage spectrum resources, flexibly adapt link parameters and allow the best possible settings for the applications running on top. The cognitive component in the system comes from the machine learning and AI algorithms which will provide the system with the capability to learn and reason. In fact, the “learning from experience” process will generate valuable knowledge that can be used in the process of optimization to improve the system performance. Introducing the learning capability and intelligence into the optimization process is a step forward in building auto-configurable and high-performance wireless networks. As shown in Figure 20.2 the decision process in the CRM is heavily dependent on the information and triggers it gathers from the relevant OSI-layers, on the one hand, and the set of available optimization and reasoning techniques on the other hand. Based on this knowledge the CRM can take appropriate actions that could, e.g., adapt MAC parameter settings or start certain tasks such as a reassignment of frequency channels. The optimization and the decision making process will be carried out using a set of advanced techniques and algorithms placed inside a toolbox. Depending on the task CRM should be able to chose among the available adequate learning, modeling or decision making techniques. While the classical cross-layer optimization takes into account only the information exchanged between the neighbouring layers in the protocol stack, the CRM framework offers more open optimization approach. The relevant data from the separate layers can be collected into one place and exposed to all other layers. Furthermore the data will be used in the optimization process, it can be stored and used for post-processing or it can be fed into the learning process. The CRM plays a central role in all these processes. It should be noted, however, that the basic principles of the OSI protocol stack operation should not be disturbed unnecessarily, meaning that the individual layers can still make decisions independently and the flow of information between adjacent layers will be possible. Having the CRM will facilitate the data exchange between non-adjacent layers and enable better system performance through cognition. Since one of the key features of the CRM is to perform multi-dimensional and multi-objective optimizations takeing into account substantial amount of data, the traditional numerical methods might not be fast and flexible enough. Due to this alternative approaches such as genetic algorithms and simulated annealing can be used. These biology and physics inspired optimization methods are interesting candidates since they are proven to have a high degree of success in problems with large number of variables and search spaces, and can work both with numerically generated or experimental data. For example, a typical multidimensional optimization problem is the throughput maximization of a single wireless link. There are several parameters that can be read from different layers of the protocol stack and tuned in order to achieve the maximum throughput. The packet size, modulation type, signal strength and TCP window size are good examples of those parameters. For optimizing such a multi-dimensional function an evolutionary algorithms (EAs) such as genetic algorithms (GAs) can be applied successfully. Speaking about optimization which is largely dependent on the measured data, the issue of the reliability of the data becomes relevant. One should bear in mind that the quality of the data is of a crucial importance in the decision process. It

402

Marina Petrova and Petri M¨ ah¨ onen

is very probable that in some cases the system has to reason and decide upon incomplete data. Additionally the system will be also exposed to a large amount of noisy data. Accordingly, e.g., data filtering techniques to handel the linear and non-linear noise are required. Techniques such as Bayesian reasoning and statistical learning theory can be deployed to deal uncertainty and ensure the reliability of the data and inference. We emphasize that soft-decision making based on estimated values under noisy environment is nothing new in wireless communication especially when one considers PHY-layer [33]. The ultimate goal in the design of the CRM framework is to make it capable of learning and reasoning. The process of learning will enable high degree of adaptivity for the system. Different parts of the system can have varying level of learning capabilities and opportunities for optimization. The reconfigurable cognitive radio itself is probably the best concise example of such sub-system. Using reinforcement learning the radio can learn in a long run how to tune its parameters (e.g., modulation type, coding rate, transmission power, frequency of operation, etc.) in order to achieve the best overall performance (maximize capacity). Rather than pre-programming these actions the radio should be able to learn to adapt to the different “environmental conditions”, the learning will provide them. In this way, by learning with a direct interaction with the environment, the radio can form a set of “learnt lessons” which will be stored in the knowledge data base for future reasoning or decision making. Other possible machine learning methods include neural networks, evolutionary algorithms (GAs), Bayesian learning, etc. Apart of learning the system can also use other mathematical tools for modeling. Building models can help the radio to predict the (radio) environment in the near future and chose an optimal state of operation by tuning its parameters. For example, the cognitive radio can build spectrum occupancy models from the sensed data using the time-series analysis. Time-series analysis is a powerful tool which will help in finding out the periodicity of having free spectrum by analyzing e.g., historical data. Relying on this model the cognitive radio will have a priori knowledge of the spectrum allocation in the near future (an hour or several hours in advance) and will be able to chose its frequency accordingly. From the description so far it should be clear that our CRM framework is a large set of components which work seamlessly together. Crucial issues in building such an architecture are the modularity and flexibility. It is important to note that the toolbox itself is not limited to enable optimizing techniques mentioned above. It is envisioned that further methods could be added in a plug-and-play fashion. In fact, from an implementation point of view, the CRM could be seen as a micro kernel with additional software modules where the scheduling and time synchronization mechanisms of different optimization and reasoning processes are carried out. This is one of the first important lessons we have learned from the early prototyping analysis. Although, a large part of our work is done with Linux and gnuRadios [1], it is clear that fully functional cognitive radio and CRM implementation needs a support kernel that looks more like a real-time operating system. In fact, we strongly believe that some PHY-layer interfaces require careful software-hardware co-design and possibly dedicated ASIC/SoC -type of solutions. The relevant information from the applications, the embedded operating system and lower OSI-layers are to be linked with the CRM through commonly defined and standardized interfaces. At present such interfaces are practically non-existent but

20 Cognitive Resource Manager

403

very much needed. In the next section we talk more about the possibilities of having well specified APIs as necessary enablers of the CRM. Overall the CRM is a multi-functional entity. Its main functionality is optimization of resource usage in terms of, e.g., communication bandwidth, spectrum, or battery power, leading to an overall enhanced system performance. Secondly, it can operate as a “connection manager” deciding upon the frequency channels as well as the type of communication technology to be used (IEEE 802.11, Bluetooth, UMTS, etc.) in case a variety of interfaces and networks are available. For example, different types of services (voice-calls, audio- and or video-streaming, etc.) can definitely benefit in quality if their different connection requirements are taken into account and accordingly the most appropriate (in terms of delay, bandwidth, bit rate, etc.) communication link is chosen. As previously mentioned, the decision making will be done in a “cognitive fashion” using both up-to-date and historical information. The information collected from each interface will be stored in a file system. For providing an easy access, the data will be sorted per parameter in different files in a knowledge and observation database. Additionally, when we have well established and fully functional CRM we foresee cooperation and information exchange among several CRMs to enrich the learning and optimization process. Of course this will bring additional complexity to the system and remains our research task for the near future.

20.3 Interfaces Well defined and unified programmable interfaces for accessing information from different layers of the protocol stack are important in order to provide easy flow of information and configuration. Most of the today’s APIs are technology specific. For example the link-layer interfaces of WLAN, Bluetooth and the cellular technologies are different from each other. Even within the same technology both APIs and the data representation can vary dramatically. Furthermore there is a lack of abstraction of the link information provided from different underlaying technologies that could be used to qualitatively compare different links in a unique manner. All this makes it difficult to develop link-layer aware protocols or applications that could run on top of different link technologies. The lack of such unified programming interfaces is not only concerning the linklayer but also the higher layers of the protocol stack. The previously defined architecture of the CRM framework would not only benefit from unified APIs but, in fact, could require the existence of such interfaces. In this section we will present the architecture and the functionalities of a Unified Link-Layer API (ULLA). Furthermore we will introduce some ideas towards designing a common interface between the CRM and the applications running on top. Eventually we will discuss possible work of building an unified network interface as an extension of ULLA for the other layers of the protocol stack.

20.3.1 ULLA The Unified Link-Layer API or ULLA is an interface we developed to retrieve data link layer and physical layer information independent of the underlaying technology

404

Marina Petrova and Petri M¨ ah¨ onen

(WLAN, Bluetooth, GSM, ZigBee, etc). Applications using ULLA do not have to be aware of the communication technology underneath since the interface offers monitoring and controlling for all technologies. The platform independence also gives us the freedom to easily extend the API for the new upcoming standards. However, the API can also provide technology- and implementation-specific details. Detailed description and the performance analysis of the API can be found in [2, 5].

Figure 20.3. ULLA architecture. The ULLA architecture is shown in Figure 20.3. It is composed of three main parts: the applications using ULLA which are also known as link users, the ULLA core and the link providers which are abstractions of the network interface cards in the device. It is important to note that a link user must not be an application in a OSI protocol stack sense per se. Link users can be applications running on the lower layers which will use the link information such as link-aware vertical handover or link-aware routing protocols. Our CRM is another possible application which will use the link-layer information for performing knowledge-based optimizations. The ULLA core is the main part of the API. Its main task is to process the requests and the updates from the link users and link providers respectively. The interaction between the link providers and the ULLA core and the link users and the ULLA core is done through two independent interfaces, namely link user interface and link provider interface. Through the link user interface the applications can issue commands to the link providers to, for example, scan the available links or request specific information (attribute values) from the link providers using queries. A ULLA Query Language, which is a subset of the SQL, was developed to specify the queries. For example, an application can request for all links that have latency smaller than 150 ms with the query: SELECT linkId FROM ullaLink WHERE txLatency < 150. Furthermore applications can subscribe for notifications if they want to be informed regularly about the updated values of the attributes or when the values reach certain threshold. The link provider interface tasks correspond to the ones of the link user

20 Cognitive Resource Manager

405

interface. It is used for asking for updates in attribute values, for issuing commands, and for registering and deregistering links and link providers. We see ULLA as an important enabling technology for the CRM. Using ULLA, the CRM can query particular parameters or information from the lower layers like bit rate, latency from particular links. On the other hand the CRM can also subscribe to ULLA for periodic or even-based notifications. The periodic notifications will provide the CRM with enough historical information of a certain parameter which will be needed in the learning process or later analysis (e.g time-series analysis of the quality of a certain link). Event-based notifications can be useful for fast interventions. Upon receiving event-based notification, CRM can modify parameters in the protocol stack in order to keep the guaranteed performance or to provide the best possible one. A typical example is the TCP-over-wireless problem, where sudden fading can trigger CRM to regulate the TCP congestion control.

20.3.2 Common Application Requirements Interface Following the concepts of the above described ULLA API we are introducing CAPRI (Common Application Requirements Interface) as an interface between the CRM and the possible applications running on top. Through the CAPRI the applications can negotiate with the CRM the QoS requirements such as delay bounds, throughput, priority etc. The CRM will use these requests to configure the lower layers of the protocol stack in order to fulfill the application demands, if possible. We emphasize that we do not foresee a strict requirement for applications to be “CRM aware”. Since the number of legacy applications is too large it is very difficult to integrate all their specifics and possible requirements without introducing large complexity. Here, we have opted to make a rather low complexity approach. Instead of designing complex interface and API (such as ULLA) or requiring SLA (Service Level Agreement) descriptions and protocols, we are just defining simple XML-coded “tags”. Application programmer can add to application task an XML-coded recommendation that defines different optimization goals (for example latency requirement) and policies (e.g., if necessary trade latency for bandwidth in joint-optimization game). The XML-tags will be located into virtual filesystem location, from which CRM can read them. If the legacy application does not have these preferences, CRM will simply skip this procedure and will use default presumptions. Additionally, CAPRI can be also used as an information provider to the applications. For example, the applications can ask, through CAPRI, information about the actual communication conditions and future estimations or get information from the historical data storage of the CRM. Furthermore operating system entities such as the scheduler can get great benefit from this information when deciding about priorities of different running applications.

20.3.3 Universal Network Interface The above described ULLA approach is naturally expendable for other protocol stack layers, too. We call the extension of ULLA concept as Universal Network Interface (UNI). The UNI is based on the same idea of providing a general, object oriented API to query and setup different protocol stack parameters. Currently, we

406

Marina Petrova and Petri M¨ ah¨ onen

are mostly interested in studying and implementing transport layer functionalities. As a case study we have been considering mostly different TCP schemes. The intention is for UNI to have a similar query engine structure as ULLA, so that they can share a part of the code base. However, there are some specific differences, which need to be commented. In the case of link-layer interface, the ULLA-CRM pair is always limited to consider one-hop links2 and CRM is mostly responsible for providing lower layer optimization decisions3 . For higher layers that are handled by UNI, we need to have a full concept of end-to-end connectivity, which naturally make decision making more complicated. Hence, the CRM should be, in the optimal case, aware of possible multi-hop connections. Because of this we require from UNI that there is a class describing end-to-end connectivity. In the case of cognitive radio networks at least, the concept of end-to-end performance is, in our opinion, an absolute necessity. The reason is that a full cognitive networking approach is impossible without recognizing full connection paths in networking sense. In principle, it is possible to trace all the link-layer information link-by-link and make a full end-to-end information based optimization decisions. The transport layer interfaces of UNI also require some specific handlers. For example in the case of TCP connections there has been a number of different proposals to make TCP more aware of wireless environment (cf. [30, 46] for review and references). The study of these have shown us that we need to have special interfaces so that CRM can directly control different timers and entities, e.g., Freeze TCP type of solutions [24]. In order to facilitate the transparent access to different times, we need to introduce them as a part of our query language and repository information framework. For example, for reading the timer information for TCP flows, where the retransmission expiry is 50 ms away, we can make a query: SELECT timers FROM streams WHERE protocol=TCP AND rt expiry < 50 ms. Similarly we are also providing functional calls in the API to freeze or rewrite the timer value (e.g., Freeze(T imerX ) or SetTimer(T imerX , −50 ms)). These kind of control and query structures mean that the concepts of timings and triggers have to be inherently embedded into UNI. The processing delays mean that e.g., FreezeTimer command needs to report the time value for which the timer was frozen and handle to restart it. As one of our early prototyping environments is to build CRM enabled IEEE 802.11 devices, this has lead us to the concept of having query capability to find interlinked timers, e.g., link-layer (MAC) and TCP timers are partially interlinked as a control loop. UNI queries can find all the timers that are related to same logical connection with a single command. Apart of timers, another concept that UNI offers uniquely is a manipulation of buffers. Finally, we need to have a capability to label different protocols, e.g., one can query TCP flavor, but here we have a problem of ontology, since getting answer “TCP Reno” does not mean much unless CRM has a knowledge on what that means. 2

3

The ULLA devolvement team has introduced a method to collect multi-hop link information, but this functionality has not been implemented. From CRM point of view it is still speculative how useful such information would be in wireless environment due to high variability of the links and the cost and latency introduced when collecting the data. Such decisions include making decisions to make handover between links and setting up link-layer and PHY layer parameters.

20 Cognitive Resource Manager

407

The concept of UNI opens, in principle, all the transport layer information for CRM and other entities. Moreover, we have a historical data repository available for the learning algorithms. Our current implementation approach is to make a historical data repository to look like files that can be accessed like a virtual filesystem (discussed later in this chapter). This opens a possibility to make different traffic and time-series classification operations very flexibly, e.g., porting our earlier work on classifying traffic automatically with neural networks is very simple in UNI/CRM context [37]. Geurts et al. [22, 23] have also proposed the use of machine learning to improve TCP congestion control with automatic detection of link-errors. This is something we are currently thinking to incorporate into our framework. The underlying principle in their methods is to use machine learning algorithms to infer from the packet transmission data the probability on wether the losses are caused by congestion or wireless errors. In their paper they have tried several methods, including decision trees, random forests [11], boosting [21], artificial neural networks and bagging [12] – going well beyond our experiments with just neural networks. The work by Geurts is one of the first to show, how the machine learning can be incorporated efficiently into the transport layer context, and UNI offers a possibility to make it without large changes to an actual TCP protocol stack.

20.4 Core Unit Aspects In the following section we highlight some CRM-core aspects and toolbox designs. Due to space limitation the description is not extensive. We have chosen few specific parts as representative examples. The selected content reports very recent on-going research work and “lessons learnt so far” in CRM context.

20.4.1 Learning and Reasoning with Genetic Algorithms Genetic algorithms (GAs) [32,34] are proven to be successful machine learning methods for solving complex optimization problems. Their main strength is the capability to efficiently find the global maximum/minimum in large search spaces using the principles of the evolutionary processes in the nature. GAs find the optimal solution of a given problem by searching a population of possible (candidate) solutions. Each member of the population is composed of set of parameters and can be represented as a bit string. In the evolutionary theory, each string is named a chromosome and the parameters are regarded as a genes of the chromosome. The first step in executing a genetic algorithm is the generation of the random population of N bit strings which represent the set of possible problem solutions. The next step is the process of evaluation of each string. The degree of “goodness” of a bit string for solving a particular problem is determined by the value of a so called fitness function f (x). The fitness function defines the criterion for ranking and probabilistic selection of the members in the next generation. The selection process is such that the better bit strings are given more chances to reproduce and survive than those which offer poor solutions to the fitness function. After the selection process has been done, an offspring can be created through a crossover as in Figure 20.4. The crossover occurs between a pair of randomly selected bit stings with a probability pc . After the crossover a mutation of the bits in

408

Marina Petrova and Petri M¨ ah¨ onen

the bit strings can happen with a probability pm . This means that some of the bits in the population will be flipped. As long as the mutation is finished, the GA has completed one interaction. The algorithm usually runs till it converges to a stable solution.

Figure 20.4. Crossover and generation of new offspring.

Genetic algorithms are able to explore the search space in several locations at the same time. Furthermore, GA search can move much more faster compared to the neural networks backpropagation algorithm, for example, replacing the parent chromosome by an offspring that may be radically different from the parent [40]. This is exactly what makes the GAs robust and powerful optimization tools. They are good in finding optimal (or close to optimal) solutions and it is less likely that they fall into local minima. In our experience GA based algorithms are particularly well suited to explore optimization possibilities, if the parameter space is large and we need to handle distributed decision making. For some problems they function less than a real optimizer and more like data mining tool. With this we mean that if the utility function and the number of parameters is relatively low then GAs find solutions, but the actual on-line parameter could be implemented much more efficiently by using, e.g., decision trees or adaptive fuzzy logic based methods. Genetic algorithms can be used to evolve radio and network parameters with a goal of optimizing the performance. One of the earliest works on using GAs in the cognitive radio context and wireless networks in general has been presented by Rieser [39].The use of GAs on the global cross-layer optimization, where the parameter space include variety of data from all OSI-layers and network topology, seems to be an extremely useful method. We argue that if the CR can afford the latency and computational power for GA optimization, it is one of the most versatile tools to include for the toolbox. Another similar method is to use Support Vector Machines [16] to find optimal solutions for complex problems. In order to carry out a multi-parameter optimization the genetic algorithm will need a well defined list of parameters that can be read and/or set through ULLAtype of interfaces and can straightforwardly define the fitness function as discussed in the previous section. Some examples of these parameters are: signal-to-noise ratio (SN R), frequency of operation, packet size or fragmentation, Packet Retry Limit, transmission power (Ptx ), contention window size, etc. A set of these parameters will be applied to the fitness function which will give an estimate how well the

20 Cognitive Resource Manager

409

parameter set meets the optimization objective. The parameter set will be further on manipulated with the GA until the optimal solution is found. Subject to optimization in a wireless system are usually the throughput, latency of the system, power consumption, packet (bit) error rate and so on. Sometimes there might be several goals that the wireless system should achieve in order to meet the QoS requirements of the applications. However one should be aware of the fact that several of these objectives might conflict with each other. Such a conflict will occur, for example, if we try, at the same time, to minimize the latency and the bit error rate (BER) because the packet size affects both objectives in an opposite way. Similarly, due to the counter influence of transmission power as a common parameter, minimizing the BER and the power consumption simultaneously will not give the wanted result unless we put more weight on one or on the optimization objective [20, 47]. Theoretically, in case of constraint multi-objective optimization the optimal solutions lie on the so called Pareto optimal front. Further improvements are not possible due to the correlation between the multiple objectives. The Pareto front trade-offs between the minimum BER fmin ber (x) and minimum latency fmin latency (x) for several error correcting coding rates are given in Figure 20.5 as an example. This example is taken from our on-going work, where we have developed an optimizing genetic algorithm which takes into account not only PHY-layer parameters, but also MAC-layer issues (e.g., the fitness functions for the throughput and latency are calculated at the MAC layer).

Figure 20.5. An example of Pareto front for trade-offs between the minimum BER fmin ber (x) and minimum latency fmin latency (x) for several error correcting coding rates. The x-axis of the graph represents the fitness function used to minimize the latency for several coding rates from R = 7/8 to R = 1/2, while the y-axis the fitness function for minimizing the probability of error. For this case we chose a

410

Marina Petrova and Petri M¨ ah¨ onen

multi-carrier system with Nc = 64 sub-carriers and SN R = 3dB. In this particular case, our parameter list was limited to packet size and coding rates. It can be clearly seen in the figure that for each curve as the fitness value that minimizes the latency drops, the fitness value for minimizing the BER increases. These curves present the key constraints for carrying out multi-objective optimization. The R/2 curve is showing the Pareto front as no parameter set on that curve can be set so that the fitness value in respect to both objectives fmin ber (x) and fmin latency (x) improves. In order to flexibly adjust the impact of each of the objectives in the fitness function a weighted sum approach can P be considered [18]. Accordingly, f (x) = Pn n w f (x) where w : 0 ≤ w ≤ 1 and i i i i i=1 i=1 wi = 1, are the weight coefficients of the objectives. By changing the weights of the objective, the different optimization goals and OoS requirements can be achieved by putting more accent on the most relevant one. Although GAs can be powerful as on-line optimization tools, in the CRM context we see them most of the time as an off-line background computational tools. This means that we position the GA modules to work in a background to explore optimization possibilities. The results are then fed as an input to our world-models and the actual dynamic adaptation is handled, e.g., through decision trees and fuzzy logic controllers. This is due to the fact that latter models provide computationally much faster way to adaptation in practical algorithms.

20.4.2 Decision Making and Utility Functions While different architectures for cognitive radios and cognitive wireless networks may conceptually be relatively simple, the devil is in the details. Most of the current algorithmic approaches are based on the assumption of the perfect information, which is in general not attainable in the case of cooperative wireless systems. This means that the decision making system in any realistic CR, in our case the CRM, has to take in to account that it is confronted with a decision situation with a number of different actions and be aware of that its information on the current state and parameters maybe only partial and valid only in a stochastic sense. One of the first researchers on stressing this point in the context of dynamic spectrum access (DSA) has been Q. Zhao with collaborators [14, 48, 49]. Although their work has been mostly focused on DSA problems, their research can in part be applied as a general decision-theoretic framework also towards a general cognition cycle in a CR. Zhao et al. adopt a Partially Observed Markov Decision Process (POMDP) approach to model DSA. Markov Decision Processes (MDP) are wellknown tools in engineering and stochastic optimization community. POMDP adds the concept of partial observation, i.e., there is a possibility that we may not be able to observe the current state, but we have some observations that gives us a hint about what the state could be. Adding partial observability to an MDP framework is not a trivial addition [13, 28, 42]. The discussion presented by Zhao et al. is quite beautifully applying the POMDP and showing how to make a decision for spectrum transitions. As such it is a valuable contribution towards building spectrum agile Cognitive Radios. The situation becomes more complex in the case of generalized CRM concept, since there is the

20 Cognitive Resource Manager

411

requirement to make cross-layer decisions (many actions based on many states) and in the case of CWNs we need to provide distributed algorithms with convergence4 . The first problem is encountered with the definition of the utility functions5 themselves. In the simplest case the subject of the optimization can be determined by a single application such as the maximization of throughput or minimization of latency. The DSA is a good example, since there reward for accessing new channel can be seen for example as a bandwidth of channel and the utility (objective) can be defined to be the number of bits that will be transmitted over such channel (compared to previous situation). One should note that this example is a rather simple model and although it has value on studying PHY-layer and DSA allocation, it is far from perfect for cross-layer optimization. From the point of view of MAClayer, for example the same packet transmission rate can lead to highly different jitter and latency situations depending how the channel is shared. In more complex scenarios, there can be several goals from different applications (running locally or in different nodes) and care should be taken to achieve as fair optimization as possible. For modeling the optimization goals we need to consider more complex utility-based approaches. For example the “goodness” of the connection for a particular application can be in general defined in terms of utility u(a1 , . . . , an ), which is a function of the various measurable attributes {ai } of the connection. Common examples of these attributes are bitrate, latency, frame error rate, and so on. Each of these attributes is itself a function ai = ai (r1 , . . . , rn ; x1 , . . . , xm ) of the parameters ri that can be configured in the (local) network stack, and some stochastic variables {xi } used to model the end-to-end connection. The cross-layer optimization problem in this context becomes the maximization of the utilities u for the applications communicating via the network stack by tuning the {ri }, subject to some fairness criteria. The utility based approach is well-suited for formulating network-wide optimization problems in the case of collaborating CRMs. The quantity to be optimized is the sum of the individual utilities ui , but this time the sum is taken over either the whole network or a subset thereof with appropriate fairness constraints. In the real CWN this multi-objective optimization is the ultimate goal we would like to achieve.

4

5

The convergence property is not always absolutely necessary, because sometimes if one is able to show permanence with good enough trajectory behaviour, that should suffice as well. The reader should note that we are using here specifically the term utility function instead of reward function. For example in the reinforcement learning the reward function is typically seen as an immediate reward for an action, and in the case of cross-correlation optimizing CR, we should not consider only immediate or shortterm consequences, but need a function that captures the long-term consequences, i.e., utility function. This is in part semantics, but as the optimization problems are studied in the different disciplines the terminology can be overlapping and confusing for some practical communications engineers. Note that we could also talk about objective function or cost function in this context instead of utility function. We refrain to use a term cost as we are using the term later to refer how much it costs to perform an action, in resource sense, in order to reach some utility (objective). POMDP is one of the good candidates to support a long-term utility optimization and in that field often the term objective function is used.

412

Marina Petrova and Petri M¨ ah¨ onen

However, there is the same challenge as with DSA with POMDP, i.e., we are not able to reliably know all the system states {Si } and parameters {xi }. Our work towards decision making core of CRM, has found that POMDP is useful in a larger context than DSA decision making. However, as we started to work with the whole protocol-stack and network approach new problems have arisen. Let us first recall the basic concepts related to POMDP following in part [48]. The dynamics of the POMDP are Markovian, but since we do not have direct access to the current state in the DSA game, our decisions require keeping track of history of the process. It is enough to have the probability distribution over all the states, and whenever we do a decision we will update the distribution. The a posteriori distribution of states (for example channel availability) is a key to build a decision making system, and the a posteriori distribution is called belief vector in the POMDP community. The belief vector Λ(t) = (p(s1 ), . . . , p(sn )), where p(si ) is the conditional probability that the system is in the state si at the time instance t, is coding our knowledge. It was shown in [42] that the belief vector is a sufficient statistics. The solutions to an (PO)MDP problem are called policies and it simply specifies the best action to take for each of the states. For example we may have a sensing policy to observe some parameters (xi ) and decision policy to take some action, π, e.g., the decision to acquire channel. In our case the key problem is the selection of the right action, as it is not always clear that we are able to define utility function in the first place. In the case of complex networks we often have a problem that we do not know a priori all the consequences of our actions. Let us now consider several actions that are performed by the several OSI-layers. As we do not have a full systems model of the whole network, this leads to the natural problem that some actions may have uncertain outcomes. More formally, we denote with S the certain outcome (new state), which is the outcome of performing an action π. In our case at the moment of decision event also S is, in fact, a random variable with n values (s1 , . . . , sn ) with the associated probabilities p(s1 ), . . . , p(sn ). The utility function u mapping from states to the numerical utilities is trivially giving us the expected utility of performing action π EU (π, u) = E[u(π)] =

n X

p(si |π)u(si ),

(20.1)

i=1

where utility function u : S → < is associating the utility u(s) with each outcome s. One should note that although we write above that the expected utility is depended on the action, it the case of CRM we naturally have to include into the function the current overall state of the system. The optimal decision for the rational decision maker is the one which is maximizing the expected utility. One can also write the above equation as a vector equation, by denoting probabilities as a vector pπ and coding different utilities ui = u(si ) as another vector, i.e., EU (π, u) = pπ u. One should note that there is a possibility that many of the actions have same end states. In practice, the situation is even more complex as we should also take into account that many of the actions have different costs6 . In CRM this can be coded into the utility function description itself. However, to do the rational decision one needs to 6

Here we want to emphasize the fact that the costs of actions will form a optimization constraints, which can be handled either as formal constraints or partially incorporated into the design of utility functions.

20 Cognitive Resource Manager

413

have complete model of actions, utilities and states. Even if we had a complete model, it is possible that it is not computationally tractable. The Bayesian networks can be extended to handle actions and utilities as decision networks (influence diagrams), see discussion, e.g., in [27]. The harder problem is often that we may have uncertainty over the precise nature of the utility functions themselves. The reason for this is often the fact that the situation itself is so complex that we are not able to define a crisp utility function for it. Also when one is considering the user preferences it is possible that the user is not able to express his utility function. There has been work in the field on developing systems that would automatically generate utility function(s) in behalf of the user in the case that it is imprecisely known. Boutilier has introduced the term and concept of expected expected utility, which formalizes well the uncertainty of utilities [8]. The idea has some similarity with path-integral techniques of particle physics. In short, the decision is determined not only by taking in account the utility functions, but also taking expectation over the space of possible utility functions. More formally we will define a probability density P over the set of utility functions and make a decision π under the uncertain conditions by reflecting the uncertainty of EU (π, u) by defining expected expected utility [8, 19, 45] Z EU (π, P ) = E[u(π, P )] = pπ uP (u)du, (20.2) where as previously a vector pπ describes all the probabilities of a reaching state by a decisions, p(si |π), and ith component of u us ui (si ). Hence, in the case of uncertainty one should use Maximization of expected expected utility (MEEU) principle max EU (π, P ).

π∈Sπ

(20.3)

We stumbled in to the MEEU principle through our practical CRM architecture and implementation work. One should note that this not a new concept by us and Boutilier has formalized it in his article in a general way. Moreover, the MEEU principle has been used in practice, e.g., in the context of POMDP [7] and incomplete information games [25] (see also [17]). Although it is a challenge to handle uncertainty with utility functions and solve tussle between competing utility functions this is also a strength of the CRM approach. The inherent capability to handle different situations and to learn the value of information and expected expected utility makes the system potentially more adaptive. It provides also a natural framework to handle CRs in the distributed fashion, since the uncertainty on the other devices states can be at least partially handled by the above mentioned methodology. It is beyond the scope of this chapter to go into more details on this, but this discussion should in part convince the reader that there is a lot of potential on using decision theoretical framework in the context of CRM. The next practical lesson we have learned in the decision making context of CRM implementation work is related to the value of information. In this context we raise the issue of the Value of Perfect Information (VPI). This concept is, of course, well known in the mathematical decision making theory. A CR has to learn through observation states and state variables (si , xi ). Some of these are random variables X, e.g., channel information parameters, and as discussed above also the states si or the utility functions ui may be only stochastically defined. The observing state information (e.g., scanning spectrum occupancy) has a resource cost for CR.

414

Marina Petrova and Petri M¨ ah¨ onen

Hence, there is a natural question on how much is it worth to pay in order observe S or X with certainty. For example, if a CR is visiting for a short duration a new environment (context), should it use its scarce resources to learn with some certainty X, or should it defer using too much resources. In the decision theory the Expected Value of Perfect Information (EVPI) is defined to be the difference between the expected value given perfect information and the expected value. Let us assume that the current knowledge of CR is K. The current best action α ∈ Π, where Π is the set of actions {πi }, is then X u(ci )p(ci |K, α), (20.4) EU (α|K) = E[u(α)|K] = max α

i

where ci means the result (consequence) of decision a, i.e., the new state si = ci . Suppose that CR could observe further information about the state variable X with some cost, should it do it? Trivially following Morgenstein, suppose that we knew X = x then we would choose a new best action α? simply by X u(ci )p(ci |K, a, X). (20.5) EU (α? |K, X) = max a

i

But as we do not know the random variable X at this stage, we have to sum over its possible values, which leads to known result that the expected value of perfect information is ! V P IE (X) =

X

p(X =

x|K)EU (a?x |K, X

= x)

− EU (a? |K).

(20.6)

x

In practice, we need to also describe how much resources it costs to find out the value of X, but VPI is giving us a formal description how much we should be willing to pay for that perfect knowledge. A lot of early CR and CWN work has been done under the assumption that we have a perfect knowledge and that collecting enough knowledge does not cost anything. Our experience is that this approach is not tenable in the case of cross-layer CWNs. One of the basic tenets in our CRM framework is to learn cost-functions for observing information and then using EVPI to decide in a new situations whether the extra observation is profitable or not. More prosaically the above outline of CRM decision making block is telling that not all learning is valuable, just like in the case of real life one has to put value for learning and information. Finally we note that in the case of cognitive radios, and particulary for large cognitive wireless networks, the values of state variables are location dependent. This has two direct practical consequences. First, for example in the case of DSA operations interference is receiver perceived, which means that only transmitted based decision can lead to wrong decisions, i.e., the well known hidden terminal problem. Even more interestingly this scales to the entire network concept. In the case of CWN, one has a multi-agent system, which is running a decentralized decision making system, where different agents (CRs) are deriving separate decisions for them that maximize some joint reward. There has been work done with the decentralized POMDP problems, although the number of practical algorithms is still limited. In the case of CWN systems one should note that individual CRs may have not only different optimization goals, but due to uncertainty of utility functions may

20 Cognitive Resource Manager

415

have different behavior even if they were trying to optimize towards some common reward. The above discussion is highly relevant in the case of CWNs, since the cost of signalling a lot of information between devices can become rapidly prohibitively high. For this purpose CRM has a capability to have a sophisticated support function to support EVPI and POMDP operations. The state information for POMDP and utility function probabilities are specific repositories as a database that need to be kept under the control of CRM. In fact, the system can use our toolbox part also to learn the details of partial EVPI. This means that although the number of different adjustable parameters (decisions) is very high, in fact some of them have only minor consequences for the overall optimization goal. This means that partial EVPI is the expected value of information on learning the true values of individual parameter or of a subset of all parameters. The toolbox can off-line parameter space optimization tools such as genetic algorithms, principal component analysis or factor analysis to find out which parameters are more sensitive for different world models – some of this can be done even through Monte Carlo sampling. However, also this approach is open for the costs, since running such algorithms will use resources. In our approach, we believe that many of such calculations could be done off-line when attached to power-supply or through opportunistic computing by asking some infrastructure to provide additional computational power. In fact, this could be also seen as an indication to have specific hardware acceleration, perhaps in the form of ASIPs (Application Specific Instruction Set), for cognitive radio cores.

20.4.3 Managing Time-Scales: CRM-core Most of the cognitive radio papers have so far focused on describing high-level functionalities and architectures. This has its own value taking in account that these systems are complex technological entities to engineer. However, many complexities are not well understood without considering the implementation and software architecture. The lack of work towards considering the implementation of such architectural designs is somewhat surprising taking in account that there exist a vast amount of literature and experience in Software Defined Radio (SDR) design [38]. Some very interesting work has been done on enabling reconfigurable protocols in the context of CRs [43]. The idea of composing different communication stacks dynamically from atomic components and including the radio components to this reconfigurability is well suited for cognitive radio design. The idea itself is not a new, and considerable amount of work has been done on this topic area by SDR research community (see for example OSSIE-platform developed in VirginiaTech [3,4] and also considered in other fields [10, 41]). Also our own work has been directed towards studying “protocol elements” as building blocks (see Figure 20.6). However, we think that the dynamic protocol stack management is not the main challenge in the implementation of the architecture. The different entities in the cognitive radio (see Figure 20.1) will have a highly different characteristic time-scales and requirements. Many PHY-layer events require microsecond granularity, but transport and application layer events and characteristic cycles may occur at the time-scales of seconds and minutes, and the overall traffic modeling and applications may need even longer time-scales to be modeled, e.g., days or weeks. The core-CRM needs to coordinate that different processes (threads) are synchronized between themselves and that resource allocation is made

416

Marina Petrova and Petri M¨ ah¨ onen

Figure 20.6. Modular protocol architecture as a part of reconfigurable CR architecture (modified from the original picture: by courtesy of K. Wehrle).

efficiently. Moreover, the capability to reconfigure different entities and changing parameters require also access control to resources (such as concepts of semaphores and interrupts). So far most of the work in CR-community has considered that the cognition cycle and reconfigurability is done by using multi-threading general purpose processors. This is a good starting point, but when one is considering some of the time-critical tasks it is clear that careful hardware-software co-design is required. This means that just like in the case of SDRs, there will be a need to consider also specialized hardware. It is indeed intriguing to consider developing and using ASIP (Application Specific Instruction Processors) based multi-processor ASICs as an execution platform. The approach to make CRM as a modular, run-time configurable combination of tools is a goal that is shared between several groups. Most of the current approaches are limited, however, to demonstration activities or have lead to development that looks a lot like a complex Radio Resource Managers (RRM) inside the current base station systems. Another relatively simplistic architectural description is to describe information exchange between different protocol layers and entities as a bus, sometimes called as a cognitive bus. This is architecturally speaking a valid view, but it hides away the complexity, the need of well defined interfaces (and ontologies) and most seriously the fact of the engineering complexity. The key difference in our emerging CRM design is that it is not anymore a monolithic resource manager software, but in fact a collection of modules, run-time libraries and independently run processes and threads. This means that CRM-core has a lot of similarities with operating system concepts. It can be seen run over the existing small operating system, or from the design point of view one can perceive CRM itself being an operating system. Here we will not discuss about the details of CRM-core, but want to point out some of the key issues that, in our opinion, are very useful general concepts. The

20 Cognitive Resource Manager Policy A

417

Policy B

Devices Device Abstractions

Statistics Parameters

CRM

Network interface

Virtual File System abstraction

Files from different network nodes

Figure 20.7. Illustration of Virtual File System access principle of CRM. According that, most of the resources and data are accessed through VFS abstraction.

first is the requirement to have a multipurpose scheduling support in the CRMcore. Some of the tasks are real-time critical and a long-term scheduler concept is required to make certain that long-term, computationally intensive, world-modeling tasks will not jeopardize the performance of other processes. Based on our present design understanding, we claim that the mid-term and short-term scheduling require also some (near) real-time scheduling support. Virtual memory and resource allocation are still difficult concepts for a resource limited terminal equipment. Due to this actual implementations need to be carefully designed, and it could be better to move some of the computationally complex tasks to be done as network services. This cooperative computing is an aspect we do not comment further in this Chapter. Apart of scheduling another important operating system concept that CRM-core needs to provide is a set of methods for restricting access to shared resources (different parameters, interfaces, storage etc.) in the environment where multiple optimization and control threads. In the prototyping environment we have opted to some relatively simple and well known solutions, namely we are providing a planning to provide a full support to use semaphors and monitors as concepts within CRM-core. The monitors are, obviously, important to avoid entering a busy waiting state as different optimization processes must be able to signal each other. The implementation of CRM or CE type of concept requires that a lot of data is exchanged between different entities, and also historically stored data needs to be accessible. This has led us to conclusion that a virtual file system (VFS) is a necessary concept to streamline design of CRM. We present most of our data as files, regardless if those are actual files, memory allocations etc. Naturally VFS concept is used just like in many modern operating systems to present also devices and

418

Marina Petrova and Petri M¨ ah¨ onen

other well defined entities (see e.g., Linux discussion in [6, 9]). For example all of our historical data (statistics) information and policy descriptions are represented internally as files. The exchange of information between different parts of CRM is often facilitated by simply passing by pointer to appropriate files (inode-pointers in our Linux prototyping sense). In fact, we are borrowing the concept of /proc File System from Linux. The CRM can provide similar virtual file system to show its inner workings for other protocol entities and optimization toolbox without highly specified and complex APIs. The final advantage on this approach is that we see it as generalization towards cognitive and cooperative wireless networks. As most of the information is presented as virtual file systems cooperating nodes can access to information logically simply in a similar manner as networked file systems. One can use for this for example distributed file system concepts, but the key concept here is that as the access is through VFS, the access methods itself are transparent for the fact whether the information is stored locally or in the another device (network). We show in Figure 20.7 schematically how CRM is accessing both networked and internal information through VFS abstraction. One should note that also the ontology descriptions of cognitive radio can be presented as files in a VFS [35].

20.5 Conclusions In this Chapter we have introduced a Cognitive Resource Manager (CRM) framework as an alternative approach for cross-layer optimization of the future wireless systems using advanced reasoning and machine learning methods. The proposed approach is complementing the emerging research paradigm of cognitive radio. The central component in this optimization framework is a new entity called Cognitive Resource Manager (CRM) which could enable optimal management of spectrum resources, scheduling and link resources, adaptive setting of protocol parameters successfully meeting the quality requirements of the user applications. All these tasks are carried out in a “smart way” using toolbox of mathematical techniques for machine learning, reasoning, modeling and decision making. We argue that an essential requirement, for building the CRM framework is the existence of well-defined programmable interfaces between the layers of the protocol stack and the CRM itself. Such interfaces will enable easy flow of information and unified communication among the entities in the framework. We also presented our design of Unified Link-Layer API (ULLA) as an example of an interface for retrieving data link layer and physical layer information independent of the underlaying technology. In this context, we also discussed ideas and early work towards designing such APIs for the rest of the protocol layers, in particular for the application and transport layer. In the second half of the Chapter we address several implementation issues, namely the problems of handling uncertain and incomplete information in the decision making process in the CRM, coping with definition of the utility functions in the optimization as well as managing different time-scale processes and give directions for possible solutions. We believe that deploying advanced optimization techniques (e.g., evolutionary optimization algorithms), inference and reinforced learning methods in combination with the information offered from the system and the environment is an promising way to go towards making the future wireless systems more intelligent and adaptive.

20 Cognitive Resource Manager

419

One of the key positions we take in this Chapter is that any sufficiently complex machine learning based optimizing entity (such as CRM or CE) should be seen as a real time operating system type of environment with a certain level of hardware/ software co-design. Thus we also recommend that expertise from more traditional operating systems and co-design research community should be considered very useful. We think that intelligent machine learning based resource management can indeed become an exciting research topic in the near future as its potential benefits are large and it offers a very fertile ground for interdisciplinary research even at the fundamental research level. The initial research results on CRM and tested enabling technologies are promising, and we believe that the further study is certainly warranted. The already existing reprogrammable hardware platforms such as gnuRadio is very tempting for the CRM framework implementation. We are currently working with a early system implementation and more detailed optimization models with the distributed approach that will lead to IEEE 802.11 based ISM band prototype in the near future. Acknowledgement. The authors would like to gratefully acknowledge their research colleagues in Cognitive Wireless Networks research group and especially Janne Riihij¨ arvi, Matthias Wellens and Alexandre de Baynast for the insights and sharing some of their results with us. This work has been in part financed by DFG, UMIC-excellence cluster initiative at RWTH Aachen University and European Commission (grants for GOLLUM and BIONETS projects).

References 1. The gnu software radio. http://www.gnu.org/software/gnuradio/. 2. Open source release of the unified link-layer api. http://ulla.sourceforge. net/ [Cited On: 15th of March 2007]. 3. C. Aguayo, J. Reed, and F. M. Portelinha. Design and implementation of an SCA core framework for a DSP platform. In SDR Forum Technical Conference, Orlando, November 2006. 4. P. Balister, M. Robert, and J. Reed. Impact of the use of CORBA for intercomponent communication in SCA based radio. In SDR Forum Technical Conference, Orlando, November 2006. 5. M. Bandholz and et al. Unified link-layer API enabling wireless-aware applications. In PIMRC ’06: Proceedings of IEEE 17th International Symposium on Personal, Indoor and Mobile Radio Communications, 2006. 6. M. Bar. Linux File Systems. McGraw-Hill, 2001. 7. C. Boutilier. A POMDP formulation of preference elicitation problems. In Proc. Eighteenth National Conference on Artificial Intelligence, pages 239–246, 2002. 8. C. Boutilier. On the foundations of expected expected utility. In Proc. of Intern. Joint Conference on Artificial Intelligence (IJCAI) 2003, pages 285–290, August 2003. 9. D. P. Bovet and M. Cesati. Understanding the Linux Kernel. O’Reilly, 2005. 10. R. Braden, T. Faber, and M. Handley. From protocol stack to protocol heap — role-based architecture. ACM SIGCOMM Comp. Comm. Review, 33:17–22, 2003.

420

Marina Petrova and Petri M¨ ah¨ onen

11. L. Breiman. Bagging predictors. Machine Learning, 24(2):123–140, 1996. 12. L. Breiman. Random forests. Machine Learning, 45(1):5–32, 2001. 13. A. R. Cassandra, L. P. Kaelbling, and L. M. Littman. Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101:99–134, 1998. 14. Y. Chen, Q. Zhao, and A. Swami. Joint design and separation principle for opportunistic spectrum access. In Proc. of IEEE Asilomar Conference on Signals, Systems, and Computers, October 2006. 15. D. D. Clark, C. Partridge, J. C. Ramming, and J. T. Wroclawski. A knowledge plane for the internet. In SIGCOMM ’03: Proceedings of the 2003 conference on Applications, technologies, architectures, and protocols for computer communications, pages 3–10, New York, NY, USA, 2003. ACM Press. 16. N. Cristianini and J. Shawe-Taylor. An Introduction to Suport Vector Machines. Cambridge University Press, 2000. 17. M. de Groot. Decision making with an uncertain utility function. In Foundations of Utility and Risk Theory with Applications (eds. Stigum B. and Wenstroep F.) Reidel, 1983. 18. Newman et al. Cognitive engine implementation for wireless multicarrier transcivers. Journal of Wireless Communications and Mobile Computing, 6:1–14, 2006. 19. P Fishburn. The Foundations of Expected Utility. Reidel, 1982. 20. C.M. Fonseca and P.J. Fleming. Multi-objective optimization and multiple constraint handling with evolutionary algorithms. IEEE Transactions on Systems, Man and Cybernetics, Part A, 28:26–37, January 1998. 21. Y. Freund and R. E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. In Proceedings of the Second European Conference on Computational Learning Theory, pages 23–27, 2005. 22. P. Geurts, I. El Khayat, and G. Leduc. A machine learning approach to improve congestion control over wireless computer networks. In Proceedings of the Fourth IEEE International Conference on Data Mining (ICDM’04), 2004. 23. P. Geurts, I. El Khayat, and G. Leduc. A machine learning approach to improve congestion control over wireless computer networks. Technical report, University of Liege, 2004. 24. T. Goff, J. Moronski, D. S. Phatak, and V. Gupta. Freeze-TCP: A true endto-end enhancement mechanism for mobile environments. In Proceedings of the INFOCOM 2000, 2000. 25. J. C. Harsanyi. Games with incomplete information played by bayesian players. part i: The basic model. Management Science, 14:320–334, 1968. 26. IEEE. Proceedings of the 2005 1st IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, November 2005. 27. F. V. Jensen. Bayesian Networks and Decision Graphs. Springer, New York, 2002. 28. L. P. Kaelbling, L. M. Littman, and A. R. Cassandra. Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101:99–134, 1998. 29. V. Kawadia and P.R. Kumar. A cautionary perspective on cross-layer design. Wireless Communications, IEEE [see also IEEE Personal Communications], 12(1):3–11, Feb. 2005.

20 Cognitive Resource Manager

421

30. K. C. Leung and V. O.K. Li. Transmission Control Protocol (TCP) in Wireless Networks: Issues, Approaches, and Challanges. IEEE Communications Surveys and Tutorials, 8(4):64–79, 2006. 31. P. M¨ ah¨ onen. Cognitive trends in making: future of networks. In PIMRC ’04: Proc. of IEEE 15th International Symposium on Personal, Indoor and Mobile Radio Communications, 2004. 32. K. F. Man, K. S. Tang, and S. Kwong. Genetic Algorithms: Concepts and Applivations. IEEE Transactions on Industrial Electronics, 43(5):519–534, October 1996. 33. H. Meyr, M. Moeneclaey, and S.A Fechtel. Digital Communication Receivers: Synchronization, Channel Estimation and Signal Processing. Wiley & Sons, 1997. 34. T. Mitchell. Machine Learning, chapter 9, pages 249–273. McGraw Hill, 1997. 35. J. Mitola. Cognitive Radio: An Integrated Agent Architecture for Software Defined Radio. PhD thesis, KTH (Royal Institute of Technology), 2000. 36. K.E. Nolan, P. Sutton, and L. Doyle. An encapsulation for reasoning, learning, knowledge representation, and reconfiguration cognitive radio elements. In Proceedings of the 1st International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM 2006), June 2006. 37. O. Raivio, J. Riihij¨ arvi, and P. M¨ ah¨ onen. Classifying and clustering the internet traffic by Kohonen network. In Proceeding of the ICSC Symposia on Neural Computation (NC’2000), Berlin, Germany. ICSC Academic Press, May 2000. 38. J. H. Reed. Software Radio: A Modern Approach to Radio Engineering. Prentice Hall, 2002. 39. C. J. Rieser. Biologically Inspired Cognitive Radio Engine Model Utilizing Distributed Genetic Algorithms for Secure and Robust Wireless Communications and Networking. PhD thesis, Virginia Tech, 2004. 40. M. N. H. Siddique and M. O. Tokhi. Training neural networks: Backpropogation vs. genetic algorithms. In Proceedings of IEEE International Joint Conference on Neural Networks, Washington DC, pages 2673–2678, 2001. 41. G. Singh and M. Narasimhan. Flexible protocol design for distributed applications, 1997. 42. R. Smallwood and E. Sondik. The optimal control of partially observable Markov processes over a finite horizon. Operations Research, pages 1071–1088, 1971. 43. P. Sutton, L. Doyle, and K.E. Nolan. A reconfigurable platform for cognitive networks. In Proc. of the 1st International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM 2006), June 2006. 44. R.W. Thomas, D.H. Friend, L.A. DaSilva, and A.B MacKenzie. Cognitive networks: Adaptation and learning tp achieve end-to-end performance objectives. IEEE Communications Magazine, pages 51–57, December 2006. 45. U. U. Chajewska and D. Koller. Making rational decisions using adaptive utility elictation. In Proc. of 17th Nat. Conference on Artificial Intelligence, Austin, 2000, pages 363–369, August 2000. 46. G. Xylomenos, G. P. Polyzos, P. M¨ ah¨ onen, and M. Saaranen. TCP performance issues over wireless links. Communications Magazine, 39(4):52–58, 2001. 47. L. Zadeh. Optimality and non-scalar-valued performance criteria. IEEE Transactions on Automatic Control, 8:59–60, 1963. 48. Q. Zhao and A. Swami. A decision-theoretic framework for opportunistic spectrum access. to appear in IEEE Wireless Communications Magazine: Special Issue on Cognitive Wireless Networks, 2007.

422

Marina Petrova and Petri M¨ ah¨ onen

49. Q. Zhao, L. Tong, A. Swami, and Y. Chen. Decentralized cognitive mac for opportunistic spectrum access in ad hoc networks: A POMDP framework. IEEE JSAC in Communications: Special Issue on Adaptive, Spectrum Agile and Cognitive Wireles Networks, (to be published), 2007.

21 The C-Cube Concept - Combining Cross-Layer Protocol Design, Cognitive-, and Cooperative Network Concepts Thomas Arildsen and Frank H.P. Fitzek Aalborg University [tha|ff]@es.aau.dk Summary. This chapter introduces the C-Cube concept combining cross-layer protocol design with cognitive and cooperative networking. As mobile devices are getting more flexible, it is envisioned that they should exploit their flexility and adapt to the current situation in the best possible manner. Adapting to a new situation could involve the selection of protocol parameters, protocol layers as such, and transmission technology with its related parameters.

21.1 Introduction This chapter advocates a combined concept merging cross-layer protocol design, cognitive and cooperative networking into a new concept referred to as C-Cube. The individual concepts are shortly described in the following: •

•

•

Cross-layer protocol design is the extension of the layered network protocol architecture by communicating between non-adjacent layers gathering information and controlling the offered flexibility of each protocol layer in order to adapt to the actual needs of communication. Cognition is understood as a network node’s awareness of flexibility, services, and opportunities for choosing the optimal operation in its surrounding network environment. Cooperation is categorized into two types of interaction between network entities: altruistic cooperation where network entities sacrifice their own benefit in favor of others and non-altruistic cooperation where all cooperating entities benefit at the same time.

These three individual concepts are now merged into the C-Cube concept. As one realization form of the C-Cube concept we introduce the Cognitive Protocol Stack referred to as CoPS. CoPS is a concept of cognitive networking focusing on the last hop in a mobile communication system. With the mobile device’s cognition about the available flexibility on its own platform and in the surrounding network and network elements, it chooses an optimal communication strategy, where optimal is seen from the mobile device’s point of view as the access point cooperates in an altruistic manner. This concept in effect makes the last hop of mobile device access active - a special case of the wider-scope active networks concept.

423 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 423–433. c 2007 Springer.

424

Thomas Arildsen and Frank H.P. Fitzek

21.2 Motivation Modern mobile phones, or maybe rather mobile devices since they contain so many features besides the plain telephone functionality, contain more and more advanced functionality. Although the device hardware may be designed more efficiently to consume less energy, the increasing functionality keeps increasing the overall energy consumption of the mobile device. An example can be seen in Figure 21.1 which shows different generations of mobile devices containing more and more functionality. Cameras are one of the features that has appeared in recent years, which adds to the energy consumption, and various other parts of the device also contribute an increasing amount of energy consumption. All this adds up to a point about now where adding more features or increasing the energy consumption of existing features (for example by increasing the capabilities of the individual components) will take the mobile device beyond the point where active cooling is required as known from larger devices such as PCs. This will be a serious limitation in a mobile device because of its physical limitations, the acceptance of the customer, and the cooling itself will require energy as well.

Figure 21.1. Increasing energy consumption of mobile phones. By courtesy of Nokia. The battery capacity is of course also an issue. In order to be able to maintain the same battery lifetime between recharges, batteries must continue to evolve, becoming more compact as the energy demands increase. Otherwise, battery sizes will have to increase as well or users will have to get used to shorter battery lifetime between recharges. Unfortunately the battery capacity has only improved by 80% in the

21 The C-Cube Concept: Bringing the Pieces Together!

425

last decade, while the processing energy is doubling every 18 months according to Moore’s law. From these two trends, it can be easily derived that the energy consumption has to be taken into consideration for the design of future wireless communication systems. In order to avoid the cooling problem and any battery capacity issues, we must look for new ways to decrease the energy consumption somewhere in the mobile devices. Cooperation among mobile devices has been identified as one possible strategy to save energy has outlined in Chapter 2. As we cannot limit or reduce the energy consumption of the mobile gadgets such as cameras and mp3 players, we will focus here on the network protocol stack that is used for communication. As more advanced multimedia applications move onto mobile devices, this will result in higher demands to the network protocol stack as those services need higher data rates and different transmission strategies than voice centric service. From the pure voice services in the beginning, over limited multimedia services relying on specialized modes in the cellular protocol stack, we will see applications known from ordinary PCs such as VoIP, video streaming from Internet servers, instant messaging applications, full-featured WWW-browsing etc. These will be applications requiring a full Internet protocol stack on the mobile device including multiple network, session, and transport layer protocols serving multiple applications, perhaps running simultaneously. An example is illustrated in Figure 21.2, showing a VoIP application running on a traditional Internet protocol stack between the mobile phone and an access point.

Voip

UI

RTP/RTCP

SIP

UDP IP Wireless Link

Wireless Link

Mobile Device

Fixed Access Point

Figure 21.2. Traditional protocol stack.

Such a significant expansion of the protocol stack found in mobile devices will surely increase the computational demands and thus the energy consumption. One way to decrease the energy consumption of a more complex protocol stack could be to find appropriate cross-layer optimizations that would allow increasing the efficiency of the protocol stack. This can be seen as a first step. The potential lies in the existing flexibility to choose transmission parameters.

426

Thomas Arildsen and Frank H.P. Fitzek

In recent years the concept of Active Networks has been discussed by researchers [6,7]. In today’s Internet, intermediate nodes in the network such as routers and switches are limited to simple actions such as forwarding packets and perhaps altering their headers. As such, todays networks are passive in the sense that they simply forward data in the packets flowing through them. The idea of active networks takes this a step further to introduce active intermediate nodes in the network. There are two sides to this active approach. On one hand, the switches, routers, etc. may actively work on the user data content of the packets. An example of this could be sensor data fusion; if data from a large number of sensors in a network could be combined under way to their destination by the nodes they pass through, transmission bandwidth could be saved because not all of the original data packets would have to be transmitted all the way to the destination. On the other hand, the routers and switches are themselves programmable such that the actions performed on the packets passing through them may be specified on-the-fly. There are two approaches to programming the actual network nodes; the discrete approach and the integrated approach, or the out-of-band and the in-band approach, one could call them. The discrete approach separates the actual programming of nodes from the transportation of data through them. In this approach, nodes are programmed explicitly by uploading code to them, and actions on packets passing through the node may then be performed based on header fields in them, utilizing the functionality that has been uploaded to the node. In the integrated approach, every single packet – or capsule as they are termed in this context – contains code to be executed on the capsule’s data, in addition to the data itself. The discrete approach is the less radical of the two, but may be easier to manage safely. For example, uploading code to nodes in the network could be made possible only through some back-door mechanism that requires authentication of the users which could be restricted to network administrators. The integrated approach is much more flexible but requires larger changes in the way we view networks today and it implies more extensive security, portability, and execution considerations regarding the active content.

21.3 Cognitive Networking in Cellular Networks Cognition in a wireless communication context is often seen as cognitive radio. Cognitive radio is however limited to the idea of a physical layer that is cognitive with respect to the available radio spectrum and utilizes the available spectrum accordingly. Cognitive networks has also begun appearing as a more general concept regarding the whole network [1, 5]. Here we introduce the concept of cognition in the protocol stack of mobile devices. Cognition here is about, at any point, being aware of what is going on around you in the scope of a particular protocol, but even more in the scope of the whole stack and even further, between individual devices. This concept is called the Cognitive Protocol Stack (CoPS). Cross-layer design is part of CoPS. CoPS is a matter of being aware of the possibilities for flexibility and adaptivity. Cross-layer design is a means by which one can get specific knowledge across the protocol stack between separate layers and thus achieve cognition in the protocol stack. Cross-layer design is also a means by

21 The C-Cube Concept: Bringing the Pieces Together!

427

which we can exploit flexibility in the protocol stack by using it to enable adaptivity – controlling specific features jointly across multiple protocols and layers. Flexibility should be understood as the possibility of changing something at some point in the protocol stack. Examples of this flexibility could be selecting different kinds of codecs in a VoIP application or using a codec that supports switching between different rates, such as AMR. It could be the possibility of choosing different modulation schemes/orders in the physical layer. It could be selection between different fixed link layer packet sizes such as in [4] where we used commercial mobile phones based on the Symbian operating system (most of the Nokia phones). On such platforms it is possible to change the transmission parameters on the fly, i.e., even for ongoing communication streams. As an example we refer to the Bluetooth technology, where information can be sent over the wireless medium with different packet types. One group of the packets (so-called DM packets) transmit user information with forward error correction, while a second group (so called DH packets) neglect the possibility of robust transmissions using all space for user information. On top of the packet type, there also exist different packet bundling facilities combing three or five packets to one big packet. Such a flexibility can be used to make multimedia communication more spectrally efficient and less energy-consuming. We have presented those results in [4]. Besides this kind of optimization on the mobile device itself, there is even more potential on the last hop between mobile phone and the related access point. Cognition in this context is the awareness of this flexibility in the protocol stack. Cross-layer design is a way of achieving this awareness by for example letting some cognition management middleware communicate directly with the different protocols at the layers of the protocol stack. Adaptivity should be understood as enabling the ability to control the flexibility present in the protocol stack. Cross-layer design is employed to enable adaptation of various features in the stack from some cognition management middleware, thus putting the flexibility to use. A new approach of abstracting the services of mobile devices from the actual devices in order to share these cooperatively among multiple devices is found in [3]. This new approach describes how the capabilities of the mobile device such as speaker, microphone, camera etc. are offered as services to other devices in order for such devices to dynamically compose a selection of required services for a given application across separate devices. The CoPS concept is built on the approach from [3] and applies it to the composition of protocol stacks in mobile devices. This new feature gives the flexibility/adaptivity issue a whole new dimension. CoPS is the concept of remote protocol components. In this framework a mobile device has the choice between running specific components of the protocol stack locally or running these components remotely on an access point through which the mobile device is connected to the network1 . In this way, CoPS takes the approach of [3] and applies it to mobile devices and their access points such that the access points can offer specific network protocols to connected mobile phones as services. This enables the mobile device to delegate some of the protocol layers’ tasks to the access point through which the device is connected which will alleviate the demands for processing power and energy consumption. Remote protocol components gives the possibility of composing the network protocol stack as needed in different situations. If the mobile device is connected to an 1

This could for example be a IEEE 802.11 WLAN or Bluetooth access point.

428

Thomas Arildsen and Frank H.P. Fitzek

access point that provides (some of) the remote protocol features, the mobile device may utilize these and avoid including the corresponding protocol component(s) in its local protocol stack. In this case we assume that the access point is cooperating with the mobile device in an altruistic manner. Then in another situation, if the mobile device is connected to a regular access point without the remote features, it may run the necessary parts of the protocol stack locally in stead. Thus it can save energy resources when it gets the opportunity to run certain parts of the protocol remotely. This remote protocol concept adds a new type of flexibility/adaptivity to the protocol stack. This requires the mobile device to be cognitive about its surrounding network environment. In this case the term cognition covers the mobile device’s awareness of the remote capabilities of the access point through which it is currently connected as well as other potential access points the mobile device might move to. Moving between different access points with different capabilities will require continuous adaptation of the mobile device’s protocol stack. Figure 21.3 shows an example where the described cognitive protocol stack has delegated some of the protocols involved in the application’s communication (exemplified here by a VoIP application) to an access point providing remote handling of the transport and network protocol layers’ tasks.

Voip

UI SIP light

RTP/RTCP

SIP

UDP IP Wireless Link Cognitive Mobile Device

Wireless Link

Fixed

Smart Access Point

Figure 21.3. Cognitive protocol stack.

In order to realize the idea of relocating network protocols to the access point, it is necessary to consider how to implement mechanisms for the mobile device’s cognition about access point capabilities. A protocol for communication related to CoPS capability discovery/awareness is needed. The protocol could as a first step be a simple mechanism for determining whether a specific access point is CoPS cooperation capable or not. If so, this will imply some standard set of protocols that can be handled remotely on the access point. This protocol could then be extended to provide mechanisms for determining exactly which protocols the access point is capable of handling remotely. In addition, a simple protocol for piping data through the link layer between the mobile device and the access point, for remote encapsulation in a protocol at the access point, is needed.

21 The C-Cube Concept: Bringing the Pieces Together!

429

When the mobile device is capable of determining under which protocols the access point offers cooperation, this opens up another class of possibilities. The interaction between mobile devices and access points could be taken further towards the active networks idea along this line. If the access point lacks CoPS cooperation capabilities for a certain protocol, the mobile device could upload the required protocol to the access point in order to exploit its services. Likewise, if the access point offers a certain protocol which the mobile device does not know beforehand, the mobile device could download the required components for interacting with this protocol on the access point. With the CoPS concept there may also emerge new opportunities for cooperation between mobile devices in addition to the cooperation between a mobile device and the access point. There are no concrete examples yet, but we imagine that there are benefits to reap when multiple mobile devices are cooperating in the CoPS sense with the same access point. These devices could locally coordinate among them via short-range communication how to set up their communication with the access point and perhaps even exchange protocol components among them in order to optimize the overall CoPS utilization of the cell.

21.4 Preliminary Results A group of students at the Department of Electronic Systems has been working on an implementation of the CoPS idea with an active networking approach on mobile devices [2]. In this particular case, an approach was tested where the Transmission Control Protocol (TCP) is run remotely on an access point by the mobile device. The mobile device was connected to an access point via Bluetooth. A dedicated light-weight protocol (Active Networking Remote Socket - ANRS) was implemented on top of the Bluetooth link as a replacement for the standard TCP in the mobile device. The ANRS component handles calls to the socket API in the mobile device and forwards these to an ANRS counterpart on the access point. The ANRS component in the access point in turn executes the socket calls through the TCP protocol in the access point. The principle, much like Figure 21.3, is illustrated in Figure 21.4. The ANRS solution has been tested and compared to a standard TCP protocol: BNEP. Ten independent tests were conducted, transferring 1MB of data across the BNEP TCP protocol vs. the students’ own ANRS protocol. The test was a comparison of the energy consumed by the BNEP protocol and the ANRS protocol, respectively. The protocol energy consumption was estimated by subtracting the baseline energy consumption measured with no network activity on the mobile device. Figure 21.5 shows a comparison of the energy consumption at comparable MTU sizes for the two protocols. The figure shows the mean values with 95% percent confidence intervals plotted around them. As the figure shows, the measurements are fairly imprecise and further investigation is needed to form a more solid statistical basis for such comparison. However, the comparison does suggest that the CoPS idea has a potential for substantial power savings in the network protocol stacks of mobile devices.

430

Thomas Arildsen and Frank H.P. Fitzek

Figure 21.4. Active Networking Remote Socket principle.

Figure 21.5. Energy consumption per transferred MB – ANRS protocol vs. reference TCP protocol (BNEP).

21 The C-Cube Concept: Bringing the Pieces Together!

431

21.5 Cross-Layer and Cognition Combination It is also necessary to consider how cross-layer designs behave when the protocol stack is made adaptive under the CoPS concept. Any cross-layer optimizations must be made cognitive in order to work together with this sort of dynamically re-composable protocol stack such that a particular optimization incorporating a certain protocol component or at least communicating through a certain layer is able to work without this component or layer locally present. Figure 21.6 illustrates what one could imagine to achieve by the CoPS concept and cross-layer design. The scenario represented by the point labeled local is the traditional approach. local xlay corresponds to adding some cross-layer optimization which might improve the performance at a slightly higher complexity. The point labeled remote corresponds to running protocol stack components remotely; the complexity is decreased while maintaining the same performance. Finally, remote xlay corresponds to adding the same cross-layer optimization, based on the situation of the point remote. Here, one could imagine that it increases the complexity as before but might not be able to improve the performance as much since certain features are now located remotely and thus not accessible to this particular optimization operation.

Figure 21.6. Expected performance versus complexity under cross-layer optimization (xlay/xlayer) and cognitive protocol stack influence.

Figure 21.7 shows an overview of the concept proposed in this chapter in combination with software defined radio (described in other chapters of this book). The figure illustrates how the different concepts come into play in the mobile device in relation to the different network layers present. The active networking part of

432

Thomas Arildsen and Frank H.P. Fitzek

the C-Cube concept is concerned with the network layer and above while software defined radio is concerned with the link and physical layers. Cross-layer design is (potentially) related to all layers in the protocol stack.

Figure 21.7. Organization of active networking, software defined radio, and crosslayer design in and among mobile devices under the C-Cube concept.

21.6 Conclusion In this chapter we combined the cross-layer protocol design, cognitive and cooperative networking into the C-CUBE concept. The main idea is to identify and exploit the flexibility in a wireless communication system to adapt to changing environments. Specifically the Cognitive Protocol Stack concept has been explored and a variant of this has been implemented and tested in practice. The empirical results obtained from these tests indicate that the idea of running some components of a mobile device’s protocol stack remotely on its access point hold potential for significant energy savings. In case non-altruistic cooperation enables even higher degrees of flexibility such as the possibility to outsource protocol layers towards the network, further improvements in terms of spectral efficiency and energy saving can be expected. Acknowledgement. This work was partially financed by the Danish government on behalf of the FTP activities within the X3MP project. Test results from a specific implementation of parts of the CoPS idea have been provided by students at the sixth semester in Communication Systems under the Department of Electronic Systems, Aalborg University. The students are: Martin Kirch Dige, Kim Højgaard-Hansen, Janus Heide Møller, Janne Dahl Rasmussen, and Kasper Revsbech.

References 1. Dragan Boscovic. Cognitive networks. Motorola technology position paper, Motorola, Inc., Schaumburg, Illinois, 2005.

21 The C-Cube Concept: Bringing the Pieces Together!

433

2. Martin Kirch Dige, Kim Højgaard-Hansen, Janus Heide Møller, Janne Dahl Rasmussen, and Kasper Revsbech. Conserving energy on a mobile device using active networking & remote socket. Technical report, Aalborg University, Department of Electronic Systems, 2007. 3. Frank H. P. Fitzek, Morten V. Pedersen, and Marcos Katz. A scalable cooperative wireless grid architecture and associated services for future communications. European Wireless 2007. 4. Tatiana Kozlova Madsen, Frank H.P. Fitzek, Gian Paolo Perrucci, Thomas Arildsen, and Shekar Nethi. Novel ip header compression technique for wireless technologies with fixed link layer packet types. In Proceedings of IEEE Global Telecommunications Conference (GLOBECOM 2006), 2006. 5. Petri M¨ ah¨ onen, Marina Petrova, Janne Riihij¨ arvi, and Matthias Wellens. Cognitive wireless networks: Your network just became a teenager. In Proceedings of IEEE INFOCOM 2006, 2006. 6. D.L. Tennenhouse, J.M. Smith, W.D. Sincoskie, D.J. Wetherall, and G.J. Minden. A survey of active network research. Communications Magazine, IEEE, 35(1):80– 86, 1997. 7. D.L. Tennenhouse and D.J. Wetherall. Towards an active network architecture. In DARPA Active NEtworks Conference and Exposition, 2002. Proceedings, pages 2–15, 2002.

22 Cellular Controlled P2P Communication Using Software Defined Radio Two is better than one, if two can work as one. Coach Mike Krzyzewski, Duke Men’s Basketball

Jesper M. Kristensen and Frank H.P. Fitzek Aalborg University {jmk|ff}@es.aau.dk Summary. This chapter introduces one realization of the cellular controlled peerto-peer (CCP2P) scenario using software defined radio (SDR) in combination with OFDMA as the preferred implementation technology. This implementation form is compared with the state of the art software controlled radio (SCR) implementation. For cellular networks, SCR has some advantages over SDR as it is less complex and the offered flexibility of SDR is not needed. This argumentation changes in case we take CCP2P under consideration, where the flexibility is needed to reduce the overall complexity. In combination with SDR, OFDM is a potential key player to realize the CCP2P concept in a very efficient manner. The chapter concludes with a short introduction to the use of the GNU radio for the CCP2P testbed.

22.1 Introduction OFDM is recognized as a potential transmission scheme for future communication systems because of its inherent capabilities of supporting high data rate transmissions by converting a high rate data stream into parallel data streams at a lower data rate and transmitting these datastreams on orthogonal sub-carriers [7]. By converting a high rate data stream into a number of parallel low rate data streams, the transmission becomes less susceptible to intersymbol interference as the symbol period versus delay spread ratio of a given communications channel becomes larger. This alleviates the need for complex channel equalization at the receiver, and together with the efficient implementation of the FFT algorithm, makes OFDM an attractive transmission scheme for the next generation communications systems. As the need for higher data rates increases due to resource demanding mobile applications, the trend will go towards increased energy consumption in the mobile terminals. Therefore there is a need to consider some of the factors influencing the energy consumption and ways to reduce the influence of these factors. Two factors are influencing the energy consumption of a mobile device. One factor is related to the processing requirement i.e., the amount of processing needed to support a certain data rate for a given application. The second factor is based on the energy used to convey the radio signals between sender and receiver.

435 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 435–455. c 2007 Springer.

436

Jesper M. Kristensen and Frank H.P. Fitzek

To address the aforementioned factors, in [5] the concept of cellular controlled peer-to-peer (CCP2P) is introduced. Traditional cellular communication is based on communication between base station and the wireless terminal, where each terminal is operating autonomously. The new concept envision cooperation among devices, which are communicating over the short range directly with each other and using the cellular communication in parallel. Such architecture offers the potential of virtual high data rate and lower energy consumption. An important precondition for the success of a cellular controlled peer-to-peer (CCP2P) scenario is the assumption of increased link quality in a short range communication due to the proximity of cooperating terminals. The assumption of improved link quality facilitates the use of higher order modulation schemes with less coding overhead and less transmit power, leading to an increased efficiency in terms of energy resources while still enabling the required performance in terms of data rate. Throughout this chapter we will focus on the realization of cooperation considering a CCP2P downlink scenario as given in Figure 22.1.

Figure 22.1. Micro cooperative downlink scenario.

22.2 Realization Forms of the CCP2P Scenario Figure 22.2 depicts a set of air interfaces, access schemes and implementation technologies for potential realization forms of the CCP2P scenario. The notion of multimode air interface is to provide leverage from the use of omnipresent cellular as well as short range technologies. The notion of a combined cellular and short range air interface in this context is to take advantage of creating an air interface with a common frequency spectrum allocated for both the cellular communication as well

22 CCP2P with SDR

437

as the short range communication and thereby creating the potential for a flexible air interface.

Figure 22.2. Realization of CCP2P.

TDMA and FDMA are indicated as possible multiple access schemes. In principle TDMA is the choice of access scheme for a multi-mode air interface where cellular and short range communication are allocated separate time slots as indicated in Figure 22.3. FDMA provides in connection with the combined cellular and short range air interface the possibility of a flexible and efficient utilization of the allocated frequency spectrum. The implementation technologies indicated are the software controlled radio (SCR) and the Software Defined Radio (SDR). The two terms have been defined by the SDR forum1 and is listed in Table 22.1.

Table 22.1. SDR forum definitions of SCR and SDR. SCR

SDR

1

Only the control functions of an SCR are implemented in software thus only limited functions are changeable using software. Typically this extends to inter-connects, power levels etc. but not to frequency bands and/or modulation types etc. SDRs provide software control of a variety of modulation techniques, wide-band or narrow-band operation, communications security functions (such as hopping), and waveform requirements of current and evolving standards over a broad frequency range. The frequency bands covered may still be constrained at the front-end requiring a switch in the antenna system.

www.sdrforum.org

438

Jesper M. Kristensen and Frank H.P. Fitzek

Figure 22.3. Multi mode through TDMA access.

In the context of this chapter, the SCR is defined as a fixed hardware implementation with multiple hardware chains for the support of cellular and short range communication. As the functionality in a SCR is largely realized in hardware, the SCR is less likely to be able to adapt to any change in functional requirements without a redesign of the hardware. This is in contrast to the SDR which is an implementation where the functionality is largely determined by flexible hardware and software design. And as such, a SDR is more likely to be able to adapt and reconfigure its functionality to changing requirements. The feature of flexibility in a SDR implementation is in principle achieved by moving the boundary between analog signal processing and digital signal processing as close as possible to the antenna of a wireless device, and by implementing the digital processing in reconfigurable hardware [4, 8]. Although flexibility is the principle goal of SDR, metrics such as complexity and efficiency need to be taken into consideration. The realization of CCP2P, will be exemplified using the scenario shown in Figure 22.1, which shows a cooperative downlink scenario of three terminals forming a cooperative cluster within a cellular network. It should be noted that the choice of three terminals is chosen for illustrative purposes although the number of cooperating terminals are to be considered dynamic and not limited to a fixed number of terminals. An essential precondition for an efficient cooperative access is a highly flexible capacity distribution between the cellular and the short range communication. The capacity needed in the different domain is a function of the number of cooperative users. For a stand alone device all spectrum should be used for the cellular communication link and no short range capacity is needed. With an increasing number of neighboring devices the capacity needed on the short range link increases. Simultaneously the capacity needed on the cellular link decreases. For the case of cooperation using the principle of multi-mode, available technologies limits the flexibility provided for this capacity distribution. This limitation motivates the introduction of a more flexible air interface. A flexible air interface based on OFDM provides flexibility in terms of sub-carrier distribution with bit as well as power loading for each subcarrier. Although OFDM offers flexibility in allocation of sub-carrier, further flexibility is needed in choice of implementation

22 CCP2P with SDR

439

technology. Considering the requirement for a flexible capacity distribution and the dynamic changing number of cooperating terminals, a given architecture must support the flexibility requirement, i.e., support of the flexible capacity distribution in terms of OFDM subcarrier allocation together with associated communication parameters for the cellular and short range links. Furthermore, as shown in Figure 22.1, the number of terminals taking part in a cooperation is dynamic and a cooperating terminal need to adapt communication parameters, and to dynamically create and close additional transceiver chains according to the dynamically changing number of cooperating terminals. We argue that considering the scenario depicted in Figure 22.1, an SDR approach will likely prove to be more advantageous in terms of complexity and flexibility whereas improved efficiency is inherent in the application of the micro cooperative network architecture. We therefore argue that although the SCR has until now been the preferred implementation technology in state of the mobile device targeted for a cellular scenario, in a cooperative wireless network scenario, the SDR is an viable implementation technology necessary to facilitate the demand for flexibility needed to achieve the gains described in [5].

22.2.1 Multi-Mode Realization CCP2P can be realized with state of the art mobile phones using multi-mode air interfaces. Using two different air interfaces, where each technology is using its own spectrum for communication, the benefits of cooperation can be shown. Nevertheless, a combined cellular / short range air interface could lead to more efficient spectrum use and less complex and therefore less energy consuming mobile communication. As mentioned earlier the principle of CCP2P using multi-mode is already implemented using the time domain access as illustrated in Figure 22.3. The figure shows the basic principle of multi-mode communication with one base station and two cooperating terminals. The base station transmit information for each of the terminals at defined time instants. When all of the cooperating terminals have received their respective information parts, they exchange that information over a short range link, thereby cooperating on receiving the total amount of information sent from the base station. The (cellular) transmission from the base station is carried out using omnipresent cellular technologies, likewise the short range links can be established using available short range technologies. From a technology perspective, multi-mode cooperation can be implemented using present state of the art mobile devices, implemented using Software Controlled Radio (SCR) hardware architectures.

22.2.2 Combined Cellular and Short Range Air Interface We now turn to the concept of a flexible frequency domain access scheme using a common frequency spectrum for cellular as well as short range communication. The motivation for introducing such an access scheme is to accommodate the flexibility needed to exploit the full potential in a cooperative scheme. In this scenario we argue the use of a multicarrier transmission scheme and in particular OFDM. From an implementation perspective, we consider two approaches, the SCR architecture, with a fixed subcarrier allocation scheme and an SDR architecture with a flexible subcarrier allocation scheme. The underlying principle of these two approaches are illustrated in Figure 22.4. For both approaches it is understood that

440

Jesper M. Kristensen and Frank H.P. Fitzek

the combined cellular and short range approach or in the case of multicarrier transmission the total set of sub-carrier, are divided between cellular communication from base station to the cooperating terminals and short range communication between cooperating terminals. The SCR approach is to be understood as cooperating terminals with SCR architectures. Considering the scenario shown in Figure 22.1, it is obvious that cooperating terminals are required to process a number of short range communication links from other cooperating terminals. An SCR architecture would need to accommodate that number of short range communication links in its architecture design. In Figure 22.4 it is indicated that because of this architecture requirement, the spectrum allocation between cellular and short range sub-carrier can only be done as fixed allocation. This is contrary to a scenario where the cooperating terminals are terminals with an SDR architecture. In such an architecture the spectrum allocation can be made flexible and dependent on the actual number of cooperating terminals. It should be noted that considering the combined cellular and short range air interface and the spectrum allocation, the number of sub-carrier allocated for short range, are smaller than the spectrum allocated for cellular communication, this is due to the assumption of better link quality resulting in more spectral efficient modulation formats and henceforth a lower number of required sub-carrier.

Frequency Domain Access with OFDM We previously introduced the concept of a combined cellular and short range air interface and the use of OFDM. In Figure 22.4 we illustrated the concept of partitioning a given frequency spectrum for cellular and short range communication and briefly introduced implementation issues. We now turn to a more detailed description of the application of OFDM in a cooperative scenario as the one shown in Figure 22.1, we describe the use of frequency domain access, FDMA, with OFDM where each cooperating terminal is assigned a set of sub-carrier. Using FDMA and OFDM provides provides the flexibility in spectrum assignment together together with power and adaptive modulation and coding (AMC) techniques. This section will motivate the need for flexible spectrum assignment together with power- and AMC, later we will turn to implementations issues and based on different scenarios of subcarrier allocation argue that an SDR architecture has the potential to be a more efficient solution compared to an SCR architecture. The concept of a combined cellular and short range air interface implies a common air interface between base station an terminals and between cooperating terminals. In such an air interface, it is assumed that the sub-carrier are freely allocated between cellular and short range communication links. Using multi-carrier transmission allows the flexible allocation of resources in terms of spectrum and the efficient use of that spectrum by proper power allocation and spectral utilization schemes, according to the prevailing channel conditions, see [6]. As indicated earlier the downlink cooperative scenario allocates only a subset of an OFDM spectrum to be processed by each cooperating terminal, this creates the potential for reducing the processing load of FFT processing by employing partial transform algorithms. In [9] it is shown that for applications requiring only a subset of FFT points the computational savings can be substantial compared to split-radix algorithm with the computational complexity of O(N logN ), with N being the length of the FFT equal to the number of sub-carrier. It is further assumed that improved link quality

22 CCP2P with SDR

441

Cellular bandwidth

(a) Cellular only communication Cellular bandwidth

Short range bandwidth

(b) Fixed subcarrier allocation of cellular and short range bandwidth using a static SCR implementation Dynamic cellular and short range bandwidth

(c) Dynamic subcarrier allocation of cellular and short range bandwidth using a dynamic SDR implementation Figure 22.4. Three subcarrier allocation scenarios illustrating subcarrier allocation using SCR and SDR implementations.

in the short range links enable the use of a higher order of modulation format at a lower transmit power and shorter transmit time as information distributed among cooperating terminals are transmitted using a higher order of M bit per information symbol than the cellular channel. To illustrate the need for flexible spectrum assignment, we consider the scenarios shown in Figure 22.5, together with the example for subcarrier assignments for those scenarios, shown in Figure 22.6. Looking at Figure 22.5(a), and Figure 22.6, data streams D1 ...D6 are assigned to chosen specific sub-carrier in the part of OFDM spectrum allocated to cellular communication. As there are only two mobile devices the datastreams are portioned with D1 ...D3 being allocated to M1 and D4 ...D6 allocated to M2 . The mobile devices

442

Jesper M. Kristensen and Frank H.P. Fitzek

(a) Two mobile device scenario

(b) Three mobile device scenario Figure 22.5. Two and three mobile scenarios indicating separation of mobile devices and related SNR. now cooperate by sharing each of their information parts with each other. Assuming channel reciprocity, and the signal-to-noise ratio (SNR) being inverse proportional to the distance between the mobile devices, the relation can be described as SN R ∝ 1/d2 assuming free space propagation. Assuming further an equal transmit power constraint for the two mobile devices and that transmit power for allocated for short range transmission are equally distributed, considering only an AWGN channel, the received SNR is assumed to be equal in both directions of transmission. As the subcarrier allocation is determined by the required rate and modulation format chosen, the subcarrier allocation shown in Figure 22.6 are considered optimal assuming the condition of similar received SNR in both directions. The condition

22 CCP2P with SDR

443

Channel Bandwidth

M1

D1

D2

D3

D1

D2

D3

M2 M3

D1

D5

Short range spectrum

D6

D1 D 2 D 3 D4

M2

M1

D4

D5

D6

D2

D4 D5 D 6

D1 D3

D4

∆f

D2 D3

D5

D6

D4 D5

D6

Figure 22.6. Example of a channel with datastreams partitioned to different mobile devices and retransmitted over short range links.

of equally received SNR was valid in the case of channel reciprocity and a equal transmit power constraint. The latter requirement is likely not to be satisfied in a scenario with mobile devices engaging in a cooperation but with unequal resources and processing capabilities. The subcarrier allocation is then determined under the constraint set by rate requirements and processing as well as resource capabilities of the cooperating mobile devices. Looking at Figure 22.5(b), and Figure 22.6, the data streams from the cellular link are now portioned among three cooperating mobile devices. As indicated in Figure 22.5(b) the degrees of freedom in allocating sub-carrier in short range links between the mobile devices have increased as the condition of equal distance and equal received SNR is unlikely to be fulfilled as the number of cooperating devices increases. Similar to the case of two cooperating mobile devices, each mobile device is allocated a number of sub-carrier for short range communication between cooperating devices. As in the case of two mobile devices, each mobile device transmits the data streams assigned from the base station. The number of sub-carrier allocated for each data stream is dependent on the actual channel conditions. While again assuming channel reciprocity between mobile devices and equal transmit power constraint for each mobile device and thereby equal SNR in both transmit directions for two communicating devices, the SNR for each short range link will in general be dependent on the distance between communicating devices as indicated in Figure 22.5(b). We consider the following scenario as an example: Setting d1 = d and d2 = d3 = 2d, furthermore assuming an ideal free space transmission, where received power relates to transmitted power using the Friis transmission formula Pr = Gt Gr Pt

λ 4πr

2 (22.1)

Assuming the wavelength of a transmission to be constant, Equation 22.1 becomes

444

Jesper M. Kristensen and Frank H.P. Fitzek 1 Pr =K 2 Pt r

(22.2)

2

λ . Using the Shannon relation for maximum rate in a Gaussian with K = Gt Gr 4π channel and inserting the relation of Equation 22.2 1 Pt R = log2 1 + K 2 2 (22.3) σ r

Here it is assumed that Pt and σ 2 are the allocated transmit power and the noise variance per subcarrier respectively. K is a constant and r is the distance between two mobile devices. Equation 22.3 expresses the maximum achievable rate per subcarrier assuming a Gaussian channel and free space transmission. It is clear that doubling the distance between the two mobile devices reduces the received power and thereby the received SNR by one fourth and that this reduction in SNR has a similar impact on the achievable rate. The achievable rate can either be increased by increasing the the transmit power, limited by the imposed power constraint, increasing the spectral efficiency by using higher order modulation formats, limited by the available SNR, or by increasing the bandwidth, that is by allocating more sub-carrier to the short range communication suffering from decreased SNR. Considering our example with reference to Figure 22.5(b), M1 and M3 are separated with a distance of d, both M1 and M2 as well as M2 and M3 are separated with a distance of 2d. Given the above considerations, the communication between the mobile devices separated with as distance of 2d would be allocated more sub-carrier to accommodate a decreased SNR. This accounts for the ∆f in Figure 22.6 which indicates that compared to the scenario in Figure 22.5(a) the spectrum usage is increased as the separation between mobile devices is increased. In this sense, the spectrum usage for a scenario with only two cooperating mobile devices can be considered optimal compared to scenarios in which mobile devices are separated at distances larger than minimum distance of d. It is evident that as the number of cooperating devices increases, so does the degrees of freedom with respect to subcarrier allocation and that different channel conditions between cooperating mobile devices requires flexible subcarrier allocation to accommodate a given rate requirement. As given in Chapter 30, the mobile devices need to understand whether the cooperation with two or three mobiles is more beneficial. I an nutshell, each additional cooperative device will decrease the number of sub-carriers needed for the cellular communication for each device. At the same ∆f may potentially increase depending on the location of the cooperating terminals too each other. Therefore whether to cooperate with two or three mobile devices depends on ∆f and the cooperative strategy explained in Chapter 30. We argued the need for flexible subcarrier allocation to accommodate different channel conditions, and though the arguments were based on a simplified assumption of free space transmission and Gaussian channels, the same principles hold in the general case of multi-path channels. We now discuss the implications on the choice of implementation technology considering the requirement on flexibility. In the following paragraphs, three different downlink scenarios of subcarrier allocations with increased requirement of flexibility is introduced. Based on these different scenarios, it is argued that in the case of a scenario with cooperation as well as dynamic carrier allocation that a SCR implementation will appear to be less feasible in all terms of architecture performance factors, complexity, cost, flexibility and efficiency. This

22 CCP2P with SDR

445

leads to the argument that SDR has the potential to be a feasible implementation alternative to the SCR in a cooperative wireless network scenario. With reference to Figure 22.7 the first scenario is similar to a traditional broadcast cellular communication between a base station and terminal using the principle of OFDM. No cooperation is formed meaning that terminals communicating with the base station are processing the whole bandwidth as indicated in Figure 22.7. All sub-carrier are assigned to each terminal. From the perspective of a terminal, the receiver architecture may need accommodate the flexibility of adapting FFT sizes according to data rate requirements and users together with adaptive modulation and coding for each subcarrier or group of sub-carrier. Similarly, forming a multicast scenario, the base station may allocate groups of sub-carrier to each terminal, i.e., FDMA with OFDM. The kind of flexibility we required in these scenarios are flexibility by adaptation. The options for adaption is limited and the degree of flexibility needed is not assumed to be high enough to justify the employment of SDR with and increased cost in complexity and efficiency.

Figure 22.7. Traditional Downlink scenario with no cooperation. Referring to Figure 22.8, the second scenario introduces cooperation, comparing to the previous scenario this scenario is a cooperative broadcast scenario, but here

446

Jesper M. Kristensen and Frank H.P. Fitzek

it is assumed that sub-carrier are statically assigned. It is apparent that the use of static carrier assignment, although providing for low complexity in system implementation, pose the risk of a group of sub-carriers being positioned in a deep channel fade thereby causing a degradation in link quality and data rate for that group of sub-carrier, this calls for the consideration of dynamic subcarrier allocation. Considering an architecture for a terminal supporting the cooperative broadcast scenario reveals the first distinction between a Software Controlled Radio based architecture and an Software Defined Radio based architecture and the concept of flexibility by reconfiguration. Considering an SCR implementation, it is apparent by looking at Figure 22.1 that the concept of cooperation involves the accommodation of a terminal having the capability of receiving and processing a number short range range links according to how many terminals that are participating in the cooperation. We now recall the definition of a state of the art SCR implementation as one that has a fixed functionality typically implemented in dedicated hardware, with different parameters being controllable in software. Accommodating the principle of cooperation in an SCR implementation would require the implementation of short range receive chains in hardware. But as the number of cooperating terminals increases so does the need for receive chains. Implementing these receive chains in hardware will likely compromise the most important reasons for SCR implementations known today, namely cost, efficiency and complexity. Furthermore having terminals supporting only a limited number of cooperating entities, to limit the hardware complexity, will limit the application and thereby the advantages gained by cooperation. Finally considering that sub-carriers might be assigned not statically but dynamically to achieve the best link quality for each cooperating terminal will impose further complexity requirements on a SCR implementation. These considerations lead to the argument that an SDR implementation offers the possibility of accommodating a flexible number of cooperating terminals by employing what we define as flexibility by reconfiguration. Flexibility by reconfiguration is to be understood as the ability of an SDR to configure an architecture to the number of short range links necessary in a given scenario. Recalling from the previous scenario that a static subcarrier assignment posed the risk of a group of sub-carrier to be positioned in deep channel fade causing a degradation in link quality. The third scenario therefore introduces cooperation with a dynamic subcarrier assignment. This dynamic subcarrier assignment is based on knowledge about the channel quality between the base station and a given terminal as well as channel quality between cooperating terminals. A possible subcarrier assignment is shown in Figure 22.9 and illustrates also that subcarrier assignments need in principle not to be contiguous in the sense that cellular and short range allocations need to be adjacent to each other. Subcarrier allocations can be freely distributed among cellular and short range carriers. Again the principle motivation is gain in energy efficiency, this involves, for each cooperating terminal, choosing a group of sub-carrier with the best link quality, enabling the minimization of transmit power and the use of higher order modulation format and less coding overhead on the short range links. From an architecture point of view, this puts a further requirement on the flexibility i.e., the capability of selecting and processing subbands or groups of sub-carrier. Considering this further requirement on flexibility again argues in favor of an SDR implementation with respect to cost, complexity and flexibility.

22 CCP2P with SDR

447

Figure 22.8. Cooperative principle with FDMA and OFDM.

22.3 SDR and SCR Architectures In the previous sections we have argued that as the requirements for flexibility increases, the advantages of an SCR implementation would decrease to the point where it is likely that an SDR implementation will be more feasible. When we are considering the feasibility of an SCR implementation versus an SDR implementation, we are considering the following parameters: Cost, Efficiency, Complexity and Flexibility. The SCR has traditionally been known for its advantages in the first three parameters due to the fact that SCR implementations are usually made up by dedicated hardware optimized for a given communication standard and application. The visions surrounding the SDR has been to implement flexibility, this has often been at the expense of cost; due to the requirement of a broadband or frequency agile RF front-end and reconfigurable baseband processing; Efficiency as the ideal SDR strives to move the ADC and DAC operations as close a possible to the antenna, putting strict requirements on the performance of ADC and DAC components and finally complexity as digital signal processing are used to support traditionally RF or analog implemented functionality. With the introduction of the concept of cooperative wireless networks, it therefore seems feasible to to argue a tradeoff between

448

Jesper M. Kristensen and Frank H.P. Fitzek

Figure 22.9. Scenario with dynamic carrier assignment.

the above parameters. Figure 22.10 summarizes the parameters and the concept of a design space determined by them. In the following is given a short introduction to the parameters and the reasoning behind them.

Figure 22.10. Architecture metrics and Design space.

22 CCP2P with SDR

449

The parameters defined in Table 22.2 make up the design space within a feasible architecture that has to be chosen. Flexibility is the main requirement, but an implementation has to be chosen that is still efficient, low cost and feasible in complexity.

Table 22.2. Parameters used for comparing SCR and SDR technologies. Cost

Cost is manifold, it can be related to cost in development, implementation and production and also the cost related use of energy resources. Efficiency Energy; Does it make efficient use of energy resources, processing; Does it make efficient use of processing resources, can resources be reused, or does it require replication Complexity Design; What are design complexity, i.e., the HW implementation complexity, Computational; What are implication on computational complexity requirements Flexibility We define flexibility by the ability to adapt and reconfigure functionality

22.3.1 SCR Architecture Versus SDR Architecture A generic SCR architecture is shown in Figure 22.11. When comparing that architecture to a generic SDR architecture it is seen that a generic SCR architecture is built upon separate hardware chains, where each hardware chain supports for example a different communication standard in case of a state of the art mobile phone. In implementation the hardware chain may partly share physical hardware processing resources as well analog RF front-end circuitry. Functionality is then changed altering parameters such a frequency range, modulation techniques etc. and the switching of key hardware components that determine the functionality of the hardware chain. This is depicted with the control from the link layer, controlling the parameters and a hardware control that activates the key hardware components needed to create the determined functionality of a hardware chain. In contrast the view of the SDR architecture as depicted in Figure 22.12 is an architecture built upon a frequency agile RF front-end and reconfigurable baseband processing. The concept of an SDR architecture therefore avoids the necessary replication of hardware components seen in the SCR architecture. Similar to the SCR architecture, the functionality of the SDR architecture is controlled from the link layer with added control for reconfiguration of the hardware chain to accommodate the change of communication standards for example, see also the section on Software Defined Radio in 33. It is furthermore shown in Figure 22.12 that an SDR architecture may accommodate separate chains for the simultaneous operation of e.g., cellular and short range communication links for example. Looking at the SDR architecture from the perspective of a cooperative scenario as describe above, Figure 22.12 shows that in a cooperative scenario an SDR architecture should accommodate the need for separate baseband processing chains according the number

450

Jesper M. Kristensen and Frank H.P. Fitzek

Figure 22.11. SCR conceptual architecture.

Figure 22.12. SDR conceptual architecture.

22 CCP2P with SDR

451

of cooperating terminals. These separate processing chains are in principle dynamically created and closed down and this differs from the SCR architecture where that number of baseband processing chains would have to be statically defined either, adding to the complexity of an implementation or limiting the architectures ability to take advantage of the full potential of a cooperative scenario.

22.3.2 SDR Receiver Architecture for Cooperating Terminal In the following we outline an architecture for an SDR receiver supporting the characteristics of cooperative wireless network scenario using FDMA and OFDM with flexibility requirements as described in the above sections. Figure 22.13 gives a conceptual schematic for a proposed receiver architecture that can accommodate the principle requirements in the downlink cooperative scenario as outlined above. Let us elaborate on the functional blocks shown in the Figure. Subchannel filters The front filters select the band of sub-carrier allocated for cellular reception and short-range reception respectively. They are flexible in center frequency and bandwidth according to requirements from the upper layers. From an implementation standpoint low complexity is essential as the complexity related to the filtering of sub-bands reduces the gain in complexity achieved by reducing the FFT size at each terminal. FFT Cellular/short-range FFT’s process selected bands of sub-carrier, they are flexible in size. Demodulation and decoding Demodulation and decoding must be flexible in the support of adaptive modulation and coding (AMC) employed in the cooperative scenario. The use of higher order modulation formats is an essential assumption in the achieved benefits for a cooperative wireless network. This assumption is however well justified as high order modulations are naturally used within the short-range clusters. Flexibility and adaptability Figure 22.13 indicates that adaptability is introduced through parameters controlling the receiver chains in terms of FFT size, modulation and coding. Furthermore flexibility is introduced through the dynamic instantiation of short-range receive chains according the number of terminals interacting cooperatively. Vertical configuration This entails the dynamic reconfiguration of receiver chains to accommodate the dynamically changing number of cooperating terminals and therefore the time varying number of short-range communication links. In principle, this requires the ability to instantly set up receive chains at base-band level according to the number of cooperating entities. From the perspective of a cognitive communication system, the ability to reconfigure the terminals according to the actual conditions is what differentiates the cognitive terminal from the non-cognitive terminal. Horizontal configuration Horizontal configuration entails the configuration or adaptation of the functional blocks of both, cellular and the short-range receive chains. For the receiver chain the immediate parameters include the filter bandwidth and center frequency for extracting a group of sub-carrier. Similarly, the FFT size is subject to dynamic configuration, and finally so is the support of adaptive modulation and coding. The parameters for the horizontal configuration depend on the vertical configuration.

452

Jesper M. Kristensen and Frank H.P. Fitzek

Figure 22.13. SDR architecture.

22.4 CCP2P Testbed with the GNU Radio This section gives an introduction to set up a testbed for a CCP2P scenario using the GNU radio.

22.4.1 Introduction to GNU Radio The GNU Radio [2] is an open source project providing a software toolkit that when combined with a hardware front-end, see Figure 22.14, provides the possibility for building radio transceivers which functionality is defined by software. The GNU radio and the toolkit therefore allows learning, building and testing wireless concepts based on a software implementation of a radio. Because of the open source nature of the GNU radio platform and therefore the low cost of implementation, the GNU radio platform provides the possibility for creating flexible software radios for real time proofing of concepts within wireless communication. The functionality implemented with the toolkit supplied by the GNU radio is executed on a standard desktop computer supporting either MS-Windows, Linux or Mac OSX. The Hardware front-end indicated in Figure 22.14, is termed the Universal Software Radio Peripheral, USRP [3] designed by Ettus Research LLC, and consists of a main board containing 4 ADC’s sampling at 64MSPS and 4 DAC’s sampling at 128 MSPS together with a programmable FPGA for preprocessing: down-conversion and decimation [1], before entering the signal processing done by the GNU radio software toolkit. Optional daughter-boards with for example RF front-ends can be easily connected through 4 expansion ports providing two independent transmit interfaces and 1, 2 or 4 independent receive interfaces depending application and chosen data

22 CCP2P with SDR

453

Figure 22.14. GNU Radio block schematic.

formats. The physical interface between the USRP and the desktop computer is a USB 2.0 interface capable of sustaining up to 32 Mbytes/s. The USB interface supports half-duplex meaning that the available data rate over the USB interface is being shared between receive and transmit. Considering a given application in the receive direction, the 32 MB/s is divided between I and Q signals using 16 bits per sample, giving a gross rate of 8MSPS (complex samples) over the USB interface. This rate will degrade in an application of both TX and RX and with the number of independent receive chains. Also the processing overhead related to the processing being carried out on the host PC will degrade maximum achievable data rate across the USB interface. The GNU radio can therefore not necessarily support high data rate applications, but provide a platform for building test beds for experimental setup’s, which is the purpose of introducing the GNU Radio platform for building a cooperative communication testbed.

22.4.2 GNU Radio Setup for a CCP2P Cooperative Scenario The test bed setup for a cooperative communication scenario is shown in Figure 22.15. The figure shows a testbed using three GNU radio platforms. The setup illustrated is shown with wired connections between nodes. The wired setup is advantageous for functionality test and experiments. The wired setup can in principle be exchanged with a wireless setup, using appropriate RF front-ends, for performance metric evaluation. Considering the setup shown, each GNU Radio platform is programmed to support two independent receive chains as well as two independent transmit chains. The USB interface supports half duplex, meaning that TX and RX directions is separated in time. Considering the achievable gross rate over the USB interface, and assuming that the TX directions account for the use of bandwidth as the RX directions, the gross data rate across the USB interface is approximately 1MSPS (complex samples).

454

Jesper M. Kristensen and Frank H.P. Fitzek

Figure 22.15. GNU radio testbed using three GNU radio platforms and with symbols showing wired connections for functionality test and experiments.

22.5 Discussion We discussed the implementation of an OFDM based cellular controlled peer to peer communication scenario. A combined cellular and short range air interface using the principle of FDMA together with OFDM was presented. It was argued that flexibility is a key factor in the implementation of CCP2P scenario. The requirement for flexibility is used to accommodate a scenario of dynamically changing number of cooperating devices. This scenario requires a flexible subcarrier allocation together with AMC. Furthermore CCP2P architecture should support dynamic reconfiguration of short range transceiver chains. This requirement lead to the argument of using Software Defined Radio over a state of the art Software Controlled Radio. As the requirement for flexibility increases, the SCR is argued to be a less efficient solution compared to a SDR solution. Based on the requirements established in previous sections, we introduced an architecture and outlined the required functionality for such an architecture. The GNU radio was presented as a platform that will be used to establish a testbed for a CCP2P scenario. The GNU radio is a cost effective and flexible implementation platform that offers the possibility of verifying functionality as well as performance in a realtime setup.

References 1. Exploring gnu radio. gnuradio.html.

http://www.gnu.org/software/gnuradio/doc/exploring-

22 CCP2P with SDR

455

2. Gnu radio web site. http://www.gnu.org/software/gnuradio/. 3. Usrp wiki. http://www.comsec.com/wiki?UniversalSoftwareRadioPeripheral. 4. E. Buracchini. The software radio concept. Communications Magazine, IEEE, 38(9):138–143, 2000. 5. F.H.P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-4020-4710-X. Springer, April 2006. 6. E. Lawrey. Multiuser ofdm. In Signal Processing and Its Applications, 1999. ISSPA ’99. Proceedings of the Fifth International Symposium on, volume 2, pages 761–764 vol.2, 22-25 Aug. 1999. 7. L. Litwin. An introduction to multicarrier modulation. Potentials, IEEE, 19(2):36–38, 2000. 8. Joseph Mitola. Software Radio Architecture. John Wiley & Sons, INC, 2000. ISBN 0-471-38492-5. 9. Charles D. Murphy. Low-complexity fft structures for ofdm tranceivers. IEEE Transactions on Communications, vol. 50(no. 12):p. 1878–1881, December 2002.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks On the Potential with Transmit Beamforming

Chenguang Lu, Frank H.P. Fitzek, and Patrick C.F. Eggers Aalborg University, Department of Electronic Systems, APNET Section, [cgl|ff|pe]@es.aau.dk Summary. In this chapter, we introduce a cooperative scheme enabling spatial reuse, namely cooperative spatial reuse (CSR), as a cooperative extension of the current TDMA-based MAC in wireless networks. In the CSR, a cooperative group is formed by the links that are willing to do spatial reuse. In the group, every cooperating link contributes its time slots for spatial reuse among the cooperating participants. Following the cooperation principle, a link joins the group only if it can benefit. Otherwise, the link will stop doing CSR and switch back to the TDMA-based MAC. As an example to illustrate the potential of the CSR, we focus on the transmit beamforming techniques on MISO (Multiple Input Single Output) links. We compared the CSR scheme using zero-forcing (ZF) transmit beamforming, namely ZF-CSR, to the TDMA-based MAC using maximum ratio combining (MRC) transmit beamforming, namely MRC-TDMA. The numerical results of a simulated two 2 × 1 MISO links scenario show the great potential of CSR to substantially increase the capacity and energy efficiency.

23.1 Introduction In recent years, wireless networks, such as WLAN (wireless local area network), have been increasingly deployed all over the world. In those networks, to reduce the system costs, there is usually no centralized access control. Each link competes for using the channel at a time and operates for its own interests. Current MAC (medium access control) layer protocols like the DCF (Distributed Coordination Function) in IEEE 802.11 MAC [5] are designed to distributedly coordinate the competition in such a way that only one link is allowed to operate at a time in its contention region. It is basically a TDMA-based MAC layer design. When one link obtains the access, the other links in its contention region keep silent during its transmission to avoid collisions. The contention region is usually conservatively specified to keep the interference from outside in a very low level and hence maintain the high link capacity of the active link. However, this design neglects the opportunity of spatial reuse allowing simultaneous transmissions of multiple links. Spatial reuse has gained a lot of attentions in ad-hoc networks as it has the great potential to increase the network capacity. The current studies on spatial

457 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 457–471. c

2007 Springer.

458

Chenguang Lu et al.

reuse (e.g., [4, 7–9]) focus on evaluating spatial reuse with respect to the size of the contention region. The analysis is usually based on an assumption of singlerate networks in which all links transmit at a fixed data rate. When one link is transmitting, the other links are allowed during its transmission if the resulted signalto-interference-plus-noise ratio (SINR) of the current link will not corrupt its ongoing packet reception. In this work, we propose a cooperative scheme to enable spatial reuse in multirate networks — cooperative spatial reuse (CSR) in which every cooperating link contributes its time slots obtained by the TDMA-based MAC to allow spatial reuse among the cooperating participants. Therefore, the CSR is a cooperative extension of the current TDMA-based MAC. Although the capacity of a link is reduced at its own time slots due to the interferences from other links, it is possible for the link to accomplish more traffic during all available time slots as it can get more time slots from others to transmit. Following the cooperation principle [3], every link doing CSR should benefit. Otherwise, it should stop doing it. In this work, we derive the conditions for each link doing CSR to gain more capacity and energy efficiency. The capacity region of CSR is defined. Furthermore, we define the availability of CSR as a measure of its occurrence probability to evaluate how easy to find cooperative partners. This is very important as the efforts should not be wasted in MAC design on something happening rarely. Lack of interference mitigation capabilities is possibly the reason for current MAC layer designs not considering spatial reuse. However, the wireless networks have been starting evolving to using multiple antenna as the next generation WLAN — MIMO (Multiple Input Multiple Output) WLAN is being standardized [6]. Multiple antenna techniques have been shown the great potential to increase the link capacity by spatial diversity and spatial multiplexing [1]. Meanwhile, it also facilitates spatial reuse by interference cancellation. Thereby, for increasing capacity, multiple antenna techniques can either boost each single link capacity at each time slot or enable spatial reuse to obtain link multiplexing gain. Which scheme is better? In this work, we focus on the transmit beamforming techniques to enable CSR on MISO (Multiple Input Single Output) links. Especially, we take a simulated twolinks scenario to show the performance of the CSR with zero-forcing (ZF) transmit beamforming, namely ZF-CSR, compared to the TDMA-based MAC with maximum ratio combining (MRC) transmit beamforming, namely MRC-TDMA. The numerical results show that the ZF-CSR scheme has the great potential to further increase the capacity and energy efficiency of the MRC-TDMA scheme. The rest of this chapter is organized as follows. In Section 23.2, we describe the proposed CSR scheme, define and illustrate the CSR capacity region, and define the CSR availability. Section 23.3 compares the MRC-TDMA scheme and ZF-CSR scheme on MISO links. The numerical results of a two-links scenario are given in Section 23.4 and we conclude the paper in Section 23.5.

23.2 Description of the Proposed CSR Scheme In TDMA-based MAC layer designs, every link exclusively uses its own time slots. In this sense, every link takes the time slots as its private resource. Doing spatial reuse requires every link to share out its personal time slots for a collective use. Therefore, spatial reuse is to use the time slots cooperatively among links. As a result, every

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

459

link loses capacity at its personal time slots due to the mutual interferences between links. However, the effective capacity of a link can be increased if the sum capacity obtained from other spatial reused time slots is more than the capacity lost at its personal time slots. Based on the cooperation principle [3], cooperation only works if every participant gains. As a cooperative scheme, a spatial reuse scheme should be designed in such a way that every link that contributes its own time slots should obtain enough time slots from the others to guarantee that it gets more capacity. Therefore, the CSR scheme is proposed to satisfy the cooperation conditions that every participant should gain. The basic idea is that the cooperating links form a cooperative group sharing all their time slots to multiplex their transmissions. The time slots are obtained by the TDMA-based MAC. A link joins the group only if it can obtain benefits from it, e.g., capacity increase. If no benefit is obtained, it leaves the group and switch back to the TDMA-based MAC. It should be noted that, with the proposed CSR, all other cooperating links should be able to recognize the ownership, starting time and duration of the upcoming time slot and be prepared to transmit together with the current link that owns the upcoming time slot. As an example, without loss of generality, Figure 23.1 illustrates a two-links scenario in a round-robin TDMA manner where ri is the link capacity with the TDMA-based MAC at the time slots owned by the ith link and ri 0 is the counterpart with CSR. ri 0 < ri due to the existence of mutual interferences with CSR. Figure 23.2 shows the power consumption situation of the transmitter and the receiver of link 1. PT M , PRM and PIM are the power consumption of a wireless transceiver in the transmitting mode, receiving mode and idle mode, respectively. When a wireless transceiver is not either transmitting or receiving, it enters into the idle mode to save power. Normally, PT M > PRM > PIM .

Figure 23.1. Illustration of cooperative spatial reuse.

460

Chenguang Lu et al.

Figure 23.2. Energy consumption of cooperative spatial reuse. Table 23.1. Comparison of the TDMA-based MAC and CSR. Effective capacity

TDMA-based MAC Ri = PkTi T ri

Energy efficiency Ei =

j

j=1 P PXM Ti +PIM ( k j=1 Tj −Ti ) ri Ti

CSR Ri 0 = ri 0 Ei 0 =

PXM ri 0

23.2.1 Cooperation Conditions In this work, we assume each link has even access probability which is the case in wireless network for the traffic with the same priority. In the following, we derive the cooperation conditions for a link to gain more capacity and energy efficiency. Without loss of generality, to compare CSR to the TDMA-based MAC, we calculate the mean effective capacity and energy efficiency of each link only within the cooperative group, during the cooperating time slots. The energy efficiency is defined as energy consumption per bit. To simplify the analysis, the length of each time slot of a link is assumed equal long and the power consumption parameters (i.e., PT M , PRM and PIM ) are assumed the same at each link. The results are summarized in Table 23.1 for the k cooperative links scenario. Ri and Ri 0 are the mean effective capacity of the TDMA-based MAC and CSR, respectively. Ei and Ei 0 are the mean energy efficiency of the TDMA-based MAC and CSR, respectively. Ti is the length of the time slot of the ith link, PXM = PT M for the transmitter case, and PXM = PRM for the receiver case. Thereby, comparing to the TDMA-based MAC, there are two conditions for the ith link to benefit from CSR with respect to the effective capacity and energy efficiency, respectively. Condition I (effective capacity): Ri 0 > Ri which can be rewritten as

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks Ri 0 > ri Xi where Xi =

T Pk i

j=1

Tj

461 (23.1)

.

Condition II (energy efficiency): Ei 0 < Ei which can be rewritten as Ri 0 > ri Xi where Y =

PIM PXM

1 Xi + Y (1 − Xi )

(23.2)

.

For Condition II, there are two cases, the transmitter case where PXM = PT M and the receiver case where PXM = PRM . Compares (23.2) to (23.1), it is easily seen 1 that Condition II is tougher than Condition I because Xi +Y (1−X > 1 with Y < 1. i) Therefore, the links satisfying Condition II must satisfy Condition I. It means that the links that obtain more energy efficiency also achieve more capacity. Furthermore, when Y = 0, Condition II is changed to Ri 0 > ri which is impossible due to Ri 0 = ri 0 and ri 0 < ri . That means it is impossible for a link to benefit energy efficiency from CSR if the idle mode consumes no power. However, in practice, for instance of IEEE 802.11 transceivers, the power consumption of the idle mode is comparable to that of the receiving mode. The transceivers can not shut off all the circuits as they have to be prepared to receive the upcoming packet. Thereby, there is some space for CSR to gain more energy efficiency.

23.2.2 CSR Capacity Region Then we define the CSR capacity region as the region when all cooperative links gain over the TDMA-based MAC. The region is defined as SCSR = { (R1 0 , R2 0 , ..., Rk 0 ) | R1 0 ∈ S1 , R2 0 ∈ S2 , ..., Rk 0 ∈ Sk } where

( Si =

{ Ri 0 | Ri 0 > Ri }, { Ri 0 | Ei 0 < Ei },

on Condition I on Condition II

(23.3)

(23.4)

Each link chooses Condition I or Condition II based on its own situation. If energy efficiency is its main concern (e.g., in the case of battery-powered devices), it should choose Condition II. Otherwise, it should choose Condition I. Figure 23.3 illustrates the CSR capacity region for the case with two links. The line DE represents the achievable capacity with the TDMA-based MAC during the two time slots with all possible time sharing ratio, i.e., T1 /T2 in this case. Point D and E are the extreme cases when one of two links occupies both time slots, i.e., T1 /T2 = 0 or ∞. In these two extreme cases, the active link achieves its link capacity, r1 or r2 . In other points on the line, the effective capacity of two links are reduced due to the time sharing. As an example, point A is an operational point for a given time sharing ratio specified by the TDMA-based MAC used. At point A, R1 and R2 represent the effective capacity of the two links. Therefore, for point A, Region I shows the capacity region when both links satisfy Condition I while Region II shows 1 the capacity region when both links satisfy Condition II where αi = Xi +Y (1−X i) following (23.2). Region II is a subset of Region I as Condition II is tougher than Condition I. To be noted, point C can be approached, for an instance, when two links

462

Chenguang Lu et al.

are separated at infinity distance. Furthermore, for the beamforming case, it can also achieved when the channels are well separated such that the mutual interferences can be completely nulled by the beams which are optimal for the reception of the desired signals on both links in the interference-free case.

Figure 23.3. Illustration of the CSR capacity region of a two-links example.

23.2.3 CSR Availability From the above discussion, it is possible for links to benefit from CSR in effective capacity and energy efficiency. Another important aspect to evaluate the usability of a CSR scheme is on how easy to form a cooperative group. Links are willing to try CSR only if it is not difficult to find cooperative partners around. Therefore, we define the CSR availability as the probability falling into the CSR capacity region ACSR = Pr [(R1 0 , R2 0 , ..., Rk 0 ) ∈ SCSR ]

(23.5)

If the availability is low, it is not worthwhile using CSR as the partner finding process will waste a lot of capacity.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

463

Figure 23.4. MISO links with transmit beamforming.

23.3 CSR with Transmit Beamforming on MISO Links 23.3.1 Transmit Beamforming on MISO Links In order to show the potential of CSR, the focus is on the transmit beamforming techniques on MISO links. It also reflects a downlink scenario with multiple-antenna access points and single-antenna terminals. Figure 23.4 shows k n × 1 MISO links including a beamforming transmitter with n antennas and k single-antenna receivers where the 1st receiver is the intended receiver. Therefore, the received signal at the ith receiver is expressed as ri (t) = (hi )T ws(t) + ni (t)

(23.6)

where (·)T denotes the transpose of a vector, hi = [h1i h2i ... hni ]T is the channel vector of the ith MISO link, w = [w1 w2 ... wn ]T is the weight vector, s(t) is the transmitted signal, and ni (t) is the noise at the ith receiver. With the TDMA-based MAC, the capacity can be enhanced by focusing the beam to the intended receiver to achieve array gain and diversity gain. The beamforming technique can also be used to cancel the mutual interferences between cooperating links to enable CSR. In the following, we will compare two transmit beamforming schemes with the TDMA-based MAC and CSR, respectively, namely MRC-TDMA and ZF-CSR.

23.3.2 MRC-TDMA Versus ZF-CSR The MRC-TDMA scheme uses the TDMA-based MAC and each beamforming transmitter applies MRC weight vector to maximize the signal-to-noise ratio (SNR) at the intended receiver and thus maximize the capacity of the MISO link. In the ZF-CSR scheme, the weight vector of each cooperating transmitter is set in such a way that the received signals at the receivers of other cooperating links are canceled (forced to be zero). Therefore, the cooperating links can transmit simultaneously without the mutual interferences. Table 23.2 gives the weight vector setting of the two schemes. For a fair comparison, the weight vectors are normalized. h1 denotes the channel vector of the desired link, (·)∗ denotes the conjugate of a vector, k·k denotes the Euclidean norm

464

Chenguang Lu et al. Table 23.2. Weight vector of MRC-TDMA and ZF-CSR. MRC-TDMA Weight vector wM RC =

∗

(h1 ) k(h1 )∗ k

ZF-CSR +

wZF =

H Ik×1

+

H Ik×1

Figure 23.5. CDF of the SNR of the ZF beamforming and the MRC beamforming. +

of a vector, H = [h1 h2 ... hk ]T is the channel matrix of the cooperating links, (·) denotes the pseduoinverse of a matrix, and Ik×1 denotes the first column of a k × k identity matrix.

SNR Comparison As the MRC transmit beamforming maximizes the SNR at the intended receiver, the SNR of ZF transmit beamforming is lower for canceling the interferences. In principle, on the i.i.d. (independent, identically distributed) complex Gaussian fading channel, the M -branch (M -antenna) MRC beamforming can provide M -fold of diversity order. For the M -branch ZF beamforming canceling the interferences at L receivers, the diversity order is reduced to M − L. As an example, we compare the SNR at the intended receiver of the two schemes in a scenario of two 2 × 1 MISO links. The MRC-TDMA scheme uses the 2-branch MRC beamforming while the ZF-CSR scheme uses the 2-branch ZF beamforming that cancels one interference. Figure 23.5 shows the CDF (cumulative distribution function) of the SNR at the intended receiver over 1,000,000 channel realizations. In the simulation, the i.i.d. complex Gaussian fading channel is assumed and the mean SNR per branch is 0 dB.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

465

Figure 23.6. CDF of the channel capacity ratio between ZF-CSR and MRC-TDMA, (CZF /CM RC ), on the desired link. SN Rb denotes the mean SNR per branch.

Link Capacity Comparison The reduced SNR at the intended receiver with the ZF-CSR scheme gives the lower data rate on the desired link than the MRC-TDMA scheme. In this work, we compare the link capacity on the desired link of the two schemes from the informationtheoretical point of view. According to information theory [2], the channel capacity of the beamforming MISO link on a AWGN (additive white Gaussian noise) channel is expressed as C = log 2 (1 + SN R) (bits/s/Hz) (23.7) where SN R is the SNR at the intended receiver. For the same two-links example shown in the SNR comparison, Figure 23.6 shows the CDF of the channel capacity ratio between the ZF-CSR scheme and the MRC-TDMA scheme showing the relative channel capacity loss on the desired link with the ZF-CSR scheme. The probability that the channel capacity ratio is less than a given value increases as the SNR per branch decreases. Therefore, the higher SNR gives the less loss in the channel capacity for using the ZF beamforming. For example, the ZF beamforming can achieve over 70% channel capacity of the MRC beamforming in about 90% cases when the SNR per branch is 30 dB. However, when the SNR per branch is 0 dB, only about 40% cases give over 70% channel capacity of the MRC beamforming. To fulfill the cooperation conditions of CSR, the low capacity loss means that the link needs less time slots from the other links to compensate its capacity loss. Therefore, the higher SNR should give the higher CSR availability. Even though the above analysis is based on the channel capacity, it is useful as the advanced coding techniques approach the channel capacity. In practice, the

466

Chenguang Lu et al.

gradient of the actual rate adaptation curve (data rate versus SNR) is even less than that of the channel capacity curve. In this case, the actual performance will be even better than the results above.

23.4 Numerical Examples In the following, we will take a simulated scenario of two 2 × 1 MISO links to illustrate the potential of the ZF-CSR scheme compared to the MRC-TDMA scheme. Especially, we will show the CSR availability, average capacity gain, and average energy efficiency saving. Assume the fading channel on each branch is the i.i.d. complex Gaussian channel and the noise is AWGN. The perfect channel knowledge is assumed available at the transmitters. Furthermore, we assume the even access probability of each link as discussed in Section 23.2. The time slot length of each link, Ti in Table 23.1, is fixed and Ti ∝ 1/ri where ri is the data rate of the ith link with the MRC-TDMA scheme. This reflects the scenario with the fixed packet length at each time slot of each link. The power consumption parameters are set such that PT M = 2 W att, PRM = 0.95 W att, and PIM = 0.85 W att. In the simulation, 1,000,000 channel realizations are simulated on each link to obtain the capacity statistics. The mean SNR of each branch on each link is from 10 to 30 dB, which are the practical values for dense deployed networks. The performance evaluation is based on the channel capacity using (23.7) from the information-theoretical point of view. Therefore, in the simulation, the channel capacity is used as the data rate of each link. As discussed in Section 23.3.2, in practice, the actual performance will be even better than the numerical results below.

23.4.1 CSR Availability Figure 23.7 shows the CSR availability versus the mean SNR per branch on the two links when both links satisfy Condition I. It corresponds the probability for the points falling to Region I illustrated in Figure 23.3. Generally, the availability decreases as the SNR of either link decreases. The decreasing rate of the availability over the SNR increases as the SNR difference of the two links increases. For the whole SNR region, the availability is fairly high from about 0.6 up to about 0.9. It means that it is fairly easy for a link to find a cooperative partner to both achieve more capacity. Figure 23.8 shows the availability when both links satisfy Condition II. Figure 23.8(a) shows the receiver case where PXM = PRM in (23.2) for both links while Figure 23.8(b) shows the transmitter case where PXM = PT M . It corresponds the probability for the points falling to Region II illustrated in Figure 23.3. As expected, the availability is lower than the Condition I case as Condition II is tougher. Furthermore, the availability is reduced less in the receiver case than the transmitter case as PRM < PT M . The availability range for the receiver case is still fairly high from about 0.55 up to about 0.9 while it is reduced to the range from about 0.35 to about 0.75 for the transmitter case. The availability reduction of the receiver case is small as PRM is just slightly higher than PIM . Therefore, it is fairly easy for two links to form a cooperative group to make both receivers more energy efficient. However, it is much more difficult to achieve more energy efficiency at both transmitters.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

467

23.4.2 Capacity Gain From above discussions, it is not difficult for a link to find a cooperative partner to both obtain more capacity. We will show the capacity gain when a cooperative group is successfully formed. Figure 23.9(a) and Figure 23.9(b) show the average capacity gain on link 1 and link 2 respectively when both links satisfy Condition I. As expected, the average capacity on link 1 and link 2 are symmetric. The link with higher SNR will obtain more gain. For example, when SNRb on link 1 is about 30 dB and SNRb on link 2 is about 10 dB, link 1 can obtain averagely over 3 times the capacity of the MRC-TDMA scheme while link 2 only achieve less than 1.4 times. It is due to the difference of the time slot length of the two links. As the packet length is fixed for each time slot, the higher SNR link has a shorter time slot than the lower SNR link as the higher SNR offers the higher data rate. As a result, the higher SNR link can get a longer time slot from the lower SNR link to compensate the capacity loss at its own time slot while the lower SNR link get a shorter time slot from the higher SNR link. For the whole SNR region, each link achieves from about 1.4 up to about 3 times the capacity of the MRC-TDMA scheme. Figure 23.10 shows the total average capacity gain of both links. The whole cooperative group can obtain the significantly increased capacity from about 1.7 to about 2.1 times the capacity before.

23.4.3 Energy Efficiency Saving We have seen that it is fairly easy for two links to form a cooperative group to achieve more energy efficiency at receivers. Figure 23.11 shows the average energy efficiency saving when both links satisfy Condition II for the receiver case. The energy efficiency saving measures how much percentage energy is saved per bit than that of the MRC-TDMA scheme. Figure 23.11(a) and Figure 23.11(b) show the results of link 1 and link 2, respectively. Similar to the capacity gain, the average

Figure 23.7. CSR availability when both links satisfy Condition I. SN Rb denotes the mean SNR per branch.

468

Chenguang Lu et al.

(a)

(b)

Figure 23.8. CSR availability when both links satisfy Condition II for (a) the receiver case when PXM = PRM and for (b) the transmitter case when PXM = PT M .

(a)

(b)

Figure 23.9. Average CSR capacity compared to that of the TDMA-based scheme when both links satisfy Condition I on (a) link 1 and on (b) link 2.

energy efficiency savings at the receivers of link 1 and link 2 are symmetric. The higher SNR link saves more energy per bit as it has the higher capacity gain as shown in Figure 23.9. For example, when SNRb on link 1 is about 30 dB and SNRb on link 2 is about 10 dB, link 1 can save averagely over 60% energy per bit than before while link 2 only saves less than 20%. For the whole SNR region, the average energy efficiency saving of each link is significant from about 20% up to about 60% than before. Figure 23.12 shows the total energy efficiency saving of both links. The whole cooperative group can save from about 35% to about 39% energy per bit than before.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

469

Figure 23.10. Average total CSR capacity on both links compared to that of the TDMA-based scheme when both links satisfy Condition I.

(a)

(b)

Figure 23.11. Average CSR energy efficiency saving when both links satisfy Condition II for the receiver case on (a) link 1 and on (b) link 2.

23.5 Conclusions and Outlook The CSR scheme performs as a cooperative extension of the current TDMA-based MAC to enable spatial reuse in wireless networks. We derive the cooperation conditions for a link to do CSR to gain more capacity and energy efficiency. To further evaluate the usability of CSR, the CSR availability is defined to measure how difficult for links to form a cooperative group. We investigate the potential of CSR with a two 2 × 1 MISO beamforming links scenario. The numerical results show the significant gain of the ZF-CSR scheme than the MRC-TDMA scheme. In dense deployed networks that have high SNR on each link, the availability results show that

470

Chenguang Lu et al.

Figure 23.12. Average total CSR energy efficiency saving on both links when both links satisfy Condition II for the receiver case.

it is fairly easy for two links to form a cooperative group to both gain more capacity and more energy efficiency at the receivers. The average capacity of each link is from 1.4 to 3 times that of the MRC-TDMA scheme while the average capacity of the whole group is from 1.7 to 2.1 times that before. The receiver of each link can save averagely from 20% to 60% energy per bit while the both receivers save from 35% to 39%. Therefore, the CSR scheme has the great potential to further increase the capacity and energy efficiency in wireless networks. In this work, the potential of CSR is shown in a simulated scenario. The further study will be on demonstrating it in real environments. Figure 23.13 shows a beamforming demonstrator setup with a four-element antenna array built on a laptop computer, which is developed at Aalborg University. Furthermore, as this work investigated the CSR on MISO links, the future investigations will be extended to MIMO links.

References 1. H. Boche, A. Bourdoux, J.R. Fonollosa, T. Kaiser, A. Molisch, and W. Utschick. Smart antennas: state of the art. In IEEE Veh. Tech. Mag., March 2006. 2. T.M. Cover and J.A. Thomas. Elements of Information Theory. New York: Wiley, 2006. 3. F.H.P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-4020-4710-X. Springer, April 2006. 4. X. Guo, S. Roy, and W.S. Conner. Spatial reuse in wireless ad-hoc networks. In IEEE VTC 2003-Fall, October 2003. 5. IEEE-802.11. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications. In ANSI/IEEE Std 802.11, 1999 Edition (R2003), 2003. 6. S.A. Mujtaba, et al. TGn sync proposal technical specification. In IEEE 802.1104/0889r7, July 2005.

23 A Cooperative Scheme Enabling Spatial Reuse in Wireless Networks

471

Figure 23.13. Photo of the beamforming demonstrator setup.

7. K. Xu, M. Gerla, and S. Bae. How effective is the IEEE 802.11 RTS/CTS handshake in ad hoc networks. In IEEE GLOBECOM 2002, November 2002. 8. F. Ye and B. Sikdar. Distance-aware virtual carrier sensing for improved spatial reuse in wireless networks. In IEEE GLOBECOM 2004, November 2004. 9. F. Ye, S. Yi, and B. Sikdar. Improving spatial reuse of IEEE 802.11 based ad hoc networks. In IEEE GLOBECOM 2003, December 2003.

24 On the Energy Saving Potential in DVB-H Networks Exploiting Cooperation among Mobile Devices You know together we will stand Every boy, girl, woman and man Oh well now, two or three minutes Two or three hours What does it matter now In this life of ours Let’s work together Come on, come on Let’s work together – Canned Heat

Qi Zhang1 , Frank H.P. Fitzek2 , and Marcos Katz3 1 2 3

Technical University of Denmark [email protected] Aalborg University [email protected] VTT [email protected]

Summary. In this chapter we focus on cooperative strategies for IP-services over DVB-H networks to save energy. It demonstrates the potential of non-altruistic cooperation between mobile devices. The envisioned cooperation is based on cellular reception of DVB-H data, which is then shared among mobile devices within each others’ proximity over short–range links. The short–range communication is realized by Bluetooth, where a cross–platform design approach is introduced to optimize the signalling communication in the Bluetooth system. Three different cooperative algorithms are designed for the short–range link (Bluetooth) communication. Numerical results show that a energy saving of over 50% can be achieved by cooperative networking of three mobile devices in fully cooperating mode.

24.1 Introduction The aim of the Digital Video Broadcasting on Handheld (DVB-H) standard is to deliver audio and video content to mobile handheld devices. As those devices are battery driven, energy is always a crucial issue for mobile application. Therefore, time slicing [2, 8] has been introduced into DVB-H to save energy. The basic idea behind time slicing is to convey data in bursts with long pause periods instead to send a low data rate stream constantly. The energy consumption with time slicing depends on the burst duration and the pause period referred to as OFF-time period. From an energy saving perspective, more energy can be saved by shortening the

473 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 473–484. c 2007 Springer.

474

Qi Zhang et al.

burst duration. However, because of receiver sensitivity performance constraint, the burst duration is often kept as certain length to alleviate the receiver sensitivity issue. Thus, the remaining factor to work on is the OFF-time period. Obviously, the longer OFF-time period results in the more energy saving. Unfortunately the OFF-time period can not be excessively long because of quality of service aspects such as the access time and zapping time1 . Therefore, there is clearly a trade-off between burst duration and OFF-time to have optimum service access time and energy consumption. IP-services over DVB-H can be transmitted in sequential elementary streams (SESs) or parallel elementary streams (PESs) [2]. Both types of streams are transmitted in either multicast or broadcast fashion. The difference is that the SESs carry one service in one burst, while PESs can carry multiple services in one burst. The reason that multiple services are bundled and transported within the same burst is that the burst needs to meet a minimum burst duration length to fit the receiver sensitivity and at the same time the DVB-H system tries to get the maximal utilization of the DVB-H bandwidth. The use of parallel elementary streams brings many benefits, for instance, zapping time reduction, bandwidth optimization, the possibility of sending message type services in parallel to the main services, etc. However, when implementing PESs, the energy leak becomes an issue because the mobile device receives parallel elementary streams of several services encapsulated in the same burst. Indeed, it only keeps the desired elementary streams, discarding the remaining ones. From the entire system or network standpoint, the elementary streams discarded by a given mobile device could be used by other devices. It is clear that discarding unwanted elementary streams leads to inefficient use of resources, particularly energy. To our knowledge, this issue has not been addressed before. Furthermore as pointed out in [6], more research is needed in terms of novel techniques aiming at reducing energy consumption. In this direction, time slicing technique has been recently introduced. End users generally expect more and more hours of streaming audio on one battery charge but, on the other hand, improvements in battery capacity develop slowly (typically 10% per year) and hence, the requirements are difficult to meet. Energy saving techniques like those considered here are a promising option to extend considerably the service time of mobile devices. In this chapter we propose a cooperative energy saving strategy for IP-services over DVB-H. The considered cooperative strategy focuses on reducing energy consumption in the PESs case, even though it can be as well applied to SESs to obtain further energy savings. The cooperative architecture is set up between mobile devices that are capable to communicate not only with a central base station but also among each other using short-range wireless technology. The essential reason of the proposed approach achieving energy saving is that by cooperative reception of data over cellular link, the OFF-time period is virtually increased. Furthermore, as the energy per bit is much lower on a short–range link than in a cellular one, the energy consumption overhead over the former is significantly lower than over the latter. A first attempt to save energy for DVB-H mobile device is introduced in [4]. However, the approach saves energy by leaving out some FEC columns in the MultiProtocol Encapsulation-Forward Error Correction (MPE-FEC) frame once the receivers have received all the error–free data packets instead of getting the full block always. The maximum energy saving for this approach is 25%. The proposed 1

Zapping time refers to the program or channel switching time.

24 Cooperation in DVB-H Networks

475

cooperative energy saving strategy has larger energy saving potential than the one in [4], furthermore both approaches can also be combined.

24.2 Cooperative Strategy for IP-services over DVB-H Cooperation is the strategy of a group of entities working together to achieve a common and/or an individual goal [5]. The proposed cooperative mechanism requires that the devices have two air interfaces: a cellular link (CL) for DVB-H packets reception and a short–range link (SRL) for exchanging packets locally. The basic idea of the cooperative mechanism is that devices cooperatively receive the DVB-H bursts. Each cooperative mobile device receives only partially the data over CL. In case of PESs, the mobile device does not discard the unwanted packets anymore, but forwards those packets to its neighboring devices. Based on reciprocity, the device gets the missing packets from those neighboring devices. As for services carried by SESs case, mobile devices simply exchange the missed packets using SRL. This scheme is beneficial to reduce the energy consumption since the energy per bit (EpB) is much less in the short–range communication system. Furthermore, it guarantees very short service access time and zapping time. It does not require any modification in current DVB-H standard. The short–range link communication is not difficult to be implemented in the mobile devices. For multi-services transmitted by PESs scenario, the cooperative strategy works as following. We assume that there are three mobile devices which are interested in three different services with the same data rate transmitted by PESs, respectively2 . The DVB-H base station transmits DVB-H bursts as usual, without being aware of any devices’ cooperation. But the mobile devices can autonomously receive DVBH bursts alternately, if they are willing to cooperate after their negotiation. For instance, after MD1 finishes reception of the first burst containing the three services, it goes into sleep on its CL. And it transmits the related packets to MD2 and MD3 on SRL, respectively. Then MD2 and MD3 wake up at the start of the second and the third DVB-H burst respectively. They deal with the packets in a similar way as MD1 does. After a cyclic period, MD1 wakes up again to receive the forth burst, and so forth. Therefore, MD1 always wakes up at the start of the (3n + 1)th burst. Figure 24.1 illustrates how this cooperative mechanism works. The ON/OFF reception characteristic of DVB-H makes it suitable for devices to cooperate, which is an inherent benefit of DVB-H. The proposed cooperative strategy exploits the inherent characteristic of DVB-H to achieve energy saving by allowing longer idle time on DVB-H link. Furthermore, it also saves the energy consumption that is spent on decoding the received MPE-FEC frame3 . In the above example, the three different services have the same data rate so each device receives one DVB-H burst every three bursts. In reality the different services can have different data rates. So the frequency of mobile device waking up can be adjusted flexibly. This could be typically negotiated among the mobile devices before cooperation starts. Thus the energy consumption of all mobile devices is balanced. 2 3

Mobile devices (MD1, MD2, MD3) are interested in service 1, 2, 3, respectively. The decoded packets are exchanged directly between mobile devices without multiprotocol encapsulation and Reed-Solomon coding

476

Qi Zhang et al.

Figure 24.1. Burst flow mechanism in a cooperative scenario exploiting the Parallel Elementary Streams technique. SESs can be regarded as the simplest case of PESs. It is obvious that the cooperative strategy can easily be adapted to the case of services transmitted by SESs. Whether to cooperate or not should be evaluated by each mobile device independently. The decision depends on the cooperative strategy and the neighboring devices. In short, cooperation should be established as soon as the individual mobile device sees it own advantages [5]. This means that the establishment and termination of cooperation between mobile devices depends on the goal of the involved mobile devices, the cooperative strategy in use, and the prevailing relationship among mobile devices. The relationships among mobile devices change with time as devices move, terminate the ongoing service, join the network, and also the associated channels change, etc. For instance, in the above scenario, they will cooperate if they are close enough to each other to get mutual energy saving. However, if a device cannot attain energy saving anymore because of mobile devices movement and the increasing energy overhead on the short-range link, it will stop cooperation right away.

24.3 Cooperative Short–Range Communication According to the above description of the cooperative communication mechanism, the SRL is required to be very flexible and transparent to the end users. It works without any infrastructure in an autonomous mode. Namely, the short–range connection is a sort of ad hoc connection based on mobile device communications. The SRL air interface and associated communication mechanism can be designed and implemented by many different approaches, only if it meets the cooperative strategy principles. We use Bluetooth technology (Bluetooth 2.0 EDR) as an example to illustrate how it supports cooperation strategies in the short–range link, though any other short–range communication technology can be also used.

24 Cooperation in DVB-H Networks

477

We define cooperation range as the range within which the mobile devices can achieve energy saving. Every mobile device has its own cooperation range. A mobile device is capable of service discovery and has possibility to cooperate with the discovered devices within its cooperation range. When the mobile device cooperates in its cooperation range, they form a cooperative piconet. In the cooperative piconet, which role (Master/Slave) to take or which cooperative approach (centralized or distributed approach) to use is dependent on the topology of the formed piconet or scatternet.

24.3.1 Topology Based Cooperative Algorithm The cooperative algorithms in the short–range link can be summarized into the following three basic approaches according to the topologies.

Piconet Based Centralized Cooperative Approach This approach is used for Topology I (see Figure 24.2(a)). In this topology all mobile devices form one piconet. The slaves within the piconet are out of each others’ cooperation range. In this topology, the master controls the slaves’ states and transmission slots as typical Bluetooth piconet. But the slaves all stay in PARK state in the most of time. When the master wishes to transmit the received DVB-H packets to its cooperative slaves, it will unpark slaves by a master-initiated unparking method (using dedicated link manager protocol unpark command with slaves PM ADDR or its BD ADDR). A slave can also unpark itself when it needs to transmit the received DVB-H packets to the master by a slave-initiated unparking method (sending access request message with AR ADDR).

Piconet Based Distributed Cooperative Approach This approach is used for Topology II (see Figure 24.2(b)). In topology II the mobile devices form one piconet and all the mobile devices are within each others’ cooperation range. Namely, they can have fully meshed connections. In such case, all the mobile devices within the piconet are capable to work as a master and they alternately take master role in the piconet. Therefore, all the mobile devices periodically switch role (Master/Slave) and only master mobile device transmits in its turn. When the first master establishes the cooperative piconet, it decides the master role switching sequence and broadcasts it to its slaves. The key technical issue here is the effective role switching which can be initiated by either slave or master. The master initiated role switching method is preferred, because it can implicitly check if the slave (i.e., the successor master) is still in the piconet or not. If the slave (i.e., the successor master) has gone, the master can timely update the master role switching sequence. Then it will switch role with another successor in the new role switching sequence list.

Scatternet Based Cooperative Approach This approach is employed for Topology III (see Figure 24.2(c)). In topology III the mobile devices form a scatternet, i.e., some mobile devices stay in more than one

478

Qi Zhang et al.

piconets. Considering signalling complexity and exchanging load, we assume that the mobile device can work at the most in two piconets. In Figure 24.2(c) MD3 can work as slave in both Piconet A and B or it can work as slave in the Piconet A and master in the Piconet B. MD3 should be able to harmonize its operation in both piconets. For instance, MD3 works as master in the Piconet B during its PARK state interval in Piconet A. It requires MD3 to have an accurate synchronization in Piconet A, otherwise it might miss the unpark message from master of Piconet A. This issue can be effectively resolved by synchronization knowledge from DVB-H system.

(a) Topology I

(b) Topology II

(c) Topology III Figure 24.2. Cooperative Topologies.

24.3.2 Signalling on the Short-Range Link In the cooperative SRL implementation, Bluetooth–supporting mobile devices will be involved into frequent role switching, SYMMETRICAL service discovery (INQUIRY/INQUIRY SCAN), connection establishment (PAGE/PAGE SCAN), connection validity check, and so on. These procedures will eventually result in frequent signalling exchange load and random SRL connection establishment delay4 . 4

The random connection delay issue is due to the symmetrical link mode in the connection establishment [3,7], i.e., any mobile device can start Inquiry or Inquiry

24 Cooperation in DVB-H Networks

479

No doubt, it is a real challenge to design an efficient and feasible short–range communication scheme supporting a cooperative strategy. It is very difficult to resolve all these critical issues in an independent Bluetooth system because of the limitations of Bluetooth’s frequency hopping physical layer. Although many researchers have explored on similar issues such as neighboring Bluetooth devices discovery, autonomous link establishment, piconet or scatternet formation [3, 7], a good engineering solution solving all these issues has not been proposed yet. However, these issues can be tackled in the cooperative system considered here. The essential reason is that the mobile device in our system exploits air interfaces diversity (i.e., DVB-H and Bluetooth). The main obstacle in resolving those issues in Bluetooth is due to the unawareness of neighboring mobile devices’ INQUIRY or INQUIRY SCAN information. It is a sort of complete autonomy. However, in a DVB-H system the mobile device has accurate synchronized time information of the DVB-H system to receive the burst at a right time. Mobile stations also have the knowledge of their target services. So in SRL, a mobile device can exploit the known cross–platform information (e.g., burst starting time and end time) from DVB-H system to effectively resolve all these issues. A good case in point is that the random service discovery delay [7] can be eliminated by mobile devices starting INQUIRY/INQUIRY SCAN state at the right time with predefined inquiry length. At the same time, mobile devices can effectively differentiate their INQUIRY/INQUIRY SCAN states to facilitate the service discovery procedure based on the different service information in the burst block. The detailed cross–platform optimization implementation is out of the scope of this chapter.

24.4 Numerical Examples for Energy Consumption Analysis This section will present two numerical examples to illustrate the energy saving by cooperation on the short–range link. The average energy consumption per burst without cooperation and with cooperation are expressed in Eq. 24.1 and Eq. 24.2, respectively. The notation of parameter expressions are summarized in Table 24.1. ! t )Pc,on + tc,of f Pc,of f + tc,i Pc,i (tc,Bd + tc,syn + c,Dj nocoop 2 Enocoop = tc,cyc tnocoop c,cyc nocoop = tnocoop c,cyc Pc

= tc,Bd + tc,syn + tc,Dj /2 + tc,of f + tc,i tnocoop c,cyc

where, Ecoop =

=

(24.1)

tcoop tc,on Pc,on +tc,of f Pc,of f +tc,i Pc,i tsr,tx Psr,tx +tsr,rc Psr,rc +tsr,i Psr,i cyc + Nburst tcoop tcoop c,cyc sr,cyc

tcoop cyc coop Pccoop + Psr Nburst

(24.2)

Scan in service discovery. In the asymmetrical link mode Inquiry or Inquiry Scan duty is predefine between mobile devices.

480

Qi Zhang et al.

where, Nburst denotes the number of bursts transmitted by the DVB-H networks in one cooperative cyclic period tc,on = tc,Bd + tc,syn + tc,Dj /2 tcoop c,cyc = tc,on + tc,of f + tc,i tcoop sr,cyc = tsr,tx + tsr,rc + tsr,i coop coop tcoop c,cyc = tsr,cyc = tcyc

In Eq. 24.2, tc,i is much longer than that in Eq. 24.1. Consequently tcoop c,cyc is much longer than the cyclic period of non-cooperative case tnocoop c,cyc . Example I: This example corresponds to fully connected mobile devices as given by Topology II in Section 24.3. In this example we assume that the SRL performs power control. Transmission power is a function of the distance between transmitter and receiver. Due to the complexity of the power control mechanism, the exact expression for the transmission power is unknown. Here we assume that transmission power is approximately proportional to the distance between mobile devices. We also assume that the SRL has very short synchronization time. The transmission data rate on the SRL for the master is 1.3 Mbps (with 3-DH5 packet symmetrical maximum rate5 ) [1]. We set the reception power consumption and idle power consumption to constant values as 10 mW (10 dBm) and 1 mW (0 dBm) [1], respectively. Transmission power consumption is varying between range of 10 mW – 100 mW (10 dBm – 20 dBm) on SRL. The values of the parameters for the CL are taken from [2] and listed in Table 24.1. With all these assumptions, the relation of average energy saving (1 − Ecoop /Enocoop ) and the transmission power is shown in Figure 24.3. It can be seen from Figure 24.3 that the achievable energy saving can be over 50% already with three cooperating mobile devices.

Figure 24.3. Energy saving by cooperation. 5

3-DH5 packet type is newly defined in Enhanced Date operation. 3-DH5 has maximum payload of 1021 Bytes, occupying five time slots.

24 Cooperation in DVB-H Networks

481

Table 24.1. Parameter list for cellular and short–range air interfaces. Cellular link Character Mean Pc,on power consumption when RF is on Pc,of f RF is shut down, but MPE-FEC is on going Pc,i DVB-H receiver waiting for next burst Pccoop ,Pcnocoop avg power consumption on DVB-H interface w/o cooperation tc,syn synchronization time tc,Bd burst duration tc,Dj delta-t jitter tc,of f duration when receiver is at RF OFF1 state tc,i idle time tnocoop frame cyclic period without cooperation c,cyc tcoop frame cyclic period with cooperation c,cyc Short–Range Character Mean Psr,tx transmission power Psr,rc reception power Psr,i power consumption for idle state coop Psr average power consumption on Bluetooth interface tsr,tx transmission time tsr,rc reception time tsr,i idle time tcoop cyclic period sr,cyc

Value 400 mW 50 mW 10 mW 120 ms 236 ms 10 ms 500 ms 2.7165 s Value 10 mW 1 mW -

Example II: This example considers a scenario with three mobile devices (MDs) which are individually receiving three different services with the same data rate. The power consumptions of transmission, reception and idle state are assumed to be constant. The positions of two mobile devices (MD1 and MD2) are fixed and they form a piconet. Another mobile device (MD3) is moving from far away towards the two mobile devices, then it moves away again. Figure 24.4 illustrates the topology of the scenario and MD3’s trace of this example. The parameters used in this example is listed in Table 24.2. Table 24.2. Parameters of example II. Parameter Velocity of MD3 Psr,tx Psr,rx Psr,i Time Value 1 m/s 10 mW 10 mW 1 mW 180 s

The calculation results of this example is shown in Figure 24.5. Two different cooperative strategies are implemented for the Topology I scenario. Figure 24.5(a) is generated by the Cooperative Strategy I which considers the selfish characteristics of mobile devices and the fairness requirement of the system. So it is based on the principle that the exchanged packets between mobile devices must be equal. In Topology I scenario, when MD1 works as master with MD2 and MD3 as slavers, MD1 can have about 54% energy saving while MD2 and MD3 only get about 26%

482

Qi Zhang et al.

Figure 24.4. Cooperative scenario of example II.

energy saving (MD1 is the final winner because its optimal location). At this situation, mobile devices get unequally pay-offs by cooperative and the difference between different mobile devices is more than 20%. It is obvious in the Cooperative Strategy I that the gain achieved by the mobile device depends on its relative position in the cooperative group. Figure 5(b) is based on the Cooperative Strategy II. This strategy tries to balance the energy saving of the mobile devices, while it is not dependent on the mobile devices’ relative location any more. Here we assume that the energy consumption on the SRL is very low. Hence, if the master receives half of all DVBH burst packets and each slave receives the remaining packets, it is very close to the optimum value to balance the energy saving of the mobile devices. By this strategy in the same Topology I scenario, MD1 gets 38.5% energy saving and MD2 and MD3 achieve 40.5% energy saving. Note that it is obtained a very good balance of energy savings for all mobile devices at a little expense of the MD1’s energy saving. Furthermore, in this case the Cooperative Strategy II saves 7.1% more energy from the standpoint of the cooperative system (10% theoretically). If one master cooperates with more slaves it can save even more energy by Cooperative Strategy II than that of Strategy I. Theoretically, it can save 20%, 26.47%, 30.77% more energy respectively, when one master cooperates with 3, 4, 5 slaves6 in one piconet. Figure 5(a) and Figure 5(b) also clearly show the corresponding energy saving changing with the movement of MD3 because the different topology based cooperative algorithms are used. It is evident that all mobile devices achieve the maximum energy saving when they are fully-connected. It is up to 54% in this example scenario.

24.5 Conclusion The proposed cooperative mechanisms are used for multi-interface DVB-H mobile devices. Multiple air interface stations are already commercially available. Results 6

In such case, the maximum cooperative slaves can only reach 5, because when more than 5 slaves are in the same piconet, two or more slaves must be in each others’ cooperation range.

24 Cooperation in DVB-H Networks

483

(a) Equal exchanged packets

(b) Balanced energy consumption Figure 24.5. Energy saving in example II.

show the strength of non-altruistic cooperation between mobile devices for IPservices over DVB-H to save energy. The numerical energy consumption analysis examples show the achievable energy by cooperative strategies using the state–of– the–art technology. The energy saving depends on the number of cooperative mobile devices and the energy consumption on the short–range link. It is expected that data rates supported by SRL will increase significantly while energy consumption in such systems will continue to decrease. Such short-range systems include UWB, Bluetooth 3.0, wireless USB and others. The development of advanced short–range air interfaces will accentuate the advantages of the proposed cooperative techniques. Therefore, with the energy consumption per bit on the SRL will decreasing, more energy saving will be achieved.

484

Qi Zhang et al.

References 1. Specification of bluetooth system, covered core package verison:2.0+edr. Technical report, November 2004. 2. Dvb-h implementation guidelines, dvb-h document a092. Technical report, July 2005. 3. Yeung K.L. Changlei Liu. Autonomous proximity awareness of bluetooth devices. In ICC, pages Vol.2, 934–938, 2005. 4. E. De Diego Balaguer, F.H.P. Fitzek, O. Olsen, and M. Gade. Performance evaluation of power saving strategies for dvb-h services usingadaptive mpe-fec decoding. In PIMRC 2005. IEEE 16th International Symposium, pages Vol.4 2221–2226, 2005 2005. 5. Frank H.P. Fitzek and Marcos D.Katz, editors. Cooperation in Wireless Networks: Principle and Applications. ISBN-10 1-4020-4710-X. Springer, 2006. 6. Cedric Paillard. Dvb-h could be the next big thing. 7. LaMaire R. Salonidis T.; Bhagwat P., Tassiulas L. Distributed topology construction of bluetooth wireless personal area networks. IEEE Journal on Selected Areas in Communications, 23(3):633–643, 2005. 8. X.D. Yang, Y.H. Song, T.J. Owens, J. Cosmas, and T. Itagaki. Performance analysis of time slicing in dvb-h. In IEEE Mobile Future, the Symposium on Trends in Communications. SympoTIC’04. Joint IST Workshop, pages 183–186, 2004.

25 Cooperative Retransmission for Reliable Wireless Multicast Services Being in the Same Boat, We Learn to Work Together!

Qi Zhang1 and Frank H.P. Fitzek2 1 2

Technical University of Denmark [email protected] Aalborg University [email protected]

Summary. Multicast services have been identified as an important key technology to increase the network and service providers’ revenue. Many data dissemination applications require even reliable multicast. Error/loss recovery for reliable multicast is different from conventional schemes taking into consideration the unreliable wireless channel, the battery driven mobile device, and the limited wireless bandwidth. To have an efficient error/loss recovery scheme for reliable multicast in wireless networks, we advocate a new communication architecture. It is referred to as cooperative wireless networking, where the mobile devices communicate directly with each other to perform retransmissions using their short–range communication capabilities in addition to their cellular links. Based on the cooperative architecture a novel retransmission scheme is proposed exploiting the short–range retransmission in this chapter. In the state–of–the–art, the non–cooperative error recovery strategies are compared with the cooperative retransmission strategy in terms of mobile device energy consumption to show the benefit of the newly introduced scheme.

25.1 Introduction Multicast communication has been identified as an effective way to disseminate information to a potential group of receivers [16] sharing the same service interest. Many data dissemination applications such as software distribution, distributed information (e.g., stock market data, sports scores, business inventory data, world news updates [15]), and mailing list delivery, etc. [1] require reliable multicast. Due to its increasing importance, it has received many researchers’ interests and attentions in the previous years [1, 6, 7, 11, 14, 16]. Another fact is that various new wireless networks are deployed and more sophisticated mobile devices are available. So it is an emerging trend that more multicast applications and services will converge into wireless networks. Comparing with the wired network, reliable multicast in wireless networks becomes more challenging due to the unreliable wireless link and the mobile device’s limitations. Reliability requirements vary depending on different applications. But in terms of general reliability, it includes error/loss recovery, ordered delivery, no duplicates,

485 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 485–498. c 2007 Springer.

486

Qi Zhang and Frank H.P. Fitzek

and isolation of independent failures aspects. Among these aspects, error/loss recovery is the most concerned designed issue. The main reasons are that first the packet error/loss rate is higher in wireless networks and second the error/loss recovery schemes have significant influences on the multicast application’s quality of service and wireless network performance, for instance, error/loss recovery latency, bandwidth utilization, for mobile device energy consumption efficiency, and others. Two well-known techniques are usually used for error/loss recovery: automatic repeat request (ARQ), which retransmits the lost packets on requests from the receivers conveyed over the feedback channel; and forward error correction (FEC), which transmits redundant parity packets with data packets and recovers error/loss directly at the receivers without any need of the feedback [11]. Pure ARQ has scalability issues such as implosion and exposure [13]. And pure FEC can not provide full reliability [11]. Better performance for reliable transmission can be achieved by combining both of them (i.e., Hybrid ARQ) [2,5,9,10]. Hybrid ARQ is usually classified into two categories, namely type I and type II schemes [8]. HARQ I scheme is used for communication systems with relatively stationary channel conditions. HARQ II is an adaptive scheme for non stationary channels. HARQ II cannot always outperform ARQ and FEC schemes, especially in wireless networks as proofed by [11]. It shows that HARQ II schemes usually outperform ARQ and layered FEC for homogeneous packet loss probability in the receivers [11]. However, for the receivers with different packet loss probabilities, the performance is almost solely determined by the receivers with high loss rate. This is true even though if the fraction of high-loss receivers among all receivers is very small. The essential reason of HARQ II performance degradation in wireless networks is due to the number of parity packets depending on the worst receivers. So the channel heterogeneity in wireless networks limits the achievable performance. Moreover, if a radio path between an access point (AP) and a mobile device is greatly deteriorated by the instantaneous channel conditions, the mobile device can not effectively help itself out by requesting retransmissions. The other devices in the multicast group consume extra energy because of receiving many useless parity packets. Additionally, the recovery latency is increased and bandwidth is wasted consequently. Some modified strategies [11] proposed that the receiver can stop receiving the parities when it has received enough parities. But network bandwidth is wasted anyway, which results in low channel efficiency. Furthermore, HARQ II also induces higher processing load at the sender and the receivers for coding and decoding. The power consumption due to heavy processing load has to be taken into account for the wireless devices. Therefore, it is worth investigating a more efficient and scalable error/loss recovery solution for reliable multicast in wireless networks taken also into account the power consumption constraints, latency and available bandwidth resources. We propose one feasible solution called cooperative retransmission strategy. In this strategy, the wireless devices can recover error by local retransmission with devices in each others’ proximity over the short–range link. Local retransmission is not a new idea in general for reliable multicast. But in the previous work of local retransmission (e.g., in Pragmatic general multicast (PGM) protocol), the local retransmission is only implemented on the network element side. It comes up one design issue of how to locate the re-transmitter and evaluate their efficiency relative to other available source. The main contribution of this chapter is that a host-side based local retransmission is designed and it is implemented by a novel and simple coopera-

25 Cooperative Retransmission for Reliable Wireless Multicast Services

487

tive retransmission scheme. It should be noted that packet error/loss happening in the multicast tree is not considered here. We only focus on the packet error/loss in the last–hop (i.e., in the wireless access networks) because most of the packet errors/losses take place here. The proposed solution can efficiently save energy consumption at the mobile device, because it essentially reduces the average number of transmissions required to receive a packet reliably at all the receivers. Consequently it can reduce retransmission delay and improve bandwidth utilization. Here we use energy consumption as metric to compare the proposed cooperative retransmission scheme with the noncooperative schemes. The corresponding delay reduction and bandwidth utilization improvements are obvious due to the average number of retransmission decreasing.

25.2 Non–Cooperative Error Recovery Strategies Traditionally in order to recover packet error/loss, the mobile devices have to either request packet retransmission or request a packet repair from the access point (AP), as the AP is the only reference for the mobile device in non-cooperative networks. In the following two subsections, the energy consumption of the representative error recovery schemes, namely ARQ and FEC/HARQ will be introduced.

25.2.1 ARQ Scheme The ARQ scheme has been widely used in several multicast protocols such as MTP (Multicast Transport Protocols), AFDP (Adaptive File Distribution Protocol), PGM (Pragmatic general multicast protocol), and SRM (Scalable Reliable Multicast protocol). In MTP and AFDP the receiver unicasts the NACKs (Negative Acknowledgements) to request retransmission and the sender retransmits the requested packets [12]. In the PGM and SRM the sender can suppress the NACKs from the different receivers that lost the same packets [12]. All these protocols have implosion and exposure issues. Implosion is a result from duplicated NACKs (or retransmission request) from many receivers. Additionally, in order to avoid loss of NACK, the receiver often continuously sends NACKs to the sender until it receives confirmation from the sender. Duplicated NACKs might swamp the sender and the network, even the other receivers. Exposure occurs when the retransmitted packets are delivered to those receivers who did not lose the packets [13]. Both implosion and exposure are fatal impediments for multicasting in wireless networks. In this subsection, we use SRM (Scalable Reliable Multicast) protocol [4] as one example of an optimal ARQ scheme and derive the corresponding energy consumption as following. We define: •

•

Pc,rx , Pc,tx , Pc,i as the power consumed in the reception process, the transmission process and the idle state for the cellular communication facility on the mobile device, respectively. tc,rx , tc,tx , tc,i as the corresponding time spent on the reception, transmission and idle state on the cellular link, respectively.

488

Qi Zhang and Frank H.P. Fitzek

The total energy consumption of a mobile device is EN oCoop = tc,rx Pc,rx + tc,tx Pc,tx + tc,i Pc,i

(25.1)

The packet size is assumed to be constant. Additionally, the reception time of one packet on the cellular link is assumed to one time unit. Average energy consumption of one valid packet reception is given in the following: EARQ = (1+f N γα) Pc,rx +f γβα Pc,tx +f γβα(N −1) Pc,i (25.2) | {z } | {z } | {z } tc,tx

tc,rx

tc,i

where • N is the number of mobile devices in the network • α is packet loss rate. 1 , which is due to retransmitted packet loss possibility • γ equal to 1−α • β is the ratio of the NACK size to the data packet size. • f is defined as the uncorrelated factor • tc,rx consists of two parts: the time for an original packet reception, i.e., one time unit, and the time for the retransmitted packet reception which depends on the number of devices losing this packet (i.e., N α), the probability of retransmitted packet loss and the uncorrelated factor. • tc,tx is the time of the mobile device transmitting NACK, which depends on the NACK message size, the probability of the mobile device fails to receive the original packet and the retransmitted packet, and also the uncorrelated factor. • tc,i is the time during which other devices send NACK. The concept of the uncorrelated factor comes from the suppression NACKs scheme from loss recovery algorithm of SRM. In SRM, the mobile device that detects packet loss waits for a random backoff time before requesting retransmission. When the access point retransmits the requested packets in a multicast fashion, some wireless devices will receive the missed packets which they have not requested retransmission yet. The reason is that the lost packets of each mobile device are correlated to some extend. The random backoff time delay is not included in our energy consumption formula, i.e., the real energy consumption and delay performance of the ARQ strategy is even worse than the calculation result from the given formula.

25.2.2 FEC/HARQ Schemes In this section, we derive the energy consumption for layered FEC and HARQ II based on E[M ] (the average number of transmissions required to transmit a packet reliably to all receivers). E[M ] has been derived in [11] (Eq. 4 in [11] for layered FEC and Eq. 5,6,7 in [11] are for integrated FEC II, i.e., HARQ II). For the heterogeneous receiving conditions, E[M ] can be calculated by Eq. 8,9 in [11]. Then using the same energy calculation methodology described in Subsection 25.2.1, the energy consumption of layered FEC or HARQ II1 scheme can be expressed as EF EC = E[M] Pc,rx +(E[M]−1)β Pc,tx +(N −1)(E[M]−1)β Pc,i . | {z } | {z } | {z } tc,rx

1

tc,tx

(25.3)

tc,i

The E[M ] of the HARQ II calculated by the formula given in [11] is the ideal lowest value. So in fact we underestimate the energy consumption of the HARQ II scheme.

25 Cooperative Retransmission for Reliable Wireless Multicast Services

489

25.3 Cooperative Retransmission Strategy The idea of cooperative retransmission strategy is based on the topology shown in Figure 25.1. Multiple mobile devices located in proximity of each other form a cooperative cluster. The mobile devices of the same cluster can communicate directly with each other using the short–range link (SRL). In contrast to SRL, the link between mobile devices and the access point is referred to as the cellular link (CL). The data rate of the SRL is much higher than that of CL. Furthermore the power consumption on the SRL is much lower, because of the shorter distance between the transmitter and the receiver, which also contributes positively to the reliability of the SRL.

Figure 25.1. Topology of cooperative groups within one central access point [3].

Exploiting the characteristics of the SRL, the wireless devices in one cluster can exchange (retransmit) the missed packets in a very short time over the SRL. This not only saves energy but also reduces the recovery latency. It consequently increases the total throughput on the CL. In case that there is still unrecoverable packet error/loss in the cooperative cluster after local retransmission, the cluster is able to send NACKs to the access point to trigger cellular retransmissions. In the following subsections, we are going to explain how to implement cooperative retransmission on both CL and SRL.

25.3.1 Frame Structure Design on Cellular Link with TDD Mode We assume that in the investigated system the mobile device has one air interface. The mobile device communicates either on the CL with the access point or on the SRL with the neighboring mobile devices. The access point controls the mobile devices to communicate alternatively on CL or on SRL. So all the mobile devices

490

Qi Zhang and Frank H.P. Fitzek

are synchronized to the clock of the AP, and then all mobile devices in one cluster are synchronized.

Figure 25.2. Frame structure in Cooperative Protocol.

The investigated system uses TDD (Time Division Duplex) for the downlink and uplink transmission. The designed frame structure on the cellular link is shown in Figure 25.2. A frame consists of a downlink subframe, an uplink subframe and a guard interval. The information in downlink subframe consists of signalling and data. The signalling is conveyed in broadcast fashion. It includes description of the downlink/uplink physical layer parameters and downlink/uplink usage assignment information. The data packets are transmitted as unicast or multicast bursts. The time slots allocation for unicast or multicast is very flexible, which depends on the current services and packet scheduling algorithm in the AP. In one downlink subframe, it can contain only unicast bursts or multicast burst or both. The cooperative retransmission slots are assigned by the AP, according to the number of mulitcast service subscribers and the number of mobile devices involved in cooperation. If mobile devices are sparsely distributed in the coverage and are unable to form cooperative clusters, there is simply no need for the AP to reserve cooperative retransmission slots. In such case, the device should be capable to recover error/loss stand alone. Hence, the cooperative retransmission is optional for the mobile devices. In stand alone mode, the devices deal with the packet loss by conventional ARQ schemes, sending the NACKs individually in the uplink unicast slots. Then the AP can retransmit the requested packets to the individual user in the downlink unicast burst slots. Cooperative retransmission is usually executed once for a batch of frames. As for stand alone mobile device, it can send a NACK message when it has lost packets. This design is very generic and flexible for the AP and the mobile devices.

25.3.2 Design Cooperative Retransmission Scheme on the Short–Range Link The cooperative retransmission is done among the cooperative devices during the cooperative retransmission slots which are reserved by the AP. We advocate a novel scheme design for the cooperative retransmission implementation.

25 Cooperative Retransmission for Reliable Wireless Multicast Services

491

First of all we assume that the cooperative retransmission is done once every M multicast packets and there are N mobile devices in one cluster. All the devices within a cluster have the short–range links to each other and all SRLs have the same data rate. The cluster membership is maintained by the Common Connectivity Table. All the devices in the cluster host a local copy of the Common Connectivity Table which is updated periodically. The proposed cooperative retransmission protocol is based on a logical token ring topology. The signalling among the devices in the cluster is carried by the token packet. The devices use the token packet to collect the lost packet information and to share the retransmission duty. Usually all the losses/errors can be recovered by K (K = 2 ∼ 4) devices called primary devices, due to packet error/loss diversities in different devices. The remaining devices, called auxiliary devices, can help to complete recovery in case the K devices are not sufficient to recover all the losses. This is implemented by an additional bit called Complete Reception Bit (CRB) and the optional field in the token packets.

Figure 25.3. Marking Lost Packet Matrix procedure.

The cooperative retransmission can basically be divided into two procedures. The first procedure is to count all the lost packets within the cluster by marking in the Lost Packet Matrix (LPM). Each mobile device generates one packet loss report vector with M elements and inserts it into LPM. The final complete matrix is composed of (N + 1) rows and M columns. If mobile device i loses the jth packet, it marks the bit LP M (i, j) as bit “1”, otherwise it sets the bit as “0”. The last additional row is called Lost Packet Information (LPI) vector which indicates all of the lost packets within the cluster. We give a concrete example here: there are four devices in the cluster and two of them work as primary devices. The marking LPM procedure is illustrated in Figure 25.3. Followed by the first primary mobile device receiving the token packet, it starts second procedure: local retransmission whose illustration diagram is shown in

492

Qi Zhang and Frank H.P. Fitzek

Figure 25.4. Local retransmission procedure (STA1 and STA2 are retransmitters).

Figure 25.4. Assuming there are L induplicated lost packets within the cluster. The first primary mobile device scans the LPI vector and knows which L packets are lost. Then it scans the first K rows and chooses the packets which are lost by the other K − 1 primary devices but are correctly received by itself. Then it retransmits such packets first. If the number of such packets is less than dL/Ke, it continues to retransmit some lost packets until the number of packets that it has transmitted is equal to dL/Ke. After finishing its retransmission duty, the first primary mobile device resets the index bits of the retransmitted packets as “0” in the LPI vector. Then it passes the token to its successor. The second primary mobile device executes the same procedure as its predecessor, and so forth. The last primary mobile device will check the CRB and the optional field, when it finishes its retransmission duty. If the CRB indicates there are unrecoverable packets in the cluster, the last primary mobile device will send a NACK to the AP requesting the missed packets over cellular link. If it knows who can complete the rest recovery task from the CRB and the optional field, it will pass the token packet directly to that auxiliary mobile device. The last case is that the last primary mobile device knows all the lost packets have been recovered and it sends an ACK to the AP. To reduce the latency of passing the LPM and the overhead load due to it, the LPM can be compressed from N + 1 rows to K + 1 rows. The compressed LPM of the aforementioned example is shown in Figure 25.5. The idea is that the auxiliary devices can mark a lost packet on a random row but with a specific column index, instead of having individual rows to indicate their lost packets. For instance, in the case of N = 128 and K = 3, the compressed LPM reduces by 95% latency. So it is a very efficient way to control latency bound in a big cluster. This cooperative retransmission scheme is very fair for all the devices to cooperate in the cluster. Furthermore, such task sharing cooperation highly meets the timely reciprocity requirements of the designing principle for cooperative wireless networks. As mentioned in [3], cooperative interaction should payoff in a timely fashion. Every cooperating party should see its benefits in doing so within the shortest possible delay. The delay of the feedback benefit in the proposed cooperative protocol is only at the order of seconds, which can be regarded as nearly instantaneous

25 Cooperative Retransmission for Reliable Wireless Multicast Services

493

Figure 25.5. Compressed Lost Packet Matrix for the example.

reciprocity. Additionally, the proposed cooperative retransmission protocol is easy to maintain the membership in a cluster.

25.3.3 Energy Consumption by Cooperative Retransmission Protocol The probability that a packet loss cannot be recovered by local retransmission is very low. Therefore, the following energy calculation only considers local retransmission within the cluster. Moreover, we can reasonably assume the SRL is very reliable and no retransmitted packet loss over it. We define: • •

Psr,rx , Psr,tx and Psr,i as the power consumed in reception, transmission process and idle state on the short–range link by mobile device, respectively. tsr,rx , tsr,tx and tsr,i is the corresponding time spent on reception, transmission and idle state on the short–range link, respectively.

The total energy consumption of mobile device is ECoop = tc,rx Pc,rx +tsr,rx Psr,rx +tsr,tx Psr,tx +tsr,i Psr,i | {z } | {z } Ec

(25.4)

Esr

where, Ec , Esr are the average energy consumption on the cellular link and the short–range link, respectively. Esr includes two parts: the energy consumption used for marking LPM and local retransmission. Marking LPM period is defined as the time of the first mobile device starting marking LPM until it receiving the complete LPM. During one marking LPM period, each mobile device receives and transmits the LPM once. The mobile device can stay in idle state during the remaining time of marking LPM period. So δ, the energy overhead due to marking LPM for each packet is given by 1 ρ(N − 2) ρ δ= (Psr,tx + Psr,rx ) + Psr,i M ω ω where, • •

ρ is the ratio of the average token packet size to data packet size ω is the data rate ratio of SRL to CL.

The average energy overhead for each packet on each mobile device due to local retransmission is derived as following. As there are K primary devices in one local retransmission, each mobile device takes retransmission task every N/K local

494

Qi Zhang and Frank H.P. Fitzek

retransmissions. Hence, we can sum up one mobile device’s energy consumption on the short-range interface during the consecutive N/K local retransmissions. Then we average the sum by N/K and M (local retransmission is done once for every M multicast packets). So Esr can be expressed by Esr =

N − 1)Eaux Epri + ( K +δ N M K

(25.5)

where, Epri and Eaux is the energy consumption of a mobile device working as primary mobile device or auxiliary mobile device in one local retransmission. Their expressions are given by Epri =

L/K L/K Psr,tx + (K − 1)Psr,rx ω ω Eaux =

L Psr,rx ω

where, L = f N M α. The explanation of Epri and Eaux expression is: after the AP transmits M multicast packets, the number of packets lost at one mobile device is equal to M α; the sum of lost packets at all devices in one cluster is N M α; due to correlation of the lost packets, the number of the unduplicated lost packets within one cluster is L = f N M α. The time a primary device spending on packet transmission is L/K and ω the time of a primary device for receiving the packets transmitted by other primary devices is L/K (K − 1). ω

25.4 Comparison of Energy Consumption Based on the previous description of the cooperative retransmission protocol, this section compares the energy consumption of the cooperative retransmission strategy against the non–cooperative ones. Table 25.1 summarizes the variables and notations that are used in the following energy analysis. The parameter values are taken from [3].

Table 25.1. Parameters assumption for Analysis. Notation α β ρ Rc Rsr ω Value 2% or 5% 0.2 0.1 6 Mbit/s 54Mbit/s 9 Notation Pc,rx Pc,tx Pc,i Psr,rx Psr,tx Psr,i Value 0.9W 2W 0.04W 0.4W 1W 0.04W

Figure 25.6 and Figure 25.7 compare the energy consumption of the non– cooperative and the cooperative strategies under the condition of homogeneous and heterogeneous packet loss rate at the receivers, respectively. Figure 25.6 and Figure 25.7 shows the energy consumption comparison based on energy units. It clearly shows that the energy consumption of the cooperative strategy is quite stable with an increasing number of devices for both, homogeneous and heterogeneous

25 Cooperative Retransmission for Reliable Wireless Multicast Services

495

Figure 25.6. Energy consumption comparison of different error recovery schemes in homogeneous packet loss rate scenario.

Figure 25.7. Energy consumption comparison of different error recovery schemes in heterogeneous packet loss rate scenario (The high packet loss is 10% or 20%. There are 5% high packet loss devices in the cluster. The rest of devices has 2% packet loss rate.)

496

Qi Zhang and Frank H.P. Fitzek

packet loss situations. However the energy consumptions of the layered FEC and the ARQ (SRM) schemes increase dramatically when the number of the devices increases. Hence the achievable energy saving gain by cooperative retransmission becomes significant when there are many devices in the system. It can be seen clearly that higher packet loss rates lead directly to a more energy saving. In these two figures the effect of ARQ (SRM) suppression scheme can also be seen. For example, due to uncorrelated factor deceasing with the number of the mobile devices increasing, the increasing trend of energy consumption of ARQ (SRM) decreases. Moreover, in Figure 25.7, the two ARQ (SRM) curves are overlapped because the packets lost by high loss rate devices are overlapped with the packets lost by normal devices and the suppression scheme in SRM highly avoids the duplicated retransmission. The HARQ II has as good energy consumption as the cooperative retransmission scheme when the receivers have the homogeneous packet loss. But the cooperative retransmission scheme outperforms the HARQ II under the heterogeneous packet loss rate assumption. For instance, it can be seen in Figure 25.7 that for 128 devices in one cluster the cooperative scheme achieves energy saving up to 40% in high packet loss rate case. All the above calculation results are based on the scenario that all devices in the system form one cluster. But considering real usage scenario, devices scatter within the coverage of the access point. It is not realistic for all of the devices to form only one cluster. However, it is possible for the devices to form small clusters. Different groups can do local cooperative retransmission concurrently by grouping, assuming that the distance between groups are far away enough to ensure no interference among groups. For example, there are 128 devices in the system and for simplicity they are assumed to be split into 2, 4, 8, 16, 64 groups. The energy saving with grouping is shown in Figure 25.8. The energy saving range is between 40% and 48% for the 5% packet loss rate case, when the number of groups varies between 1 and 64. It means that the average cooperative energy saving is within the range of [40% ∼ 48%] when there are 128 devices in the system. Furthermore, it indicates that the smaller cluster results in the higher energy saving. The reason is that the smaller cluster has less energy overhead on the short–range link.

25.5 Conclusion In this chapter, we propose a novel and generic cooperative retransmission scheme for the wireless reliable multicast services to reduce energy consumption and to minimize packet loss recovery latency. It exploits the higher data rate and better reliability characteristics of the short–range link and can recover almost all packet errors/losses locally in a very short time. It is robust to not only homogeneous but also to heterogeneous channel conditions. Comprehensive comparison and analysis of energy consumption between the cooperative and the non–cooperative strategies have been given, based on the advocated communication architecture and scheme. The analysis results show that the proposed cooperative retransmission scheme is more efficient and suitable for the reliable multicast services than the non–cooperative ones.

25 Cooperative Retransmission for Reliable Wireless Multicast Services

497

Figure 25.8. Cooperative Energy Saving Gain with Grouping.

References 1. J. William Atwood. A classification of reliable multicast protocols. IEEE Nextwork, 18:24–34, 2004. 2. E. R. Berkekamp. Algebraic Coding Theory. McGrawHill, 1968. 3. Frank H. P. Fitzek and Marcos D. Katz, editors. Cooperation in Wireless Networks: Principle and Applications. ISBN-10 1-4020-4710-X. Springer, 2006. 4. S. Floyd, V. Jacobson, C.-G. Liu, S. McCanne, and L. Zhang. A reliable multicast framework for light-weight sessions and application level framing. Networking, IEEE/ACM Transactions on, 5(6):784 –803, 1997. 5. Jr. G. C. Clark and J. B. Cain. Coding for Error Control. 1981. 6. Li-wei H. Lehman, Stephen J. Garland, and David L. Tennenhouse. Active reliable multicast. Proceedings - IEEE INFOCOM, 2:581–589, 1998. 7. J. C. Lin and S. Paul. Rmtp: A reliable multicast transport protocol. In ACM SIGCOMM, August 1996. 8. S. Lin, D. J. Costello, and M. J. Miller. Automatic-repeat-request error-control schemes. IEEE Commun. Mag., 22:5–17, 1984. 9. F. J. MacWilliams and N. J. A. Sloane. Theory of Error Correcting Codes. North-Holland, 1977. 10. D. M. Mandelbaum. Adaptive-feedback coding scheme using incremental redundancy. IEEE Trans. Infom. Theory., pages 388–289, May 1974. 11. Jorg Nonnenmacher, Ernst W. Biersack, and Don Towsley. Parity-based loss recovery for reliable multicast transmission. IEEE/ACM Trans. on Networking, 6(4), August 1998. 12. Katia Obraczka. Multicast transport protocols: a survey and taxonomy. IEEE Communication Magazine, January 1998. 13. P. Radoslavov, C. Papadopoulos, R. Govindan, and D. Estrin. A comparison of application-level and router-assisted hierarchical schemes for reliable multicast. IEEE/ACM Trans., 12:469–482, June 2004.

498

Qi Zhang and Frank H.P. Fitzek

14. B. Whetten, T. Montgomery, and S. Kaplan. A high performance totally ordered multicast protocol. Theory and Practice in Distributed Systems, LCNS938, 1994. 15. Tina Wong, Thomas Henderson, and Randy H. Katz. Tunable reliable multicast for periodic information dissemination. Mobile Networks and Applications, 7(1):21–36, 2002. 16. M. Yamamoto. Multicast communication in next-generation internet. 10th IFAC/IFORS/IMACS/IFIP Symposium on Large Scale Systems: Theory and Applications, pages 639–645, July 2004.

26 IP Header Compression for Cellular-Controlled P2P Networks Or how to Reduce Bandwidth Requirements for Delay-Sensitive Applications

Tatiana K. Madsen1 , Qi Zhang2 , and Frank H.P. Fitzek1 1 2

Aalborg University [email protected] Technical University of Denmark [email protected]

Summary. As the demand for high quality IP services for mobile users is increasing, the problem of bandwidth limitation and unreliability of wireless links is faced. Additional requirements are imposed for real-time audio and video over cellular networks. Applications such as Voice over IP or video conferencing are delay sensitive. To use the precisions bandwidth efficiently, the voice payload and IP packet headers are compressed. However, using the conventional IP header compression schemes, the stringent latency budget of real-time applications can not be met. It happens due to the low responsiveness of the omnipresent cellular networks where the compressor entity is located at the concentrator node and not at the access point itself. Delayed context update leads to appearance of the error burst that can not be hidden from a user by CODECs. To overcome the above-mentioned problems and to make it attractive to employ header compression in cellular networks, we propose a new context repair mechanism that is based on the concept of micro cooperation.

26.1 Introduction and Motivation The boundary of wireless services has been continuously expanded. A typical user expects that all the communication services known from the wireline communication are also provided in a wireless format. This, however, poses formidable challenges, especially in case of delay-sensitive applications as e.g., real-time audio and video [9]. Voice over IP (VoIP), video conferencing, audio and video feeds from live events constitute real-time applications. These services are quite sensitive to latency. Speaking about VoIP, a user notices round-trip delays when they exceed 250-300 mSec. Therefore, one-way latency budget is about 150 mSec. 150 mSec is also specified in ITU-T G.114 recommendation as the maximum desired one-way latency to achieve highquality voice. High overhead in VoIP communications poses an additional challenge for its implementation in wireless networks. Depending on the CODEC used, data is compressed down to 6.4 Kbps (G.723 MP-MLQ), 8 Kbps (G.729 CS-ACELP), 16 Kbps (G.728 LD-CELP) and it is sent every 20 mSec [2]. If G.729 CODEC is applied, the

499 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 499–511. c 2007 Springer.

500

Tatiana K. Madsen et al.

resulting voice payload is just 20 bytes for each packet. However, to transfer these voice packets over a network, we should add IP header of 20 bytes, UDP header of 8 bytes and 12 bytes for RTP header. In some cases, such an overhead is fine. However, for low-bandwidth links this overhead should be reduced. One solution is to increase the packet size. By deciding to send packets every 40 mSec, it is possible to increase the bandwidth efficiency. This approach will affect the delay budget that in case of wireless networks is quite tight beforehand. Another solution is to employ header compression. This chapter is fully devoted to investigation of this approach in wireless settings. The concept of IP header compression has been around for over a decade. One of the first algorithms was Van Jacobson TCP/IP Header Compression [11] described in RFC 1144 and designed to improve TCP/IP performance over slow serial links. Introduction of compression methods for TCP/IP packet headers was followed by algorithms for compression of other headers, e.g., RTP. Header compression is popular with some vendor’s equipment, especially on slow links such as PPP or ISDN. For these links Compressed RTP [4] is typically used. IP header compression is crucial for low-bandwidth links, both wired and wireless, since it achieves significant header size reduction, typically from 40 bytes for IPv4 to 2-4 bytes. The achieved savings are even more prominent if the payload of the self packet is small. Nowadays, delivery of IP-based services over cellular links is not something extraordinary. Here the benefit of header compression is obvious. However, despite of the research efforts, IP header compression in wireless networks remains a challenging task. This is due to high bit error rates of the wireless environment that are harmful for the IP header compression. Packet losses lead to the de-synchronization of the compressor and decompressor and the conventional header compression algorithms fail to provide robustness towards channel errors. The necessary enhancement of the classical approaches should be introduced considering error proneness and long round-trip delays of cellular links. One of the solutions proposed for the cellular networks is Robust Header Compression (ROHC) [3]. ROHC has become a part of the Third-Generation Partnership Project - Universal Mobile Telecommunications Systems (3GPP-UMTS) specification. IP header compression can be considered as a special case of data compression. It exploits statistical redundancy in such a way as to represent the IP packet headers more concisely. Intrapacket redundancies are derived from the various headers (IP, UDP, RTP) of a single packet. Interpacket redundancies correspond to the marginal differences between contiguous packets of a given flow. For successful decompression both sender and receiver should have the same decompression state, referred to as context. Losses of packets in a flow lead to the failure of the decompression procedure and, upon detection of the missing packets, the context repair mechanism is triggered. The full context update is requested from the compressor entity that is typically located at the concentrator node in the network and not at the access point. Due to propagation and processing delays, responsiveness of the networks on the channel error is low and a single packet error causes the appearance of the error burst. All conventional header compression scheme will suffer from this problem. Typically, the application layer can deal with single packet losses, but bursty errors degrade significantly Quality of Service (QoS) perceived by a user. RTP data is transferred over the unreliable UDP layer. For this reason many CODECs are designed to account for the possibility of packet losses [2, 16]. E.g. interpolation

26 IP Header Compression in CCP2P Networks

501

to eliminate interruptions in the audio stream will be performed by CODEC and an occasional packet loss will be unnoticeable to the user. However, packet loss starts to be a real problem when the percentage of the lost packets exceeds a certain threshold, or when packet losses are grouped together in large packet bursts. In those situations, even the best CODECs will be unable to hide the packet loss from the user, resulting in degraded voice quality. Therefore, due to the low responsiveness of the omnipresent cellular networks it is not feasible to apply conventional header compression schemes for applications with stringent real-time requirements, such as wireless VoIP. To improve efficiency of header compression schemes there is a clear need to reduce the time required for the re-synchronization of the compressor and decompressor. So far there has been two approaches to solve this problem: •

•

To extract the missing information from other protocol headers. As an example of this approach, we can mention a Link-Layer Assisted Profile [12] for ROHC that utilizes functionality provided by the lower layers to increase compression efficiency. Another example is partial header compression. In some cases, when it is known that network links are highly error-prone, some headers, e.g., RTP, will not be compressed at all and the context state information is updated based on uncompressed fields. The compression gain in this case is low. To train the system to make an educated guess and predict the change of the context in case of packet losses. The well-known example is TWICE algorithm [5]. Applying the delta of the current segment (or multiples of the delta), the decompressor can compensate for packet losses in case the the pattern of the stream headers fields does not change. Using TWICE algorithm from 53 to 99% of the losses can be repaired, but processing complexity increases.

In this chapter we describe yet another approach for context repairing that avoids the bursty errors problem. The proposed mechanism helps to keep the decompressor operational without the need of excessive signalling to an access point. It does not introduce any additional delays in the latency budget and results in the improved QoS. The presented algorithm is based on the concept of micro cooperation [7]. The omnipresent architecture in wireless communication consists of autarky terminals that individually receive services from an access point. If the terminals have the capabilities to establish peer-to-peer (P2P) connections with each other as well, e.g., using short-range links, cooperative groups can be formed. We refer to the network architecture formed in this way as cellular-controlled P2P networks. It has been noted that using short-range links for data transmission is less costly compared with the cellular links since higher data rates and lower powers for transmission and reception can be achieved over close distances [7]. Based on this observation, we ague for cooperative behavior of wireless terminals that exploit P2P connections for information exchange in order to keep their decompressor states updated. The benefit of micro cooperation comes from the multi–path diversity: if a radio path between an AP and a terminal is greatly deteriorated by the instantaneous channel conditions, the whole IP packet is lost. A neighboring user might be experiencing good channel conditions and might be able to deliver information for the context update. The cooperative users are rewarded by keeping their decompressors operational.

502

Tatiana K. Madsen et al.

26.2 Autonomous and Cooperative Header Compression in Cellular-Controlled P2P Networks The quality of the received service in cellular networks highly depends on the available system capacity. All terminals are directly connected to an access point and all traffic goes to and from the access point. In highly populated cells or when the traffic demands are high, quality of service degrades. Alternative network architectures have been considered aiming at improving the performance of cellular data networks. These include hybrid architectures where both the centralized (e.g., cellular) and distributed (e.g., peer-to-peer) topologies are combined [10]. Recently, a new approach to bridge cellular and peer-to-peer architectures, referred to as Cellular Controlled Peer-to-peer networks [7, 8] has been proposed. Besides being able to communicate with the base station using cellular interfaces, terminals have the capability to establish direct peer-to-peer connections over short-range links. A group of terminals, typically in close proximity, form a cooperative cluster that is a peer-to-peer network in its full right. The base station works as a service entry point and administrator for instance for authentication and billing purposes. Peer-to-peer connections can be potentially used for content distribution, error healing and retransmissions. In this chapter we demonstrate how information exchange over short-range links among the cooperative terminals can facilitate efficient operation of IP header compression algorithms. The system under investigation follows the architecture presented in Figure 26.1. A number of wireless terminals connected to access points are interested in data services, e.g., VoIP. In general case, the last wireless hop can be presented by the same or different wireless technologies, e.g., WLAN IEEE 802.11 or GPRS. The considerations given below will hold as long as they share the same concentrator.

Figure 26.1. System under investigation.

Traditionally, header compression schemes are designed to work on an individual link. Header compression is typically performed on the headers of the network layer

26 IP Header Compression in CCP2P Networks

503

and above. The packet headers are compressed at the sending node. The decompressor at the receiver side reconstitutes the original header before delivering the packet to the higher protocol layers. Despite the fact that there exist over a dozen different schemes for different types of packet streams, they all rely on the same principle: redundancies between contiguous packets of a given flow allow significant reduction in size by using differential encoding [6]. Random fields of packet headers are transmitted unchanged, whereas delta fields are compressed by reference to the previous packets. The context is known and maintained at the receiver side as well and used for decompression of the incoming packets. M packets with full headers are required in order to establish the context at the decompressor; afterwards packets with compressed headers are sent. The context is updated with every new packet. Packet losses lead to inconsistencies in the context state at the decompressor and failure of the decompression procedure. Upon an update request from the receiver, packets with full headers will be transmitted in order to repair the context. However, in the omnipresent system architectures, compressor is located not at the access point but somewhere in the network (see Figure 26.1). For example, in GPRS system compression is performed in Serving GPRS Support Node (SGSN). The system can not be highly responsive to the update request and it can take d packets before the context will be re-established due to propagation and processing delays. We refer to the parameter d as responsiveness of the network on the channel errors. If a number of wireless terminals establish data connections with an access point; all entities posses header compression capabilities; and compression of packet streams will be performed individually for each active connection, then we refer to this situation as autonomous or non cooperative operation mode. All conventional compression schemes are non-cooperative in their nature. Let us assume that we can model packet losses between an access point and a wireless terminal as i.i.d. We denote the probability to loose a packet as p. Let the random variable γ denote the number of packets within the cycle, i.e., the number of compressed headers that can be sent before the decompression procedure suffers from erroneous transmission and the context update takes place. The expectation of γ can be calculated as 1 N = E[γ] = + d (1) p In order words, we expect that in case of available feedback channel the context update will be repeated every N packets on average. Let now assume that J terminals form a cooperative group (Figure 26.2). We assume that each wireless terminal has the capability of communicating with the access point and simultaneously with other terminals (by using either the same or different air interfaces). Each terminal receives a data stream with compressed packet headers together with some additional information intended to the neighboring terminals. This information, refereed to as additional information container (AIC), can be used to update the context state of the decompressor - note that we are speaking about the decompressors of other terminals within the cooperative group. Each terminal will receive J − 1 AICs. AICs exchange is performed on demand using P2P connections. In order to facilitate the context maintenance, an AIC should contain differentially encoded fields of a compressed header. For example, if CRTP scheme is used

504

Tatiana K. Madsen et al.

Figure 26.2. Example of a cooperative group consisting of two terminals.

for compression, the size of an AIC is 2 bytes (total size of a compressed header of 4 bytes minus 2 bytes reserved for random fields, such as UDP checksum). Since the context update from the access point is required only in case when all wireless terminals fail to receive packets correctly at the same time instant, the average cycle length can be found as N = E[γ]coop =

1 +d pJ

(2)

From formulas (1) and (2) one can see that the average cycle length in cooperative scenario is longer compared with the conventional schemes. It means that more packets with compressed headers can be sent resulting in better compression gain and better bandwidth efficiency.

26.3 Design of Information Exchange over Short Range Connections In order to limit the costs of AICs exchange, it should be performed on demand, i.e., by a request issued by a terminal. Additionally, we assume that the packet streams for different terminals have the same rate. In this case the AICs on different channels are received in tact by terminals. In case of asynchronous streams, the AICs sending rate should be adjusted to the rate of the packet flows. We will illustrate the concept of peer-to-peer information exchange using shortrange connectivity provided by Bluetooth technology. The widespread of mobile phones with Bluetooth capabilities makes this technology an obvious choice to support local AICs exchange. Bluetooth has a star topology with a central coordinating unit-master and up to 7 active slaves. One terminal is assigned a role of a master and all communication flows go through this terminal. This master-slave mode of communication within a piconet is not very efficient for peer-to-peer information exchange. Additionally, power burden of the communication lies on the master of the piconet. To alleviate

26 IP Header Compression in CCP2P Networks

505

such unfairness, a master-slave switch should be performed periodically where all members of the cooperative group take a role of the master in turn. Alternatively, a flexible network structure can be introduced where each wireless terminal creates its own piconet as a master, admits the other nodes as slaves in that piconet and parks them. Thus, if there is 3 terminals in a cooperative group, then 3 overlapping piconets consisting of the same nodes but with different masters are formed (see Figure 26.3). It is beneficial to use park mode to minimize power consumption of terminals. Park mode has the lowest duty cycle comparing with hold and sniff modes. If no AICs exchange is needed, the only nodes’ activity is periodical listening to the beacon signal from the master for synchronization purpose. If a terminal does not receive a packet over a cellular link, it initiates the procedure of AIC retrieval over Bluetooth connection. It sends unparking request that in this situation is considered as an indication that the context is missing and the corresponding AIC should be sent. If another terminal has the AIC, it unparks the first node and the data is sent. After information exchange is done, the terminal is parked again. This operation is especially suited for traffic patterns where the data is generated in bursts from time to time. For a more detailed description of flexible piconet structure readers are referred to [15].

Figure 26.3. System under investigation.

Let assume that 1-slot DM1 packets are used for transmission (the choice of DM packet over DH is due to its high error-resilience). The payload of DM1 packet contains up to 17 information bytes. This corresponds to the 8 AICs 2 bytes each. Choosing the number of terminals for the cooperative group, we should ensure that the context healing can be done and at the same time the overhead connected with the cooperative behavior is reasonably small. It is unrealistic to have more than 8 terminals in a group due to increase in overhead. Actually, as we will show in the next section, the number of two or three cooperative terminals is the optimal one for the wide range of system parameters. Therefore, one packet can contain AICs for all neighbors in the cooperative group. In order to reduce the number of messages sent, instead of sending AICs individually to each slave, the master can broadcast one packet containing all requested AICs. What is more, according to the Bluetooth specification [14], unparking of the slaves can be avoided if general

506

Tatiana K. Madsen et al.

broadcast messages should be carried to the parked slaves. In the beacon train, the master can support broadcast messages to the slaves. Using this approach, only a few slots will be required for AICs exchange, namely 3 slots (see Figure 26.4). If a node does not receive broadcast packet, it assumes that no healing of the context can be done by using short-range communication and it sends a full context update request to the access point.

Figure 26.4. Communication flow over Bluetooth link.

One should note that the proposed scheme of the AICs exchange using Bluetooth connections is just one of the many possible algorithms and some other methods might be found to be more efficient .

26.4 Evaluation of CCP2P Header Compression First of all, we will demonstrate that the proposed header compression for cellular controlled peer-to-peer networks can successfully avoid appearance of bursty errors. Probability of the error burst is derived for the case of a conventional header compression based on differential encoding and for the cooperative header compression. Secondly, the performance of the cooperative approach is investigated in terms of bandwidth efficiency. Thirdly, energy efficiency of the described approach is studies since the minimization of the energy consumption is a desired property for small battery-driven devices.

26.4.1 Probability of Error Burst Let assume that a conventional header compression scheme can sustain loosing m packets in a row without de-synchronization between compressor and decompressor. But when m + 1st packet is lost, the next d packets are lost as well due to delay in obtaining the context update and the burst of errors appears. Let denote probability of this event as q. Let now assume that there are J terminals in a cooperative group. A certain amount of packet losses can be healed provided that they do not occur simultaneously for all the terminals. Therefore, the probability of bursty errors in case of cooperative compression can be found as q J . Table 26.1 shows that even if we have only 2 cooperative terminals, the corresponding probability is significantly reduced. In case of J = 3, the probability of a burst is negligible small. The values in the table are given under assumption of a perfect short-range communication channel. Considering that loss of packets can also occur over short-range link and

26 IP Header Compression in CCP2P Networks

507

the fact that there is place only for a limited number of retransmissions, the values for the error burst probability for J = 2 and J = 3 can be expected to be slightly larger.

Table 26.1. Probability of error burst. J=1 1% 2% 5% J=2 0.01% 0.04% 0.25% J=3 0.0001% 0.0008% 0.0125%

26.4.2 Bandwidth Savings and Energy Efficiency The expected bandwidth savings due to header compression can be calculated as a ratio of the reduction in file sizes due to compression over the total amount of information sent in the uncompressed case. Calculating energy consumption, we should take into account both the energy consumed by a terminal in reception over cellular link and the energy spent on the communication over short-range link. Table 26.2 gives the considered values for power for cellular and short-range connections [15,17]. Furthermore, energy efficiency is founded as a ration of the amount of transmitted useful data over the total energy spent for the complete transmission (including all overheads). The detailed derivations can be found in [13].

Table 26.2. Typical power values for cellular link and short range link (Bluetooth) communication. Power Type Pc,rx Pc,i Psr,tx Psr,rx Psr,i Psr,sleep mill Watt 1254 125 100 78 0.486 0.054

Figures 26.5-26.8 present comparison of non-cooperative and cooperative approaches for header compression in terms of bandwidth savings and energy efficiency. All figures are plotted as a function of the network responsiveness, d. The case d = 0 corresponds to the ideal situation when the context update is received immediately after a de-synchronization between the decompressor and compressor appears. The results are given for the cooperative groups consisting of two and three terminals (J = 2 and J = 3) and for four different values of individual packet losses (packet loss rate P LR = 0.5%, 1%, 2% and 5%). The performance of non-cooperative compression strategy decreases with increase in d. However, the curves for cooperative method remain almost constant. This can be explained by the large cycle length in cooperative case. Due to the large values of N , the average values for bandwidth savings and energy efficiency are not affected by the relatively small changed in d. The higher packet loss rate results in higher payoff of micro cooperation. In the case of P LR = 2% and 5%, the cooperative strategy outperforms the noncooperative one. However, in the case of good channel conditions (low values of PLR), the context update will be required seldom and, therefore, there is no need to

508

Tatiana K. Madsen et al.

Figure 26.5. Bandwidth Saving Comparison. P LR = 0.5%, 1%.

Figure 26.6. Bandwidth Saving Comparison. P LR = 2%, 5%.

Figure 26.7. Energy Efficiency Comparison. P LR = 0.5%, 1%.

Figure 26.8. Energy Efficiency Comparison. P LR = 2%, 5%.

invest the resources in cooperation. Terminals that experience very low PLR should sustain from cooperation and do not join cooperative groups. The number of three cooperative terminals (J = 3) shows the lowest probability of error burst. However, the bandwidth and energy efficiency for J = 2 is higher compared with J = 3. The later case corresponds to the larger compressed header size and, thus, more resources per bit of useful information are spent. We can conclude that the number of two cooperative terminals are optimal. Considering the energy efficiency graphs, we observe very low efficiency in the case when no header compression is applied. This emphasizes the importance of compression for the battery-driven hand-held devices.

26.5 Discussion on Cooperation Strategies It has always been understood that performance of peer-to-peer, ad hoc and hybrid networks depends on the level of cooperation of the participants. While most existing

26 IP Header Compression in CCP2P Networks

509

peer-to-peer networks are built on the assumption that participants are generally cooperative, there is a growing evidence suggesting the opposite. The problem of how to effectively engage a selfish rational user to contribute with his own resources and how to encourage a user to help others is still an open issue. Staring from [1], there have been many contributions addressing this problem. The proposed mechanisms include trusted third parties, usage of reputation information or application of reciprocal punishment. It is a challenging task to facilitate cooperation in networks with a large population. Any accountability mechanism demands high memory requirements. However, as we have shown, the gain from cooperative behavior used to enhance compression algorithms is maximized for clusters with a small number of terminals. The best performance has been observed for groups of size just two terminals. This greatly simplifies the task of guaranteeing reciprocity and detection of cheaters. It can be achieved by a simple tit-for-tat strategy: individuals store the result of the last interaction made by those they interact with and return the same if they meet again in the future. In practice, the tit-for-tat strategy can be realized with a counter-based algorithm involving threshold values. When forming a cooperative group, each user assigns a threshold value (payoff margin) for all other members of the group. The threshold represents the delay that the user can tolerate in receiving a pay-off. Each user maintains a table that records information on previous interactions with other terminals and contains current status of a transaction history. The current payoff value P ayof fi,j can be calculated as follows: P ayof fi,j = Rewardj,i − Costi,j + σ where σ is a payoff margin, Rewardj,i is the service provided to the jth user and Costi,j is the service received from the jth user. The payoff margin and service are measured in the number of context updates received over the short-range link. The P ayof f value is updated after each interaction. If P ayof fij < 0, then terminal i stops providing service to the terminal j. In case there are personal relationship among members of a cooperative group, it is observed that a user is willing to provide service without being “payed back” immediately. Indeed, in such situation repeated interactions with the same user are likely and “pay-off” can be received another time. The threshold σ should just be put very large for such terminals (e.g., people from your address book). When designing algorithms for cooperation among wireless devices, additional issues should be taken into account. Indeed, a terminal might not have the requested content to share with other terminals in a group due to the high bit error rates caused by the wireless channel as well as unreliable transmission. Thus, this terminal can be mistakenly assumed to be a cheater. A terminal will request a context update from a neighboring terminal only in case it has experienced a packet loss. A request will not be fulfilled in two cases: • •

another terminal is a cheater; channel errors experienced by two terminals are correlated in time.

In both cases cooperation should be stopped, since there is no benefit of cooperation in case of correlated errors and one should not help a cheater. Therefore, the simple tit-for-tat strategy can be applied without need for any changes.

510

Tatiana K. Madsen et al.

26.6 Conclusions In this chapter we have considered cellular controlled peer-to-peer communications, a dynamic approach to bridge cellular and peer-to-peer networks. We have demonstrated its potential to overcome some limitations of the omnipresent cellular networks focusing on the example of IP header compression. The performance of IP header compression suffers from the unreliability and error-proneness of wireless links. Forming a cooperative group, wireless terminals can provide and receive “first aid” information to heal the decompressor state in case of packet losses on the cellular link. This leads to significant increase in robustness of the compression algorithm. The investigation was based on the assumption of a feedback channel to react on lost packets and reinitialize the decompressor state. The feedback channel responsiveness was varied from low to large delays. This is needed to reflect the architecture of existing 2.5G and 3G, and future 4G networks. For a varying value of the responsiveness and packet error rates, we could show that header compression schemes (non-cooperative and cooperative) outperform the uncompressed communication in terms of bandwidth savings and energy efficiency. For large values of responsiveness the cooperative header compression should be preferred since it can successfully avoid appearance of packet errors bursts.

References 1. R. Axelrod. The Evolution of Cooperation. Basis Books, 1984. 2. Y. Boger. Fine-tuning Voice over IP services. VP Business Development. RADCOM, 2000. 3. C. Bormann, C. Burmeister, M. Degermark, H. Fukushima, H. Hannu, L-E. Jonsson, R. Hakenberg, T. Koren, K. Le, Z. Liu, A. Martensson, A. Miyazaki, K. Svanbro, T. Wiebke, T. Yoshimura, and H. Zheng. RObust Header Compression: ROHC: Framework and four profiles: RTP, UDP, ESP, and uncompressed. Technical report, Request for Comments 3095, July 2001. 4. S. Casner and V. Jacobson. Compressing IP/UDP/RTP Headers for Low-Speed Serial Links, Request for Comments 2508, February 1999. 5. M. Degermark, B. Nordgren, and S. Pink. IP Header Compression, Request for Comments 2507, February 1999. 6. F.H.P. Fitzek, S. Hendrata, P. Seeling, and M. Reisslein. Wireless Internet – Header Compression Schemes for Wireless Internet Access, ISBN 0849316316 Chapter 10, pages 1–24. Electrical Engineering & Applied Signal Processing Series. CRC Press, March 2004. 7. F.H.P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-40204710-X. Springer, April 2006. 8. F.H.P. Fitzek, M. Katz, and Q. Zhang. Cellular Controlled Short-Range Communication for Cooperative P2P Networking. In Wireless World Research Forum (WWRF) 17, Heidelberg, Germany, 2006. 9. F.H.P. Fitzek, A. K¨ opsel, A. Wolisz, M. Krishnam, and M. Reisslein. Providing Application-Level QoS in 3G/4G Wireless Systems: A Comprehensive Framework Based on Multi-Rate CDMA. IEEE Personal Communications - Special issue on 4G Technologies and Applications, 9(2):42–47, April 2002.

26 IP Header Compression in CCP2P Networks

511

10. H.-Y. Hsieh and R. Sivakumar. On Using Peer-to-Peer Communication in Cellular Wireless Data Networks. IEEE Trans. on Mobile Computing, 3(1):57–72, 2004. 11. V. Jacobson. Compressing TCP/IP Headers for Low-Speed Serial Links, Request for Comments 1144, February 1990. 12. L-E. Jonsson and G. Pelletier. RObust Header Compression (ROHC): A LinkLayer Assisted Profile for IP/UDP/RTP, Request for Comments 3242, April 2002. 13. T. Madsen, Q. Zhang, F.H.P. Fitzek, and M. Katz. Design and Evaluation of IP Header Compression for Cellular-Controlled P2P Networks. In IEEE International Conference on Communications 2007, Glasgow, GB, June 2007. 14. Specification of the Bluetooth system. Std., Rev. 2.0. [Online]. Available: http://www.bluetooth.com. 15. P. Popovski, H. Yomo, G. Kuijpers, T. K. Madsen, and R. Prasad. Blue-Park: Energy-Efficient Operation of Bluetooth Networks using Park Mode. Elsevier Computer Communications Journal, 29(17):3416–3424, 2006. 16. H. Sanneck, N. Le, A. Wolisz, and G. Carle. Intra-Flow Loss Recovery and Control for VoIP. In ACM Multimedia 2001, Ottawa, September 2001. 17. E. Shih, P. Bahl, and M. Sinclair. Wake on Wireless: An Event Driven Energy ‘Saving Strategy for Battery Operated Devices. In MOBICOM’02, Atlanta, USA, September 2002.

27 Cluster Based Cooperative Uplink Access in Centralized Wireless Networks One for All and All for One!

Qi Zhang1 , Frank H.P. Fitzek2 , and Villy B. Iversen1 1 2

Technical University of Denmark [qz|vbi]@com.dtu.dk Aalborg University [email protected]

Summary. In this chapter we propose the one4all cooperative access strategy to introduce a more efficient uplink access strategy for cellular networks. The one4all scheme is based on the cellular controlled peer-to-peer network architecture. The basic idea is that mobile devices form a cooperative cluster using their short–range air interface and one device contends the channel for itself and all neighboring devices within the cluster. This strategy reduces the number of mobile devices involved in the collision process for the wireless medium resulting in larger throughput, smaller access delay, and less energy consumption. Based on an analytical model, the proposed strategy is compared with the two existing strategies RTS/CTS (request to send/clear to send) and packet aggregation. The results show that the proposed cooperative scheme has better throughput performance than packet aggregation and has much higher throughput than the conventional RTS/CTS scheme. Furthermore, the newly introduced scheme outperforms packet aggregation in terms of channel access delay and energy consumption.

27.1 Introduction Current wireless local area networks (WLANs) suffer from an inefficient wireless access strategy. For example, the IEEE 802.11 WLAN standard product can provide up to 54 Mbps transmission rate at the physical layer. The recent IEEE 802.11n proposals aim at providing physical layer transmission rates up to 600Mbps. But theoretical throughput limits exist [13,15] due to the medium access control (MAC) and the physical layer (PHY) overhead, the backoff time in case of contention, the inter-frame space (IFSs) and others. Therefore, to achieve high throughput values at the network layer, research should focus not only on higher physical layer data rates but also on more efficient MAC strategies to reduce the aforementioned overhead. So far the most popular and effective strategy to enhance WLAN throughput is packet aggregation [6–9, 12]. But packet aggregation has inevitable drawbacks, for instance the throughput gain is highly dependent on the arrival traffic pattern. Packet aggregation does improve the throughput for bursty traffic such as video streaming, but it would not improve throughput so much for non–bursty traffic such

513 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 513–529. c

2007 Springer.

514

Qi Zhang et al.

as VoIP traffic. Furthermore, it also may cause longer channel access time which in turns leads to higher energy consumption and unfair channel usage between mobile devices. In this chapter, we propose a new cooperative uplink access strategy named one4all. It is based on the CCP2P architecture described in Chapter 2. The basic idea of the proposed strategy is that every mobile device within a communication cell is not contending for the channels for itself anymore, but benefit from cooperation among neighboring mobile devices. Before we explain the one4all strategy more in detail, we first explain the two existing channel access strategies in WLANs.

27.2 CSMA/CA Based MAC Strategies This section presents two existing MAC strategies using carrier sense multiple access with collision avoidance (CSMA/CA), namely the conventional RTS/CTS strategy and the packet aggregation strategy. Based on the analytical model of CSMA/CA in [2, 16], we start to analyze the throughput, channel access delay and energy consumption of conventional RTS/CTS strategies. Afterwards we extend the existing analytical model and apply it to a system with packet aggregation. The corresponding performances of packet aggregation scheme are analyzed.

27.2.1 RTS/CTS Strategy An analytical model for CSMA/CA to analyze the throughput and delay performance in case of saturation is developed in [2, 16]. In this model, it is assumed that the network consists of n contending mobile devices and each device has an arrival packet for transmission immediately after its completion of a successful packet transmission. In [16] it has already shown that CSMA/CA with RTS/CTS outperforms the basic CSMA/CA under given packet size assumption. Therefore we only address the RTS/CTS case within this chapter. The RTS/CTS access mechanism illustration diagram is shown in Figure 27.1.

Figure 27.1. The RTS/CTS access mechanism [2].

In the following the throughput and channel access delay are analyzed and their mathematical expression are derived. The notations in the equations are listed in Table 27.1.

27 Cooperative Uplink Access

515

Table 27.1. Parameter list. Notation n W m pb p Ps τ Ψ Nc Db σ Cbk F Nf r Ts Tc To Tp Tint η Dc RTS/RTSt CTS/CTSt P /Pt PHYh /PHYht MACh /MACht ACK/ACKt δ SIFS DIFS R

Meaning the number of contending mobile devices min (initial) contention window backoff stage (Wmax = 2m W ) the channel busy probability collision probability of a transmitted frame the probability that a transmission is successful the probability that a mobile device transmits during a slot time The number of consecutive idle slot times before a transmission takes place the number of collisions of a frame until its successful transmission the backoff delay one unit slot time the value of the backoff counter the duration that a counter is in freezing state before a counter reaching zero the number of times that a counter freezes before a counter reaching zero transmission time of a single successful frame transmission the average transmission time of transmission with collision the time that a mobile device has to wait when its frame transmission collides, before sensing the channel again. the time used for successful transmission of a payload the time interval between two consecutive transmissions throughput the channel access delay size of a RTS frame/transmission time of a RTS frame size of a CTS frame/transmission time of a CTS frame size of the PDU payload/ transmission time of a payload size of a physical layer header/transmission time of a PHY header size of a physical layer header/transmission time of a MAC header size of a block ACK frame/transmission time of a block ACK frame propagation delay short interframe space distributed interframe space channel data rate

516

Qi Zhang et al.

Throughput Analysis For the model presented in [2, 16], it is assumed that each transmission is a renewal process, no matter whether it is successful or not. Therefore, the saturation throughput η can be calculated according to the payload transmitted during a single renewal interval between two consecutive transmissions. The expression of saturation throughput η is given in [16] as following: η= =

E[Tp ] E[Tint ] Ps E[Pt ] E[Ψ ]σ + Ps Ts + (1 − Ps )Tc

(27.1)

Based on the model in [2, 16], the unknown variables in Eq. 27.1 are obtained by using the equations given in [16] which defines three new variables p, pb and τ (see Table 27.1). The relations of these three variables are given in the following: p = 1 − (1 − τ )n−1

(27.2)

pb = 1 − (1 − τ )n

(27.3)

2(1 − pb )(1 − 2p) 2(1−pb)2(1−2p)(1−p)+(pb +p(1−pb))(1− 2p)(W +1)+pW(pb +p(1−pb))(1−(2p)m) (27.4) Substituting Eq. 27.2 and 27.3 to Eq. 27.4, we can obtain the probability τ . Afterwards p and pb are calculated. The probability of a successful transmission Ps is given by [16] τ=

Ps =

nτ (1 − τ )n−1 1 − (1 − τ )n

(27.5)

The average number of consecutive idle slots before a transmission takes place E[Ψ ] is given by [16] 1 E[Ψ ] = −1 (27.6) Pb Then based on the RTS/CTS mechanism (see Figure 27.1), the transmission time of a single successful frame transmission Ts and Tc can be calculated as following: Ts = RTSt +δ+SIFS+CTSt +δ+SIFS+Ht +Pt +δ+SIFS+ACKt +δ+DIFS (27.7) where, Ht = PHYht + MACht Tc = RTSt + δ + DIFS

(27.8)

After getting all these unknown variables in Eq. 27.1, the throughput η can be easily obtained.

27 Cooperative Uplink Access

517

Channel Access Delay Analysis The channel access delay is defined as the time duration starting when a mobile device contends the channel to transmit a packet for the first time until the instant where it can start to transmit the packet successfully. It includes the backoff delay (Db ), which is the time a mobile device chooses to wait before accessing the channel under busy channel condition; the time (Tc ) during which the channel is captured by mobile devices and collision happens and the time (To ) that a mobile device has to wait if the transmitted frame collides [16]. The channel access delay also depends on the number of collisions before it finally access the channel successfully. The average channel access delay can be expressed as E[Dc ] = E[Nc ](E[Db ] + Tc + To ) + E[Db ]

(27.9)

where the average number of collision before a mobile device finally access the channel is taken from [16] 1 E[Nc ] = −1 (27.10) Ps The backoff delay depends on the product of Cbk (the value of the backoff counter) and the slot time, and the time duration F during which each mobile device freezes the counter before the counter reaching zero. The average of the backoff counter is a random variable depending on the initial contention window and the backoff stages. The total duration of the counter in freezing state depends on the number of times that the mobile device freezes the counter and also on the duration of each freezing. So the average backoff delay is given by [16] E[Db ] = E[Cbk ]σ + E[F ] = E[Cbk ]σ + E[Nf r ](Ps Ts + (1 − Ps )Tc )

(27.11)

where, E[Cbk ] −1 max(E[Ψ ], 1) The time To can be calculated according to the standard as E[Nf r ] =

To = SIFS + CTStimeout

(27.12)

Thus, the average channel access delay can be calculated by substituting Eq. 27.8, 27.10, 27.11, 27.12 into Eq. 27.9.

Energy Consumption Analysis The energy consumption depends on the energy consumed in different communication phases of the mobile device and the time that the mobile device staying at the corresponding state. There are four different possible states that a device can stay: transmission state, reception state, listening state and idle state. Their corresponding power consumptions are denoted by Ptx , Prx , Pli , Pi , respectively. The mobile device is in the transmission state only when it sends RTS messages or data frames. The device is in reception state when it receives CTS messages or ACK messages from the access point or receives RTS message from other mobile devices. During DIFS and SIFS the device is in listening state. We assume that

518

Qi Zhang et al.

the mobile device performs a smart energy saving strategy, i.e., the mobile device can switch to the idle state when it adjusts its network allocation vector (NAV). The energy consumption of each packet is calculated by summing up all the energy consumption from a mobile device starting to contend until it has finished sending the packet. Hence, the energy consumption for each packet can be expressed by Eng = Ptx Ttx + Prx Trx + Pli Tli + Pi Ti

(27.13)

where, Ttx = RTSt +PHYht +MACht +Pt + E[Nc ]RTSt | {z } | {z } successful transmission time transmitted RTS collides Trx = (E[Nc ]+1)E[Nf r ](Ps (RTSt +CTSt )+(1−Ps )RTSt )+CTSt +ACKt | {z } others’ RTS&CTS Tli = E[Nc ](Tlibk + To + δ + DIFS) + Tlibk where, Tlibk is the mobile device’s listening time during a backoff delay and is given by Tlibk = E[Cbk ]σ+E[Nf r ](Ps (2δ+ 2SIFS+ DIFS)+(1−Ps )(δ+DIFS)) Ti = (E[Nc ] + 1)E[Nf r ]Ps (PHYht +MACht +Pt +δ+SIFS+ACKt +δ) | {z } NAV data time

27.2.2 Packet Aggregation Strategy Packet aggregation is a popular strategy to improve throughput in wireless networks based on CSMA/CA. It has been addressed in many research works [6–9, 12] and it will be included in IEEE 802.11n standard. Aggregation can be performed on different sub-layers. There are two main categories of packet aggregation [1,3,5,10,11, 14]: Aggregation of multiple MAC Protocol Data Units (A-MPDU) and aggregation of multiple MAC Service Data Units (A-MSDU). A-MPDU is also called packet concatenation [14]. The idea of A-MPDU is to concatenate multiple MAC PDUs into a single physical PDU. MAC PDUs can be concatenated if they are available and have the same physical source and destination address. The length of concatenation should not exceed a given threshold. The A-MSDU is also referred to as packet packing in some documents [14]. The idea is to combine multiple MAC SDUs from a higher layer into a big MAC PDU. MAC SDUs have the same MAC addresses. The detailed description of these two aggregation schemes can be seen in proposals from TGnSync or WWiSE [1, 11]. A comprehensive performance comparison of AMPDU and A-MSDU considering the channel error rate and the number of packets in an aggregated frame is given in [10]. One conclusion drawn in [10] is that the A-MPDU approach has better performance than A-MSDU approach considering the channel error impact. Hence, we use the A-MPDU strategy as a representative packet aggregation method here. Additionally, block ACK is also used in the AMPDU strategy. The A-MPDU frame structure diagram is shown in Figure 27.2.

27 Cooperative Uplink Access

519

Figure 27.2. Frame structure in A-MPDU packet aggregation scheme. Table 27.2. Parameter list. Notation TsAgg Na Namax Ta

Meaning transmission time of an aggregated frame transmission the number of packets in an aggregated frame the threshold of the number of packets in an aggregated frame the duration of the previous aggregated frame starting to contend until its sucessful completion transmission ηa throughput with packet aggregation MD/ MDt size of a MPDU delimiter/ transmission time of a MPDU delimiter λ packet arrival rate

Throughput & Channel Access Delay Analysis To calculate throughput and channel access delay in CSMA/CA with packet aggregation, we can extend the model developed by [16]. First of all, when the A-MPDU packet aggregation strategy is used in RTS/CTS mechanism, Ts has to be calculated in a different way. Based on the frame structure of A-MPDU (Figure 27.2), the average transmission time of an aggregated frame transmission TsAgg is a function of E[Na ], which is the average number of the packets in an aggregated frame. E[Na ] depends on the number of available packets in the buffer when a mobile device just finishes transmitting an aggregated frame. Hence, similarly as the expression of Ts , TsAgg is given as TsAgg = RTSt+δ+SIFS+CTSt+δ+SIFS+PHYht+E[Na ]Ut+δ+SIFS+ACKt+δ +DIFS (27.14) where, Ut = MDt +MACht +Pt. The model in [16] assumes that each mobile device has an arrival packet for transmission immediately after its completion of a successful packet transmission. But this assumption has to be extended considering the packet aggregation strategy. Packet aggregation is performed under the condition that the packets are available in the buffer and the number of packets in one aggregated frame must be smaller than the threshold, which means the number of packets in one aggregated frame is not fixed. Furthermore, to calculate the saturation throughput, the arrival traffic should meet the condition that there is at least one arrival packet in the buffer immediately

520

Qi Zhang et al.

after the mobile device completes an aggregated frame transmission. We generate such arrival traffic by a poisson process with additional constraints. The arrival rate of this poisson process has to meet that the probability of no arrival packet during time Ta is close to zero, denoted by ε (e.g., ε = 10−4 ). Under this condition, it can be regarded as that there is at least one packet in the buffer. Hence, the assumed poisson arrival process P (i, t) with constraints is expressed by P (0, Ta ) = e−λTa = ε (27.15) then λTa = −lnε

(27.16)

The time Ta based on its definition can be given by Ta = E[Dc ] + TsAgg = E[Nc ](E[Db ] + Tc + To ) + E[Db ] + TsAgg

(27.17)

where, E[Db ] = E[Cbk ]σ + E[Nf r ](Ps TsAgg + (1 − Ps )Tc )

(27.18)

According to the packet aggregation mechanism, the average number of the packets in an aggregated frame can be given by max Na

∞ X (λTa )i −λTa (λTa )i −λTa + Namax e e i! i! max+1 i=1 i=Na  max  Na ∞ i X (−lnε)i X (−lnε) max  = ε i Na + i! i! i=1 i=N max+1

E[Na ] =

X

i

(27.19)

a

So the throughput of CSMA/CA with the packet aggregation strategy similarly as Eq. 27.1, can be expressed as ηa =

Ps E[Na ]E[Pt ] E[Ψ ]σ + Ps TsAgg + (1 − Ps )Tc

(27.20)

By Eq. 27.14, 27.17 and 27.19, we can solve E[Na ] and TsAgg . Then the throughput η a can be calculated by Eq. 27.20. The average channel access delay of each mobile device can be calculated by substituting Eq. 27.18 into Eq. 27.9. So the average channel access delay for each packet is the average channel access delay experienced by each mobile device divided by E[Na ]. From Eq. 27.20, it is also clear that the achieved throughput gain depends on E[Na ] which is a function of Namax and the arrival rate of poisson process. So when the arrival rate is low, the packet aggregation can not enhance the throughput anymore. Another drawback of packet aggregation is its longer backoff delay (see Eq. 27.18) which causes longer channel access delay. Longer channel access delay results in larger energy consumption for the mobile devices.

27 Cooperative Uplink Access

521

Figure 27.3. The power level changing of mobile device 1 diagram.

Energy Consumption Analysis Energy consumption of packet aggregation scheme is calculated by the same methodology described in Subsection 27.2.1. According to the packet aggregation strategy, the power level changes as shown in Figure 27.3. Based on the Figure 27.3, the energy consumption of each packet is calculated by summing up all the energy consumption from a mobile device starting to contend until it finishing sending the aggregated frame; then the sum is averaged by the number of packets in one aggregated frame. Hence, the energy consumption for each packet can be expressed by Eng a =

Ptx Ttx + Prx Trx + Pli Tli + Pi Ti E[Na ]

(27.21)

where, Ttx = RTSt +PHYht +E[Na ](MDt +MACht +Pt )+ E[Nc ]RTSt | {z } | {z } successful transmission time transmitted RTS collides Trx = (E[Nc ]+1)E[Nf r ](Ps (RTSt +CTSt )+(1−Ps )RTSt )+CTSt +ACKt | {z } others’ RTS&CTS Tli = E[Nc ](Tlibk + To + δ + DIFS) + Tlibk where, by

Tlibk

is the mobile device’s listening time during a backoff delay and is given

Tlibk = E[Cbk ]σ+E[Nf r ](Ps (2δ+ 2SIFS+ DIFS)+(1−Ps )(δ+DIFS)) Ti = (E[Nc ] + 1)E[Nf r ]Ps (PHYht +E[Na ]Ut +δ+SIFS+ACKt +δ) | {z } NAV data time where, Ut = MDt +MACht +Pt

522

Qi Zhang et al.

27.3 The One4all Strategy As described above, packet aggregation has three drawbacks. (i) the achievable throughput depends on the arrival traffic pattern. It is good for bursty traffic but not for smooth traffic. (ii) long channel access delay in packet aggregation might lead to unfair media usage between mobile devices if delay sensitive service such as VoIP does not have higher priority. (iii) longer channel access delay also causes higher energy consumption in the mobile device. To overcome these drawbacks of packet aggregation, we propose the one4all strategy for CSMA/CA. The proposed strategy is based on the cellular controlled peer-to-peer (CCP2P) network architecture [4]. The mobile devices in the network are assumed to have two air-interfaces: one for the cellular link and the other for the short–range link. The mobile device is capable to form a cluster with the mobile devices in its proximity by the short-range link. The idea of one4all strategy is that the mobile device cooperates with the other devices in its clusters and only one device in a cluster contends to access the channel instead of all of them contending to access the channel. The contending device also receives the block ACK and distributes the block ACK over the short–range link. The advantage of the proposed strategy is that: first the collision probability of the transmitted frames by the contending mobile devices is reduced and second the remaining devices in a cluster can access the channel free of contention. The contention duty and transmission sequence in the cluster can be maintained by a logical token ring topology. The signalling between the devices in a cluster is exchanged over the short–range link. The proposed strategy also has its drawback because the achievable throughput gain depends on the number of the cooperative mobile devices in one cluster. But it can be integrated with packet aggregation strategy to exploit the advantages of both strategies. The device contention and transmission procedure of packet aggregation can be seen in Figure 27.3. Figure 27.4 shows the integrating of the one4all scheme and packet aggregation.

Figure 27.4. Contention and transmission procedure in one4all scheme.

The throughput and channel access delay of the proposed strategy are analyzed in the following subsection. The additional notations are given in Table 27.3.

27 Cooperative Uplink Access

523

Table 27.3. Parameter list. Notation cm Tsc ηc pcb pc Psc

Meaning the number of the mobile devices in a cluster transmission time of a cooperative cluster throughput with one4all strategy the channel busy probability with one4all strategy collision probability of a transmitted frame with one4all strategy the probability of a success transmission with one4all strategy

27.3.1 Throughput & Channel Access Delay Analysis The throughput of the one4all strategy can be calculated by extending the model described in Subsection 27.2.1. We define one contention and transmission period as the duration that the representative mobile device starts to contend until the available packets in the clusters being transmitted. In the period each mobile device sends one packet and mobile device also has one packet available in the buffer immediately after the period. So the throughput is expressed by ηc =

cm E[Pt ]Ps E[Ψ ]σ + Psc Tsc + (1 − Psc )Tc

(27.22)

The related variables in the model can be given as following: n

pc = 1 − (1 − τ ) cm pcb = 1 − (1 − τ ) Psc =

n τ (1 cm

− τ)

−1

n cm

n cm

(27.23) (27.24)

−1

n

1 − (1 − τ ) cm

(27.25)

Tsc = RTSt +δ+SIFS+CTSt +δ+SIFS+cm Ut +δ+SIFS+ACKt +δ+DIFS (27.26) where, Ut = PHYht + MACht +Pt In the proposed strategy, only one device contends to access the channel in one contention and transmission period; and the remaining mobile devices access the channel free of contention. The mobile devices in the cluster alternately take the role as contending mobile device. So the average channel access delay per device is the channel access delay experienced by the contending mobile device averaged by cm . It is given by E[Dc ] =

E[Nc ](E[Db ] + Tc + To ) + E[Db ] cm

where, 1 −1 Psc E[Db ] = E[Cbk ]σ + E[Nf r ](Psc Tsc + (1 − Psc )Tc ) E[Nc ] =

(27.27)

524

Qi Zhang et al.

27.3.2 Energy Consumption Analysis Like device working with other strategies, with one4all scheme the device also has four power levels. The difference is that when one representative mobile device contends the channel (i.e., sending RTS, receiving CTS, listening the channel states and backoff. etc.), the other mobile devices in the cluster are all in idle state. After the representative has caught the channel successfully, the remaining mobile devices alternately wake up to transmit their own frames. They switch to idle mode right away after completion of transmission. So the average energy consumption per packet is calculated by summing up all the energy consumption of the mobile devices in one cluster in one contention and transmission period; and then the sum is averaged by cm . Hence, the energy consumption for each packet can be given by Eng c =

Ptx Ttx + Prx Trx + Pli Tli + Pi Ti cm

(27.28)

where, Ttx =

RTSt + cm (Ht + Pt ) + E[Nc ]RTSt | {z } | {z } successful transmission time transmitted RTS collides

Trx = (E[Nc ]+1)E[Nf r ](Psc (RTSt +CTSt )+(1−Psc )RTSt )+CTSt +ACKt | {z } other’s RTS&CTS Tli = E[Nc ](Tlibk + To +δ+DIFS) + Tlibk where, by

Tlibk

is the mobile device’s listening time during a backoff delay and is given

Tlibk = E[Cbk ]σ+E[Nf r ](Psc (2δ+2SIFS+DIFS)+(1−Psc)(δ+DIFS)) Ti = (E[Nc ]+1)E[Nf r ]Psc (cm (Ht +Pt +δ+SIFS)+ACKt +δ) + | {z } contending MD’s NAV data time (cm −1)(Ht +Pt)+cm(SIFS+δ)+(cm −1)((cm −1)(Ht +Pt)+cm(SIFS+δ)+ACK) | {z } | {z } callout1

• •

callout2

callout1 the contending mobile device’s idle time during other mobile devices of the cluster transmitting callout2 the sum of the non-contending mobile devices’ idle time during other mobile devices of the cluster transmitting

Here Ttx includes the time of the representative sending RTS, its own frame and the time of the remaining mobile devices sending their frames. Trx includes the time of the representative receiving CTS and block ACK for the cluster, the RTSs and CTSs of the other clusters. Tli is the time of the representative listening the channel. Ti is the sum of the time that all mobile devices are in idle states during one contention and transmission period. It is clearly seen in Figure 27.5.

27.4 Numerical Results To illustrate the proposed cooperative access strategy outperforming other existing MAC strategies, some numerical results are presented in this section. The assumption of the parameters is listed in Table 27.4.

27 Cooperative Uplink Access

525

Figure 27.5. Active and idle switching diagram of the mobile devices in one cluster. Table 27.4. Parameter list. Notation W m n P MACh (incl. FCS) PHYh RTS (incl. PHYh ) CTS (incl. PHYh ) Pt ... CTSt SIFS

Value 32 3 [4 – 60] 1023 34 16 20 14 P/R ... CTS/R 10

Unit Notation DIFS δ Namax byte cm bytes σ bytes R bytes Ptx bytes Prx us Pli us Pi

Value 50 1 4 4 20 11, 54 2 0.9 0.9 0.04

Unit us us

us Mbps W W W W

In this section we compare the performance of three different MAC strategies (i.e., the one4all strategy, packet aggregation strategy and conventional RTS/CTS strategy) in CSMA/CA. We focus on the performance of throughput, channel access delay and energy consumption. It should be mentioned here that these three strategies are not completely independent. For instance cooperative access strategy and packet aggregation strategy are both built on RTS/CTS scheme. More generically, cooperative access strategy can also be built on top of packet aggregation scheme, but in the example they are implemented independently. The throughput comparison of three different MAC strategies in CSMA/CA is shown in Figure 27.6. It is similar as the conclusion in [16] that with RTS/CTS scheme, the throughput is insensitive to the number of the mobile devices in the network. The difference is that when packet aggregation and one4all strategies are employed in CSMA/CA, both of them can greatly enhance the throughput. The one4all strategy slightly outperforms packet aggregation scheme. The more throughput gain can be achieved under the condition of the higher channel data rate. As for the channel access delay performance, we define the average per device channel access delay, which means the delay experienced by a mobile device starting to contend until it successfully catching the channel. We also define the average per packet channel access delay, which is the per device channel access delay averaged by the number of PDUs the mobile device(s) transmitting after the channel is obtained. So in conventional RTS/CTS and the one4all strategies, the average per packet channel access delay is same as per device channel access delay. The average per device channel access delay comparison is shown in Figure 27.7, from which

526

Qi Zhang et al.

Figure 27.6. Throughput comparison of different MAC strategies in CSMA/CA.

we can see the average per device channel access delay of the packet aggregation scheme is much larger than the other two schemes. The reason is that it takes longer time for a device to transmit a big aggregated frame, which prolongs all the other devices’ channel access delay. Figure 27.8 shows the average per packet channel access delay comparison. It shows that the average per packet channel access delay of packet aggregation scheme is a little shorter than conventional RTS/CTS scheme, but it is much longer than one4all scheme. Furthermore, the channel access delay increases with an increasing number of devices. The longer channel access delay has great impact on fairness of media usage between different mobile devices, if they have different types applications. It might also affect the QoS of the delay sensitive services such as VoIP, if there is no priority differentiation between applications. The energy consumption performance comparison the three strategies is shown in Figure 27.9. Because the conventional RTS/CTS and packet aggregation strategies have much longer channel access delay, more energy is wasted in the channel contention duration, even though we have assumed both of these two schemes have smart energy saving scheme. The energy comparison clearly shows that in case the number of mobile devices in the network exceeds 30, the conventional RTS/CTS consumes as over three times energy as the one4all scheme does. Packet aggregation strategy uses approximately double the energy as the one4all cooperative scheme.

27.5 Conclusions Based on the CCP2P network architecture, we propose one4all strategy for MAC protocol design in WLANs. The proposed scheme significantly enhances the throughput compared with conventional RTS/CTS scheme and packet aggregation scheme. It overcomes the drawback of packet aggregation where the throughput performance

27 Cooperative Uplink Access

Figure 27.7. Average per mobile device channel access delay comparison.

Figure 27.8. Average per packet channel access delay comparison.

527

528

Qi Zhang et al.

Figure 27.9. Average energy consumption for one packet successful transmission. is sensitive to the arrival traffic pattern. Hence it is suitable for both video streaming and VoIP traffic. Furthermore, it outperforms the packet aggregation scheme in terms of channel access delay and energy consumption. The proposed scheme can also be built on top of the packet aggregation scheme to enhance the throughput in case there are not enough mobile devices in the cluster. The one4all strategy is a good illustration underlining the great potential of cooperation among mobile devices to solve the low uplink throughput issue in wireless network.

References 1. IEEE 802.11n WWiSE. Wwise proposal: High throughput extension to the 802.11 standard. Technical report, Jan 2005. 2. G. Bianchi. Ieee 802.11-saturation throughput analysis. IEEE Communications Letters, 2(12):318–320, 1998. 3. Enhanced Wireless Consortium. Ht mac specification. Technical report, Enhanced Wireless Consortium, Jan 2006. 4. Frank H.P. Fitzek and Marcos D.Katz, editors. Cooperation in Wireless Networks: Principle and Applications. ISBN-10 1-4020-4710-X. Springer, 2006. 5. Matthew S. Gast. 802.11 Wireless Networks: The Definitive Guide. ISBN-10: 0596100523. O’Reilly Media, 2 edition edition, April 25, 2005. 6. Seongkwan Kim, Youngsoo Kim, Sunghyun Choi, Kyunghun Jang, and JinBongChang. A high-throughput mac strategy for next-generation wlans. World of Wireless Mobile and Multimedia Networks, 2005. WoWMoM 2005. Sixth IEEE International Symposium on a, pages 278–285, 2005. 7. Youngsoo Kim, Sunghyun Choi, Kyunghun Jang, and Hyosun Hwang. Throughput enhancement of ieee 802.11 wlan via frame aggregation. Vehicular Technology Conference, 2004. VTC2004-Fall. 2004 IEEE 60th, 4:3030–3034, 2004.

27 Cooperative Uplink Access

529

8. Dzmitry Kliazovich and Fabrizio Granelli. On packet concatenation with qos support for wireless local area networks. ICC 2005 - 2005 IEEE International Conference on Communications and IEEE International Conference on Communications, 2:1395–1399, 2005. 9. Tianji Li, Qiang Ni, D. Malone, D. Leith, Yang Xiao, and T. Turletti. A new mac scheme for very high-speed wlans. World of Wireless, Mobile and Multimedia Networks, 2006. WoWMoM 2006. International Symposium on a, page 10 pp., 2006. 10. Yuxia Lin and Vincent W. S. Wong. Frame aggregation and optimal frame size adaptation for ieee 802.11n wlans. In GlobeCOM, San Francisco, Nov 2006. 11. IEEE 802.11 n TGn Sync. Tgn sync proposal technical specification. Technical report, May 2005. 12. Y. Nagai, A. Fujimura, Y. Shirokura, Y. Isota, H. Ishizu, F.and Nakase, S. Kameda, H. Oguma, and K. Tsubouchi. 324mbps wlan equipment with mac frame aggregation. Personal, Indoor and Mobile Radio Communications, 2006 IEEE 17th International Symposium on, pages 1–5, 2006. 13. Shao-Cheng Wang and Ahmed Helmy. Performance limits and analysis of contention-based ieee 802.11 mac. In 31st IEEE Conference on Local Computer Networks (LCN), Tampa, Florida, U.S.A., November 2006. 14. Yang Xiao. Efficient mac strategies for the ieee 802.11n wireless lans. Wireless Communications and Mobile Computing, 6(4):453–466, 2006. 15. Yang Xiao and J. Rosdahl. Throughput and delay limits of ieee 802.11. IEEE Communications Letters, 6(8):355–357, 2002. 16. Eustathia Ziouva and Theodore Antonakopoulos. Csma/ca performance under high traffic conditions: throughput and delayanalysis. Computer Communications, 25(3):313–321, 2002.

Part V

Methodologies and Tools

28 Cooperation for Cognitive Networks: A Game Theoretic Perspective Cristina Comaniciu Stevens Institute of Technology [email protected]

Summary. The next generation of wireless networks is evolving towards networks of small, smart devices, which opportunistically share the spectrum with minimal coordination and infrastructure. This evolution is motivated by the technological advances in software defined radio (SDR) technology which promote high adaptability and flexibility for radio transmissions. The new emerging generation of networks consists of cognitive terminals, which intelligently and autonomously adapt to the channel environment to optimize their transmission parameters. In this chapter, we address the problem of designing and analyzing such networks, in the context of improving the overall network performance through cooperation. We discuss how the protocols run by these networks can be modeled as games, and what it takes to enforce cooperation in such scenarios. Our perspective is to take a fresh look at the design of upper layer protocols, and the interactions between them, in the context of enforcing cooperation for more efficient spectrum sharing. The questions we will address are related to the necessary protocol modifications to support cooperation, as well as to the viability of such cooperative network protocols.

28.1 Future Generation of Wireless Networks: Opportunities and Challenges Wireless networks are evolving towards networks of small, smart devices, which opportunistically share the spectrum with minimal coordination and infrastructure. This evolution is motivated by the technological advances in software defined radio (SDR) technology which promote high adaptability and flexibility for radio transmissions. The new generation of smart terminals, the cognitive radios, build on the latest SDR capabilities and add intelligence to the devices. This intelligence is incorporated into a cognition cycle, which allows the terminal to gather information about its environment (“learn”) and make decisions (“act”) regarding their transmission parameters and possible access strategy. The evolution towards smart networks is further motivated by the new paradigm shift in the FCC’s spectrum management policy [13, 14] which promotes opportunistic spectrum sharing to balance the spectrum utilization, in both licensed and unlicensed bands.

533 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 533–554. c 2007 Springer.

534

Cristina Comaniciu

The performance of such networks is ultimately limited by interference, and therefore efficient and adaptive interference management techniques, such as power control, access control and, for end-to-end connectivity, routing, become key elements in the network design. An added difficulty arises because terminals with possible heterogeneous capabilities and requirements, independently measure the channels, and independently take decisions to maximize their own benefit. Such actions affect not only their performance but also that of the entire network (or at least their local neighborhood). What kind of performance and network behavior can result from the above scenario? Can the overall performance be improved by enforcing a certain level of cooperation among the network nodes? How can this cooperation be achieved: pricing, reputation schemes, etiquette? Several previous works have shown that cooperation in wireless networks can improve the performance by exploiting some form of multiuser diversity. For example, capacity studies for ad hoc networks have determined that the throughput per node in a network with N terminals can be improved from the initial pessimistic result of √ O(1/ N ) in [23] to O(1) in [24], for a particular example of a specific ad hoc network configuration with a multiple transmit-receive antenna architecture. In some other related references [7, 21, 22, 46], the mobility of the nodes was exploited to improve the capacity at the expense of very large to moderate transmission delays. A form of receiver cooperation through multiuser detection was also shown to improve spectral efficiency in a CDMA ad hoc network in our previous work in [11]. An increasingly large body of literature is focused on quantifying the benefits of transmitter and receiver cooperation (e.g., [5, 6, 25, 27, 28, 31, 42]), or network coding [1, 52]. While sufficient evidence exists that various forms of cooperation may improve the network performance, the research focus on developing higher layer distributed protocols for enabling and exploiting cooperation is very recent, and opens new perspectives for network design. In this book chapter we will discuss various aspects of cooperation for network protocols design, starting from modeling, analysis, and design examples that illustrate potential performance gains versus complexity tradeoffs. We will also discuss future opportunities and challenges in this emerging area of research.

28.2 A Game Theoretic Framework for Cooperation 28.2.1 Modeling the Cognition Cycle Starting with the general model of a cognitive radio network, we see that this network is comprised of smart terminals capable of measuring and analyzing their environment, and making decisions in terms of their transmission parameters to maximize their own defined quality of service (QoS) parameters. Using game theory terminology [18], these generic QoS metrics are called utilities, and can include any combination of performance measures, such as throughput, delay, energy, interference related measures (e.g., SIR - signal to interference ratio), etc. The cognitive radio network can be naturally modeled using a game theoretic framework, in which the players of the game are the radio terminals, their actions are their choices of transmission parameters (e.g., transmission powers, access probability, backoff interval, relaying nodes), and their utilities are their defined performance measure that each radio is trying to maximize. The players (radios) choose their actions independently, but their choice impacts on all the users in the network. The

28 Game Theory and Cooperation for Cognitive Networks

535

players are also assumed to be rational, i.e., they act in their best interest, to maximize their own utility. We illustrate the correspondence between the cognition cycle proposed by Mitola in [40] and a game theoretic based modeling of this process in Figure 28.1 (see also [41]). To analyze the cooperative play of such users, two different avenues can be pursued. The cooperation viability can be analyzed by modeling the game using coalitional game theory, or cooperative protocols can be designed based on a non-cooperative game framework, for which the selfish utility of each player is modified by a pricing function, such that the player acts cooperatively.

Figure 28.1. A Game Theoretic Model for the Cognition Cycle.

28.2.2 Coalitional Game Theory For coalitional games, users cooperate by forming coalitions [44]. We consider a class of coalitional games for which every coalition can be described by a single number, interpreted as the payoff available to the coalition. The share of the payoff received by players in a coalition is called a payoff vector, which can be arbitrarily shared among the members of a coalition (transferable payoff) or might be strictly determined by the nature of the game (non-transferable payoff). In analyzing if a certain upper layer protocol with cooperation is viable, we essentially need to determine if the grand coalition (the coalition formed by all the users in the network) is stable, and this can be achieved using the notion of core. As

536

Cristina Comaniciu

we will exemplify using a coalitional game framework in the next section as games with non-transferable payoff, we give the definition of the core for these games. We note that a similar definition exists for games with transferable payoff. Definition: The core of a non-transferable payoff game is the set of all payoff profiles for players, for which there is no coalition S ∗ ⊂ S that can achieve a higher payoff vector than S: Ri (S ∗ ) > Ri (S), ∀i ∈ S ∗ . Another important measure to characterize the value of each player in a coalition is the Shapley value. The model to derive the Shapley value is that, as a player enters the coalition, he receives the amount by which his entry increases the value of the coalition he enters. The amount a player receives by this scheme depends on the order in which the players are entered. The Shapley value represents the average payoff to the players if the players are entered in completely random order. Theorem: The Shapley value is given by Φ = (φ1 , ..., φn ), φi (v) =

X S⊂N, i∈S

(|S| − 1)!(n − |S|)! [v(S) − v(S − i)], n!

(28.1)

where the summation is over all coalitions S that contain user i. The quantity, v(S) − v(S − i), is the amount by which the value of coalition S − i increases when player i joins it.

28.2.3 Non-Cooperative Games In a non-cooperative game framework, users act selfishly such as their actions maximize their own utilities. The game behavior is characterized by the Nash equilibrium, which is defined as a set of strategies for players, such that no player has incentive to unilaterally change its action. The existence of a Nash equilibrium for a game is related to mathematical properties of the utility function. As such, there are known classes of games, characterized by certain mathematical properties of the utility functions, which can be shown to have a pure strategy Nash equilibrium (e.g. submodular and supermodular games, potential games, some class of congestion games, etc.). Every finite strategic form game has a mixed strategy equilibrium [18]. As opposed to a pure strategy Nash equilibrium, a mixed strategy equilibrium is characterized by an equilibrium probability distribution for each player, over its pure strategies (possible actions). The concept of Nash equilibrium is especially important to show convergence for distributed iterative algorithms, in which users learn the equilibrium point by using a best (or better) response strategy at each step of the game. However, the convergence to a Nash equilibrium will not guarantee the best operating point for a network. The concept of Pareto efficiency brings us one step closer in characterizing the equilibrium quality. By definition, a strategy profile is Pareto optimal if some players must be hurt in order to improve the payoff of other players. The Nash equilibrium convergence point of an algorithm can be changed by appropriately modifying the utility function. Theoretically, any network behavior can be enforced by appropriately modifying the utilities for users. A possible limitation for the above described games is the requirement that users have complete knowledge of the game. This assumption can be relaxed by defining games with imperfect and incomplete information (e.g., Bayesian games) [18]. Furthermore, almost “model free” games can be designed by using learning

28 Game Theory and Cooperation for Cognitive Networks

537

techniques (e.g., fictitious play, minimum regret learning, etc.). Minimum regret learning algorithms [16, 17, 19, 20] have been very recently shown to be able to learn mixed strategy equilibria in games, and also to converge to a pure strategy Nash equilibrium if one exists for the given game [20]. Applying learning techniques for games has also the advantage that games with heterogeneous users (having different utility functions) can also be analyzed. Furthermore, they also provide alternative performance measures, such as minimum regret, in conjunction with the classic Nash equilibrium concept. These new measures may be more valuable for certain classes of highly dynamic games which cannot reach convergence before the game changes (due to dynamic traffic and mobility).

28.3 Cooperative Protocols for Cognitive Networks: Some Examples In the previous section, we have discussed a game theoretic framework that can be used for protocol design and analysis for cooperative cognitive networks. In this section, we discuss the current state of the art in the literature for cooperative protocols, and we present a couple of design examples for various layers of the protocol stack: physical layer, MAC layer and network layer.

28.3.1 Implicit Cooperation in Protocol Design Network protocols in general can be seen as exhibiting some form of implicit cooperation, that should lead to fairness and good performance for the entire network [15]. This implicit (or passive) cooperation is usually transparent for the users, which simply follow a protocol/policy for using the network. However, when talking about networks of cognitive radios, we expect that a certain protocol (or policy) will be viable if all the users will have incentives to follow it. The reason for this, is that every user acts to maximize its own QoS metrics (utilities), and sometimes that might be detrimental for the overall network performance. As we have mentioned before, the viability of using a certain protocol in the network can be proved by showing that no user has incentives to cheat and cannot do better by forming small coalitions within the network. The earliest work on this aspect has discussed the viability of multi-hop ad hoc network routing protocols [45, 48–50]. More recently, protocol viability analysis in the context of rate allocation for multi-access Gaussian channels, and for coalitions involving receivers cooperation have been proposed in [30] and [38], respectively. To illustrate the concepts involved in proving the viability of a network, we discuss in a little more detail the simple example of coalitional games in linear multiuser detectors proposed in [38]. Multiuser detection has been shown to increase the performance of users, by jointly decoding their signals from noise and interference [55]. In the model proposed in [38] the authors take a fresh look at this problem, by investigating if using linear multiuser detectors (MUDs) is a viable solution for the network. To show the viability of the MUD solution, the authors consider using linear multiuser detectors for forming coalitions between users in a MAC. The coalitions are formed among the receivers in order to maximize their payoffs. It is assumed

538

Cristina Comaniciu

that a communication link is available for receivers to communicate to each other. Two cases are considered: the decorrelating receiver and the MMSE receiver, and the performance measure is the achieved SIR. The network model is a MAC channel with M users communicating with one base station (BS) in a BPSK modulated, synchronized CDMA system with no power control. The received signal at each users is filtered using a matched filter matched to the desired signature sequence for the user. The outputs of the matched filters for users participating in coalitions can be combined (filtered) to obtain the decoded signals. The filter further applied is L = R−1 for the decorrelator receiver, and L = (R + σ 2 A−2 )−2 for the MMSE receiver, where R is the cross-correlation matrix for the signature sequences of the users, A is a diagonal matrix containing the received amplitudes for all users, and σ 2 is the noise power. To show that the MUD receivers scenarios are viable, the authors in [38] show that the grand coalition1 is stable and sum-rate maximizing, i.e., no user has incentive to break from the coalition and decode by itself. We illustrate the steps of such a proof for the simplest case of decorrelating receiver. A property of the decorrelating receiver is that is nullifying all interference from jointly decoded users (users that participate in the coalition), at the price of enhancing the noise. The interference from users outside the coalition can be seen as noise. The SIR (payoff) of a user participating in a decorrelating detector coalition G can be expressed as: xi (G) =

Pi σ 2 1+ρ(|G|−2) 1−ρ 1+ρ(|G|−1)

+

h

ρ 1+ρ(|G|−1)

i2 P

j∈Gc

,

(28.2)

Pj

where Pi is the received power of user i, ρ is a cross-correlation coefficient for signature sequences between users inside and outside the coalition, G is the set of users participating in the coalition, and Gc is the set of users not in the coalition. The SIR achieved by the grand coalition can be shown to be (when G = M ): xi (M ) =

Pi σ 2 1+ρ(|M |−2) 1−ρ 1+ρ(|M |−1)

.

(28.3)

For the grand coalition to be stable, we need to prove that the payoff of the grand coalition is always higher. This is proven in [38] for the high SIR regime: lim xi (G) < lim xi (M ).

σ 2 →0

σ 2 →0

(28.4)

From the above relation, we see that, in the high SNR regime, all users achieve the highest payoff in the grand coalition and therefore have no incentive to leave the coalition. Furthermore, since each user achieves its greatest payoff in the grand coalition, it follows that the sum of the rates achieved is the greatest among all coalition structures, given the property that maximizing the SIR leads to maximizing the rate. As a word of caution, we note that there is no guarantee that the grand coalition of users should form or that the stable coalition structure should be the one that maximizes the sum-rate. In [38], the authors give several examples of possible stable coalitions, and show that the only stable coalition that maximizes the sum-rate is the grand coalition. 1

The coalition formed by all the users in the network.

28 Game Theory and Cooperation for Cognitive Networks

539

28.3.2 Incentivizing Cooperation for Non-Cooperative Games In the previous section we have discussed the viability of certain protocols that implicitly cooperate to achieve a common goal. In this section, we focus on discussing cooperative frameworks for protocol design at various layers of the protocol stack. The most common tool to establish cooperation in a network of otherwise selfish users is to provide incentives for users that change their optimization goals and thus modify their behavior. The most commonly used incentives are pricing and reputation. In this context, most of the existing literature has focused on incentivizing forwarding in ad hoc networks with selfish nodes, using both pricing (see for example [9, 26, 47, 58]), and reputation/punishment schemes in repeated game models (e.g., [8, 37, 39]). Another strong research focus in this area is on designing efficient pricing schemes for power control in wireless networks (see for example [10, 12, 53, 54]). The MAC layer is less represented in the body of research, with a few cooperative schemes being proposed, most of them focusing on designing implicit cooperation protocols that support cooperation at the physical layer (e.g., MAC algorithms for cooperative diversity support [32, 33, 57, 60]). In what follows, we will focus on presenting a framework for cooperation in cognitive networks, while detailing some example designs for physical, MAC and network level protocols. The model that we have in mind for the cognitive networks is that they are formed by independent terminals, with a “mind of their own”, which independently adjust their transmission parameters based on their own perceived QoS metrics (measured on the channels). This scenario leads to distributed resource management implementations. We assume that the terminals have incentives to be selfish because of the scarcity of resources (e.g., bandwidth and energy). In this context, there are three questions that we are addressing: • • •

How does a network of selfish users behave? Is there any gain from cooperation? Can users be persuaded to cooperate?

We will address the above questions for three different design scenarios: cooperative distributed MAC protocols, distributed channel allocation for bandwidth sharing, and interference aware routing for near-far effect mitigation. Cooperative Distributed MAC Protocols: Slotted Aloha Game theoretic formulations for analyzing MAC protocols [2], [3], and in particular slotted Aloha, were recently proposed in the literature, including the work in [34], its extension for multipacket reception models in [35], and a pricing strategy in [59] for an Aloha network of heterogeneous users with inelastic bandwidth requirements. Of particular interest is the work in [34, 35] which shows that a distributed Aloha based MAC protocol for selfish users is viable and stable, contrary to the previous belief that selfish users running their own MAC strategies could lead to protocol failures by constantly colliding in an attempt to maximize their individual throughput. The approach in [34, 35] was to model a transmission cost c for the users and to show that this transmission cost influences the equilibrium for the network. It

540

Cristina Comaniciu

was shown that the achievable throughput for this network of selfish users is less than the one obtained by the classic centrally controlled Aloha protocol, and it depends on the value of c. Following the work in [34, 35], we have recently proposed a differentiated pricing scheme for users, in order to modify their behavior and thus to change the network equilibrium throughput to the optimal one [56]. The design steps for incentivizing cooperation are as follows: 1. Introduce pricing function for access in the non-cooperative game model. According to the utility/price model described above, the Aloha matrix for costs for a player i can be summarized in the following table.

Table 28.1. Payoff functions. T and S T but F W 1 − ci − µi −1 + µi −(1 − ci − µi )

In the above table, T = transmit, W = wait, S = successful transmission, F = failed transmission (collision), ci = transmission cost for user i (e.g., percentage of energy spent), and µi = price charged by the network for user i for a successful packet transmission. 2. Determine the value of the price for optimal throughput and revenue. If the optimal throughput and revenue can be determined for the centrally controlled slotted Aloha, then a condition can be imposed to select the access probability for each user p to be equal to the value that maximizes the throughput or the revenue (in case of revenue optimization). We note that the design imposes p to be the same for all users in order to enforce fairness in the network. We illustrate the derivation for the simplest case of throughput optimization for collision models. Detailed derivations for revenue optimization and for multipacket reception (MPR) models can be found in [56]. It is well known that a centrally controlled slotted Aloha algorithm achieves its maximum throughput for p = 1/N , where N is the number of users accessing the system. Imposing that the same throughput should be achieved in the cooperative network, we have the condition: 1 . (28.5) N However, the above value should also be an equilibrium for the access game. The equilibrium probability expression p can be obtained by using the “indifference principle”, i.e., determine p such that a user is indifferent between transmitting and waiting at a given one shot game. The access probability, can be thus determined to be p=

(−1 + µi ) 1 − (1 − p)N −1 + (1 − ci − µi )(1 − p)N −1 = −(1 − ci − µi ) ⇒

28 Game Theory and Cooperation for Cognitive Networks ⇒p=1−

p N −1

ci /(2 − 2µi − ci ).

541 (28.6)

By equating (28.5) and (28.6), we can determine the optimal pricing strategy that leads to the same optimal throughput as the centrally controlled slotted Aloha: µi = 1 −

(1 − 1/N )N −1 + 1 ci . 2(1 − 1/N )N −1

(28.7)

We illustrate the design tradeoffs for the cooperative slotted Aloha framework by plotting the achievable throughput and revenue comparisons for both the collision and the MPR models (Figures 28.2, 28.3, 28.4 and 28.5). In the figures, β represents the capture parameter that characterizes the MPR channels. A smaller β means a better capture capability for packets, with β → ∞ being the case of collision channels. Also, the total cost c represents the sum of all transmission costs for all users in the system. We can see that the pricing scheme can be determined by optimizing either the achievable throughput, or the achievable revenue (a more realistic measure for the service provider). However, for the collision model, the revenue given by the two pricing schemes is identical, which suggests that optimizing the throughput is better. For the MPR model case, this is no longer true, as throughput and revenue based pricing optimization lead to different results.

Figure 28.2. Throughput comparison: collision models.

Distributed Channel Allocation for Bandwidth Sharing A classic problem in cognitive radio networks is the distributed channel allocation scenario: users measure the available spectrum and dynamically decide which

542

Cristina Comaniciu

Figure 28.3. Throughput comparison: MPR models.

Figure 28.4. Revenue comparison: collision models.

28 Game Theory and Cooperation for Cognitive Networks

543

Figure 28.5. Revenue comparison: MPR models.

frequency should they use for transmission, based on their current measurements. The channel measurements that the users perform is usually related to their desired QoS metrics. In this example, we consider that there are N users in the system, all active, which are sharing K available channels, with K < N . As their quality of service metric, users require the highest possible throughput, which is related to the achievable SIR for the link via adaptive modulation and/or adaptive coding. Consequently, the users will be searching for channels that will give them the highest possible SIR, in an attempt to build the best frequency reuse maps in a distributed fashion. Once the channel is selected, the maximum achievable throughput is used, given a target SIR constraint (BER constraint). The distributed channel allocation problem can be modeled as a game as follows [43]: • • •

players: users; actions: selection of one of the K channels; utility: related to the achievable SIR, U 1i (si , s−i ) = −

N X

pj Gij f (sj , si ).

(28.8)

j6=i,j=1

∀i = 1, 2, ..., N. For the above definition, we denoted P=[p1 ,p2 ,...,pN ] as the transmission powers for the N radios, S=[s1 ,s2 ,...,sN ] as the strategy profile for the users, Gij as the link gain between transmitter i and receiver j, and f (si , sj ) as an interference function:   1 if sj = si , transmitter j and i choose the same strategy (same channel) f (si , sj ) =  0 otherwise

544

Cristina Comaniciu The above utility function (U1) accounts for the case of a “selfish” user, which values a channel based on the level of interference perceived on that particular channel. To enforce cooperation, a second term must be added to the utility definition, in the form of a pricing function. The new utility, U 2 accounts for the interference seen by a user on a particular channel, as well as for the interference this particular choice will create to neighboring nodes. Mathematically we can define U 2 as: U 2i (si , s−i ) =

−

N X

pj Gij f (sj , si ) −

j6=i,j=1

N X

pi Gji f (si , sj )

(28.9)

j6=i,j=1

∀i = 1, 2, ..., N Comparing the above two utility functions, we can see that U 2 requires more information to be gathered at the terminal, which entails a protocol implementation based on information exchange among terminals. This will increase the implementation complexity and the required overhead. On the other hand, we conjecture that by accounting for the interference users create to their neighbors, a fair channel allocation pattern can be achieved. We will verify this shortly via simulation results. Based on the utilities defined in (28.8) and (28.9), the game consists of users selecting channels to minimize their utility2 . As for designing any distributed algorithm, the issue of convergence is very important, otherwise the users will continuously change channels in an attempt to improve performance. Is the algorithm converging, and how fast is the convergence? Is the equilibrium achieved a desirable operating point? The game theoretic formulation can easily answer the first question based on mathematical properties of the utility function. As such it can be shown that, due to the symmetry property of U 2, the game can be formulated as a potential game, which it is known to yield a pure strategy Nash equilibrium, when the users update their choices sequentially. A more general formulation can also be used, by designing the distributed channel allocation algorithm as a minimum regret learning game, which is shown to converge in general to a mixed strategy Nash equilibrium, and to a pure strategy Nash equilibrium, if one exists. Since the selfish users formulation does not have a symmetric utility function, this more general formulation needs to be used for this case, and the algorithm converges to a mixed strategy equilibrium. The speed of convergence can only be verified by simulations. Furthermore, the game theoretic framework will not guarantee that the achieved equilibrium point is unique or optimal. Additional modifications to the utility function (e.g., introducing or modifying the pricing function) may change the equilibrium point for the system. We compare the performance results of the two schemes (cooperative and noncooperative) by means of simulation results. We have simulated a fixed wireless ad hoc network for which N = 30 transmitters and their receivers are randomly distributed over a 200 m × 200 m square area, and share K = 4 available channels. A random channel assignment (RND) is selected as the initial assignment and, for a fair comparison, all simulations start from the same initial channel allocation. 2

Here the utility has a meaning of cost.

28 Game Theory and Cooperation for Cognitive Networks

545

Here we present a comparison for the selfish and unselfish behavior based on three different metrics: total average throughput in the network, average throughput per user, and variance of achievable throughput per user (Figure 28.6). The variance performance measure quantifies the fairness, with the fairest scheme achieving the lowest variance. We can see that cooperation improves the overall performance of the network, mostly in terms of fairness. As the average achieved throughput is not significantly changed for the three tested scenarios, we can infer that some of the users that had a very good performance in the selfish case will see a somewhat reduced throughput when cooperating. Thus, strong incentives should be provided for these users to convince them to alter their behavior. We can envision that the added term in the utility function U 2 may represent a monetary penalty for selfish behavior (e.g., the price charged for a user may go up proportional to the interference it creates to its neighbors). Another interesting observation is that the convergence point achieved depends also on the adaptation algorithm selected. Even with the same initialization point, and using the same utility function definition (U 2), the mimimum regret learning algorithm and the potential game formulation lead to different pure strategy Nash equilibrium points, reflected in a better performance for the potential game formulation. Consequently, when designing resource allocation games, the mathematical properties of the game play a significant role in the achieved performance. Interference Aware Routing: A Cooperative Approach Another example to illustrate the benefits of cooperation is interference aware routing at the network layer. Interference aware routing can also be seen as a resource management tool, as pairs of source-destination nodes are selecting routes to better distribute the interference in the network. The first question that arises is: Is routing selfish? Traditionally, routing has been seen as involving a degree of selfishness (or rather as requiring a degree of cooperation), since relaying nodes may not be willing to cooperate in forwarding, in order to preserve their own resources. As we have mentioned before, this problem have led to a rich literature on incentivizing forwarding for multi-hop networks. Here we look at a different form of selfishness, in the context of interference aware routing protocols. For the general approach, the routes are optimized based on a cost per route, which is related to the interference seen on the participating links. However, the routes are cross-coupled via interference, and by selecting a particular route, the interference seen by the others significantly changes. The idea of cooperation in this case arises from selecting individual routes by somehow accounting also for the interference created to others. To illustrate this cooperative scenario, we follow the approach in [36], in which every relaying node is considered as a facility that can be shared by multiple flows3 . When establishing a route, the source node considers a collection of facilities (relaying nodes), such that the cost per route is minimized. The cost per route is the cost 3

A flow is the traffic exchanged by a source-destination pairs of nodes, on a given route path.

546

Cristina Comaniciu Total Average Throughput overall Users 18

Average Throughput per User 0.7

Variance of the Throughput per User 0.12

16 0.6

0.1

14 0.5 0.08

12 0.4

10

0.06 8

0.3

6

0.04 0.2

4 0.02

0.1 2

0

U1 U2 POT RND

0

U1 U2 POT RND

0

U1 U2 POT RND

Figure 28.6. Total Average Throughput, Average Throughput per user and Variance Throughput per user (U1 = selfish users with minimum regret learning algorithm implementation; U2 = cooperative users with minimum regret learning algorithm implementation; POT = cooperative users with potential game implementation; RND = equal probability random channel allocation).

of all facilities enclosed in a route. The cost of one particular facility could be for example the interference measured at that facility for the case of selfish routing. To enforce cooperation, the cost of one facility should depend on both the interference seen by the current flow, as well as on an estimate of the interference impact of the relaying node on the other neighboring flows. Estimating the real impact a relaying node has on its neighboring receivers is a difficult task, as this continuously changes with establishing new routes. In [36], a very rough estimate for the interference impact is given, as being proportional with the density of nodes in a neighborhood. The rationale is that a relaying node is expected to have a higher interference impact in a more densely populated area, compared with a sparse region. The above model with shared facilities as relaying nodes is very suitable for a congestion game formulation of the routing problem. In a congestion game, there are a finite set of players which select facilities that are considered as common resources [29]. These players are decision makers and can act according to finite set of strategies, associated with each player [4, 51]. More specifically, they will select a set of facilities such as to maximize their overall utility. Characteristic for a congestion game model is that the cost/benefit associated with sharing each facility should depend only on the number of users, and not on their identity (anonymity condition), which means that the cost is unrelated

28 Game Theory and Cooperation for Cognitive Networks

547

to specific characteristics of users. The payoff of a player is the sum of the benefits associated with each facility in his strategic choice, given the choices of other players. Formally, a congestion game model is defined as: < N, F, (Xi )i∈N , (Wf )f ∈F >,

(28.10)

where N = nonempty, finite set of players, F = nonempty, finite set of facilities, and for each player i ∈ N , its collection of pure strategies Xi is a nonempty, finite family of subsets of F . For each facility f ∈ F , we can define wf : 1, · · · , n → R as the benefit of facility f , with wf (r), r ∈ 1, · · · , N equal to the benefit of each of the users of facility f , if there is a total of r users sharing that facility. For the above game we define the cost of a facility (the cost associated to using a node i for relaying) to be [36]: i wf =i (m) = Pmax mη,

(28.11) i Pmax

is the transmission where m is the number of flows sharing the facility (node i), power for a given flow (assumed equal for all flows) at node i, and η is the number of nodes in a neighborhood grid (for a given area grid, this approximates the nodes’ density). Modeling the routing game as a congestion game, has the advantage that convergence to a Nash equilibrium can be readily proved by showing that the game is isomorphic with a potential game. In [36], a cross-layer hierarchical power control and routing has been proposed for interference aware mitigation, and it was shown to converge to a Nash equilibrium, by following a three step convergence requirement: at the physical layer (convergence of the power control), at the network layer (convergence of routing game), and across layers (the joint algorithm should not lead to oscillatory behavior). A simplified flowchart for the algorithm can be seen in Figure 28.7.

Figure 28.7. Cross-layer interference mitigation algorithm.

The advantages of cooperation can be illustrated by simulation results. In [36], an ad hoc network with N nodes uniformly distributed in a square area of dimension

548

Cristina Comaniciu

100 × 100 m2 was simulated. Each relaying node was using multi-codes in a CDMA systems to simultaneously transmit multiple relayed flows. Two performance metrics were selected: energy per bit consumption and achievable throughput. Two different neighborhood delineations were considered: square grid and circular grid. To determine the square grids the entire network area should be divided into grids of equal sizes. The number of nodes within these areas determines the density of nodes in the grid. This approach is more costly, as it requires global knowledge on the network topology. By contrast, circular regions surrounding the transmitting nodes (approximating the interference range areas) can be easily estimated. In Figure 28.8 we show the energy gains for the proposed joint power control and routing algorithm (JPCR), compared with a classic energy aware routing scheme with fixed transmission powers. We infer that most of the energy gains come from the power control implementation.

Figure 28.8. Energy per bit consumption: JPCR versus minimum energy routing c (Reproduced with permission from [36] IEEE 2006). More importantly, in Figure 28.9 we see significant increase in the achievable throughput, when cooperation is enforced by modifying the routing metrics. As a tradeoff, the game theoretic cooperative scheme results in an iterative procedure which leads to higher overhead and higher implementation complexity. It is expected that the overhead and complexity increase proportional with the number of iterations required, which is relatively low (see Figure 28.10).

28.4 Conclusions and Open Issues The emergence of new generation of networks with smart terminals capable to adjust to their environments in an independent fashion, raises new research challenges

28 Game Theory and Cooperation for Cognitive Networks

549

Figure 28.9. Normalized throughput: JPCR versus minimum energy routing. (Rec produced with permission from [36] IEEE 2006).

c Figure 28.10. JPCR convergence. (Reproduced with permission from [36] IEEE 2006).

550

Cristina Comaniciu

related to the efficiency of spectrum reuse and fairness. At their very basic behavior, these cognitive terminals are essentially selfish, as they will try to maximize their own performance in the network. They can be persuaded to “cooperate” by either imposing a rigorous protocol/etiquette that they should abide to (passive cooperation), or by using incentivizing schemes that can be designed to influence their behavior. As we have discussed in this chapter, the behavior of such networks can be analyzed using a game theoretic framework. There are several main questions of interest that need to be resolved: • • •

What incentives have users to cooperate in cognitive networks (passively or actively)? How should we design these incentives to optimize the network performance? What are the tradeoffs involved?

While the game theoretic framework will provide the designer with the right analytical tool for devising convergent distributed resource management protocols for these networks, there are a lot of open issues unresolved regarding the quality of the operating point achieved, and there is a lot of room for creativity on how to drive these equilibrium points towards optimal solutions. Furthermore, the tradeoffs involved in the design usually highlight a higher complexity solution for the cooperative schemes. How to quantify the effective gains achieved by cooperation, and what is the best design: an optimal operating point, or a better tradeoff performance/complexity? These remain open issues for a rich and challenging new research area. Acknowledgement. This work was supported in part by the ONR, grant number: N00014-06-1-0063, and by the NSF, grant number: CNS-0435297.

References 1. R. Ahlswede, N. Cai anf S.-Y.R. Li, and R.W. Yeung. Network information flow. IEEE Transactions on Information Theory, IT-46, 2000. 2. E. Altman, R.E. Azouzi, and T. Jimenez. Slotted aloha as a game with partial information. Computer Networks: The International Journal of Computer and Telecommunications Networking, 45(6):701–713, August 2004. 3. E. Altman, D. Barman, R.E. Azouzi, and T. Jimenez. A game theoretic approach for delay minimization in slotted aloha. In IEEE International Conference on Communications (ICC), pages 3999–4003, June 2004. 4. V. Anantharam. On the nash dynamics of congestion games with player-specific utility. In Decision and Control,CDC, 43rd IEEE Conference on, pages 4673– 4678, December 2004. 5. Salman Avestimehr and David Tse. Outage capacity of the fading relay channel in the low SNR regime. IEEE Transactions on Information Theory, 2006. submitted. 6. K. Azarian, H.E. Gamal, and P. Schniter. On the achievable diversitymultiplexing tradeoff in half-duplex cooperative channels. Information Theory, IEEE Transactions on, 51(12):4152–4172, 2005.

28 Game Theory and Cooperation for Cognitive Networks

551

7. N. Bansal and Z. Liu. Capacity, delay and mobility in wireless ad hoc networks. In Proceedings of the IEEE INFOCOM, April-May 2003. 8. S. Buchegger and J.Y.L. Boudec. Performance analysis of the CONFIDANT protocol (cooperation of nodes: Fairness in dynamic ad-hoc networks). In ACM MobiHoc, 2002. 9. L. Buttayan and J-P. Hubaux. Nuglets: a virtual currency to stimulate cooperation in self-organized mobile ad hoc networks. Technical report, Institute for Computer Communications and Applications, Department of Communication Systems, Swiss Federal Institute of Technology, January 2001. 10. D. Goodman C. Saraydar, N. Mandayam. Efficient power control via pricing in wireless data networks. IEEE Trans. Communications, 50(2), February 2002. 11. C. Comaniciu and H. V. Poor. On the capacity of mobile ad hoc networks with delay constraints. IEEE Transactions on Wireless Communications, 2006. to appear. 12. D. Famolari, N.B. Mandayam, D.J Goodman, and V. Shah. Wireless Multimedia Network Technologies, chapter A new framework for power control in wireless data networks: games, utility and pricing, pages 289–310. Kluwer Academic Publishers, 1999. 13. FCC. Notice of proposed rule making and order (NPRM&O), ET Docket No. 03108. “http://hraunfoss.fcc.gov/edocs public/attachmatch/FCC-03-322A1.doc”, December 2003. 14. FCC. Secondary Markets Report and Order and FNPRM, FCC 03-113, October 2003. 15. F.H.P. Fitzek and M. Katz, editors. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! Springer, April 2006. 16. Y. Freund and R. Schapire. A decision-theoretic generalization of online learning and an application to boosting. In Computational Learning Theory: Proceedings of the Second European Conference, Springer-Verlag, pages 23–27, 1995. 17. D. Fudenberg and D. Levine. The Theory of Learning in Games. MIT Press, 1998. 18. Drew Fudenberg and Jean Tirole. Game Theory. MIT Press, 1991. 19. A. Greenwald and A. Jafari. A general class of no-regret algorithms and gametheoretic equilibria. In Proceedings of the 2003 Computational Learning Theory Conference, pages 1–11, August 2003. 20. A. Greenwald, A. Jafari, D. Gondek, and G. Ercal. On no-regret learning, fictitious play, and nash equilibrium. In Proceedings of the Eighteenth International Conference on Machine Learning, June 2001. 21. M. Grossglauser and D. Tse. Mobility increases the capacity of ad-hoc wireless networks. IEEE/ACM Transactions on Networking, 10(4):477–486, August 2002. 22. N. Gupta and S. R. Das. A capacity and utilization study of mobile ad hoc networks. In Proceedings of the 26th Annual Conference on Local Computer Networks, pages 576–583, November 2001. 23. P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Transactions on Information Theory, 46(2):388–404, March 2000. 24. P. Gupta and P.R. Kumar. Towards an information theory of large networks: an achievable rate region. IEEE Transactions on Information Theory, 49(8):1877– 1894, August 2003.

552

Cristina Comaniciu

25. Host-Madsen and A.J. Zhang. Capacity bounds and power allocation for wireless relay channels. IEEE Transactions on Information Theory, 51(6):2020–2040, June 2005. 26. O. Ileri, S-C. Mau, and N.B. Mandayam. Pricing for enabling forwarding in self-configurating ad hoc networks. IEEE Journal on Selected Areas in Communications, 23(1):151–162, January 2005. 27. N Jindal, U Mitra, and A Goldsmith. Capacity of ad-hoc networks with node cooperation. In Proc. IEEE Int. Symp. Inform. Theory, 2004. 28. G. Kramer, M. Gastpar, and P. Gupta. Cooperative strategies and capacity theorems for relay networks. IEEE Trans. Inform. Theory, 51:3037–3063, September 2005. 29. Nikolai S. Kukushkin. Congestion games revisited. Game Theory and Information, EconWPA, February 2006. 30. R. La and V. Anantharam. A game-theoretic look at the Gaussian multiaccess channel. In Proceedings of DIMACS workshop on Network Information Theory, pages 37–42, November 2003. 31. J. N. Laneman, D. Tse, and G. W. Wornell. Cooperative diversity in wireless networks: Efficient protocols and outage behavior. IEEE Transactions on Information Theory, 50(12):3062–3080, December 2004. 32. R. Lin and A.P. Petropulu. A new wireless network medium access protocol based on cooperation. IEEE Transactions on Signal Processing, 53(12):4675– 4684, December 2005. 33. P. Liu, Z. Tao, and S. Panwar. A cooperative MAC protocol for wireless local area networks. In Proceedings of IEEE International Conference on Communications, pages 2962–2968, May 2005. 34. A.B. MacKenzie and S.B. Wicker. Selfish users in Aloha: A game-theoretic approach. In IEEE Vehicular Technology Conference (VTC), pages 1354–1357, Atlantic City, NJ, October 2001. 35. A.B. MacKenzie and S.B. Wicker. Stability of multipacket slotted Aloha with selfish users and perfect information. In IEEE INFOCOM, pages 1583–1590, March-April 2003. 36. H. Mahmood and C. Comaniciu. A cross-layer game theoretic solution for interference mitigation in wireless ad hoc networks. In IEEE Military Communications Conference (MILCOM), October 2006. 37. S. Marti, T.J. Giuli, K. Lai, and M. Baker. Mitigating routing misbehavior in mobile ad hoc networks. In ACM MobiCom, 2000. 38. Suhas Mathur, Lalitha Sankaranarayanan, and Narayan B. Mandayam. Coalitional games in receiver cooperation for spectrum sharing. In Proceedings of CISS (Conference on Information Sciences and Systems), Princeton, NJ, 2006. 39. P. Michiardi and R. Molva. Core: A collaborative reputation mechanism to enforce node cooperation in mobile ad hoc networks. In Sixth IFIP conferece on security communications and multimedia, 2002. 40. J. Mitola and G.Q Maguire. Cognitive radio: making software radios more personal. IEEE Personal Communications, 6(4):13–18, August 1999. 41. J. Neel, R. Menon, J. H. Reed, and A.B. MacKenzie. Using game theory to analyze physical layer cognitive radio algorithms. “http://quello.msu.edu/ conferences/spectrum/papers/neel-reed-menon-mackenzie.pdf”. 42. C.T. K. Ng and A.J. Goldsmith. Capacity gain from transmitter and receiver cooperation, 2005.

28 Game Theory and Cooperation for Cognitive Networks

553

43. N. Nie and C. Comaniciu. Adaptive channel allocation spectrum etiquette for cognitive radio networks. ACM MONET (Mobile Networks and Applications), 2006. to appear. 44. G. Owen. Game Theory. Academic Press, 1995. 45. Min Kyoung Park and Volkan Rodoplu. Core capacity region of portable wireless networks. In IEEE GLOBECOM ’04, volume 6, pages 3783 – 3787, Dec. 2004. 46. E. Perevalov and R. Blum. Delay limited capacity and ad hoc networks: asymptotically optimal transmission and relaying strategy. In Proceedings of IEEE INFOCOM, volume 2, pages 1575–1582, March-April 2003. 47. Y. Qiu and P. Marbach. Bandwidth allocation in ad hoc networks: A price-based approach. In Proceedings IEEE INFOCOM, 2003. 48. V. Rodoplu and T.H. Meng. Computation of core capacity of wireless ad hoc networks. In Proc. of 36th Asilomar Conf. on Signals, Systems, and Computers, volume 2, pages 1543 – 1547, November 2002. 49. V. Rodoplu and T.H. Meng. Core capacity of wireless ad hoc networks. In Proc. of the 4th Internation Symposium on Wireless Personal Multimedia Communications (WPMC), volume 1, pages 247 –251, October 2002. 50. V. Rodoplu and T.H. Meng. Core capacity region of energy-limited wireless ad hoc networks. In Proc. of IEEE Inform. Theory Workshop, 2002, page 207, October 2002. 51. R. W. Rosenthal. A class of games possessing pure-strategy nash equilibria. International Journal of Game Theory, 2:65–67, 1973. 52. Y.E. Sagduyu and A. Ephremides. Joint optimization of mac and network coding for cooperative and competitive wireless multicasting. In 12th International Conference on Parallel and Distributed Systems (ICPADS’06), pages 69–76, 2006. 53. C. Saraydar, N.B. Mandayam, and D. Goodman. Pricing and power control in a multicell wireless data network. IEEE Journal on Selected Areas in Communications, 19(10):1883–1892, October 2001. 54. V. Shah, N.B. Mandayam, and D.J.Goodman. Power control for wireless data based on utility and pricing. In Proceedings of the IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, volume 3, pages 1427–1432, September 1998. 55. S. Verd´ u. Multiuser Detection. Cambridge University Press, Cambridge UK, 1998. 56. D. Wang, C. Comaniciu, and U. Tureli. Cooperation and fairness for slotted Aloha. Special Issue of Springer Wireless Personal Communications on Cooperation in Wireless Networks, 2006. to appear. 57. Wing-Hin Wong, J.M. Shea, and T.F. Wong. Cooperative-diversity slotted ALOHA. In Second International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks, pages 35–43, August 2005. 58. Y. Xue, B. Li, and K. Nahrstedt. Price-based resource allocation in wireless ad hoc networks. Technical report, University of Illinois at Urbana-Champaign, Urbana-Champaign, Illinois, April 2003. 59. Y.Jin and G. Kesidis. A pricing strategy for an aloha network of heterogeneous users with inelastic bandwidth requirements. In Proceeding of 39th Annual Conference on Information Sciences and Systems (CISS), 2002. 60. Haitao Zheng, Yan Zhu, Cong Shen, and Xiaodong Wang. On the effectiveness of cooperative diversity in ad hoc networks: a MAC layer study. In Proceedings

554

Cristina Comaniciu of IEEE International Conference on Acoustics, Speech, and Signal Processing, pages 509–512, March 2005.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios Mohammad Hossein Manshaei1 , M´ ark F´elegyh´ azi1 , Julien Freudiger1 , Jean-Pierre Hubaux1 , and Peter Marbach2 1

2

EPFL, Switzerland [hossein.manshaei|mark.felegyhazi|julien.freudiger| jean-pierre.hubaux]@epfl.ch University of Toronto, Canada [email protected]

Summary. Since the beginning of the 20th century, the wireless frequency spectrum has been carefully controlled by government regulators. In response to the recent advances in radio technology, the spectrum regulators have opened some parts of the available spectrum for unlicensed usage. In addition, they have reformed the traditional command and control regulation policies and have allowed more opportunistic transmissions over unused spectrum bandwidth in licensed bands, for certain times and locations. This paradigm shift can lead to a more flexible and efficient spectrum sharing in the near future. In this chapter, we address the problem of spectrum sharing between network operators and cognitive radios. Because of the dynamic nature of spectrum sharing, it is difficult to analyze and to provide sound spectrum management schemes. Several researchers rely on game theory that is an appropriate tool for modelling strategic interactions between rational decisionmakers (e.g., spectrum sharing in wireless networks). We present a selected set of works to highlight the usefulness of game theory in solving the main problems in this field.

29.1 Introduction Wireless communications rely on the frequency spectrum as a fundamental resource. As the number of wireless communication technologies and the number of wireless networks using them kept increasing, the regulation of the access to the available frequency spectrum, i.e., controlled spectrum sharing, has become unavoidable. A straightforward solution for the spectrum sharing problem is to let government agencies, such as the FCC in the USA, allocate communication frequencies to different wireless networks. This was first practiced, and basically still is, on a first-comefirst-served basis and then by auctions [5]. The allocated right, called the spectrum license, grants an exclusive usage of a given frequency band to a certain company for a given purpose. The main problem with the licensed spectrum is that the licenses are typically established for long periods of time. Recent performance studies [28,29] have shown that this significantly affects efficiency.

555 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 555–578. c 2007 Springer.

556

Mohammad Hossein Manshaei et al.

Around 1980, government agencies realized that the available spectrum was scarce and reserved certain frequencies as unlicensed bands for common use. Unlicensed bands eliminated the lengthy process of spectrum licensing thus allowing companies to enter into the communication market quickly. In spite of the advance of technologies, unlicensed bands are useless when used arbitrarily. Hence, government agencies limited the transmission power of wireless devices in unlicensed bands (this limit can vary among technologies). Yet, unlicensed bands can be quickly saturated, which also means that, in contrast to licensed usage, the quality-of-service (QoS) is hardly guaranteed by these networks. Although unlicensed bands have improved the overall spectrum utilization, they still do not solve the inflexibility caused by the licensing process. Cognitive radio [9, 16, 20] is an emerging technology that enables devices to determine which of the available frequencies are unused, and to use them even if they are licensed to others. Cognitive radio devices can adapt to the actual frequency utilization, thus increasing the efficiency of wireless communication. One fundamental requirement of these devices is that they should not hamper the communication of the primary users, who obtained the license for the given frequency band. There exist several techniques that are appropriate for studying the behavior of this new networking environment. On the one hand, game and auction theory are useful tools to study the strategic behavior of network participants; on the other hand, graph coloring techniques can be used to assess the system optimum solution in many problems, such as the channel allocation problem. In Section 29.2 we provide a short introduction to these analytical tools. Many researchers are currently engaged in designing spectrum sharing schemes using these analytical tools. Our goal in this chapter is to present a selected set of contributions in this field and to provide a better understanding of the current research efforts in this field. So we give only a high-level overview of the schemes. It is, of course, impossible to mention all the results of the selected papers in the field. Hence, we will briefly express the game model and the main results of each scheme. Table 29.1 shows that we can divide the spectrum sharing games into three main groups, according to the players of the games: licensed band operators, unlicensed band wireless systems, and cognitive radios. In Section 29.3, we focus on the interaction between wireless operators in shared spectrum competing for users. The second set of scenarios, presented in Section 29.4, addresses the problem of unlicensed spectrum sharing. Finally, the scenarios presented in Section 29.5 are related to spectrum sharing by cognitive radios. For each game, we identify the key ideas and the game model. We discuss the main results of the games in each subsection as well.

29.2 Theoretical Background Game theory, auction design, and graph coloring are the main tools for the analysis of the spectrum sharing schemes presented in this chapter. In this section, using a practical example, we introduce the fundamental concepts of non-cooperative game theory, such as Nash equilibrium (NE) and Pareto-optimality. The interested reader can find a comprehensive tutorial on game theory for wireless networks in [2,6]. Then, we examine the benefits of using auctions in spectrum assignment and spectrum sharing design. Finally, we briefly discuss graph coloring techniques.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

557

Table 29.1. Spectrum sharing games presented in this chapter divided to three main groups: licensed band, unlicensed band, and cognitive radio. Section 29.3.1 29.3.2 29.3.3 29.4.1 29.4.2 29.5.1 29.5.2 29.5.3

Spectrum Sharing Players Strategy Results Game Scheme Asymmetric Network WAN and Operator Nash Equilibrium Operators [30] WiFi Operators Selection National Border Cellular Operators Pilot Power Nash Equilibrium Spectrum Sharing [7] with Convergence Network Operators’ Cellular Operators Pilot Power Nash Equilibrium Spectrum Sharing [8] Heterogeneous High and Low Power Power Spectral Pareto Wireless Systems [4] Wireless Systems Density Optimality WiFi Operators’ WiFi Channel Nash Equilibrium Spectrum Sharing [13] Operators Selection Opportunistic Spectrum Cognitive Radios Channel Equilibrium Sharing [3, 26, 31, 32] Selection Auction-Based Cognitive Radios SNR and Nash Equilibrium Spectrum Sharing [17] Power and Pareto Optimal Multi-Cell OFDMA Cognitive Radios Rate Nash Equilibrium Spectrum Sharing [14] by Virtual Referee

29.2.1 Game Theory Game theory [10, 12, 25] is a discipline for modelling situations in which decisionmakers have to make specific actions that have mutual, possibly conflicting, consequences. There is a significant amount of work in wireless networking that makes use of game theory. The basic elements of a game G are the players, the strategies, the payoffs, and the knowledge and can be shown in strategic form by G = (P, S, U). P, S, and U are the set of all players, the joint set of the strategy spaces, and the set of payoff functions of all players, respectively. Considering a player i ∈ P, −i represents all the players belonging to P except i himself, they are often designated as being the opponents of i. Si corresponds to the strategy space of player i and the set of chosen strategies constitutes a strategy profile s (e.g., s = {s1 , s2 } for two players). The payoff ui (s) is the difference of the benefit b and the cost c of player i given the strategy profile s (i.e., u(s) = b(s) − c(s)). Note that, we refrain from using the word “utility” in this chapter to avoid confusion: in game theory, the utility usually corresponds to what we call the payoff in this chapter, whereas in many computer science papers, the utility corresponds to what we call here the benefit. In the following, a game that we call the Multiple Access Game is used to illustrate the fundamental concepts of game theory. We choose this game as a simple illustration of spectrum sharing games. In this game, two players, p1 and p2 , share the wireless medium and want to send one packet to their respective receivers, r1 and r2 . The players have a packet to send in each time slot, and they can either transmit (T ) or stay quiet (Q). When player p1 transmits, it incurs a transmission cost of 0 < c << 1. The packet transmission is successful if p2 does not transmit in the same time slot, otherwise we say that there is a collision and each players must pay the transmission costs c while the two packets are lost. If there is no collision, player p1 gets a benefit of 1 for the successful packet transmission. We call this game a static game or one-shot game because the players have only one move to act.

558

Mohammad Hossein Manshaei et al.

Best Response The Multiple Access Game can be represented in a strategic form as shown in Table 29.2. If player p1 transmits, then the best response of player p2 is to be quiet. Conversely, if player p2 is quiet, then p1 is better off transmitting a packet. We can write the best response of player i to an opponent’s strategy vector s−i as follows. Table 29.2. The Multiple Access Game in strategic form. (1 − c, 0) means that player p1 ’s payoff is 1 − c, while player p2 gets nothing. p1 \p2 Q T Q (0,0) (0,1-c) T (1-c,0) (-c,-c)

Definition: The best response bri (s−i ) of player i to the profile of strategies s−i is a strategy si such that: bri (s−i ) = arg max ui (si , s−i ) si ∈Si

(29.1)

Nash Equilibrium If two strategies are mutually best responses to each other, then the players have no reason to deviate from the given strategy profile. In our example, two strategy profiles have this property: (Q, T ) and (T , Q). To identify such strategy profiles in general, Nash introduced the concept of Nash equilibrium (NE) in his seminal paper [23]: Definition: The pure strategy profile s∗ constitutes a Nash equilibrium if, for each player i, ui (s∗i , s∗−i ) ≥ ui (si , s∗−i ), ∀si ∈ Si (29.2) This means that in a NE, none of the users can unilaterally change his strategy to increase his payoff. Alternatively, a NE is a strategy profile comprised of mutual best responses of all the players. Hence, the system is stable. (Q, T ) and (T , Q) are two NE for Multiple Access Game. The first step towards solving a game is to investigate the existence of NE. It is worth mentioning that Nash [12,23,24] proved that every finite strategic-form game has a NE. Once we have verified that a NE exists, we have to determine whether it is a unique equilibrium point. If the players have identified various Nash equilibria, it still might be difficult for them to coordinate on which one to choose. For example, in the Multiple Access Game both players know that there exist two Nash equilibria, but each of them tries to be the winner by deciding to transmit. Hence, their actions will result in a profile that is not a NE.

Pareto-optimality and Price of Anarchy One method to assess the efficiency of the equilibrium point in a game is to compare the strategy profiles using the concept of Pareto-optimality. To introduce this concept, we first define Pareto-superiority.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

559 0

Definition: The strategy profile s is Pareto-superior to the strategy profile s if for any player i: 0 0 ui (si , s−i ) ≥ ui (si , s−i ) (29.3) with strict inequality for at least one player. In other words, the strategy profile 0 s is Pareto-superior to the strategy profile s , if the payoff of a player i can be 0 increased by changing from s to s without decreasing the payoff of other players. Note that the players might need to change their strategies simultaneously to reach the Pareto-superior strategy profile s. Based on the concept of Pareto-superiority, we can identify the most efficient strategy profile or profiles. Definition: The strategy profile spo is Pareto-optimal if there exists no other 0 strategy profile s that is Pareto-superior to spo . In a Pareto-optimal strategy profile, one cannot increase the payoff of player i without decreasing the payoff of at least one other player. Thereby, using the concept of Pareto-optimality, we can eliminate poor Nash equilibria by selecting those with a Pareto-superior strategy profile. In the Multiple Access Game, both pure strategy profiles (T , Q) and (Q, T ) are NE and Pareto-optimal. Finally, a metric to measure the quality of a given NE is the Price of Anarchy (PoA). First defined in [21], the PoA is the ratio between the worst NE of the game and the Pareto-optimal.

29.2.2 Auction Design In economics, an auction is a method to determine the value of a commodity that has an undetermined or variable price. In progressive auctions for example, the players propose increasing bids for a good, and the highest bid wins the auction. For the first time in July 1994, the FCC allocated the commercial spectrum via competitive auctions instead of the previous best public use method. Auctions are a suitable way to assign the spectrum licenses because it is the player who values the most the spectrum who obtains it. Still, the process can be long and can lead to an overestimated price due to strong competition (e.g., UMTS spectrum). Hence, the rules for designing and conducting spectrum auctions has evolved towards more efficient mechanisms inspired by principles of game theory. Vickrey introduced a type of sealed-bid auction in which the highest bid wins, but the price paid is the second highest bid. This mechanism provides an incentive to bidders to declare their true evaluation to maximize their payoff. An auction is called efficient if it maximizes the total payoffs of all players (bidders). When multiple divisible goods are sold individually, a generalization of the Vickrey auctions [22] maintains the incentive to bid truthfully, the Vickrey-Clarke-Groves (VCG) auctions. The generalized VCG mechanism is an efficient auction scheme achieving the socially optimal (i.e., maximum of the total payoff; which is also a Pareto optimal, but the reverse is not true) allocation for players with quasi-linear payoff functions. The idea is that the bidders pay the opportunity cost that their presence introduces to all the other players. In a generalized VCG auction, players report their payoff to the auctioneer. The auctioneer then computes the optimal allocation that maximizes the aggregated payoff and allocates the resource accordingly. The auctioneer also solves other optimization problems: How to calculate, for each bidder, the price that each bidder should pay. The VCG auction is truthful, which means that each players reports her true valuation independently of the report of the other players.

560

Mohammad Hossein Manshaei et al.

Each player only needs to know his own payoff function, and then the social optimal solution can be found in one-iteration. Despite their advantages, VCG auctions have several shortcomings when applied to spectrum sharing. VCG auctions generate a large communication overhead and require high computational resources for the auctioneer to calculate the optimal payments. Hence, in Section 29.5.2, we will discuss an alternate auction scheme that is simpler to implement in wireless networks.

29.2.3 Graph Coloring Graph coloring consists in assigning a color to the vertices of a bidirectional graph G = (V, E), where V is a set of vertices and E a set of edges. The coloring is a mark that defines the category of the vertex. Marking each vertex of a graph with a finite set of k colors is equivalent to partitioning the vertices into k categories. In the following, graph coloring will refer to the coloring of the vertices of a graph. A coloring that uses at most k colors is called a k-coloring. Channel allocation

{A} 2

A

A,B

A 1

A 2

{A,B} 3

4

A,B

{A,B}

{A,B}

(a)

3

B 5 {B}

5 1

4 A

B

B

(b)

Figure 29.1. Example of graph coloring algorithm: (a) Conflict Graph representation of a wireless network with two channels A and B. The edges are labeled to indicate interfering channels. (b) Resulting proper 2-Coloring of the Conflict Graph. Note that node 3 could not be assigned a color.

problems, for example, can be solved by graph coloring algorithms. Let us assume that interferences in a wireless network with two channels are modeled as a conflict graph where each vertex is a mobile user and nodes are connected if they interfere as shown in Figure 29.1(a). The conflict graph can be reduced to a colored conflict graph representation in which colors map to channels. To minimize interference, the same channel must not be assigned to adjacent nodes and the problem is solved by finding a proper k-coloring of the conflict graph. A coloring is proper if no two adjacent vertices are assigned the same color (e.g., Figure 29.1(b). Traditional graph coloring algorithms minimize the number of colors used to mark each vertex. The best strategy consists in coloring the most difficult vertices first. As graph coloring algorithms are NP-Complete, optimal graph coloring solutions are obtained via approximations [11]. In graph coloring approximation, the idea is to prioritize the graph coloring process, instead of exhaustively testing all color assignments. Labels are computable values satisfying a number of predefined conditions known as rules. A label represents the importance of a vertex in the coloring process.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

561

A simple sequential heuristic solution of a graph coloring problem assigns to each vertex a unique label and colors the vertices starting from the highest label with the lowest indexed color, while not violating the constraints (e.g., interference). The algorithm then removes the colored vertex and the associated edges from the graph and repeats the procedure until all vertices are colored. The prioritization of coloring via labels is a good approximation for solving graph coloring problems and it offers flexibility to solve the problem. Indeed, by varying the rules definitions, the algorithm can be used to minimize the number of colors for example, or, maximize the fairness of the graph coloring.

29.3 Network Operator Games In this section, we address the game models that are proposed to design the spectrum sharing mechanisms for network operators with a part of the licensed band spectrum. Section 29.3.1 presents an asymmetric game that considers both licensed and unlicensed bands in order to design efficient networks by a single operator. In the games presented in Section 29.3.2 and 29.3.3, the network operators control the transmission power of the pilot signal at their Base Stations (BSs). Their objective is to attract as many users as possible to their BSs.

29.3.1 WAN-WiFi Competition Zemlianov and De Veciana [30] study the problem of competition or cooperation in a multi-operator network where WAN BSs operating in a licensed band and WiFi hotspots in an unlicensed band coexist. They investigate how the decision mechanisms of the mobile users affect the ability of networking entities in such heterogeneous networks. In their scenario, the users can choose to connect to a wireless WAN with full coverage or a WiFi with limited coverage if available in their location. They use a stochastic geometric model that allow them to model channel diversity, the interference, and the spatial load fluctuations of the wireless environment.

System Model Zemlianov and De Veciana define three point processes Π a , Π h , and Π w in order to represent the location of mobile users, WiFi hotspots, and WAN BSs respectively. They suppose that the wireless WAN operator can cover the whole spatial location whereas the WiFi operator has limited coverage. Two strategies are considered for the mobile users to choose between the WAN and WiFi. With the first strategy, called proximity-based, the users connect to the nearest service provider (i.e., WAN BSs or WiFi AP) whereas in the second one, called utility-based, they can select the operator based on their payoff function. The authors assume that each mobile makes a decision to choose its operator periodically. In the utility-based mechanism, a mobile user switches to WAN BS wm w − from hotspot hk if and only if she was connected to hk at t− and uw i (Nm (t ) + 1) > h h − w w h ui (Nk (t ) + c ), where ui and ui are the payoffs of user i when it is served by BS w wm and hotspot hk respectively, Nkh is the number of users connected to hk , Nm

562

Mohammad Hossein Manshaei et al.

is the number of user connected to wm , t− refers to the time immediately before t, and cw is the cost of switching to WAN BS. A similar condition can be written for a mobile user who wants to switch from WiFi to WAN operator.

Results Zemlianov and De Veciana prove that if the payoff function of mobile user is a w continuous and monotonically decreasing function of Nkh or Nm , given any initial configuration of the agents’ choices at time t, the system converges to an equilibrium configuration at t → ∞. This equilibrium configuration is obtained by constructing a feasible path for the chain evolution that hits an equilibrium state with positive probability, starting from any initial configuration. Note that this equilibrium might not be unique. But, by appropriately selecting the payoff function, the set of equilibria could be made tight. They also show that the class of payoff functions that are congestion dependent, provide much better performance to users on average than the simple proximity-based decision strategy. The above results can help the operators, with both wireless WAN infrastructure (e.g., WiMAX) and a set of WiFi hotspots, to design an efficient network. They show that by making use of the proposed optimal joint design with utility-based strategy at the mobile users, the operator can achieve a target performance and significantly reduce the resource costs in the same time.

29.3.2 National Border Spectrum Sharing F´elegyh´ azi et al. [7] study a spectrum sharing game between two cellular operators along the national border of two countries. They consider the problem of strategic behavior in CDMA networks. Spectrum licenses are allocated within each country, thus leading to possible conflicts along national borders. There exist many examples where cities reside close to a national border (e.g., Geneva in Switzerland).

Game Model Spectrum sharing on the border is modeled as a two-player non-cooperative power control game. The players are two cellular network operators (e.g., A and B) with one BS each, which provide wireless access in the same frequency band on the two sides of a national border. Hence, they share the spectrum and cause interference to each other. The strategies of the operators determine the pilot transmission power of their base stations. Player i’s strategy is si = Pi , where 0W < Pi < 10W is the pilot signal power of BS i. The standard value of the UMTS pilot signal is P s = 2W . There is a set of users M equipped with wireless devices that access the communication network. These users select the operator with the highest pilot signal quality or carrier-to-interference ratio (CIR) [27] and pay for the service to the selected operator. The CIR is a function of pilot and traffic signal powers, the distance between user and BS, own-cell and the other-cell interferences, and the processing gain for the pilot and traffic signal from BS to user. If Mi represents the set of users connected to BS i and θu is the expected income obtained by serving user u of a certain traffic type, then the payoff function of operator i can be calculated by:

29 Spectrum Sharing Games of Network Operators and Cognitive Radios 563 X ui = θu (29.4) u∈Mi

The operators then define their strategies in order to maximize this payoff function. The main goal of this study [7] is two-fold: to establish whether the operators have an incentive to be strategic and to characterize the Nash equilibria and the Pareto-optimal strategy profiles in the game.

Game Results In [7], F´elegyh´ azi et al., present a numerical simulation study to evaluate the above game in UMTS system using Matlab. They distribute the mobile users according to the uniform distribution and calculate the number of users that attach to each of the BSs based on the CIR requirements over several simulations. This defines an experimental payoff matrix for the two players. Figure 29.2(a) shows the payoffs of players A and B, as well as the sum of their payoffs as a function of the pilot signal power PA . There are, on the average, 10 data traffic users in the simulation area who pay 50 CHF/month for a data traffic service. Player A is strategic by adjusting its pilot signal power, whereas player B operates his BS according to the standard pilot power P s . The payoff function of operator A has a unique maximum point that requires a higher pilot power than P s . The increase in payoff of player A means the decrease of the payoff of the non-strategic player B. This shows that the operator A has an incentive to be strategic. Figure 29.2(b) shows the payoff surface for operator A as a function of the pilot power values of the two operators (i.e., when two operators are strategic). One interesting observation from this figure is that uA has a unique maximum point for PA . Moreover, this maximum point depends on the pilot power of the other BS, PB .

(a)

(b)

Figure 29.2. Payoff of player A as a function of his pilot power when (a) only c IEEE, player A is strategic and (b) when both operators are strategic. From [7], 2007. Using the two payoff surfaces, one can derive the best response functions for the operators as shown in Figure 29.3(a). Based on the concept of best responses introduced in Section 29.2.1, the NE in the power control game can be identified.

564

Mohammad Hossein Manshaei et al.

This NE point is unique for any user density and the NE pilot powers decrease as the number of users increases. The reason is that the BSs can serve enough users by using a relatively small power and hence there is no incentive for them to go above these pilot power values. Figure 29.3(b) shows the achieved payoffs as a function of the pilot power values PA and PB . We observe that in this case the Pareto boundary defines a straight line, because in a Pareto-optimal strategy profile each user in the system is attached to one of the BSs. Furthermore, the standard pilot powers and the NE strategy profile result in the same payoffs for the players, and they both lie on the Pareto boundary. This means that the players achieve a desirable state, from the system point of view.

(a)

(b)

Figure 29.3. Best response and NE for border games: (a) Best response functions for the two players with 10 data users. (b) The payoff region with all possible payoffs for 10 data users. The NE, the payoff of the standard powers and all Pareto-optimal c IEEE, 2007. points are highlighted. From [7],

It is shown that the NE pilot powers are higher than the standard value. In [7], the authors extend the payoff function to include the cost of a high pilot power. The P extended payoff can be defined as u ˆi = ( u∈Mi θu ) − C ∗ , where C ∗ is the power cost. Hence, the players have the choice between the standard (P s ) and the NE strategies. F´elegyh´ azi et al. highlight the connection between this extended game and the well-known Prisoner’s Dilemma.

29.3.3 Network Operators Spectrum Sharing F´elegyh´ azi and Hubaux [8] study the spectrum sharing problem in a scenario where the mobile users can freely roam across the base stations (BS) of different network operators, attaching to the one offering the most favorable signal quality (i.e., the one with the strongest pilot signal) and bandwidth. This free roaming could be beneficial for both operators and users, because the former could serve an increased set of users, and the latter could enjoy various services across several operators.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

565

Game Model The above problem is modeled as a two-player, nonzero-sum game. The players are two cellular network operators. The strategy of the operators is to define the radio range of their BSs (which is determined by their pilot signal strength). F´elegyh´ azi and Hubaux assume that the players apply the same radio range for each of their BSs. Accordingly it is assumed that they use the following radio ranges: rH (i.e., heavy player with larger radio range) and rL (i.e., light player with smaller radio range). The authors consider a scenario where BSs are symmetrically placed on a grid with a minimum distance d. this means that each BS of a given operator has four neighboring BSs that belong to the other operator. Consequently, the game can be analyzed considering two neighboring base stations, as shown in Figure 29.4. The useful coverage area (Oi ) for any BS i is its Voronoi calculated from the radio ranges of BS i and the radio ranges of its neighbors. The interference area (Yi ) for a BS i (i.e., the area, which is in its radio range, but it does not cover eventually) can be expressed as: Yi = ri2 · π − Oi (29.5) where ri denotes the radio range of BS i (i.e., the player i’s strategy). The players want to maximize the area they cover with their pilot signals, while minimizing their interference area. Hence, the authors express the payoff function of player i (i.e., the payoff of her BS) as follows: ui (ri , rj ) = Oi − γi · Yi = (1 + γi ) · Oi − γi · ri2 · π

(29.6)

where γi ≥ 0 is a cooperation parameter that defines how much player i cares about the size of its interference area. Two cases can be distinguished for the coverage range of heavy and light player. In the first case, both players have a non-empty coverage area as presented in Figure 29.4(a). In the second case, the light player is overwhelmed by the heavy player, meaning that the pilot signal of the heavy player is the strongest everywhere, as shown in Figure 29.4(b). The authors assume that there exists a maximum transmission power Pmax defined by the regulator of the wireless spectrum, which determines the maximum radio range Rmax . They also assume that the operators want to cover the total service area. If the radio ranges of all base stations are equal, the√minimum radio range for which there is full coverage can be calculated by Rmin = 22 d. Considering the limit case, in which the operators cover only the service area, one can write the following bounds on rL and rH : q √ 2 d2 − 2drL + rL ≤ rH ≤ Rmax (29.7) √ q 2 2 max{0, )} ≤ rL ≤ rH (d − −d2 + 2rH 2

(29.8)

Game Results F´elegyh´ azi and Hubaux first study the results of a single-stage game (i.e., both players simultaneously choose their radio range once and for all.). They show that there exist a variety of Nash equilibria depending on cooperation parameters (i.e., γi and γj ) and maximum radio range (see Table III in [8]). They also identify the Pareto optimal solutions as:

566

Mohammad Hossein Manshaei et al.

(a)

(b)

Figure 29.4. Coverage and interference area of a base station, illustrated with two base stations: (a) both BSs have a coverage area; (b) the BSs of the light player are overwhelmed by the BSs of the heavy player and thus the light player has no c IEEE, 2006. coverage area at all. From [8],

• •

• •

If the operators are cooperative, they should play the radio range with which they are able to cover the service area (i.e., the lower limit in Equation (29.7)). If one of the players does not cooperate and the other does, then the noncooperative player can increase its radio range to force the cooperative player out of the game (i.e., ri = d, rj = 0). If neither of the players cooperates, then they both will end up in playing the maximum radio range (i.e., ri = rj = Rmax ). In a fair solution, they should both play the minimum radio range (i.e., ri = rj = Rmin ).

The authors also prove a condition for which the socially desirable NE (i.e., ri = rj = Rmin ) exists and that it can be enforced using punishments in a repeated game. In the proposed repeated game, if operator i uses the Punisher strategy, it plays Rmin in the first step. For any further time step, it plays: (i) Rmin in the next time step if the other player played Rmin in the previous time step; or (ii) Rmax for the next ki time steps, if the other player played anything else. The parameter ki (also called the punishment interval ) defines the number of time steps for which player i punishes the other player. The Punisher strategy is similar to the well-known Tit-For-Tat (TFT) strategy [1], but it retaliates any defection by playing Rmax instead of copying the same behavior. It is also different from the Trigger strategy defined in [22] (or infinite punishment), because the Punisher strategy imposes a punishment that is comparable to the amount of misbehavior and thus it is able to recover from erroneous defections. Figure 29.5 illustrates the average per time slot payoff of a player for both cooperation and defection. One can observe that cooperation is more beneficial, because defection is quickly retaliated by the other player. Finally, the authors show that the solution of the power control problem is NP-complete for a general topology of base stations.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

567

Figure 29.5. Average payoff of player i for d = 1 km, γj = 0.1 and Rmax = 1.5 km, if player j applies a punishment. One-time defection is quickly retaliated and hence cooperation is the best choice. The Trigger strategy stabilizes in infinite punishment, c IEEE, 2006. and the Punisher strategy returns to the cooperative state. From [8],

29.4 Games in Unlicensed Bands As the name suggests, the unlicensed bands are the radio spectrum that can be freely used without obtaining a license. In 1986, the FCC provisioned for the first time unlicensed bands for Industry, Science, and Medicine (ISM) applications based on spread-spectrum technologies in the 915 MHz, 2.4 and 5.7 GHz spectrum bands. In the 90s, the FCC allocated additional unlicensed bands at 2, 5, and 59-64 GHz for wireless applications which require small coverage. Spectrum sharing in unlicensed bands suffers from two main problems: (i) devices accessing unlicensed band may experience severe interference as they do not have exclusive access to the spectrum, or worse, (ii) spectrum sharing in unlicensed bands may result in the tragedy of the commons [15] as there are no inherent incentives to efficiently use the radio band [19]. In the following we study these two problems in the context of spectrum sharing among heterogeneous wireless systems [4] and spectrum sharing among WiFi operators [13].

29.4.1 Spectrum Sharing among Heterogeneous Wireless Systems We first consider the situation where heterogeneous wireless systems (e.g., Bluetooth and IEEE 802.11 WiFi) share the spectrum of an unlicensed band where each system behaves selfishly and tries to maximize its transmission rate. Etkin et al. model and study in [4] the resulting interaction among the systems as a non-cooperative game, and proposed various spectrum sharing rules and protocols to allow the wireless devices to share the bandwidth in a fair and efficient way.

One-Shot Game Model Suppose that M wireless systems, each consisting of a single transmitter-receiver pair, share an unlicensed band of W Hz. Let pi (f ), f ∈ [0, W ], be the power spectral density of the transmitted signal in system i, i = 1, . . . , M , where the total power for each system can not exceed Pi , i.e.,

568

Mohammad Hossein Manshaei et al. Z

W

pi (f )df ≤ Pi .

(29.9)

0

Each system decides on a power allocation in order to maximize its transmission rate as follows. Given the power allocations of all other systems, system i chooses a spectral density pi (f ), f ∈ [0, W ], that maximizes ! Z W |hi,i |2 pi (f ) log 1 + df (29.10) Ri = P n0 + i6=j |hj,i |2 pj (f ) 0 where hj,i is the channel gain between the sender of system j and the receiver of system i, and n0 is the background noise power. Assuming that each system has perfect information, i.e., each system knows all the channel gains hi,j , i, j = 1, . . . , M , and all power constraints Pi , i = 1, . . . , M . Etkin et al. show in [4] that frequency-flat allocation given by pi (f ) =

Pi , W

i = 1, . . . , M 2

is always a NE for the above game. Furthermore, if

|hj,i | j6=i h | i,i |2

P

< 1, then the

frequency-flat allocation is the only NE. The rate allocation under the frequencyflat allocation is generally not Pareto efficient [4], and the outcome of the game may lead to a poor overall system performance.

Repeated Game Model To improve system performance, Etkin et al. then study the above situation as a repeated game. At time step t, t ≥ 0, each system i decides on its spectral allocation pi (f ), f ∈ [0, W ] subject to the constraint that the total power can not exceed Pi . The payoff ui of system i is then given by ui = (1 − δ)

∞ X

δ t Ri (t)

(29.11)

t=0

where Ri (t) is the maximal rate that system i can achieve at step t (see Equation (29.10)), and δ ∈ (0, 1) is a discount factor that accounts for the delay sensitivity of the system. Consider the achievable rate region R for the above system, i.e., R is the set of rate vectors (R1 , . . . , RM ) for which there exist spectral density functions pi (f ), i = 1, . . . , M , that achieve (R1 , . . . , RM ) by using the relation given by Equation (29.10). Furthermore, consider the following strategy for system i. Let (R1 , . . . , RM ) be a rate vector in the achievable rate region R and let pi (f ), i = 1, . . . , M , be the corresponding spectral densities. (i) Then at t = 0: system i uses the above power allocation pi (f ), and (ii) at time t = t0 , if at time t = t0 − 1 every system j, j = 1, . . . , M , uses the pj (f ) then system i uses pi (f ) at time t = t0 ; otherwise system i uses the frequency-flat allocation, i.e., uses Pi /W for f ∈ [0, W ]. Etkin et al. show in [4] that for every rate vector (R1 , . . . , RM ) ∈ R there exists a threshold δ0 < 1 such that if the discount factor δ is larger than δ0 , then the above strategy is a NE for the repeated game. Under the NE, the systems will

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

569

cooperate as long as no systems deviates from (p1 (f ), . . . , pM (f )) corresponding to the rate vector (R1 , . . . , RM ). A deviation by a system, will trigger a “punishment” where all systems adapt the frequency-flat power allocation leading to a poor NE. In particular, the above strategy can be used to select a Pareto efficient power allocation as the NE.

29.4.2 Spectrum Sharing among WiFi Operators Next, we consider the situation where several WiFi operators share a common unlicensed band that is sub-divided into a fixed number of channels. Each WiFi operators owns several Access Points (AP) and has to decide on the channel that each AP uses. If two APs (of the same or two different WiFi operators) are within a (sufficiently) small distance of each other, then they will interfere if both are assigned the same channel. Therefore, in order to ensure an acceptable level of service to their mobile users, neighboring APs must be assigned different channels. Halldorsson et al. [13] model the above channel assignment problem as a game between WiFi operators where each operator decides on a channel assignment for its own APs in order to maximize the total number of mobile users that it can serve. The outcome of the game is evaluated by means of the Price of Anarchy (PoA) (see Section 29.2.1). The PoA measures how far the outcome of the game is from a social optimal channel allocation (i.e., a channel allocation that maximizes the total number of mobile users that are served by APs).

Game Model Consider a set V of APs that are owned by several WiFi operators. Let d(u, v), u, v ∈ V , be the distance between the two APs u and v, and let Rt (u) and Rs (u) be the transmission and sensing range of AP u. Note that Rt (u) and Rs (u) depend on the transmission power of AP u. Let G = (V, E) be the corresponding interference graph where there is an edge between AP u and v if they are located close enough (i.e., if d(u, v) ≤ Rt (u) + Rt (v) + max{Rs (u), Rs (v)}). We say that AP u and v are neighbours if there is an edge in G between u and v. There are k channels available in the unlicensed band. Operators set up (activate) APs sequentially where the order with which APs are set up is given by an exogenous process (i.e., operators do not decide when to activate an AP). Whenever an AP is set up, the corresponding operator must choose a channel that does not interfere with any of the previously set up APs. In addition, an operator is allowed to change the channel assignments of the APs that it controls as long as it does not cause any interference with APs of other operators. When an operators sets up an AP, it only has knowledge about the channel used by neighbouring APs that have already been set up. It does not have any knowledge about the channels used by previously set up APs that are outside the interference region of the AP nor does it have any knowledge about the order with which APs are set up. The payoff that an operator receives for setting up an AP is equal to the expected number of mobile users that it can serve with the AP, where different APs can have different payoffs. The goal of each operator is to choose channels in order to maximize the overall payoff of its APs.

570

Mohammad Hossein Manshaei et al.

Channel Allocation Results For the above game, each NE corresponds to a maximal k-colored subset of the graph G. This allows to use graph theory results to compute the price of anarchy for the above game. In particular, Halldorsson et al. derive in [13] the following results. For the general case where APs can have different transmission powers and different payoffs, the price of anarchy is potentially unbounded, i.e., P oA = ∞ (see Figure 29.6). The price of anarchy is also unbounded for the case where APs have different transmission powers, but all APs have the same payoff. If all APs have the same transmission power (i.e., the interference graph is a unit disk graph) and the same payoff, then the price of anarchy is at most 5 + max(0, 1 − 5/k) and at least 5.

Figure 29.6. Network operators A and B with APs ai , bi ∈ V provide the Internet access to mobile users. The mobile users do not have Internet access because they are in the proximity of operator B, while operator A controls the channel. The PoA increases with the number of mobile users and is potentially infinite.

Local Bargaining The above results show that if operators are forced to decide on a channel as soon as it is set up and are only allowed to reassign channels among the APs they control, the above game can lead to a poor coverage. An approach to improve performance is to allow operators to negotiate changes to the channel assignments of the APs that they control. Halldorsson et al. consider in [13] such an approach where operators can use channel bargaining to locally optimize their total payoff. In particular, they consider two bargaining schemes called local 2-buyer-1-seller bargains and local 1-buyer-multiple-seller bargains. Figure 29.7 illustrates the local 2-buyer-1-seller bargain.

Game Results Halldorsson et al. show in [13] that for the general case where APs have different transmission power and different payoffs, the price of anarchy is unbounded even if

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

571

Figure 29.7. Example of 2-buyer-1-seller bargain: (a) a1 controls the channel (i.e., is colored), but the sum of b1 and b2 payoffs (i.e., the number of mobile users) is greater than a1 payoff. (b) b1 and b2 bargain with a1 to acquire the channel and improve the system payoff.

local 2-buyer-1-seller bargains or local 1-buyer-multiple-seller bargains are allowed. For the case where all APs use the same transmission power and have the same payoff, the price of anarchy under local 2-buyer-1-seller bargains is at most 3 + max(0, 1 − 3/k) and at least 3. For local 1-buyer-multiple-seller bargains the price of anarchy is at most 5 + max(0, 1 − 5/k) and at least 5 for the case where all APs use the same transmission power but have different payoffs. Halldorsson et al. also show that the above bargaining schemes will converge to a NE after a polynomial number of steps as a function of the number of APs, given that the payoffs are integers bounded by a polynomial in the number of APs. Halldorsson et al. also consider in [13] more general bargaining schemes than local 2-buyer-1-seller bargains or local 1-buyer-multiple-seller bargains. They prove that generally local bargaining may still lead to a poor performance unless the channel assignment of a large number of APs can be changed (i.e., global bargaining) at each bargaining step.

29.5 Cognitive Radio Games Cognitive radios can detect whether a certain radio band or channel is currently used, as well as sense the amount of interference (interference temperature) within a given radio band or channel [16]. In addition, they are able to control the transmission powers and dynamic spectrum management with the help of software defined radios [16, 20]. These capabilities open the possibility of a flexible sharing of the wireless spectrum [20]. In this section, we focus on the scenarios where cognitive radios are used to efficiently share the available spectrum. We first consider the situation where a primary user (operator) acquires and owns a licensed radio band. If the primary user does not fully utilize this band, then it can be accessed by secondary users (cognitive radios), as long as they do not create any (or a sufficiently small) interference to the primary user. In particular, we consider the the following two cases: (1) secondary users can freely utilize the radio spectrum as long as they do not interfere with the primary user (Section 29.5.1) and (2) the primary user sells access to the radio band

572

Mohammad Hossein Manshaei et al.

through an auction mechanism (Section 29.5.2). In Section 29.5.3, we consider an OFDM network where several cognitive radios share the available OFDMA channels.

29.5.1 Opportunistic Spectrum Sharing We first consider the situation where several primary users (operators) acquire their own radio band and where each radio band is further divided into several channels. We focus on an opportunistic spectrum sharing where secondary users (cognitive radios) are free to utilize channels as long as they do not interfere with the primary users [3, 26, 32]. Here, it assumed that secondary users will cooperate with each other to obtain a channel allocation with a maximum utilization subject to a given fairness criteria. Figure 29.8 provides an example of opportunistic spectrum sharing where the unused spectrum from a TV broadcast channel is utilized to provide WiFi connections to a residential community. Secondary user 2 cannot make use of channel A as it would interfere with primary user X. Secondary users 1 and 3 can emit on channel A as long as they control their transmit power not to interfere further than ds (1, A) or ds (3, A) respectively.

Figure 29.8. Secondary users 1 and 3 exploit channel A with their WiFi APs without interfering with the base station of the primary user.

System Model Consider the situation where N secondary users share M channels. Different channels allow different secondary users to transmit at different rates. Let bn,m be the throughput that user n can achieve on channel m, where n = 1, . . . , N and m = 1, . . . , M . The interference constraints are modeled by an interference graph (see for example [26]) which defines on which channels a given secondary user does not interfere with the primary user, and which secondary user can simultaneously transmit on a given channel m without causing interference among themselves. In addition, it is assumed that the maximum number of channels assigned to a secondary user can not exceed a given threshold Cmax . For a given feasible channel allocation (i.e., a channel allocation that does not violate any interference constraints and any of the constraints on the maximum number of channels allocated to a secondary user), the network utilization is defined as the sum of the throughput over all secondary users.

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

573

Results Using the above model, Peng et al. formulate in [26] the channel allocation problem as a graph coloring problem where the goal is to maximize network utilization, or to maximize network utilization subject to a given fairness criteria such as max-min fairness and proportional fairness [18]. Finding such an optimal channel allocation is NP-hard and Peng et al. propose several heuristic graph coloring algorithms for each of the above performance objective. Peng et al. derive lower bounds for the performance of the centralized as well distributed implementation of these algorithms. The algorithms proposed by Peng et al. in [26] require coordination and frequent information exchange among secondary users, and may impose substantial overhead on the network. As an alternative approach, Zheng and Cao in [32] propose a socalled device-centric management scheme where each secondary user accesses channels based on simple rules that require only local information. The authors propose several allocation rules and derive lower bounds for their performance in terms of the poverty line (i.e., the minimum number of channels each secondary user is guaranteed to obtain), An important property of the proposed algorithms is that they reach a stable channel allocation in a finite number of iterations. In [3], Cao and Zheng allow secondary users to be mobile. Instead of computing the channel allocation at each network topology change, secondary users negotiate channel allocations with their neighborhood via local bargaining. For the proposed local bargaining mechanism, a theoretical bound on the lower bounds for their performance in terms of the poverty line is provided.

29.5.2 Auction Based Spectrum Sharing Next, we consider the situation where a primary user (operator or government agency) lets secondary users access its spectrum subject to a given power constraint, i.e., the total interference created by the secondary users at fixed measurements points has to be below a given threshold. For this situation, Huang et al. propose in [17] an auction-based spectrum sharing where secondary users submit bids. Based on these bids, the primary user decides on the transmission power allocated to each secondary user, as well as the cost per unit transmission power that secondary users are charged. The goal of each user is to submit bids in order to maximize its payoff minus cost, where the payoff is a function of the received signal-to-interference plus noise ratio.

System Model Consider M secondary users and let pi be the transmission power of user i, where i = 1, . . . , M . The primary user then allocates transmission power to secondary users such that the total Preceived power at a given measurement point is less than a threshold Pmax (i.e., M i=1 pi hi0 ≤ Pmax , where hi0 is the channel gain from user i’s transmitter to the measurement point). Let γi = n0 + κ

pi hii P

j6=i

pj hji

574

Mohammad Hossein Manshaei et al.

be signal to noise and interference ratio (SIN R) at secondary user i’s receiver, where κ is a positive constant that depends on the spectrum bandwidth, hji is the channel gain from secondary user j to secondary user i’s receiver, and n0 is the background noise power. The payoff function of secondary user i is a function of γi given by ui (γi ) = θi ln(γi ), where θi is a user-dependent parameter.

Auction Based Allocation For the above situation, Huang et al. consider the problem where the primary user wants to allocate transmission power to secondary users in order to maximize the social welfare [17] subject to the interference constraint at the measurement point. As discussed in Section 29.2.2, a VCG auction could be used to achieve this. However, a VCG auction might not be suitable for this situation due to (a) the overhead of communicating the payoff function to the primary user and (b) the computational complexity of computing an optimal allocation. As an alternative approach, Huang et al. propose the following auction based power allocation scheme. The primary user decides on a reserve power p0 and announces a reserve bid β ≥ 0 and a price π s > 0. After observing β and π s , each secondary user i submits its bid bi . The primary user then allocates transmission power pi to secondary users so that the received power at the measurement point is proportional to the bids, i.e., we have bi Pmax βPmax hi0 pi = PM , and p0 = PM , (29.12) b + β j j=1 j=1 bj + β and charges each secondary user a price Ci = π s γi where γi is the SIN R of secondary user i under the above power allocation. The goal of each secondary user is to submit a bid such that the resulting power allocation maximizes its payoff ui (γi ) minus cost Ci . Assuming complete information, Huang et al. model this situation as a nons cooperative game, and prove that for β > 0, there exists a threshold price πth >0 s s such that a unique NE exists if π s > πth ; there does not exist a NE if π s ≤ πth . For this game Pareto optimality and stability (NE) are conflicting, however an -Pareto optimal NE can be achieved. The -Pareto optimal allocation is the Pareto-optimal solution for the -system in which the total received power at measurement point is less than (1 − )Pmax . The assumption that secondary users have complete information when deciding on their bids is unrealistic, and Huang et al. propose an iterative bidding algorithm that requires each secondary user to have access only to his local information (i.e., its own payoff function and its local channel gains) and therefore can be implemented in a fully distributed manner. It is shown that this algorithm converges to the NE of the complete information game.

29.5.3 Spectrum Sharing in OFDM Networks In this section we study spectrum sharing among cognitive radios in OFDMA networks [14]. OFDMA is a transmission technique which divides the available spectrum into sub-carriers and hence can support different QoS by assigning different number of sub-carriers to the users. OFDMA has recently used in the WiMAX (IEEE 802.16 Wireless MAN) uplink. In the following, we consider the situation where several cognitive radios, each having its own QoS constraint in terms of throughput, compete

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

575

for access to the available sub-channels in an OFDMA network. Note that, there is not any spectrum manager (auctioneer) in this OFDMA network.

System Model Consider an OFDMA network consisting of L sub-channels that are shared among K users (cognitive radios). Each user has a given QoS constraint in terms of throughput, (i.e., let Ri , i = 1, . . . , K, be the required transmission rate of user i). Consider a given power allocation given by the matrix P where [P ]il = Pil is the transmission power of user i on sub-channel l, l = 1, . . . , L. For a given power allocation P , Let γil (P ) =

n0 +

P l hl Pi ii j6=i

Pjl hlji

,

l = 1, . . . , L,

(29.13)

be SIN R for user i on sub-channel l, where hlji is the channel gain from user j to user i’s receiver (e.g., user i’s BS) on sub-channel l, and n0 is the background noise power. The rate ril (P ) at which user i can transmit on sub-channel l under the power allocation P is then given by ril (P ) = W log2 (1 + cγil (P )), where W and c are known positive constants that depend on the system parameters. The goal of an OFDMA network provider is to minimize the overall transmission power subject to the given QoS constraints, i.e., min P

 PL l  Pl=1 ri (P ) − Ri ≥ 0, L s.t. P l − Pmax ≤ 0,  ll=1 i Pi ≥ 0,

K X L X

Pil

(29.14)

i=1 l=1

i = 1, . . . , K, i = 1, . . . , K, i = 1, . . . , K, and l = 1, . . . , L.

where Pmax is a constraint on the maximal power at which a user can transmit. For the above optimization problem, we denote with Ω the set of feasible sub-channel l rate allocations. In other words, Ω is the set of all sub-channel rate allocations PL l ri , i = 1, . . . , K and l = 1, . . . , L, which satisfy the QoS constraints that is l=1 ri (P )− Ri ≥ 0 for i = 1, . . . , K and for which there exists a feasible power allocation P such that, ril (P ) = ril for i = 1, . . . , K and l = 1 . . . , L. Note that the set Ω might be empty. Necessary conditions on the QoS constraints Ri , i = 1, . . . , K, for Ω to be non-empty are given in [14].

Game Model The optimization problem given by Equation (29.14) is a generalized knapsack problem and finding an optimal power allocation is NP-hard [14]. Rather than relying on the network operator to decide on a power allocation, suppose that each user i is free to choose its own power allocation with the goal to satisfy the rate constraint Ri with a minimal transmission power. The resulting interaction among users leads to the following non-cooperative game. Let P−i be the sub-matrix indicating the power allocation over all users except user i and let Pi = (Pi1 , . . . , PiL ) be the power allocation of user i. Furthermore, let

576

Mohammad Hossein Manshaei et al. ril (Pi , P−i ) = W log2 1 + cγil (Pi , P−i )

be the transmission rate of user i on sub-channel l, where γil (Pi , P−i ) is the SIN R for user i on sub-channel l under the power allocation given by Pi and P−i as defined above. Given an allocation P−i of all other users, user i chooses a power allocation Pi to minimize its own transmission power subject to the QoS on its transmission rate by solving the following optimization problem, min Pi

L X

Pil

(29.15)

l=1

 PL l  Pl=1 ri (Pi , P−i ) − Ri ≥ 0, L s.t. P l − Pmax ≤ 0,  ll=1 i l = 1, . . . , L. Pi ≥ 0,

Game Results For the above non-cooperative game, Han et al. show in [14] that there always exists a NE if the set Ω is non-empty, i.e., if there exists at least one feasible power allocation that leads to the rate allocation that satisfies the QoS constraints Ri , i = 1, . . . , K. However two difficulties arise in this context: (1) if the required rates Ri , i = 1, . . . , K, are too large, then the set Ω might be empty and no NE exists, and (2) there might exist several NE, some of them with low system and individual performances. To overcome these difficulties, Han et al. consider the use of a virtual referee that is implemented by the network operator. Below we briefly describe the role of the virtual referee. Suppose that when the set Ω is empty and no NE exists, then a user i who is not able to satisfies the QoS constraint Ri without violating the power constraint will decide on a power allocation as follows. Given a power allocation P−i by all other users, user i chooses a power allocation Pi that solves the following optimization problem, L X max ril (Pi , P−i ) (29.16) Pi

PL s.t.

l l=1 Pi l Pi ≥ 0,

l=1

− Pmax ≤ 0, l = 1, . . . , L,

i.e. user i maximizes its throughput subject to the given power constraint. Han et al. show in [14] that there always exists a NE for the non-cooperative game based on the two optimization problems given by Equation (29.15) and (29.16), respectively. However, the game may result in an undesirable NE with low system and individual performances. To improve system performance, the network operator implements a virtual referee that selectively restricts for some users the set of sub-channels that they are allowed to access. Han et al. provide in [14] the rules that the virtual referee uses to limit sub-channel access, one rule being that each user must have access to at least one sub-channel. Through numerical results, Han et al. illustrate in [14] that the resulting spectrum sharing mechanisms can significantly outperform two benchmark mechanisms where each sub-channels is allocated to at most one user (no sub-channel sharing) and where all users access all sub-channels (complete sub-channel sharing).

29 Spectrum Sharing Games of Network Operators and Cognitive Radios

577

29.6 Conclusion In this chapter, we have provided a detailed description of several research contributions in the area of spectrum sharing games. More specifically, we have focused on two kinds of players: network operators and cognitive radios. The reason of this choice is that we believe that the described scenarios will be among the most relevant ones in the coming years. They provide insights on the possible consequences of greedy behavior in the either. As we have explained, most of the cases that we have described are amenable to modeling by game theory. In this way, it is possible to predict potential outcomes of the observed conflicting situations. But as we have mentioned, in order to make the problem tractable, all authors have made relatively drastic assumptions, notably in terms of information possessed by the players. Much research is still needed in this field, in particular to better capture the perception that each of the players has of the context in which it operates.

References 1. R. Axelrod. The Evolution of Cooperation. Basic Books, New York, 1984. 2. L. Buttyan and J.-P. Hubaux. Security and Cooperation in Wireless Networks, Thwarting Malicious and Selfish Behavior in the Age of Ubiquitous Computing. Cambridge University Press, 2007. 3. L. Cao and H. Zheng. Distributed Spectrum Allocation via Local Bargaining. In SECON, 2005. 4. R. Etkin, A. P. Parekh, and D. Tse. Spectrum Sharing in Unlicensed Bands. IEEE JSAC on Adaptive, Spectrum Agile and Cognitive Wireless Networks, April 2007. 5. Federal Communications Commission. Amendment of the Commission’s Rules to Establish New Personal Communications Services. Technical Report 90-314, Federal Communications Commission, 1994. 6. M. F´elegyh´ azi. Non-cooperative Behavior in Wireless Networks. PhD thesis, EPFL – Switzerland, April 2007. 7. M. F´elegyh´ azi, M. Cagalj, D. Dufour, and J.-P. Hubaux. Border Games in Cellular Networks. In IEEE INFOCOM, 2007. 8. M. F´elegyh´ azi and J.-P. Hubaux. Wireless Operators in a Shared Spectrum. In IEEE INFOCOM, April 23-29 2006. 9. B. Fette. Cognitive Radio Technology. Newnes, 2006. 10. D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991. 11. M. R. Garey and D. S. Johnson. Computers and Intractability - A Guide to the Theory of NP-Completeness. W. H. Freeman, 1979. 12. R. Gibbons. A Primer in Game Theory. Prentice Hall, 1992. 13. M. M. Halldorsson, J. Y. Halpern, L. E. Li, and V. S. Mirrokni. On Spectrum Sharing Games. In ACM PODC, 2004. 14. Z. Han, Z. Ji, and K. J. R. Liu. Non-Cooperative Resource Competition Game by Virtual Referee in Multi-Cell OFDMA Networks. IEEE JSAC, Non-cooperative Behavior in Networking, (2nd Quarter), 2007. 15. G. Hardin. The Tragedy of the Commons. Science, 162:1243–1248, 1968. 16. S. Haykin. Cognitive Radio: Brain-Empowered Wireless Communications. JSAC, 23(2):201–220, February 2005.

578

Mohammad Hossein Manshaei et al.

17. J. Huang, R. Berry, and M. L. Honig. Auction-based Spectrum Sharing. ACM/Springer (MONET), 11(3):405–418, June 2006. 18. X. Huang and B. Bensaou. On Max-min Fairness and Scheduling in Wireless Adhoc Networks: Analytical Framework and Implementation. In ACM Mobicom, April 2001. 19. R. E. Hundt and G. L. Rosston. Spectrum flexibility will promote competition and the publicinterest. IEEE Communications Magazine, 33, December 1995. 20. J. Mitola III. Cognitive Radio Architecture: The Engineering Foundations of Radio XML. Wiley, 2006. 21. E. Koutsoupias and C-H. Papadimitriou. Worst-case equilibria. In 16th Annual Conf. Theoretical Aspects of Computer Science, 1999. 22. A. Mas-Colell, M. D. Whinston, and J. R. Green. Microeconomic Theory. Oxford Univ. Press, 1995. 23. J. Nash. Equilibrium Points in N -person Games. Proceedings of the National Academy of Sciences, 36:48–49, 1950. 24. J. Nash. Non-Cooperative Games. The Annals of Mathematics, 54(2):286–295, 1951. 25. M. J. Osborne and A. Rubinstein. A Course in Game Theory. The MIT Press, Cambridge, MA, 1994. 26. C. Peng, H. Zheng, and B. Y. Zhao. Utilization and Fairness in Spectrum Assignement for Opportunistic Spectrum Access. In Mobile Networks and Applications, 2006. 27. T. S. Rappaport. Wireless Communications: Principles and Practice (2nd Edition). Prentice Hall, 2002. 28. Shared Spectrum Company SSC. Dynamic Spectrum Sharing. In Presentation to IEEE Communications Society, 2005. 29. T. A. Weiss and F. K. Jondral. Spectrum Pooling: An Innovative Strategy for the Enhancement of Spectrum Efficiency. Communications Magazine, IEEE, 42, 2004. 30. A. Zemlianov and G. de Veciana. Cooperation and Decision Making in Wireless Multi-provider Setting. In INFOCOM, 2005. 31. J. Zhao, H. Zheng, and G. H. Yang. Distributed Coordination in Dynamic Spectrum Allocation Networks. In DySPAN, 2005. 32. H. Zheng and L. Cao. Device-centric Spectrum Management. In DySPAN, 2005.

30 Introduction to NetLogo A Powerful Programming Tool for Modeling Cooperative Interactions in Wireless Networks

Federico Albiero1 , Frank H.P. Fitzek2 , and Marcos Katz1 1

2

VTT - Technical Research Centre of Finland [Federico.Albiero|Markos.Katz]@vtt.fi Aalborg University [email protected]

Summary. NetLogo is a multi-agent modeling environment originally conceived by the complex systems community for simulating natural and social phenomena. Despite its simplicity, NetLogo is a powerful tool in many fields of research. It is especially well suited for modeling large collections of independent agents developing over time, thus being a promising solution for simulating and analyzing distributed systems such as modern wireless networks. In this chapter we illustrate the feasibility of this approach, by providing a detailed tutorial about the application of NetLogo to wireless communications.

30.1 Why NetLogo Based on a JAVA platform, NetLogo is a multi-agent modeling language developed at the Center for Connected Learning (CCL) and Computer-Based Modeling of the Northwestern University of Evanston, Illinois (US), and freely available on the web [11]. This programming language was originally conceived in order to provide a simple but powerful tool for modeling and exploring emergent phenomena [12], namely large-scale patterns that arise out of the complex interactions of numerous interdependent micro-“agents” at a lower level scale. Many examples of these emerging patterns can be found in nature and social sciences. Although NetLogo was not designed specifically for telecommunications, in principle a countless number of potential applications can be found. As a matter of fact, the environment appears very suitable for modeling and analyzing rich and dynamic interactions in wireless networks. Specifically, this chapter discusses the use of NetLogo to model distributed (e.g., ad hoc) and centralized (e.g., cellular) networks. Furthermore, it can be used to study wireless networks from a game-theoretic perspective. Basically, NetLogo is particularly convenient for the analysis of any complex system developing over time, as the programmer can give instructions to thousands of independent agents all operating concurrently. Such a scenario is perfectly matching with the picture of a wireless AHN (Ad Hoc Network) as a collection of terminals, either static or mobile, without any fixed infrastructure. In this perspective,

579 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 579–602. c 2007 Springer.

580

Federico Albiero et al.

NetLogo may be used for simulating complex network systems according to specific settings in order to provide an estimation of the parameters of interest (i.e., troughput, power consumption, spectral efficiency, etc.), whereas on the other hand a theoretical analysis could be highly complicated.

30.2 Main Features The agents in NetLogo are called turtles. Each turtle holds some predefined variables such as position, direction, label, etc.; additional inner data fields can be defined by the user. Turtles can be placed at any position in a Virtual World and let them interact according to user-implemented subroutines which define their behavior. Then, the results of the simulation can be collected and displayed in real time with monitors and charts in the program graphic interface. An example is given in Figure 30.1.

Figure 30.1. Example of NetLogo Graphic User Interface (GUI). The Figure shows a library model of Gas-in-a-Box.

As it can be noticed, NetLogo’s layout is formed by a main graphic window (the Virtual World ) together with a collection of buttons, sliders, monitors and charts, all in the same interface. In the upper-left corner, the user can set the model parameters and run the simulation. On the right side of Figure 30.1, is the graphic window showing NetLogo’s agents interacting in the Virtual World. In addition, to the bottom side, is a command-center toolbar for interactively running commands and debugging (not displayed in figure). Finally, the user can easily handle data and visual results of the simulation, with the aid of monitors and dynamic charts.

30 Introduction to NetLogo

581

The model illustrated in Figure 30.1 was taken from the Models Library included in NetLogo. In fact, a major key-feature of the programming environment is its extensive documentation and tutorial section. An exhaustive user’s manual [5] is included in the package, providing a general overview, some tutorials, and several guides. Particularly useful is the Programming Guide, explaining the most important features, and the Primitive Dictionary, including a detailed description of all commands. Moreover, the embedded Models Library provides a large collection of pre-written simulations and Code Examples which can be used and modified. With this “extensible modeling” approach, even beginners can easily get familiar with the language and acquire skills in order to build their own models. Finally, the significant advantage of NetLogo for the wireless communications engineer can be summarized as the following: it is simple and easy-to-use, but it is advanced enough to serve as a powerful tool for analysis and simulation of complex distributed systems such as ad hoc mobile networks. In this way, the researcher can quickly develop his models without being excessively concerned on the programming and debugging practice. Although a complete description of the environment is beyond the scope of this chapter, here is a list of the most interesting features (for further details, refer to the User Manual [5]): • • • • • • •

Simple language structure Unlimited number of agents and variables Large vocabulary of built-in language primitives Extensive Models Library with Code Examples Friendly and multi-purpose Model’s Interface Monitors for inspecting and controlling agents Info tab for annotating variables and debugging

We have seen so far the major characteristics of NetLogo. However, how it can be applied in the field of communication technology? After a brief introduction given in Section 30.3, this topic will be addressed in Section 30.4 and the following.

30.3 Getting Started Perhaps the most noticeable aspect of NetLogo is that it belongs to the class of so-called multi-agent modeling languages (also known as object-based parallel modeling), recently developed by the complex systems community. In the following, we will give some insights to the concept of parallel modeling through a simple illustrative example. As already mentioned, NetLogo’s agents are called turtles. The first thing we may need when starting a model is to create these agents populating the Virtual World. For instance, if we type on the command-line (interface bottom side) the statement: c r e a t e −t u r t l e s 100 hundred turtles will be created and placed in the middle of the screen, as shown in Figure 30.2(a). The default shape of a turtle is a triangle; however, several shapes are available, included user-defined ones. Note that the gray borders of the screen have been added for visual purposes: the reader can skip this detail for the moment.

582

Federico Albiero et al.

(a) create-turtles (mobile phones)

100.

(b) ask turtles [forward 10].

(c) ask turtles [back random 10].

Figure 30.2. Example of a simple command sequence.

Now that we have our turtles, we can start to “orchestrate” the new agents by defining their behavior. Since the world of NetLogo is modeled by the interaction between independent individuals, this can be accomplished by “gently” asking the turtles to do a specific task, rather than giving system commands. For example, if we type: ask t u r t l e s [ f o r w a r d 1 0 ] the 100 turtles move forward by 10 screen units, laying on a circle as displayed by Figure 30.2(b). The reason is that each single agent is initiated by default with a different inner variable heading (i.e., angle): therefore, even though the agents receive the same command, they are oriented in different directions and move as a consequence. The resulting circle provides a simple illustration of emergent global phenomenon arising out of the behavior of single individuals, thus providing more insight on the concept of agent-based modeling. In this case a coherent circle has been formed as the number of turtles is large enough to let them uniformly distribute in all directions. We underline that the same result could be obtained by using the circle equation x2 + y 2 = R2 to set the agents coordinates, but the implementation turns out to be significantly more complex. As a matter of fact, the real potential of NetLogo comes from the point that each turtle is conceived as an independent agent. As a consequence, the response of turtles to the same command statement can vary markedly, depending on the inner features of each individual entity. In this perspective, the user can model the agents and their related behavior in a fully programmable object-oriented way. Then, once the rules are set, the resulting simulation is determined by the turtles’ interaction. Moreover, NetLogo turtles are not only independent agents but they also act in parallel, in the sense that each turtle does its own separate computations at the same time. As an example, if we state now: ask t u r t l e s [ back random 1 0 ] turtles will go back with a random value of steps between 0 and 10. Since each turtle makes its own calculation, each one gets a different number of back steps, and the result is that they spread randomly inside the circle of radius 10. An illustration is provided by Figure 30.2(c).

30 Introduction to NetLogo

583

After this short introduction, we start modeling a wireless AHN, whereas our agents are mobile terminals and communication links. This goal requires some more programming practice and will be addressed step by step. An introduction to a game-theoretic model for cooperation in AHNs will be provided in Section 30.4. Then, a basic NetLogo library specifically designed for wireless communications will be presented and discussed in Section 30.5.

30.4 A Model for Cooperation in Wireless Networks In this section we will explore the potentials of NetLogo for the field of wireless communications. In particular, the possibility to give parallel instructions to a large number of independent agents, all operating concurrently, makes the environment especially suited for modeling ad hoc networks; and, in general, each network of terminals operating without central authority. More specifically, a good way is provided for analyzing complex systems in which the topology is not well defined and may change over time, such as a dynamic mobile network. We will show here that NetLogo can be used as a powerful tool for developing models of wireless networks based on specific purposes. Simulations can be used in order to provide an estimation of the performance figures of interest (as, e.g., throughput, power consumption, spectral efficiency, etc.) according to initial predefined settings. The importance of this approach has to be stressed for all scenarios characterized by high mobility within the users and/or network topology evolving over time. In fact, in those conditions a theoretical analysis is not straightforward, involving onerous mathematics and calculations. Furthermore, the task of finding a good model likely to predict the outcome of such complex systems is also rather troublesome in many cases. This is especially true when dealing with emerging realities such as wireless AHN, where terminals form a network with no fixed infrastructure. In order to perform proper communication, nodes need to forward traffic to the others: therefore some kind of agreement between terminals is required, by means of cooperation. Although cooperation in wireless networks is receiving increasingly attention in the recent literature, the results are usually derived from a theoretic standpoint. The aim of this Chapter is to motivate researchers and communication engineers to consider the analysis of network issues from a different point of view, by using NetLogo for agent-based simulation.

30.4.1 The PD Iterated Example Cooperation in AHNs can be modeled as a game involving independent decisionmaking. According to this perspective, game theory has gained growing importance in the telecommunication research over the last few years, as it provides a good theoretical framework for analyzing distributed networks and the involved decisionmaking strategies. Furthermore, as one can notice, NetLogo’s world of individual agents matches perfectly with the picture of an AHN as a collection of nodes all operating independently. Thus, we may choose to model the original agents (turtles) as wireless terminals and exploit the agent-based properties of the language in order to simulate the

584

Federico Albiero et al.

network operation, heavily dependent on the interaction among individuals. In this way, the resulting performance of the simulation (i.e., cooperation) may arise out of a countless number of single decisions that nodes are making at any time, according to one’s specific perspective on the current network state. Therefore, NetLogo can provide a very good way for modeling cooperation from individual decision-making. However, as we may need a mathematical background to substantiate our analysis, how the concepts of game theory can be applied in designing a model for wireless AHNs? Some useful guidelines can be found in the Models Library, providing the basic know-how for further developments.

Figure 30.3. NetLogo’s Library Model of Iterated Prisoner’s Dilemma.

A very good example for getting started with game theory is the classic game of the Prisoner’s Dilemma (PD). As we are not going into details, the interested reader may refer to [8] and [9]. We take a glance here at its N -Person Iterated version (IPD) as it gets closer to our final objective of modeling cooperation in wireless AHNs. An illustration of the model’s layout in NetLogo is given in Figure 30.3 (both models of PD and IPD are readily available at the Models Library, selected from the menu File1 ). Specifically, the IPD simulation shows the results analyzed at first by Axelrod [6] in 1984, within his contest for the best cooperative strategy. The NetLogo model allows the user to set the number of units for several different strategies. Then, the 1

In the Models Library, the appropriate path for a quick search is: Sample Models –> Social Science –> Unverified –> Prisoner’s Dilemma.

30 Introduction to NetLogo

585

agents are placed at random positions (setup) and the simulation can start (go). One interesting possibility is to run the program step-by-step (go-once) in order to see the players’ payoffs at each iteration. Notice that, although the overall number of players is N, only two-persons games are considered in the model. As a matter of fact, every time two units get in close proximity, a classic PD game is played according to the strategies of the players and the corresponding payoffs are set (see the numbers on top of turtles). Then, the units move randomly one step and the procedure is repeated. The resulting average payoff per strategy can be checked in real time by means of a dynamic chart. A very interesting test can be performed by selecting different strategies and let the model run for a while. In this way, one can easily experience the well known outcome of tit-for-tat being the most effective strategy for cooperation (the strategy simply reciprocates the opponent’s last move). A more detailed analysis of the model performance has been given in [7]. A good tutorial in order to achieve familiarity with NetLogo and game theory might be to implement a new strategy, as suggested in the model itself. For instance, the reader may choose a slightly more complicated strategy, as tit-for-two-tats. In this case, the players’ decisions whether to cooperate or defect are based on the opponent last two moves, instead of the only last move: therefore a mechanism is needed to store some additional information for each agent. This task can be accomplished by modifying the turtle inner variable partner-history. In the standard IPD model, the partner-history list allows the agents to keep track of the last interaction with every other turtles (indexed by “who”2 values), if previously encountered. As a consequence, tit-for-tat is straightforward, being a simple repetition of the opponent’s last move. On the other hand, the implementation of tit-for-two-tats is a bit more challenging. One possible solution, though not unique, can be performed by using list of lists. Basically, we can think at the variable partner-list as a list of references to other lists move-list, which store the past interactions of current player with other players previously encountered. In this way it might be easy to try out even more complex submissions based on the current and past state of the system. One pitfall to be aware is the NetLogo programming syntax when managing lists, as the language is mainly thought for controlling agents. For instance, a code example taken from the IPD model is listed below: l e t default −l i s t [ ] r e p e a t num−t u r t l e s [ set default −l i s t ( fput f a l s e default −l i s t ) ] ask t u r t l e s [ s e t p a r t n e r −h i s t o r y d e f a u l t − l i s t ] Notice the particular structure wherein, at first the temporary variable defaultlist is initialized (let is used to define local variables) and modified (with primitive fput), then every agent is given a copy of this list which is stored in the partnerhistory internal field. If we use list of lists, a similar operation must be accomplished every time access is needed to the move-history of each player. For further details about lists, see the Programming Guide and the Primitive’s Dictionary from the user’s manual [5]. In conclusion, as the source code of IPD model is fully available from the Models Library, the reader shall find it rather easy to modify in order to implement his 2

Turtle ID number, a built-in variable of NetLogo.

586

Federico Albiero et al.

personal strategies. Presumably, the resulting scenario will still lead to tit-for-tat being the most effective strategy for cooperation.

30.4.2 Modeling Cooperation in Wireless Networks The IPD model was illustrated here in order to familiarize the reader with the multi-agents world of NetLogo and, at the same time, to get some insights into a classic result of game theory, i.e., cooperation between independent individuals. Now we should have a proper background to start the development of new models for simulating ad hoc networks. Our proposed goal is to model the decision-making process between the single terminals, whereas each one has a unique perspective on the current network state and each one makes its own decisions about related issues as transmit power, packet forwarding, delay time, and so on. With such a model we should be able to provide an answer to the fundamental question: is cooperation really possible in a world of individuals? And, more specifically: how cooperation can emerge as the optimal solution in the multi-shaped reality of AHNs?

Figure 30.4. Modeling cooperation in wireless communications.

As a matter of fact, it can be easily argued that the IPD framework considered so far is widely insufficient for the purpose of modeling cooperation in wireless networks. The main reason lies behind the concept itself of a network as a collection of nodes sharing resources to achieve a common goal, i.e., perform efficient communication among connected entities. In real applications, thousands of nodes are likely to use

30 Introduction to NetLogo

587

the network services at the same time, sending and receiving data among each others and through the base station. In addition, we take into account here the cooperative framework discussed in [7]. Theoretic analysis and tests in different scenarios have shown the benefit of cooperation to be increasing with the number of units in a cluster. Therefore, in a model for cooperation in AHNs, multi-player games rather than two-person games shall be considered. In general, the natural objective that we pursue in our model of wireless cooperation is: given network topology and nodes preferences, find the best cooperative clusters for a predefined goal (for instance, minimum power consumption). An illustration of modeling with NetLogo according to these principles is given in Figure 30.4. Perhaps, considering this scenario, one may ask if we are not simply dealing with a problem of optimization. This concern is very well addressed in [3], an excellent tutorial and summary on game theory for wireless engineers. As a matter of fact, it shall be pointed out that the introduced game-theoretic approach completely holds its validity here, as we are dealing with individual decision-making with no central authority. Furthermore, even in the assumption that nodes are sharing objectives (which can be possible in our case), they each have a unique perspective on the current network state. Therefore, the tools of game theory can provide a useful background in order to properly model the nodes decision-making process towards cooperation.

30.5 NetLogo Libraries for Cooperation As we have seen so far, NetLogo offers the appropriate “bricks” to build our model for cooperation in AHNs with a game-theoretic framework. This section explains how this goal can be achieved by means of a detailed tutorial, which can be divided in three parts. The first part, given in Section 30.5.1, is focused on the development of a simple model in order to create nodes and links, and make them move and update dynamically. Then, in Section 30.5.2, cooperative games among nodes are designed with the ultimate objective of analyzing the system performance related to a chosen network parameter (here we refer to power consumption, but several choices are possible). Finally, in Section 30.5.3, we explain how to collect important data and results from the simulation and how to present them on charts or plots. A few subroutines are not implemented here, but the full code is available at ().

30.5.1 Model Setup Initialization and Breeds When writing a program in NetLogo, first of all some declarations are needed in order to specify typology of agents, global variables, and agents’ internal variables. In our model for wireless AHNs, we implement two different kinds of agents: nodes and links. To define these agentsets, the keyword breed is used. The statement: b r e e d [ nodes node ] breed [ l i n k s l i n k ]

588

Federico Albiero et al.

Figure 30.5. Main features of the NetLogo model for analyzing cooperative power savings. The figure shows the agents internal implementation and the algorithm flow-chart with related programming subroutines.

defines the two races above. Notice that the first input defines the name of the agentset associated with the breed, while the second defines the name of a single member of the breed. In this way, agents belonging to different breeds can be created with specific internal features (as will be explained later on). The primitive ask can be used for giving commands only to the members of a certain breed. For instance, the sequence: c r e a t e −nodes < number−o f −nodes > c r e a t e −l i n k s < number−o f −l i n k s > ask nodes [ s e t c o l o r g r e e n ] creates a specified number of agents, either nodes or links (which internal features are not defined yet); then, only the turtles belonging to the class “nodes” set their color to green. The use of agentsets is very common in NetLogo. We underline that the chosen typologies of agents are arbitrary: the programmer may choose to implement any breeds according to particular goals (such as APs, USB devices, PDAs, and so on). After the breeds declaration, the keyword globals is used to define global variables which are visible to the observer (i.e., the user of the model, as often referred in NetLogo). For example: globals [ n−o f −t e r m i n a l s

30 Introduction to NetLogo

589

n−o f −l i n k s iteration ] defines three global variables that can be accessed and modified through the usual primitive set. Moreover, the keyword -own in conjunction with a specific agentset allows to define the turtles inner variables. A possible choice for the design of wireless nodes might be: nodes−own [ n e i g h b o r −nodes c u r r −l i n k s power ] The intended meaning of each field is listed below: neighbor-nodes agentset of surrounding nodes within a maximum predefined range; curr-links list of the links from the node to the current reachable terminals (i.e., neighbor nodes); power instantaneous normalized power consumption of the terminal. It shall be noticed that, differently from other languages, the variable types needs not to be declared at this point (e.g., in nodes-own: first field is an agentset, second is a list and third a floating-point number). In fact, NetLogo allows to define variables on-the-fly when they are firstly initialized.

Screen Layout and Nodes Setup Now we can make the screen layout and setup the nodes in NetLogo’s Virtual World. To address the first concern, the designer shall think about the real situation to be modeled. Since in this tutorial we aim to develop a dynamic model for cooperation in a wireless mobile network, we are not concerned about layout details. Therefore we choose here a basic setting represented by a square box (modeling, for instance, people walking in a room with wireless handheld devices). The screen edges, limiting the square box, can be easily created by using NetLogo patches 3 . For example, we can start to write our first simple routine, setup: to setup ca ask p a t c h e s with [ abs pxcor = max−p xcor o r abs p ycor = max−p ycor ] [ s e t p c o l o r gray ] end Notice the use of keywords to and end in order to begin and close the procedure, respectively (ca is used to clear screen and variables). A very important feature is the use of with in conjunction with a specific agentset. The structure: 3

NetLogo’s world is two dimensional and is divided up into a grid of patches. Each patch is a square piece of “ground” over which turtles can move.

590

Federico Albiero et al. ask < a g e n t s e t > with [ r e p o r t e r ] [ command ]

is perhaps the most common expression in NetLogo. In fact, by using with, the programmer can select for an action all and only those agents reporting a specified feature, either being this feature a built-in variable or a user-defined one. For instance, the ask statement in subroutine setup calls only the screen patches having maximum values for the x and y coordinates and set their color to gray. As a consequence, a boundary box is drawn as soon as the procedure is executed. Normally, this task is accomplished by pushing a button with the same label of the subroutine in the program interface: hence we need to add such a button into our model. Note the strong connection existing between the visual interface and the hidden implementation, which can be developed almost in parallel. Therefore the user has a high level of freedom in choosing his personal way of modeling. However, we argue that a good programming methodology for NetLogo is probably working at the interface and the code mostly at the same time, so that changes are immediately mirrored from one side to the other, and errors are easier to detect. Let us now create the wireless nodes and place them inside our square box. Here we choose an initial random distribution, implemented by the following routine setnodes. t o s e t −nodes s e t c o l o r green s e t c u r r −l i n k s [ ] s e t x y max−p x cor max−p ycor w h i l e [ p c o l o r = gray ] [ s e t x y random−f l o a t world−width random−f l o a t world−h e i g h t ] end A simple way to setup the nodes at random positions in the box is provided by using the command setxy together with random-float (the while loop ensures that nodes are not placed on the gray borders). Additionally, the variable curr-links is initialized as an empty list. Again, a further button “set-nodes” is required for starting the routine in the program interface. Notice that, if one tries to run the program, at this point a run-time error occurs. This is because we are attempting to run an agent-only routine in a observer (“external”) standpoint. In fact, the considered routine set-nodes does not manage global variables but it’s rather intended to modify the turtles internal fields. Therefore, it has to be called in a turtle (“internal”) context by using the ask statement. For instance we can add to the setup function the following lines: c r e a t e −nodes 20 ask nodes [ s e t −nodes ] which result in creating 20 agents and making them execute routine set-nodes. Figure 30.6 provides an illustration of the resulting model layout achieved so far. In general, it would be more desirable to have a variable number of nodes which can be set by the user as an initial parameter for the simulation. We simply need to add a slider n-of-terminals to the user’s interface and define a global variable with the same label in section “Procedures” (already present in previous code example). Now we just have to type create-nodes n-of-terminals, and we got the desired effect. This example shows the simplicity of managing data I/O with NetLogo.

30 Introduction to NetLogo

591

Figure 30.6. Initial screen layout of model for wireless cooperation, with 20 nodes placed at random positions.

Links and Data Rates As it can be found in NetLogo user’s manual, links is a new item currently under development. However, we have already implemented this feature in [4], and a brief explanation is provided here as it can give some useful insights for the programming practice. A link is conceived as a special turtle connecting two other turtles (nodes). The inner variable “size” of a link is always equal to the distance between the two nodes. Its “heading” is always equal to the direction from one node to the other. Its location is always halfway between the two nodes. In principle, there are two different sorts of links: directed and undirected. Since we assume full duplex communication is available, we choose to model undirected links. For this purpose, each node is given a list of the current available links stored in variable curr-links. A good model of a mobile network should take into account that not all the terminals can communicate with each others at any moment, due to statistical behavior at the associated radio channel. Specifically, in this tutorial we assume that a rate-adaptive protocol-based mechanism is available, as suggested by standards IEEE11.a/g [1], [2] for WLAN technology. As a consequence, different data rates are associate to the links depending on distance among terminals. If the distance is too big, the connection is not available. Although we consider here only the reciprocal distance between terminals, extending the model with fading probabilities is rather straightforward. The implemented mechanism to check the link availability is accomplished by node routine find-neighborhood. The subroutine provides a simple way to find the current reachable terminals in maximum predefined range max-range. An illustration is provided in Figure 30.7. t o f i n d −n e i g h b o r h o o d l e t c u r r e n t −node s e l f

592

Federico Albiero et al. l e t n e i g h b o r h o o d nodes in−r a d i u s −nowrap max−r a n g e l e t n e i g h b o r −nodes n e i g h b o r h o o d with [ s e l f != c u r r e n t −node ] i f ( n e i g h b o r −nodes != nobody ) [ ask n e i g h b o r −nodes [ form−l i n k c u r r e n t −node ] ] end

By calling find-neighborhood (with ask statement) in the main program, each node looks for other nodes in close proximity and then set the wireless links if any reachable terminal is found. The last operation is accomplished by function formlink.

Figure 30.7. Example of routine find-neighborhood. Current-node C finds three neighbor-nodes within maximum range. Since procedure find-neighborhood is executed by all nodes at the same time, a check is needed in order to avoid setting multiple connections. Therefore routine form-link is just a simple call to function check-connection; which, in turn, exploits the service provided by connected-to?. We omit the details of this implementation. If no links are already present, a new link is created with the following procedure, connect-to. t o connect −t o [ c u r r −node ] hatch 1 [ s e t breed l i n k s s e t c o l o r gray s e t a c u r r −node set b myself s e t ( c u r r −l i n k s −o f a ) l p u t s e l f ( c u r r −l i n k s −o f a ) s e t ( c u r r −l i n k s −o f b ) l p u t s e l f ( c u r r −l i n k s −o f b ) reposition ]

30 Introduction to NetLogo

593

end Some explanations are in place here. The command hatch is used to create a new agent “link” associate with the agent “node” which is currently calling the procedure. Then, the ending points of the link, a and b, are defined as curr-node (the node looking for neighbors) and myself (the neighboring node currently executing the procedure)4 . Finally, the new link, identified by keyword self, is added to the list curr-links of both connecting nodes and routine reposition is called. An example is shown in Figure 30.8.

Figure 30.8. Example of routine connect-to. Since no connections are present, neighbor-nodes {N1 , N2 , N3 } are asked to set the links to current-node C. Here, for instance, N2 executes the routine and the link l2 = {C, N2 } is created. Then, node lists of current links are updated. We underline the difference between self and myself. “Self” is simple; it means “me”: i.e., “the agent who is executing this command”. Instead, “myself” means “the agent who asked me to do what I’m doing right now”. In other words, when an agent has been asked to run some code, using myself in that code reports the agent that did the asking. Such distinction is very important, as many times the NetLogo programmer is facing these structures. Now routine reposition can be used to set and display the new links according to the given definitions about coordinates, heading and size. to r e p o s i t i o n s e t x y ( xcor−o f a ) ( ycor−o f a ) s e t h e a d i n g towards−nowrap b s e t s i z e d i s t a n c e −nowrap b 4

Notice that routine form-link is executed by neighbor-nodes, with curr-node as input parameter.

594

Federico Albiero et al. jump ( d i s t a n c e −nowrap b ) / 2 end

If we assume that the nodes are moving, it is very easy to update heading, size, and location of the links by simply calling the procedure in this way: ask l i n k s [ r e p o s i t i o n ] and in one line of code we achieve the proposed goal. Thus, if from one side the programming structures related to the agent-based properties of NetLogo are not straightforward, on the other hand the language is shown to be significantly powerful when controlling at the same time large collections of agents (such as the number of nodes and links in AHNs can be). As a matter of fact, a case of particular interest arises when the number of entities involved in the simulation becomes large. In this view, NetLogo is a flexible tool for the researcher in order to test new ideas in the complex multi-shaped reality of wireless systems, whereas intuition provides no more clues on the situation and the possible implications of interaction among individual nodes.

Figure 30.9. Screen layout of model for wireless cooperation including links and data rates according to WLAN specifications.

The implementation of routine set-rates, which assigns different data rates to the communication links depending on distance, is left as an exercise to the reader. This operation can be done by using the link labels. An illustration of the visual result is provided in Figure 30.9, assuming the communication between terminals is performed with WLAN technology.

Mobility Model While designing a dynamic simulation, the chosen mobility model shall be carefully analyzed as it can deeply influence the final results. However, these considerations

30 Introduction to NetLogo

595

are beyond the scope of this chapter (an analysis of the impact of mobility when simulating AHNs is provided in [10]). Here we choose to focus on a simple model, leaving the details about mobility for further developments. The following routine move is used to perform a random movement of the terminals in NetLogo’s Virtual World. t o move r t ( random−f l o a t 10 − random−f l o a t 1 0 ) w h i l e [ p c o l o r −o f patch−ahead 1 = gray ] [ r t random−f l o a t 360 ] f d random−f l o a t 0 . 1 end From the code listing, it can be noticed that the movement comes from the combination of two actions performed by the commands rt and fd (rotate and forward, respectively). In fact, first a random angle is selected in range [−10◦ , +10◦ ] and the units rotate of that amount. Then, they move forward of a random distance chosen in range [0; 0.1]. The while loop ensures that no units are crossing the boundary box: they simply bounce back when finding a wall, by turning around of a random angle. Once again, at this point we can make the units move by simply adding an ask statement into the main program, that we call go. Here a code-listing is provided. t o go set iteration iteration + 1 ask nodes [ move ] ask nodes [ f i n d −n e i g h b o r h o o d ] ask nodes [ s e a r c h −c l u s t e r s ] ask nodes [ play−coop ] c o l l e c t −data d i s p l a y −r e s u l t s end The global variable iteration keeps track of the number of times the program is repeated. The last four operations, not considered yet, will be described in the following sections.

30.5.2 Cooperative Games In this section, a basic methodology is given in order to apply the game theoretic approach previously introduced to the world of wireless communications by using NetLogo. Specifically, the object of our analysis is cooperation in wireless networks and the proposed goal is power saving. Notice that we rely on the multicast TDMA scenario discussed in [7], thus the cooperative games we model are not related in principle with the concepts of cooperative game theory. In this context, the natural question arises: how the node’s actions can be modeled as an independent decisionmaking process? Here a countless number of choices are available, at least as many as the works in progress in this field at the moment. However, we aim to provide the reader with some useful guidelines in order to build his own models. As pointed out in

596

Federico Albiero et al.

Section 30.4.2, for analyzing cooperation in wireless networks we need to model multi-player games. As we choose to focus on power consumption, each node is given a variable power representing the instantaneous normalized power consumption. In the visual interface, this value is displayed by the agents’ labels (small numbers on top of each terminal, set to unit in case of normal non-cooperative operation). For every iteration of the program illustrated so far (first part of subroutine go) nodes move around randomly, find current “neighborhood”, and set the communication links. At this point the terminals should have enough information to group in cooperative clusters in order to achieve lower power consumption, given the payoffs of cooperation are defined and accessible to every player5 . This is a very important assumption to be remarked, as we are not dealing in this scenario with malicious nodes and not even “blinded” terminals that would not cooperate even though they can get some benefit (only reliable and smart terminals are considered). In this game, only players with full knowledge about the benefits of cooperation are considered. Of course, many other solutions are possible. In these assumptions, we can start to design cooperative games. An exhaustive description of such a modeling task is beyond the scope of this Chapter, thus we focus here on the programming structure. In principle, the framework we choose to develop for the players’ decision-making is very similar to the one provided in the IPD model, but the fact that we are dealing with multiple decisions. Even though the current scenario of investigation is completely different, NetLogo is perfectly suited to take inspiration from this model in order to accomplish our needs. First of all, we assume our terminals to be “smart”, in the sense that they are able to consider all possible groups of connected units within the maximum predefined range6 . This operation is accomplished by algorithm search-clusters, which stores all potential clusters in the internal variable clusters-list. As usual, the routine is performed by every unit. Once the available clusters are found, the units start to play in order to establish cooperative groups. Here comes the game-theoretic approach, as the simulation is based on the independent decision-making of single individuals. Nevertheless, in this scenario, games are modeled in order to avoid players’ simultaneous decisions7 as it makes non sense for our quest of cooperative power savings. Instead, each time a game is performed (i.e., a cooperative group is tested out), the current player knows the decisions of the other players involved in the game. In our model, the current player sends an invitation to the other units of the potential group whereas a good opportunity is found. After that, a new cooperative cluster is established only if all considered terminals give a positive response about the current opportunity. Figure 30.10 provides an illustration of this process. Notice that some terminals might be already cooperating in a group and can decide to join

5

6 7

For example, this goal might be achieved by means of a global payoff matrix reporting the power consumed by a single terminal cooperating in a cluster of given number of units at a given data rate. We leave behind further assumptions related to the way these data are derived from analytic calculations. In general, this parameter is mapped by the technology used for communication. Notice that we use simultaneous by means of “not knowing what the other players do at the same time”.

30 Introduction to NetLogo

597

Figure 30.10. Model of invitation for cooperative games.

a new group depending on the specific circumstances. Therefore cooperative clusters can be formed and dismissed dynamically, improving the system performance. In practice, when terminals are asked to join a new group, the implemented mechanism for decision-making is analogous to the one of the IPD: first players select an action depending on strategy (another internal variable strategy can be easily added to nodes-own) and then they perform the selected action. In our model, the internal field “answer” is set according to the current individual decision about cooperation. As an example, given the payoff for the considered cooperative cluster is available, a basic strategy might be “Selfish-Cooperation”. The implementation with NetLogo is shown by the following routine selfish: t o s e l f i s h [ c u r r −p a y o f f ] i f e l s e ( c u r r −p a y o f f < l a b e l ) [ s e t answer t r u e ] [ s e t answer f a l s e ] end The strategy is very easy. The unit simply checks its label, representing the actual power consumption, and compare it to the current opportunity of cooperation. This represents the rational, self-regarding behavior - “If I gain I will join you!” - of individuals which are trying to maximize their interests in a selfish way. Naturally, many other strategies are possible, and can be easily tested out once the model has been completed. For instance, one may take into account further elements than only the current payoff and develop a more sophisticated routine

598

Federico Albiero et al.

based on the current power consumption of neighbor-nodes. Of course in the end it has to be considered how close is the model to real life situations of wireless systems. As ultimate step, the cluster negotiation mechanism is finalized by subroutine agree-on-coop?. The procedure reports a true value if and only if every terminals of current cluster have given a positive response to the invitation (set to “true” the internal variable “answer”). to−r e p o r t a g r e e −on−coop ? [ c u r r −c l u s t e r ] l e t num−pos−ans count c u r r −c l u s t e r with [ answer = true ] l e t num−c l u s t e r count c u r r −c l u s t e r i f e l s e ( num−pos−ans = num−c l u s t e r ) [ report true ] [ report f a l s e ] end Notice the structure to-report for a procedure with output argument, and the use of ifelse statement. Now, once a cooperative game has been played, if agree-on-coop? reports a true value we need to update the screen with a new cooperative cluster and, in case, dismiss old clusters that are no more operating. The latter can happen as the model allows already cooperating terminals to leave old groups in order to join a new cluster, if they can get a better payoff (see, e.g., selfish strategy). Therefore routine make-new-cluster is needed. t o make−new−c l u s t e r [ new−c l u s t e r new−l a b e l ] l e t new−l i n k s l i n k s with [ member? a new−c l u s t e r and member? b new−c l u s t e r ] l e t new−r a t e l a b e l −o f min−one−o f new−l i n k s [ l a b e l ] ask new−c l u s t e r [ s e t c l u s t e r new−c l u s t e r s e t r a t e new−r a t e s e t l a b e l new−l a b e l ] ask new−l i n k s [ s e t c o l o r y e l l o w ] end Let us analyze this procedure in detail. First, the routine takes two input parameters new-cluster and new-label representing the new cluster agentset (set of nodes) and the new power consumption for each terminal in the group (cooperative payoff), respectively. Then, the links new-links are set from the global link agentset as those ones connecting the nodes of new-cluster (i.e., having both ending points in new-cluster ). We underline the use of structure member? < agent >< agentset > which reports a boolean as a result of the belonging relation. Then, the new data rate used for transmitting in the cluster, new-rate, is set as the minimum of all rates currently supported by the considered communication links new-links 8 . Notice the joint use of commands label-of < agent > and min-oneof < agentset > [reporter] in order to address this task. In particular, the latter expression is a very interesting feature of NetLogo: it reports the agent belonging to

8

In this model data rates are reported by the links built-in variable label.

30 Introduction to NetLogo

599

a specific group with the minimum value for a chosen variable (reporter). The basic construct -of is probably the most common statement in the programming practice. Finally, terminals of the new cluster update the internal fields (additional data can also be stored, if necessary) and communication links change color to yellow (which is the chosen color for displaying cooperative groups). We underline that the nodes instantaneous power consumption and the links data rates are automatically displayed in NetLogo by the respective turtles labels. Figure 30.11 provides an illustrative example of the model described so far, for an initial setting of 30 mobile terminals with selfish strategy.

Figure 30.11. Layout of dynamic model for wireless cooperation. Figure shows two cooperative clusters of four units each. The implementation of routine clear-old-clusters is a bit more complex, and is left as a training exercise to the reader. One useful hint: use a node variable cluster to store the agentset of current cooperating cluster (nobody, in case the node is not taking part in any group).

30.5.3 Displaying Data Finally, once cooperative games are properly modeled and the payoffs are set, we need to collect the resulting data from the simulation and calculate the performance figures. Then, the results can be displayed in real time with NetLogo’s monitors and graphic charts, which are updated at every iteration of the program. Some basic parameters can be, for instance: the number of iterations, the overall number of terminals, the number of terminals for each implemented strategy. If these data are defined through global variables (such as iteration) we just need to add a “monitor” with the same label in the model’s interface. Other important results can be derived from the simulation. For example, in the model considered so far, we can be interested in analyzing the overall power consumption of the network as cooperation is enabled. Once more, NetLogo is shown to be very powerful in controlling agentsets. In fact, the simple one-line statement:

600

Federico Albiero et al. s e t power−system sum ( v a l u e s −from nodes [ l a b e l ] )

set the variable power-system as the total sum of all nodes labels, representing the current individual power consumption. Clearly, our interest is focused on the average power consumed by one node when cooperation is active, avg-power-nodes. This parameter can be easily calculated by dividing power-system by the total number of terminals n-of-terminals. In this way we get an estimation of the system performance in terms of power savings. Collecting data and displaying parameters with NetLogo is really straightforward. Furthermore, dynamic plots can be added in order to provide a better understanding of what is going on in the simulation. Some very good tutorials about plotting can be found in the IPD model and in the Models Library, thus we are not going into details here. In addition, an further possibility is to grab a small movie from the simulation model. This is a very attractive feature especially in the case of a dynamic simulation, as the movie could be used for presentations and visual purposes. An example of making a movie with NetLogo can be found in the folder “Code Examples” of the Models Library. To summarize, NetLogo provides a good way also for displaying data and results. However, it has to be considered that the language it is not the best for technical analysis. Mathematical tools and plotting features are rather limited from the point of view of engineering research. Therefore, some dedicated software shall be used to address these tasks. For instance, Matlab could be used for processing the program output, and clear graphs to be included in scientific papers could be created by using Gnuplot.

30.6 Conclusion The goal of this chapter was to introduce NetLogo programming language and encourage the use of this tool in the field of wireless engineering. The main reason is the possibility to control large collections of independent agents, all operating concurrently, at the same time: thus making it especially suited for analyzing decentralized wireless networks with recent game-theoretical insights. After a brief introduction presented in the first section, the potentials of NetLogo for wireless communications have been illustrated through a detailed tutorial about creating a model for exploring cooperative power savings. The development has been described step-by-step, listing the most common programming structures, in order to provide the reader with the basic skills in order to build his own simulation models of AHNs. Finally, it has been shown how to collect data and display results, underlying some related drawbacks. Figure 30.12 completes the picture of the potentials of NetLogo applied to wireless communications. In conclusion, a first-order analysis of large distributed systems such as wireless AHNs is perfectly suited with NetLogo, thus driving this tool to become an attractive solution for modeling cooperation in the field of telecommunications research.

30 Introduction to NetLogo

601

Figure 30.12. Screenshot of NetLogo model for analyzing power-savings in cooperative wireless networks [4].

References 1. IEEE Std 802.11a. Wireless LAN Medium Access Control (MAC) and the Physical Layer (PHY) specifications - High-speed Physical Layer in the 5 GHz Band., 1999. 2. IEEE Std 802.11g 2003. Amendment to IEEE Std 802.11, 1999 Edn. (Reaff 2003) as amended by IEEE Stds 802.11a-1999, 802.11b-1999, 802.11b-1999/Cor 1-2001, and 802.11d-2001., 2003. 3. L. DaSilva A. MacKenzie. Game Theory for Wireless Engineers. ISBN 1-59829016-9. Morgan & Claypool, 2006. 4. Federico Albiero. Power Savings in Cooperative Networks. A Game-theoretic Approach. Master’s thesis, Universita’ agli Studi di Padova (Italy) and Aalborg University (AAU - Denmark), 2006. 5. Various Authors. NetLogo User Manual. Center for Connected Learning and Computer-Based Modeling. Northwestern University of Evanston, IL, September 2006. Available at http://ccl.northwestern.edu/netlogo/docs/. 6. R. Axelrod. The Evolution of Cooperation. Basic Books, 1984. 7. F.H.P. Fitzek and M. Katz. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-4020-4710-X. Springer, April 2006. 8. W. Pundstone. Prisoner’s Dilemma. Doubleday, 1992. 9. J.M. Smith. Evolution of the Theory of Games. Cambridge University Press, 1982. 10. Frank H.P. Fitzek Tatiana K. Madsen and Ramjee Prasad. Simulating Mobile Ad Hoc Networks: Estimation of Connectivity Probability. Dept. of Communication Technology, Aalborg University, 2004. Available at http://kom.aau.dk/ ∼ff/documents/wpmc2004tatiana.pdf.

602

Federico Albiero et al.

11. U. Wilenski. NetLogo. Center for Connected Learning and Computer-Based Modeling. Northwestern University of Evanston, IL, 1999. Available at http: //ccl.northwestern.edu/netlogo. 12. U. Wilenski. Modeling Nature’s Emergent Patterns with Multi-agent Languages. Center for Connected Learning and Computer-Based Modeling, Northwestern University of Evanston (IL), 2001. Available at http://ccl.northwestern. edu/papers/MEE/.

31 Analysis of Cooperative Power Saving Strategies with NetLogo Simulating Cooperation in Mobile Wireless Networks

Federico Albiero1 , Frank Fitzek2 , and Marcos Katz1 1

2

VTT - Technical Research Centre of Finland [Federico.Albiero|Markos.Katz]@vtt.fi Aalborg University [email protected]

Summary. After a detailed introduction to NetLogo programming language, given in the previous chapter, here we aim to present and discuss some results obtained with this tool. Our focus is a model for analyzing cooperative power saving strategies in a mobile wireless network. The promising outcomes of the simulation in terms of highly reduced power consumption underline both the success of cooperative strategies for a specific fundamental network application and the effectiveness of using NetLogo agent-based approach for modeling decentralized communication systems.

31.1 Scenario of Investigation Power consumption of mobile communication devices is a major subject of concern in the wireless domain. The current trend for new services as well as new transmission techniques supporting various and rich data contents may result in the close future in a troublesome energy trap. One possible solution to address this issue is given by means of a novel cooperative network architecture, exploiting the combined data reception/transmission among the terminals over a central (or cellular, C) and a short-range (SR) communication link [5], [6]. Although several applications of cooperative techniques can be found in the recent literature (e.g., in the field of relaying), we will show this approach to cooperation to be a promising solution in order to improve energy efficiency. In the previous chapter we provided the reader with the basic know-how of NetLogo programming language, describing the main programming structures and providing a detailed tutorial about building models for cooperation in a dynamic wireless network. In this chapter we aim to stress the validity of this approach, by showing some concrete results of cooperative strategies with the goal of power saving. First of all, as required by the agent-based modeling practice, we need to spend a few words about the simulation testbed. The cooperative scenario under investigation is described by the following assumptions:

603 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 603–620. c 2007 Springer.

604 •

•

•

•

Federico Albiero et al. we consider a TDMA multicast transmission scenario, whereas a number of wireless terminals are distributed under the coverage of a central access point (AP) and they are interested in receiving the same data stream (such as, e.g., a video signal); we assume that every terminal can access two networks, i.e., it supports two air interfaces (WNIC, Wireless Network Interface Cards) in order to perform simultaneously the communication over the cellular and the SR link (parallel cooperation scenario); we assume a rate adaptive protocol-based mechanism is available as suggested by standard IEEE802.11a/g for WLAN (Wireless Local Area Network) technology [1] [2]. Specifically, as the remote AP is considered to be relatively far from the terminals, the cellular rate RC is set to the lowest specification of 6Mbit/s. Instead, the short-range rates RSR are dynamically assigned depending on distance among the terminals1 ; the power levels [W] used to perform the communication among the terminals are chosen according to specifications of typical present-day devices for WLAN equipment.

Figure 31.1. Scenario of investigation for cooperative power saving strategies. In figure some SR clusters are shown.

We underline that the efficiency of our cooperative approach for power saving is heavily dependent on the ratio between the cellular and the SR data rates ( RRSR ≥ C 1), and the capability of exploiting the low-power mode of wireless devices during 1

We consider the situation of a number of users located in close proximity one to each other, receiving data from a base station (BS) located at a bigger distance. Therefore we motivate the implicit assumption for the success of cooperative techniques: RSR > RC .

31 Analysis of Cooperative Power Saving Strategies with NetLogo

605

idle periods. In these assumptions, the objective of our model is finding the best cooperative clusters in terms of power saving gain. An illustration of the considered scenario is given in Figure 31.1. At this point, a mathematical review of the potential power savings of cooperation is required in order to clearly define the starting point of our analysis. This task will be accomplished in the following Section 31.2. After that, cooperative strategies will be discussed in Section 31.3, and the results of the NetLogo simulations will be illustrated in detail in Section 31.4. All the materials presented in this chapter have been developed by the author in his MS thesis work [3].

31.2 Theoretic Analysis In the previous chapter, we explained how a model for wireless power saving can be developed in NetLogo with a game-theoretical framework, thus taking into account the independent decision-making process among individual nodes. We consider, next, the potential payoffs of cooperation for the scenario described in Section 31.1. Assuming a multicast service is provided, we also assume that the service can be split into a given number J of substreams, equal to a number of terminals in close proximity, forming a cooperative cluster. The basic idea is that the users agree on a cooperative technique for exploiting the more efficient communication over the SR link. In this way, the power consumption could be reduced by switching off the receiving devices during idle periods (for more details, refer to [5]). The power consumption in such a scenario is given by:

PCoop (Z, J) = +

1 Prx,C + J

1−

1 J

Pi,C +

J −1 1 Ptx,SR + Prx,SR + J ·Z J ·Z

1−

1 Z

Pi,SR

(31.1)

where: • • •

Z is the ratio between the data rates used for the short-range and the cellular communication: Z = RRSR ; C J is the number of units in the cooperative cluster; the power levels Prx , Ptx , Pidle are defined in Table 31.1.

Thus, an estimation of the power consumption of a cooperative wireless network PCoop can be easily obtained by replacing in 31.1 the power levels used for reception (Prx ), transmission (Ptx ) and idle (Pidle ) by different technologies. Specifically, the power levels considered in this investigation are motivated by measurements of real systems. The actual values are taken from a report of Atheros Communications [4] about several products and technologies for WLAN, and listed in Table 31.1. As it can be seen, a very low level of power is needed by the devices during idle times, in comparison to the power levels required for transmitting/receiving data. Newer technologies are expected to be even more energy efficient and thus promising for this approach. However, a detailed analysis of 31.1 and the related data about power consumption is beyond the scope of this chapter.

606

Federico Albiero et al.

Table 31.1. Power levels and data rates provided by Atheros [4] for Wireless LAN technology. Link Description C Receiving power from central AP Power for central rx while idle SR Receiving power over short-range Transmitting power over short-range Power for short-range while idle

Name Value Unit Prx,C 900 mW Pi,C 40 mW Prx,SR 900 mW Ptx,SR 2000 mW Pi,SR 40 mW

Regarding the cooperative games we model, we assume that every player/terminal is given a payoff-matrix reporting the normalized power consumption of cooperation for all possible settings (see the previous chapter). This matrix is given by the following Table 31.2, where the reported values are calculated from 31.1 using power levels and data rates for WLAN technology.

Table 31.2. Payoff matrix of NetLogo model for cooperative power savings. Number of terminals 2

3

4

5

6

7

8

9

10

54 0.667

0.503

0.421

0.372

0.339

0.316

0.298

0.285

0.274

48 0.686

0.520

0.437

0.387

0.354

0.330

0.312

0.298

0.287

36 0.745

0.571

0.484

0.432

0.397

0.372

0.354

0.339

0.328

24 0.863

0.673

0.579

0.522

0.484

0.457

0.437

0.421

0.409

[Mbit/s] 18 0.980

0.776

0.673

0.612

0.571

0.542

0.520

0.503

0.489

12 1.215

0.980

0.863

0.792

0.745

0.711

0.686

0.667

0.651

9 1.450

1.184

1.052

0.972

0.919

0.881

0.853

0.830

0.813

6 1.920

1.593

1.430

1.332

1.267

1.220

1.185

1.158

1.136

Rate

In the Table, the element P (i, j) represents the power consumed by one terminal when cooperating in a group of j units (j = J) at rate i (i = RSR , directly related to Z). For instance, the payoff for a unit cooperating in a group of 3 terminals at 24 Mbit/s is 0.673 in terms of normalized energy consumption. Notice in this chapter we always refer to cooperative payoffs as terminal power consumption 2 , as these values can be directly compared to the power consumed by the units during normal non-cooperative operation, always set to unit by definition. We underline that the considered scenario WLAN/WLAN for the cellular and the SR communication is not the best in terms of power saving potentials. Nevertheless, cooperation is shown to be significantly beneficial, involving a power saving gain for most of the cases.

2

In other words, differently from the usual game-theoretic approach, we consider a problem of power minimization instead of gain maximization.

31 Analysis of Cooperative Power Saving Strategies with NetLogo

607

31.3 Strategies Overview As it can be noticed from the payoff-matrix of Section 31.2, the scenario we are dealing with is very peculiar. In fact, the situations where cooperation does not contribute to the units in terms of reduced power consumption are rather unlikely in our scheme. Moreover, we consider a framework where every participants gain from the cooperative exchange. In these conditions, cooperation is the preferred solution of rational, selfregarding individuals. Here we need to underline a major difference in comparison to the IPD example considered in the previous chapter. In fact, frequently in our scheme the players gain an immediate benefit in choosing cooperation. Notice this scenario is significantly different from the typical frameworks of cooperation in AHNs considered by several works of research. For instance, referring to the field of relaying, cooperation (i.e., packet-forwarding) normally involves a cost in terms of resources and the units might be tempted to drop the packets. Conversely, in our approach, the units have no incentive to cheat on themselves as they can instantaneously save power: cooperation provides an immediate profit to every participants. Still, strategies are fundamental for modeling the network operation as a process of independent decision-making. First of all, it can be easily argued that the behavior always cooperate is not productive even in our scheme. In fact, though not many, there are a few cases where cooperation is not beneficial to the units (see, for instance, the case of two units at 9 Mbit/s from the payoff matrix). In those situations, rational players would choose the normal self-sustaining reception. Then, a first effective strategy for our investigation is selfish cooperation. The strategy expresses the basic attitude of rational, self-regarding individuals trying to maximize their own gain (i.e., minimizing their power consumption). This simple strategic profile pursues the double goal of building the cooperative clusters aiming at power savings and, at the same time, preventing the units from losing energy whenever profitless conditions are met. As we will show later, the results of the numerical analysis are very promising. However, selfish cooperation might not always be the best choice for the purpose of saving power, as we are facing a heterogeneous scenario [5] where different short-range rates are used depending on distance among the terminals. Figure 31.2 introduces a challenging dilemma through a key-example taken from our model. Three wireless terminals are connected with different data rates. On the left side of the picture, all terminals receive the whole data stream from the AP (non cooperative reception). On the right side, the terminals exchange information through the SR communication links (cooperative reception). Now assume that unit T1 is searching for cooperative clusters (according to the procedure described in the previous chapter): at this point, we are in front of two choices. The first solution, Scenario A, leads to a cooperative cluster of two units with the lowest individual power consumption for terminals T1 and T2 (payoffs P1,2 = 0.6667), while terminal T3 is forced to receive the whole data stream from the AP in a stand-alone fashion (hence P3 = 1, non cooperative reception). The second solution, Scenario B, leads to a larger group of three cooperating terminals: in this case the individual payoff is slightly less advantageous for units T1 and T2, but is much more profitable for unit T3. In addition, all the considered terminals take part

608

Federico Albiero et al.

Figure 31.2. Terminal T1 is looking for clusters and finds two neighbors (nodes T2 and T3). What is the best strategy for cooperation?

in the cooperative exchange, significantly reducing their power consumption (payoffs P1,2,3 = 0.6733) in comparison to the normal autarchic operation (PN oCoop = 1). At this point a natural question arises: which alternative is better? Of course, the answer is a matter of point of view. Clearly, if the players are rational, units T1 and T2 would choose solution A while unit T3 would prefer by far solution B. On the other hand, from the overall system perspective, Scenario B is preferable as all the units are gaining out of cooperation. In fact, let us calculate the overall power consumption of the three terminals for the two cases: • •

A Ptotal = 2.3334, for Scenario A; B Ptotal = 2.0199, for Scenario B.

The total power consumed before the decision about cooperation is Ptotal = 3, as the units are initially operating alone. Therefore, in principle, the best choice in terms of power consumption is solution B (notice that, instead of total power, in the followings we will always refer to the average power saving gain per unit). However, selfish terminals would choose alternative A, leading to an elite cluster of two cooperating terminals at the highest possible data rate3 . In this case, unit T3 would have no chance to share the benefit of cooperation. Hence we motivate the introduction of a further strategy, that we call wise cooperation. In this strategy, we assume that the units are given a further knowledge regarding the actual power consumption of every other terminals in the 3

This scenario has been underlined in [5] by the example of the exposed terminal.

31 Analysis of Cooperative Power Saving Strategies with NetLogo

609

current range (for instance, the units might receive this information when looking for reachable terminals as a feedback to their searching request). In this way, wise nodes are able to make calculations about the average power consumption of the local subnetwork and to consider this data in their decision towards cooperation. For instance, refer to the example of Figure 31.2 and assume that node T1 features wise strategy. Then, the player can evaluate the two different alternatives A and B, and the consequent average power consumption per unit, before making the action. After that, it could sacrifice its best individual profit in order to improve the performance of the local subnetwork, by choosing solution B. We motivate this decision as an attempt to establish a better cooperative behavior that is more beneficial to all units in the long run. We will show that the results confirm this thesis. Nevertheless, there is a drawback in the strategy, as a new cluster is made only if all the terminals agree on the proposed solution (that might not happen if T2 and T3 are not wise). At this point, we need to make an important clarification. The wise nodes we model never incur to any power loss because of cooperation. They cooperate only when gaining respect to the normal operation, just in the same manner of selfish nodes (i.e., terminals will never cooperate if they get a payoff greater than 1). Nevertheless, the major difference is that the units feature a larger knowledge so that they can decide over a bigger set of choices. Of course, they are designed in order to renounce at their best profit for the local network optimization. However, this decision only involves an opportunity cost, namely a small gap from the best individual solution. Therefore, wise strategy is not exactly an altruistic behavior, being rather a way to implement a sort of improved network-wise terminals. On the other hand, one may wonder why nodes should have a wise behavior, giving up their best profit, even though they can still achieve a good power saving gain respect to the normal operation. An opportunity cost is still a cost for the user. What is the incentive, then, to choose wise over selfish? We motivate this decision as cooperative games are repeated multiple times (as in the IPD example, previous chapter) and the units are facing the same situations again in the future. Therefore the nodes may decide to look forward and be clever in order get a better average power saving gain performance in the long run. As a matter of fact, in Section 31.4.1 we will show that wise strategy achieves a better long-run payoff than selfish strategy in our model. Anyway, this result holds only if the strategies are tested separately in a scenario where all the units feature the same behavior one to each other. As soon as selfish and wise terminals are let to interact in a mixed scenario the situation dramatically changes and selfish strategy outperforms wise. The reason comes from the fact that we are dealing with independent decisionmaking. As already mentioned while analyzing Figure 31.2, even in the assumption that node T1 is wise it is not sure that solution B will be chosen at all. The other terminals might not agree to cooperate in such cluster, as they could possibly get a better payoff in another group (selfish) or find a better optimizing solution from their specific knowledge on the current network state (wise). In fact, when the units are playing cooperation, they can just evaluate one’s best potential opportunity according to strategies and ask the other nodes to participate in the cluster. However, the final result depends on the decisions of single individuals. This process is managed by the mechanism of invitation discussed in the previous chapter.

610

Federico Albiero et al.

The vulnerability of wise strategy to exploitation of selfish nodes has shown the need to develop further strategies. A latter strategy, that we called safe-wise cooperation, will be introduced and analyzed in detail Section 31.4.3. The basic idea is to play “wise” with wise terminals and “selfish” with selfish. We will show that this strategy is capable of preserving the optimal power saving gain of wise strategy in presence of selfish nodes. Finally, we point out that the units in our scenario are fully aware of the potential power savings of cooperation given in Section 31.2, thus trying to get the best advantage out of it. We do not consider here non-cooperating terminals (e.g., terminals that might not be aware of power saving potentials of cooperation) or malicious nodes which may try to harm the network operation.

31.4 Performance Evaluation Once the simulation testbed and the input parameters have been set, we aim to present the main results of our model. The program output in NetLogo is formed by a graphic window, showing the dynamic network of mobile terminals, and several monitors and charts displaying the performance figures. In this way, the user can observe how cooperative clusters are formed, preserved and dismissed depending on the particular circumstances. Furthermore, the power consumption of the terminals can be controlled by checking the units’ labels. The results of the model are collected at every iteration and estimation of the following parameters, in current and average values, is provided: • • • • • • • • • • •

number of cooperating terminals NT,Coop ; number of cooperating clusters NCl ; frequency distribution of number of clusters vs. number of units NCl (Nunits ); frequency distribution of number of clusters vs. data rate NCl (R); probability density function of number of clusters vs. number of units; probability density function of number of clusters vs. data rate; power consumption per terminal P ; power consumption only-cooperating terminals POnlyCoop 4 ; power consumption per strategy PStrategy ; power saving gain per terminal G; power saving gain only-cooperating terminals GOnlyCoop .

The numerical values (number of terminals or clusters, power consumption, power saving gains) are displayed in real time by several monitors, while the parameters involving multiple data (frequency distributions and probability functions) are visualized through histograms and plots. Moreover, a dynamic chart displays the average power consumption of cooperation compared to normal operation, and another one shows the payoffs of the two strategies. Although the average values are more significant for our analysis, the instantaneous values can provide a detailed picture of the system state at a specific iteration. The number of cooperating terminals and cooperating clusters per iteration can be directly derived from the model. Other parameters have been calculated through 4

Currently non-cooperating units are not considered.

31 Analysis of Cooperative Power Saving Strategies with NetLogo

611

the following equations, where I is a large number of program iterations (typically, I > 10000 in this analysis): PI

NCl,i (Nunits ) I PI i=1 NCl,i (R) ¯Cl (R) = N I

¯Cl (Nunits ) = N

i=1

(31.2) (31.3)

for the average frequency distribution of clusters versus number of cooperating terminals or data rate, respectively. NCl,i (x) represents the number of clusters reporting feature x at a specific iteration i; PI pdf (x) =

i=1 P I

NCl,i (x)

i=1

NCl,i

(31.4)

for the probability density function (pdf ) related to the number of clusters versus number of terminals or data rate. It is obtained by dividing the overall number of clusters of given feature x for the overall number of clusters; PI

i=1 Pi (31.5) I · NT for the average (normalized) power consumption per terminal, where Pi is the total power consumed by the units at a given iteration i and NT the total number of terminals;

P¯ =

¯ = 1 − P¯ G

(31.6) for the average power saving gain of cooperation. Notice that the values of P¯ ¯ refer to the overall set of terminals involved in the simulation, NT . The and G power consumption and gain of only-cooperating terminals can be easily obtained by replacing NT with NCoop in 31.5 and 31.6.

31.4.1 Power Saving Gain of Cooperation First of all, the two strategies discussed in Section 31.3 have been tested separately for an initial setting of 50 mobile terminals. The resulting performance of the two strategies in the long run is shown by the following Table 31.3.

Table 31.3. Power saving gain of the two strategies in a homogeneous scenario of 50 mobile terminals. Percentage values are calculated over I > 10000 iterations. ¯ G

¯ OnlyCoop G

Selfish

34.4

38.2

Wise

37.3

39.1

As it can be noticed, the results are quite promising, involving percentage power saving gains about 35 − 37% respect to the normal autarchic network operation.

612

Federico Albiero et al.

8

8

7

7

6

6 Number of Clusters

Number of Clusters

The first outcome to be stressed is that selfish strategy is an efficient way to build cooperation in our scheme. Clearly, the model is based on the theoretic framework of Section 31.2, which in turn relies on several assumptions. However, the potentials of this approach to address power saving are significant. A second important result is that wise strategy achieves a better performance out of cooperation in a homogeneous scenario. Referring to the same testbed, a further comparison between the two strategies can be traced by analyzing the frequency distribution of cooperative clusters. Figures 31.3 and 31.4 shows the histograms concerning average number of clusters versus number of terminals or data rate, respectively.

5 4 3

5 4 3

2

2

1

1

0

0 1 2 3 4 5 6 7 8 9 Number of terminals/cluster selfish

1 2 3 4 5 6 7 8 9 Number of terminals/cluster wise

Figure 31.3. Cluster distribution vs. number of cooperating units for the two strategies.

First, from Figure 31.3 we notice a significant difference in the behavior of the two strategies. As a matter of fact, selfish nodes are more likely to cooperate in small elite groups of two units (peak value on the chart shows > 6 clusters of 2 units per iteration). On the other hand, wise terminals tends to form larger groups where more entities can share the power saving benefit of cooperation (the peak value shifts to clusters of 3 units). For both strategies, the number of clusters of big size is rather small, as large groups of interconnected nodes5 (say, > 5 units) are more difficult to be found. An interesting monitor is provided by the first column of the histograms, showing the average number of units that are not cooperating at a given iteration (clusters of 1 unit are in fact alone terminals). While this value is about 5 units over a total of 50 for selfish strategy, it drops as only 2 units for wise. To summarize, the histograms show that the latter strategy is capable to bring the profit of cooperation to a larger number of units and it is less likely to leave isolated terminals out of cooperative clusters. 5

We consider a framework where all the terminals in a cluster are connected to each other, and thus can participate in the cooperative SR exchange.

8

8

7

7

6

6 Number of Clusters

Number of Clusters

31 Analysis of Cooperative Power Saving Strategies with NetLogo

5 4 3

5 4 3

2

2

1

1

0

613

0 54 48 36 24 18 12 9 6 Data Rate [Mbit/s] selfish

54 48 36 24 18 12 9 6 Data Rate [Mbit/s] wise

Figure 31.4. Cluster distribution vs. data rate for the two strategies.

Figure 31.4 also provides an interesting comparison. In fact, it can be noticed that wise clusters are more uniformly distributed over different data rates. An exception is the number of wise clusters at high data rate of 54 Mbit/s which is lower than selfish in the charts, while the number of wise clusters at low data rate of 18 Mbit/s is rather larger (cooperating groups at lower rates are not present as the terminals are out of the maximum range chosen for the simulation). This different behavior is deeply related to the strategies features analyzed in Section 31.3 and gives more insights to the introduction of wise strategy. In the considered dynamic testbed the probability of finding connected groups of wireless nodes is proportional to the distance among the units, hence inversely proportional to the SR data rate. For this reason, cooperative clusters with more than two units at 54 Mbit/s are quite unlikely to be found as the terminals hardly get close enough. However, even though large connected groups at lower data rates are rather common (e.g., at 18 Mbit/s), they are preferentially discarded by selfish nodes as the terminals can often get a better individual profit by cooperating in smaller clusters at higher data rate. Therefore we needed a different strategy in order to bring the power saving gain of cooperation to more units. This task has been achieved by wise strategy, leading to a more leveled cluster distribution over data rates. A final term of comparison between the two strategies in this scenario can be obtained by analyzing Figure 31.5. The plots show the average power consumption per terminal of the overall network versus the power consumption of only-cooperating units (i.e., without considering the isolated terminals). As suggested by intuition, the latter value is lower, involving a higher gain for both strategies and representing an upper bound for cooperation. The comparison shows that the gap between the two curves is notably reduced with wise strategy. This provides an additional evidence of the superior efficiency of wise over selfish in terms of average network power saving gain.

Federico Albiero et al.

0.7

0.7

0.68

0.68 Normalized energy consumption

Normalized energy consumption

614

0.66 0.64 0.62 0.6 0.58 0.56 0.54 0.52

0.66 0.64 0.62 0.6 0.58 0.56 0.54 0.52

0.5

0.5 0

2000

4000 6000 Number of iterations

Power Network

8000

10000

0

Power Coop

2000

4000 6000 Number of iterations

Power Network

(a) selfish

8000

10000

Power Coop

(b) wise

Figure 31.5. Average power consumption per terminal considering the overall network and only-cooperating units for the two strategies.

31.4.2 Mixed Scenario Another major result concerns the interaction of the two strategies in a mixed scenario. It has been shown in Section 31.3 that wise strategy achieves a better payoff in terms of power saving gain. Nevertheless, the strategy is effective only if all the terminals involved in the simulation are wise.

Normalized energy consumption

0.68 0.67 0.66 0.65 0.64 0.63 0.62 0.61 0.6 0

2000

4000 6000 Number of iterations wise

8000

10000

selfish

Figure 31.6. Payoffs of the the two strategies in a mixed scenario of 25 units each.

In fact, after calculating the payoffs separately, we tested the two strategies interacting in a heterogeneous scenario of 25 units each. As shown by Figure 31.6,

31 Analysis of Cooperative Power Saving Strategies with NetLogo

615

in these settings selfish strategy wins. Indeed, the wise strategy is implicitly conceived to get the best result out of cooperation in a world of wise individuals. As a consequence, if other terminals with a different behavior are operating, the strategy fails. In this particular case, wise nodes are somehow exploited by selfish, which participate in cooperative groups with wise units but at the same time make the usual elite clusters among themselves. In other words, they take advantage of the benefit of wise strategy without giving anything in return. Therefore wise terminals perform poorly, as they sometimes give up their best individual profit. However, this fact does not lead to a system optimization due to the presence of selfish nodes. The NetLogo model considered so far has been slightly modified in order to address this issue.

31.4.3 Two-Box Model The main idea is to include in our model a square box, placed in the middle of the screen, in order to split the virtual world into two parts: internal and external (notice the chosen box shape is arbitrary). In this way, the terminals are forced to move within one’s specific side, only occasionally interacting with the nodes on the other side as they approach the box boundaries. Different strategies can be selected for the behavior of the units inside or outside the box and the resulting payoffs can be carefully analyzed. An illustration of the Two-Box model is given in Figure 31.7.

Figure 31.7. Layout example of the Two-Box model. Figure shows the interaction between internal and external nodes, forming a cooperative cluster of three units (brighter links corresponds to cooperative links).

616

Federico Albiero et al.

The model has been conceived for studying in detail the interaction between the two strategies. More specifically, we aimed to understand how wise strategy could be improved in order to preserve its power saving gain in presence of selfish nodes. For this purpose, the model has been tested in different configurations. The major figures involved in the simulation are: NI number of terminals inside the box; NO number of terminals outside the box; SI strategy for terminals inside the box; SO strategy for terminals outside the box; PI payoff terminals inside the box; PO payoff terminals outside the box; PA payoff inside terminals when acting alone (no interaction with the “external world”). A first numerical analysis has been performed with the following parameters: • • • •

NI = 10; NO = 20; SI = selfish; SO = selfish/wise.

While keeping constant the number of units and choosing selfish strategy inside the box, we tried out the two strategies for the outside terminals. In this scenario, we noticed that the payoff of the internal nodes is not dependent on the strategy of the units outside PI,S ' PI,W 6 . Furthermore, this result is almost independent from the presence of external interaction PI,S ' PI,W ' PA . This can be explained by the fact that cooperation finds fruitful conditions inside the box, being the units often close to each others and capable of clustering in large groups at high data rates. Therefore the presence of some terminals outside, which in addition are not likely to cooperate in good conditions, does not impact the average result of inside units in this case. Instead, the situation dramatically changes if we choose the wise strategy for the terminals inside. A second test has been performed with the same number of units inside/outside the box. While the payoff of the internal units does not change significantly in presence of wise nodes outside PI,W ' PA , as we place selfish nodes in the external world the performance drops: PI,S > PI,W (increased power consumption). The Two-Box model has been designed in order to focus on this latter outcome. As a matter of fact, a more detailed numeric analysis has been performed by choosing wise strategy for the internal network and selfish for the external one. The simulation has been carried out according to the following steps: •

6 7

first, a number of units NI are placed inside the box and the payoff7 PA is calculated with no interaction to the external world; The subscripts stand for the outside strategy SO : S for selfish, W for wise. Notice that, differently from the terminology of game theory, we refer to cooperative payoff P as the normalized average power consumption per terminal, related to the power saving gain by G = 1 − P .

31 Analysis of Cooperative Power Saving Strategies with NetLogo •

•

617

then, we keep the same number of units inside and place a constant number of units outside the box, NO = 20. The payoffs of the internal/external subnetwork, PI and PO , are derived; the test is repeated for several numbers of inside terminals: NI ∈ {2, 4, 6, 8, 10}.

The results over a large number of iterations (I > 10000) are summarized in Table 31.4. Table 31.4. Payoff table for the Two-Box model with wise strategy. Payoff Number of terminals PA

PI

PO

2

0.836

0.859

0.765

3

0.741

0.793

0.761

4

0.689

0.752

0.759

6

0.619

0.698

0.758

8

0.581

0.662

0.754

10

0.545

0.634

0.755

While the payoff of the terminals outside the box PO does not change significantly with the number of units inside, an increasing gap can be noticed between the performances of the internal units, PA and PI . An illustrative plot is provided by Figure 31.8. Specifically, in this scenario the power consumption of wise terminals inside the box increases because of the interaction with external selfish nodes, thus the payoff curve lifts up. In fact, as the terminals approach the box borders, they try to cooperate and some mixed clusters are formed as in Figure 31.7. A larger number of inside terminals implies a higher probability of finding mixed groups. In those groups wise behavior is exploited by selfish, as the external units take advantage of cooperating in wise clusters (thus often involving an opportunity cost for wise units) without providing any benefit in return (i.e., being themselves “wise”). This result also comes as a consequence that wise strategy has not been conceived in order to interact with other strategies. Therefore we modify the strategic profile as described in the following procedure: Case 1 the unit is looking for clusters: • if all terminals in the potential cluster are wise, play wise; • if terminals with other strategies are present, send the invitation (see the previous chapter for the details): – if all other players answer positively, make the cooperative cluster; – if one or more non-wise terminals do not agree on cooperation, leave them out and try to group with the remaining wise units; Case 2 the unit is asked to cooperate in a group: • if all terminals in the potential cluster are wise give a positive answer (trust units of the same strategy);

618

Federico Albiero et al.

Normalized energy consumption

1 0.9 0.8 0.7 0.6 0.5 0.4 2

3

4

5 6 7 Number of terminals

Payoff A

8

9

10

Payoff I

Figure 31.8. Two-Box : performance of wise strategy. •

otherwise play selfish.

The strategy illustrated above has been called safe-wise, as it pursues the double goal of achieving the best power saving gain of wise strategy and preventing the units from exploitation of selfish terminals. Despite its long description, the strategy is very simple. If the current player finds wise terminals, it plays wise. Otherwise, in case other terminals are encountered, first it tries to be wise expressing the aim to make the best cooperation, then it may decide to cluster with only-wise terminals whereas its request is denied. Furthermore, safe-wise nodes protect themselves by playing selfish when asked for cooperation by selfish nodes. Note that we make the implicit assumption that players have cognition about the other players’ strategies. The results of our model with safe-wise strategy for the terminals inside the box are quite significant. Figure 31.9 shows the new payoff curves in this scenario (clearly we refer to the latter simulation testbed considered so far). As it can be noticed, differently from Figure 31.8 the curve PI almost coincides with PA , namely to the average power consumption of the terminals when no external interaction is present at all. Moreover, the interaction of safe-wise terminals with other nodes is shown to be even beneficial for small numbers of units (values of PI are slightly less than PA in the left side of the chart) as the terminals take advantage of the presence of external nodes. This is a numerical example of the meaning of the sentence: “real egoistic behavior is to cooperate!” [5]. To summarize, we demonstrated the strategy to be capable of preserving the high power saving gain of wise cooperation in close groups of trustful units and, at the same time, preventing the terminals from potentially harmful interaction with selfish nodes. This is an encouraging result, showing a successful way to improve

31 Analysis of Cooperative Power Saving Strategies with NetLogo

619

Normalized energy consumption

1 0.9 0.8 0.7 0.6 0.5 0.4 2

3

4

5 6 7 Number of terminals

Payoff A

8

9

10

Payoff I

Figure 31.9. 2-Boxes: performance of safe-wise strategy.

the winning strategy of our analysis of cooperative power saving (wise) in order to preserve its efficiency in presence of different strategies involved in the independent decision making process.

31.5 Conclusion In this chapter, we presented and analyzed cooperative strategies for wireless networks. Strategies were implemented in NetLogo, an agent-based programming tool discussed in the previous chapter. Specifically, a numerical analysis for the power saving gain of two cooperative strategies has been performed and illustrated in details, leading to the following major results. •

•

•

The analytic framework described in Section 31.1 and 31.2 has been successfully applied in the development of a dynamic model to explore cooperative power savings in a wireless ad hoc network, based on independent node-decision making. Two cooperative strategies, selfish and wise, have been developed and tested separately leading to considerable average power saving gain of about 35% in comparison to the normal autarchic (non cooperative) network operation. The strategic interaction among wireless nodes has been carefully analyzed, underlying potential vulnerability of the winning strategy (wise) in situations of interdependence.

620 •

Federico Albiero et al. An improved strategy, safe-wise, has been conceived in order to address this drawback, providing a successful methodology to prevent the nodes from the exploitation of other strategies.

To summarize, by showing the outcomes of the simulation illustrated throughout this chapter we pursue a double objective: • •

We show significant potentials of cooperative techniques for the goal of power saving in mobile wireless networks. We show the effectiveness of agent-based modeling in the field of wireless communication by means of building and improving strategies for cooperation in a world dominated by individual interaction.

In conclusion, we provided an overview of how NetLogo agent-based programming language can be successfully applied by the communication engineer for the purpose of modeling and analyzing cooperation in the complex dynamic world of wireless networks.

References 1. IEEE Std 802.11a. Wireless LAN Medium Access Control (MAC) and the Physical Layer (PHY) specifications - High-speed Physical Layer in the 5 GHz Band., 1999. 2. IEEE Std 802.11g 2003. Amendment to IEEE Std 802.11, 1999 Edn. (Reaff 2003) as amended by IEEE Stds 802.11a-1999, 802.11b-1999, 802.11b-1999/Cor 1-2001, and 802.11d-2001., 2003. 3. Federico Albiero. Power Savings in Cooperative Networks. A Game-theoretic Approach. Master’s thesis, Universita’ agli Studi di Padova (Italy) and Aalborg University (AAU - Denmark), 2006. 4. Atheros Communications. Power Consumption and Energy Efficiency Comparison of WLAN products. Technical report, Atheros Communications, 2003. Available at www.atheros.com/pt/whitepapers/atheros power whitepaper.pdf. 5. F.H.P. Fitzek and M. Katz. Cooperation in Wireless Networks: Principles and Applications – Real Egoistic Behavior is to Cooperate! ISBN 1-4020-4710-X. Springer, April 2006. 6. Marcos Katz, Frank H.P. Fitzek and Qi Zhang. Cellular Controlled Short-Range Communication for Cooperative P2P Networking. In WWRF 17, 2006.

Part VI

Visions, Prospects and Emerging Technologies

32 Cooperation in Optical Wireless Communications Dominic O’Brien University of Oxford [email protected]

Summary. RF wireless communications will be unable to meet all the demands for wireless capacity that will be faced in the future, and some of these will be met using optical frequencies. In this chapter we outline the key characteristics of Optical Wireless (OW) communications, contrast it with Radio Frequency (RF) techniques and discuss how cooperation between the two techniques might improve the capabilities of both.

32.1 Introduction The communications data rates available with RF systems have grown rapidly in recent years; the use of Multi-Input-Multi-Output (MIMO) techniques [11], and higher order modulation schemes [46] have allowed a rapid rise in the capacity available for services such as wireless Local Area Networks (LANs). Similarly Ultra Wideband (UWB) [10] provides very high speed short distance connectivity. Over wider areas WiMax [12] can provide provide mobile broadband, and the data rates available using cellular communications will continue to rise. There is therefore no doubt that highly mobile users will continue to be served by RF for the foreseeable future. As user mobility decreases the demand for bandwidth increases. Nomadic and fixed users will ultimately demand bandwidths approaching that available using fibre networks, but from a wireless connection. In these situations using optical frequencies for communications has some compelling advantages, with the potential to; • • • •

provide fibre-like data rates deliver very high bandwidth to a small region in space use simple baseband transmission schemes, and low cost components operate in an unregulated, unlicensed region of the electromagnetic spectrum

The cooperative use of Optical Wireless (OW) and RF communications is described in this chapter, which is organised as follows; a brief review of optical wire-

623 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 623–634. c 2007 Springer.

624

Dominic O’Brien

less systems is given, together with a comparison with RF communications. Potential cooperative strategies and some of the potential enabling technologies are then discussed.

32.2 Overview of Optical Wireless Communications There are a number of more detailed overviews of OW [1] [22] [13] [14]. In this section a brief description of the key characteristics of such systems is given. Figure 32.1 shows several different types of OW configurations. Diffuse OW systems (Figure 32.1 (a)) use the high diffuse reflectivity of walls and surfaces in the room to create an optical ‘ether’ within the coverage space, much as is the case for RF communications. There has been extensive work on predicting the characteristics of the diffuse channel, including [40] [25] [6] [19], as well as several measurements [23] [41] [31]. The Inter-Symbol Interference (ISI) impairment together with the high path loss offers little advantage over RF approaches, at least in the near to medium term, so this chapter focuses on the use of Line of Sight (LOS) propagation.

Figure 32.1. Optical wireless configurations (after [39]). (a) Diffuse system. (b) Wide LOS system. (c) Narrow LOS system with tracking. (d) Narrow LOS system using multiple beams to obtain coverage. (e) Quasi diffuse system. Line Of Sight (LOS) (Figure 32.1(b,c,d)) systems control illumination to varying degrees to ensure that there are no intermediate reflections between transmitter and receiver. These systems are categorised by the directivity of the transmitted beam.

32 Cooperation in Optical Wireless Communications •

•

•

625

Figure 32.1 (b) shows a Wide LOS OW configuration. A single source with a broad beam illuminates the coverage area, and care is taken to minimise reflections from intermediate surfaces. The power available at the receiver is limited by the high path loss. Narrow LOS systems (Figure 32.1 (c)) use a highly directed beam from transmitter to receiver in order to minimise dispersion and path loss. These channels are limited by the performance of the transmitter and receiver components rather than dispersion. In order to provide coverage a mechanism to track the receiver with the transmitted beam is required. Cellular architectures (Figure 32.1 (d)) use a number of narrow beams to provide both coverage and high bandwidth.

The LOS channel is necessarily subject to blocking, and quasi-diffuse systems aim to overcome this by minimising the number of received multipaths. In these systems an array of illumination spots is projected onto the surfaces of the coverage area, and a suitable receiver selects one, or a small number of transmission paths only (Figure 32.1(e)). Several variants of such systems have been reported [49] [18].

Input electrical signal

Input radiation Source Optics Optical filter Optical system Photodetector

Output radiation

Amplifier

Output electrical signal Transmitter

Receiver

Figure 32.2. Schematic of optical wireless transceiver.

32.3 System Components Figure 32.2 shows a representative OW transmitter and receiver. These are described below.

626

Dominic O’Brien

32.3.1 Transmitter An electrical signal containing modulated data is used to drive a source, or array of sources, in order to transmit the required data. Various types of optics are then used to control the radiation emitted from these sources, and in some cases to change emission characteristics to ensure that the transmitter is eyesafe. Most systems use laser diodes as sources, and their transmission power is limited by eye safety regulation. This is wavelength dependent and determined by regulation [4]. For point sources in the near Infra-red regime this is approximately less than 1mW, rising to ten times this at longer wavelengths. Large transmission powers are available by increasing the apparent size of the source [4] using a diffuser such as a ground glass plate. Holographic [26] and reflective [2] elements can also be used to create transmitters that are eyesafe. (The device reported in [2] has been incorporated in a commercial optical link [21].) There is rapidly growing interest in using white LEDs for lighting applications, and these sources can be modulated to provide communications [44]. The large source area and divergence of typical lighting fixtures allows many times more optical power to be transmitted than is permitted in the near infra-red region of the optical spectrum when laser sources are used. Typically the level of light required for the occupants of the illuminated space ensures sufficient power to provide a high-quality optical communications channel [27].

32.3.2 Receiver Figure 32.2 shows a typical OW receiver. Light entering the receiver passes through an optical filter that rejects optical noise, a lens system or concentrator that focuses light onto a photodetector, and an amplifier to amplify the resulting photocurrent. Ambient light is generally the largest source of interference in OW [3]. A narrowband optical filter is usually used to reject as much of the broadband optical noise as possible and to let through the desired radiation. In addition an electrical high-pass filter can be used in the amplifier to minimise any AC electrical noise from fluorescent lighting [35] [34] [40] and block DC photocurrent from ambient illumination [42]. Imaging or non-imaging optical systems [47] can be used to concentrate radiation onto the receiver photodetector. Increasing detection area can be achieved using multiple receivers, each with their own optical system, that point in different directions-so called angle diversity systems [5]. Alternatively an imaging optical system [37] [24] that focuses light from a particular angle to a particular detector element in an array can be used. These types of receiver can select the strongest optical signal, or to reject unwanted signals from a particular direction [24]. The performance of OW systems is largely determined by receiver performance, and this is a strong function of the capacitance of the detector and the preamplifier tolerance to this. Techniques such as bootstrapping [33] and equalization [32] can be used to improve performance, as well as the careful co-design of detector and front-end [37] [29] [15].

32 Cooperation in Optical Wireless Communications

627

32.4 Cooperation between RF and Optical Wireless (OW) Systems The gap between wired and wireless data rates continues to grow, with RF systems lagging by a significant factor. This mismatch will continue to grow, and meeting the demands of ‘Ambient intelligence’ and seamless connectivity will be difficult to meet with RF alone. The success of WiFi [48] is due to the availability of unlicensed spectrum in the same band worldwide, and such regions are being filled at a rapid rate. Higher data rates will necessitate a move to higher carrier frequencies, as is the case with the growing interest in systems operating in the 60 GHz band [43]. The cost of transceivers operating in these high frequency bands is likely to follow the similar cost reductions that have occurred for WiFi, given the continued ‘Moore’s law’ development of Integrated Circuit processing. However, the propagation characteristics of higher frequency RF carriers become less amenable to providing robust coverage with simple low-cost antennas, so similar challenges in providing coverage exist for OW and RF approaches.

Table 32.1. Complementary properties of OW and RF communications. LOS OW communications Highly directive. Possible to use low cost optical elements to create highly directed radiation Interference Light confined to single space, and directivity reduces possibility of interference Channel Subject to Rayleigh and Ri- Stable cian fading Typical transmit power +10dBm +10 − 30 dBm Receiver sensitivity (for High (Typically −70 dBm- Low (−30 dBm) 1G b/s capacity) using 80 dBm) 60 GHz carrier Receiver complexity High Low Directivity

RF communications Low directivity except at high frequencies. Requires complex structures to engineer directivity Crosstalk and co-channel interference limit performance

Table 32.1 summarises the properties of an RF and optical channel. The major disadvantage of the RF channel is that it is a complex amplitude channel, so is subject to coherent fading processes. However, associated with the coherence is high receiver sensitivity and excellent link margin. The LOS optical channel does not fade as the detector and emitter structures are very large compared with the wavelength of light, so any spatial effects are integrated. This creates a channel where information is transmitted using intensity. This leads to a poor link budget for OW, where the margins are 20-30 dB worse than for RF at 60 GHz. Highly directive channels must therefore be used to compensate for this. Offsetting this disadvantage is the simplicity of the transceiver architectures. Typical optical transceivers are simple, and have low power consumption compared with their RF counterparts [39].

628

Dominic O’Brien

32.5 Potential Scenarios for Cooperative Working 32.5.1 Hotspots Future RF systems will require low (<10 GHz) frequencies to provide coverage, with zones of high frequency coverage (60 GHz at present) to provide capacity. Cooperation with OW may allow a simpler implementation of this ‘dual frequency’ approach. Optical wireless hotspots [13] [38] use OW to augment the capacity provided by RF. Figure 32.3 shows a schematic of such a system. Outside the hotspot RF communications is used to provide coverage, and optical communications is used to provide capacity within it. Additionally RF communications can be used to control the optical channel, and protect the line of sight channel from blocking. In [17] such a system is investigated: a 100 Mb/s optical bidirectional link is used cooperatively with a 10 Mb/s RF LAN. The optical link is periodically blocked and if these events last for a sufficiently long time the system switches to RF communications. The decision to make this switch is made using a fuzzy inference engine. When the optical link is blocked 10% of the time the combined system is a factor of approximately 12 times as efficient as the RF only LAN over a wide range of traffic loads. Figure 32.4 shows a similar approach, using a Visible Light Communications (VLC) transmitter to create a broadcast only downlink. Analysis of this type of system, (where there is no optical uplink), was also undertaken in [17] and this shows that there is still significant advantage. For similar parameters as the full bidirectional system the VLC example is still nine times as efficient as the RF only example. This configuration has the advantage that VLC may be present within a building anyway, and the additional infrastructure might be very low cost, with the lighting units fed by Power Line Communications (PLC) [28].

32.5.2 Cooperative Transceivers The future ‘cognitive radio’ will support multiple standards and have significant baseband processing capability. This is required for the complex modulation, spatial multiplexing and other techniques used in modern wireless communications. In contrast OW uses an Intensity Modulated/Direct Detection (IM/DD) baseband communications channel. A cooperative receiver could support IM/DD OW channels by connecting the optoelectronic transmitter (consisting of source and optics) and receiver (consisting of optics, detector and amplifier) to the RF baseband signal processing. The relatively simple modulation schemes suitable for optical channels would require a small proportion of the hardware complexity typically used in RF systems for transmission and reception, and this would be reflected in lower power consumption, and the ability to reallocate processing hardware to another function. A key challenge for OW is that of improving receiver sensitivity, and coherent detection has been proposed as a potential solution to this [16], although this is only likely to be achieved in the long term. In such systems an Intermediate Frequency (IF) in the GHz frequency range is produced by mixing the incoming optical wave with an optical local oscillator. There are significant challenges to achieving this, including the need for an optical local oscillator, and the processing of the intermediate frequency. However, with the increase in speed of integrated circuits and the

32 Cooperation in Optical Wireless Communications

629

Figure 32.3. Potential scenarios for high-speed optical wireless communications and RF cooperation. Bidirectional OW communication is combined with RF communications. IF processing associated with high frequency RF approaches, a future OW receiver might use the IF subsystems used for RF coverage.

32.5.3 Enabling Developments Improving the link budget of OW systems is the key to deploying them widely in the hotspot configurations shown above. Potential methods of achieving this include 1. Using narrow LOS channels. This requires tracking [20] transmitters, and for maximum benefit imaging receivers [37] [24] [8]. Mechanical tracking such as that used in [20] is often undesirable, and a possible alternative is detailed in [30]. This uses a simple passive base station, consisting of diffractive elements and beam-shaping optics. The wavelength of the source controls the angle of diffraction within the base station, thus steering light to a particular receiver. 2. Using solid-state lighting to provide high transmission power, and using a hotspot as shown in Figure 32.4. Solid-state lighting is likely to be used in many commercial and domestic buildings, and there is a growing interest in using these sources for communications [27] [45] [7]. Such systems might offer approximately 30-40dB more power than would be typical for infra-red OW transmitters. The major challenge for Visible Light Communications (VLC) is to develop means to transmit at high data rates using high-power, low modulation bandwidth LED sources.

630

Dominic O’Brien

Figure 32.4. Potential scenarios for Visible Light Communications (VLC) and RF cooperation. VLC is used for a high-speed optical downlink, and combined with RF communication that provides channel control and uplink for the hotspot.

3. Improved receiver sensitivity is dependent on optimised design of optoelectronic components, as detailed in [37], in the near term. Optical amplification at the receiver is a potential gain mechanism, and studies of the use of Erbium Doped Fibre Amplifiers (EDFAs) in LOS systems [9] have been made . However, simple analysis suggests that the coupling losses either completely or largely offset the available gain [36] for both semiconductor and fibre amplifiers. Coherent systems are an alternative means to provide ‘gain’, and the linewidth control that is required for Wavelength Division Multiplexed (WDM) optical fibre communications systems has meant that sources with sufficient stability are widely available, although still costly for consumer applications. However, there are also significant problems with wavefront and polarisation control in order to ensure optimum performance, so widespread adoption of such techniques may only occur in the longer term.

32.6 Conclusions There are a growing number of standards for RF communications, and terminals now incorporate at least a few of these. Future cognitive radios will have to switch between these seamlessly depending on factors such as cost of communications, power consumption and channel reliability. A fundamental challenge for RF will be to provide very high data rate communication, as the frequencies where bandwidth

32 Cooperation in Optical Wireless Communications

631

is available do not provide the robust and reliable coverage characteristic of the low frequency RF wireless channels used at present. Within this context OW can provide very high bandwidth with simple baseband components, both to augment and to substitute these high frequency approaches. Such a solution requires the development of new and low-cost approaches, but success would allow the overall capabilities of transceivers to be substantially enhanced. Acknowledgement. The contributions of present and past members of the Optical Wireless Communications group at the University of Oxford are gratefully acknowledged, and valuable comments from Colleagues in the Wireless World Research Forum have especially helped to shape this contribution. Much of the work in the group at Oxford has been funded by the UK Engineering and Physical Sciences Research Council.

References 1. J. D. Barry. Wireless infrared communications. Kluwer, Netherlands, 1994. 2. P. Benitez, J. C. Minano, F. J. Lopez, D. Biosca, R. Mohedano, M. Labrador, F. Munoz, K. Hirohashi, and M. Sakai. Eye-safe collimated laser emitter for optical wireless communications. In Optical Wireless Communications V, volume 4873, pages 30–40, Boston, 2002. SPIE. 3. A. C. Boucouvalas. Indoor ambient light noise and its effect on wireless optical links. IEE Proceedings-Optoelectronics, 143(6):334–338, 1996. 4. British Standards Institution. Safety of laser products part 1. Technical Report IEC 60825-1, British Standards Institution, 2001. 5. J. B. Carruthers and J. M. Kahn. Angle diversity for nondirected wireless infrared communication. IEEE Transactions on Communications, 48(6):960– 969, 2000. 6. J. B. Carruthers and P. Kannan. Iterative site-based modeling for wireless infrared channels. IEEE Transactions on Antennas and Propagation, 50(5):759– 65, 2002. 7. Visible Light Communications Consortium. www.vlcc.net. 8. P. Djahani and J. M. Kahn. Analysis of infrared wireless links employing multibeam transmitters and imaging diversity receivers. IEEE Transactions on Communications, 48(12):2077–2088, 2000. 9. Dong-Yiel-Song, Yoon-Suk-Hurh, Jin-Woo-Cho, Jung-Hwan-Lim, Dong-WooLee, Jae-Seung-lee, and Youngchul-Chung. 4x10 Gb/s terrestrial optical free space transmission over 1.2 km using an EDFA preamplifier with 100 GHz channel spacing. Optics Express. 9 Oct. 2000; 7(8), 2000. Opt. Soc. America. 10. J. Foerster, E. Green, S. Somayazulu, and D. Leeper. Ultra-wideband technology for short- or medium-range wireless communications. Intel Technology Journal. 2001; (2), 2001. 11. D. Gesbert, M. Shafi, Da shan Shiu, P. J. Smith, and A. Naguib. From theory to practice: an overview of mimo space-time coded wireless systems. IEEE Journal on Selected Areas in Communications, 21(3):281–302, 2003. 12. Ghosh-A, Wolter-Dr, Andrews-Jg, and Chen-R. Broadband wireless access with WiMax/802.16: current performance benchmarks and future potential.

632

13. 14.

15.

16.

17.

18. 19.

20. 21. 22. 23.

24.

25.

26.

27.

28.

29.

Dominic O’Brien IEEE Communications Magazine, 43(2):129–36, 2005. IEEE USA JournalPaper doi:10.1109/MCOM.2005.1391513 English. D. J. T. Heatley, D. R. Wisely, I. Neild, and P. Cochrane. Optical wireless: The story so far. IEEE Communications Magazine, 36(12):72–+, 1998. D. J. T. Heatley, D. R. Wisely, I. Neild, and P. Cochrane. A review of optical wireless - what is it and what does if offer? British Telecommunications Engineering, 17:251–261, 1999. D.M. Holburn, V.A Lalithambika, V.M. Joyner, R. Samsudin, and R. J. Mears. Integrated CMOS transceiver for indoor optical wireless links. In Optical Wireless Communications IV, volume 4350, Denver, 2001. SPIE. M Jafar, S.J. Sheard, and D.C. O’Brien. Optimisation of coverage area for indoor optical free-space communications employing coherent detection. In 9th OptoElectronics and Communications Conference; 3rd International Conference on Optical Internet, Japan, 2004. IEEE. Jindong-Hou and D. C. O’Brien. Vertical handover decision-making algorithm using fuzzy logic for the integrated radio and ow system. IEEE transactions on wireless communications, 5(1):176–185, 2005. S. T. Jovkova and M. Kavehard. Multispot diffusing configuration for wireless infrared access. IEEE Transactions on Communications, 48(6):970–8, 2000. V. Jungnickel, V. Pohl, S. Nonnig, and C. von Helmolt. A physical model of the wireless infrared communication channel. IEEE Journal on Selected Areas in Communications, 20(3):631–40, 2002. Jvc. http://www.jvc.com/ds2/f prod.htm. JVC. Luciole link, 2003. J. M. Kahn and J. R. Barry. Wireless infrared communications. Proceedings of the IEEE, 85(2):265–298, 1997. J. M. Kahn, W. J. Krause, and J. B. Carruthers. Experimental characterization of non-directed indoor infrared channels. IEEE Transactions on Communications, 43(2-4):1613–1623, 1995. J. M. Kahn, R. You, P. Djahani, A. G. Weisbin, B. K. Teik, and A. Tang. Imaging diversity receivers for high-speed infrared wireless communication. IEEE Communications Magazine, 36(12):88–94, 1998. S. H. Khoo, W. Zhang, G.E. Faulkner, D.C. O’Brien, and D. J. Edwards. Receiver angle diversity design for high-speed diffuse indoor wireless optical communications. In Optical Wireless Communications IV, volume 4350. SPIE, 2001. S. H. Khoo, E. B. Zyambo, G. Faulkner, D. C. O’Brien, D. J. Edwards, M. Ghisoni, and J. Bengtsson. Eyesafe optical link using a holographic diffuser. In IEE Colloquium on Optical Wireless Communications, volume 128, pages 3/1–3/6. IEE, 1999. T. Komine and M. Nakagawa. Fundamental analysis for visible-light communication system using led lights. IEEE Transactions on Consumer Electronics, 50(1):100–7, 2004. Komine-T and Nakagawa-M. Integrated system of white led visible-light communication and power-line communication. In Proceedings of 13th IEEE International Symposium on Personal Indoor and Mobile Radio Communications, volume 4, pages 1762–6, 2002. V. A. Lalithambika, V. M. Joyner, D. M. Holburn, and R. J. Mears. Development of a CMOS 310 Mb/s receiver for free-space optical wireless links. In Proceedings of the SPIE The International Society for Optical Engineering, volume 4214 of Proceedings of the SPIE, pages 133–43. SPIE, 2001.

32 Cooperation in Optical Wireless Communications

633

30. K Liang, H Shi, S.J. Sheard, and D.C. O‘Brien. Transparent optical wireless hubs using wavelength space division multiplexing. In Free-Space Laser Communications IV, volume 5550, pages 80–87. SPIE, 2004. 31. D.P. Manage, S. H. Khoo, G.E. Faulkner, and D.C. O’Brien. Novel system for the imaging of optical multipaths. In High speed photography and detection, volume 5210, pages 47–54, San Diego, 2003. SPIE. 32. G. W. Marsh and J. M. Kahn. Performance evaluation of experimental 50Mb/s diffuse infrared wireless link using on-off keying with decision-feedback equalization. IEEE Transactions on Communications, 44(11):1496–1504, 1996. 33. M. J. McCullagh and D. R. Wisely. 155 Mb/s optical wireless link using a bootstrapped silicon apd receiver. Electronics Letters, 30(5):430–432, 1994. 34. A. J. C. Moreira, R. T. Valadas, and A. M. D. Duarte. Performance of infrared transmission systems under ambient light interference. IEE ProceedingsOptoelectronics, 143(6):339–346, 1996. 35. R. Narasimhan, M. D. Audeh, and J. M. Kahn. Effect of electronic-ballast fluorescent lighting on wireless infrared links. IEE Proceedings-Optoelectronics, 143(6):347–354, 1996. 36. D. C. O‘Brien. Improving the coverage and data rate in optical wireless communications. In Free space optical communications V, volume 5892, pages 58920X1– 9, San Diego, 2005. SPIE. 37. D. C. O’Brien, G. E. Faulkner, E. B. Zyambo, K. Jim, D. J. Edwards, P. Stavrinou, G. Parry, J. Bellon, M. J. Sibley, V. A. Lalithambika, V. M. Joyner, R. J. Samsudin, D. M. Holburn, and R. J. Mears. Integrated transceivers for optical wireless communications. IEEE Journal of Selected Topics in Quantum Electronics, 11(1):173–83, 2005. 38. D. C. O’Brien and M. Katz. Optical wireless communications within fourthgeneration wireless systems. Journal of optical networking, 4(6):312, 2005. 39. D. C. O’Brien, M. Katz, P. Wang, K. Kalliojarvi, S. Arnon, M. Matsumoto, R. Green, and S. Jivkova. Short-range optical wireless communications. In Technologies for the Wireless Future: Wireless World Research Forum (WWRF), Volume 2. Wiley, 2006. 40. T. O’Farrell and M. Kiatweerasakul. Performance of a spread spectrum infrared transmission system under ambient light interference. In Proceedings of Ninth International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC’98), volume vol. 1574, page 3, Boston, MA, USA, 1998. 41. M. R. Pakravan and M. Kavehrad. Indoor wireless infrared channel characterization by measurements. IEEE Transactions on Vehicular Technology, 50(4):1053– 73, 2001. 42. K. Phang and D. A. Johns. A 3-V CMOS optical preamplifier with dc photocurrent rejection. In ISCAS ’98. Proceedings of the 1998 IEEE International Symposium on Circuits and Systems, 1998. 43. P. F. M. Smulders. 60 GHz radio: prospects and future directions. In Proceedings Symposium IEEE Benelux Chapter on Communications and Vehicular Technology, Eindhoven, 2003. 44. Y. Tanaka, S. Haruyama, and M. Nakagawa. Wireless optical transmissions with white colored led for wireless home links proceedings of 11th international symposium on personal, indoor and mobile radio communication. In 11th IEEE International Symposium on Personal Indoor and Mobile Radio Communications. PIMRC 2000, pages 2 vol.xxxii+1603, London, UK, 2000.

634

Dominic O’Brien

45. Y. Tanaka, T. Komine, S. Haruyama, and M. Nakagawa. Indoor visible light data transmission system utilizing white led lights. IEICE Transactions on Communications, E86-B(8):2440–54, 2003. 46. R. van Nee and R. Prasad. OFDM for Wireless Multimedia Communications. Artech House, 2000. 47. W.T. Welford. The optics of nonimaging concentrators : light and solar energy. Academic Press, New York, 1978. 48. Carey Williamson. Wireless Internet: Protocols and Performance, pages 118– 142. Springer, 2004. 49. G. Yun and M. Kavehrad. Spot-diffusing and fly-eye receivers for indoor infrared wireless communications. In IEEE International Conference on Selected Topics in Wireless Communications, 1992.

33 Evolution of Digital Radios From Analog to Cognitive Features

Friedrich K. Jondral and Volker Blaschke Institut f¨ ur Nachrichtentechnik, Universit¨ at Karlsruhe (TH) [fj|blaschke]@int.uni-karlsruhe.de

Summary. Starting with some historic remarks about radio communications, this chapter first of all highlights the limitations on mobile communications, induced by the physics of electromagnetic wave propagation, and the importance of standards to meet technical as well as commercial goals. The development from past analog over digital and software defined to future cognitive radios is reviewed. Special attention is dedicated to the parameterization of software defined radios. The software communications architecture as a framework for the reconfigurability of radio hardware and the portability of waveforms is briefly described. Cognitive radio features in current and future communication systems are discussed.

33.1 Introduction Electronic information transmission over long distances became feasible around 1830 when amongst others Gauss and Weber did their transmission experiments over a copper line. Radio communication is not possible without the Maxwell equations that were first published in 1873. But it took another 13 years until in 1886 Heinrich Hertz succeeded in proving that electromagnetic waves are virtually able to travel through the free space. In 1901 the first transmission over the North Atlantic was realized by Guglielmo Marconi. Immediately after his success, the commercialization of radio transmission started in the UK with his Marconi Company and for example in Germany with Telefunken (in 1903). Since then radio communication became a major means of the modern society performing broadcast, emergency, police and military communication as well as mobile radio. Moreover, radio communications industry became a large economic player and employer in many countries.

33.2 Transmission Physics and Standards Maxwell’s equations provide us with the theoretical background of radio wave propagation. The optimum situation is given if there is a line of sight (LOS) between transmitter and receiver and if the geometry is static. Unfortunately, we usually have

635 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 635–655. c 2007 Springer.

636

Friedrich K. Jondral and Volker Blaschke

to consider no line of sight (NLOS) transmission and radio stations that move relatively to one another. This causes that the receiver simultaneously sees several paths of the transmitted signal which are due to reflections, scattering or diffraction. This observation results in the association that radio channels may be mathematically modeled by wide sense stationary uncorrelated scattering (WSSUS) processes [1,11]. Moreover, mobile radio transmissions according to the relative movement between transmitter and receiver encounter Doppler shift and Doppler spread [8]. Most of the physical phenomena influencing wave propagation (e.g., attenuation in free space and also by materials) prove to be frequency dependent. As a consequence, the range of electromagnetic waves decreases while the available transmission bandwidth increases with increasing radio frequency. So, we have to accept that video transmission is impractical for short wave radios and at the same time that micro wave links do not reach beyond the horizon. In daily life we use radio links for many different applications. This leads to a classification of radio networks and associated radio communication standards:

Wireless Personal Area Network (WPAN) WPANs are used in a relatively small area like offices or living rooms. Their range is relatively short (about 10 meters) but their data rate may be quite high (e.g., 1 Mbit/s). A typical representative of WPANs is Bluetooth that is mainly used to connect printers and scanners to computers or ear sets to mobile phones.

Cordless Phone DECT (Digital Enhanced Cordless Telecommunications) provides a cordless connection of handsets to the fixed telephone system for in-house applications. Its channel access mode is FDMA/TDMA and it uses TDD. The modulation mode of DECT is Gaussian minimum shift keying (GMSK) with a bandwidth (B) time (T) product of BT = 0.5. The transmission is protected only with a cyclic redundancy check (CRC).

Wireless Local Area Network (WLAN) Today, IEEE 802.11b/g has many installations all around the world. Also IEEE 802.11a systems are in operation. If IEEE 802.11a/g is to be implemented it should be recognized that its modulation mode is OFDM. It should be pointed out here that there are major efforts towards the development of joint UMTS/WLAN systems which use the software defined radio (SDR) approach.

Cellular Systems GSM (Global System for Mobile Communication) is presently the most successful mobile communication standard worldwide. Channel access is done via FDMA/TDMA and FDD/TDD is used. The modulation mode of GSM is GMSK with a bandwidth time product of BT=0.3. Error correction coding is done by applying CRC as well as a convolutional code. GSM was originally planned to be a voice communication system, but with its enhancements HSCSD, GPRS or EDGE it became more and more a data transmission system, too. In Europe, GSM systems are operating in the 900 MHz (GSM 900) as well as in the 1800 MHz (GSM 1800) bands. The North American equivalent of GSM is IS-136. Also GSM 1900 as well

33 Evolution of Digital Radios

637

as IS-95, a second generation CDMA system, are widely used in the US. UMTS (Universal Telecommunication System) is the European version of the third generation family of standards within IMT-2000. One of the differences with respect to second generation systems is that third generation systems are mainly developed for data (multi media) transmission. UMTS applies two air interfaces: UTRA-FDD and UTRA-TDD according to the duplex mode used. The channel access mode is CDMA. CRC, convolutional codes as well as turbo codes [14] are employed for error protection. The basic data modulation is QPSK. Furthermore it should be mentioned that one user within an UTRA-FDD cell can occupy up to seven channels (one control and six transport channels) simultaneously.

Professional Mobile Radio (PMR) PMR standards are developed for police, fire fighters and other administrative applications. The main difference to cellular systems is that they allow direct handheld-to-handheld communication. The main PMR systems in Europe are TETRA (recommended by ETSI) and TETRAPOL.

Location and Navigation One important feature of mobile terminals is their ability to determine their own location as well as to track location information. Today many location dependent services rely on the global positioning system (GPS). Currently the European satellite location and navigation system Galileo is under construction.

Digital Broadcast There is a possibility that digital broadcast systems may be used as downstreaming media within future communication infrastructures. The main developments in Europe in this area are digital audio broadcast (DAB) and digital video broadcast (DVB). In order to provide an overview of different radio applications and standards the most important parameters of selected air interfaces are summarized in Table 33.1.

33.3 Radio Usually, a radio is an audio broadcast receiver for the short wave or the ultra short wave region. Here, we name radio a combination of transmitter (Tx) and receiver (Rx), that is sometimes also called transceiver. A radio is used for information exchange by electromagnetic waves and may not only support audio but eventually also video and multi media transmissions. Radio waves are generated in the transmitter using an oscillator that produces a sine wave that then may be modulated with the information bearing signal (e.g., audio, video). The modulation product is shifted in frequency domain by mixing it to the transmission frequency and is then fed, after suitable processing (like filtering or amplification) to the transmitter antenna. The radiated electromagnetic waves travel through the channel and arrive at the receiver’s antenna. From there the signals are mixed downwards in order to make them audible by a loudspeaker

638

Friedrich K. Jondral and Volker Blaschke Table 33.1. Parameters of selected air interfaces. (a) Part I - Bluetooth, DECT, GSM, UTRA-FDD

Parameter

Bluetooth

DECT

GSM

UTRA-FDD

Frequency range

2.4 GHz (ISM-band)

1900 MHz

900, 1800, 1900 MHz

2 GHz

Channel bandwidth 1 MHz

1728 kHz

200 kHz

5 MHz

Access mode

TDMA

FDMA/TDMA

FDMA/TDMA DS-CDMA

Duplex mode

TDD

TDD

FDD

Users per carrier

8 maximum

12

8

-

Modulation

FH synchronized GMSK to master station, GFSK with modulation index between 0.28 and 0.35

GMSK

QPSK

Channel coding

-

No (CRC)

CRC, convolutional code

CRC, convolutional, turbo code

Bit- (chip-)rate

1 Mbps

1152 kbps

270.833 kbps

3.84 Mchip/s

Number of bits (chips) per burst (slot)

625

424

156.25

2560

Frame duration

-

10 ms

4.165 ms

10 ms

Number of bursts (slots) per frame

-

24

8

15

Burst (slot) duration

0.625 ms

0.417 ms

0.577 ms

0.667 ms

Maximum cell radius

5 - 10 m 300 m (1mW Tx power)

35 km (10 km)

Few km

Spreading sequences

-

-

-

User specific OVSF codes, call specific scrambling

Spreading factor

-

-

-

2k, k=2, 3,..., 8; 512 downlink only

Bit (chip) pulse shaping

Gauss (BT=0.5) Gauss (BT=0.5) Gauss (BT=0.3) RRC, rolloff factor 0.22

Data rate

1 Mbps

26 kbps

13 kbps

Evolutionary concepts

UWB

-

GPRS; HSCSD, HSDPA, HSUPA EDGE

Comparable systems

-

PHS, PACS

IS-136, PDC

FDD

8 kbps to 2 Mbps

UTRA-TDD, cdma2000

33 Evolution of Digital Radios

639

(b) Part II - TETRA, IEEE 802.11a, GPS, DVB-T Parameter

TETRA

IEEE 802.11a

GPS

Frequency range

400 MHz

5.2 und 5.5 GHz

1200, 1500 MHz VHF, UHF

Channel bandwidth 25 kHz

20MHz

-

7 (VHF) or 8 MHz (UHF)

Access mode

TDMA

FDMA/TDMA

DSSS

FDMA

Duplex mode

FDD/TDD

Half duplex

-

-

Users per carrier

4

-

-

-

Modulation

π/4-DQPSK

OFDM with subcarrier modulation BPSK/QPSK/ 16QAM/64QAM

BPSK, QPSK

OFDM with subcarrier modulation QPSK/16QAM/ 64QAM

Channel coding

CRC, Reed-Muller, RCPC

Convolutional code -

Reed-Solomon, convolutional code

Bit- (chip-) rate

36 kbps

6/9/12/24/36/ 48/54 Mbps

50 bps

9.143 Msamples/s for a 8 MHz channel

Number of bits (chips) per burst (slot)

510 (255 symbols)

52 modulation symbols per OFDM symbol

-

2k mode: 2048 + guard interval 8k mode: 8192 + guard interval

Frame duration

56.67 ms

Packets of several 15 s 100 µs

68 OFDM symbols

Number of bursts (slots) per frame

4

Variable

68

Burst (slot) duration

14.167 ms

1 OFDM-Symbol 30 s of 3.2 µs + 0.8 µs guard time

2k mode: 224 µs + guard time 8k mode: 896 µs + guard time

Maximum cell radius

14 km

Some 10 m

-

-

Spreading sequences

-

-

Gold- or PRN code

-

Spreading factor

-

-

1023 or 10230

-

Bit (chip) pulse shaping

RRC, roll-off factor 0.35

-

-

Rectangular, other filtering possible

Data rate

Up to 28.8 kbps Up to 25 Mbps

-

49.8 - 131.67 Mbps

Evolutionary concepts

-

IEEE 802.11n

Galileo

-

Comparable systems

TETRAPOL

HiperLAN/2

GLONASS

DAB

5 subframes

DVB-T

640

Friedrich K. Jondral and Volker Blaschke

or visible via a screen. Initially, the signal processing was done by analog means like resonant circuits or vacuum tubes.

33.4 Digital Radio (DR) Just after World War II many new ideas emerged, that effectively influenced the evolution of digital radio technology. In the field of applied mathematics Claude E. Shannon published the basic ideas of information theory, especially coding [27]. In his paper [28] he formulated the sampling theorem stating that an analog lowpass signal and its sampled version are equivalent if and only if the sampling rate fs is higher than twice the highest frequency fmax contained in the signal and that one of the signals may easily be computed from the other. This theorem is a fundamental prerequisite of all digital signal processing. Another idea is of technological nature: The transistor that was invented 1947 by William B. Shockley, John Bardeen and Walter Brattain also in the Bell Laboratories. Of course it took some time to develop the beauty of digital signal processing [26], but then in the second half of the 1970s time had come to develop the first digital short wave radios. Figure 33.1 shows a block diagram of a digital transmitter. The signal processing (e.g., coding and modulation) is done in the complex baseband by digital means. The resulting inphase (I) and quadrature (Q) components are digital-to-analog (D/A) converted, lowpass filtered and mixed to the radio frequency (RF), combined to the (real) bandpass signal, amplified, bandpass filtered and then radiated by the antenna.

Figure 33.1. Transmitter.

The first digital receivers used the superhet principle sketched in Figure 33.2(a). The incoming signal, bandpass filtered after the antenna, is mixed to a first intermediate frequency (IF), filtered again in order to suppress mirror frequencies and mixed to the second IF. After a third bandpass that simultaneously is used as anti-aliasing filter the signal is analog-to-digital (A/D) converted and subsequently processed in a digital signal processor (DSP). An alternative to the superhet is the direct mixing receiver (c.f. Figure 33.2(b)) that after bandpass filtering mixes down the signal to the complex baseband, anti-alias filters and then converts it to the digital domain in I and Q branches before digital signal processing takes place. Advantages and drawbacks of superhet and direct mixing receivers are opposed in Table 33.2.

33 Evolution of Digital Radios

641

(a) Superhet

(b) Direct mixing (Zero IF) Figure 33.2. Receiver structures. Table 33.2. Comparison of receiver architectures. Architecture

Advantages

Drawbacks

Superhet

- High sensitivity

- Monolithic integration Applicable in impossible base station - Compromising of am- receivers only

- High selectivity

SDR

- No I/Q mismatch if bandpass subsampling plification, noise figure, stability and power conis applied sumption necessary - A/D converter: High resolution on high sampling rate necessary, aperture jitter Direct mixing - No IF processing

- Direct current (DC) Applicable for mobile terminals - No mirror frequencies offset Local oscillator (LO) - Low noise amplifiers leakage simply realizable - Monolithically inte- - I/Q mismatch grable

642

Friedrich K. Jondral and Volker Blaschke

The most critical building block of a digital receiver is its A/D converter. It realizes Shannon’s sampling theorem. The performance of the A/D converter is mainly determined by its amplitude resolution and sampling rate. The A/D converter’s resolution determines the receiver’s dynamic range as well as the quantization noise while the sampling rate tells us the bandwidths of the signals that can be received. An excellent overview over the A/D converters’ performances currently available is provided in [2]. The main advantage of digital radios over its analog predecessors is that digital processing may easily be adapted to the received signal. If, for example, filters like pulse shapers in the transmitters and the associated matched filters in the receivers are realized as finite impulse response (FIR) systems (c.f. Figure 33.3) changing the filter order N as well as the filter coefficients {hn ; n = 0, 1, ..., N − 1} is quite simple: Doing so, the filter characteristics can be optimized to a signal. Currently, the use of FIR filters is very popular in digital radios because they are always stable and linear phase behavior is realized by choosing the filter coefficients according to a simple symmetry [26].

Figure 33.3. FIR filter.

During the 1980s many digital receiving systems, especially for radio intelligence applications were developed. These systems included multi channel filter bank receivers (and direction finders) based on the fast Fourier transform (FFT) for the detection and tracking of frequency hopping signals as well as adaptive RAKE receivers for detection of direct sequence spread spectrum (DSSS) transmissions. Some of these systems were not only able to detect specific signals but could recognize, demodulate and decode a plethora of different air interfaces. From the present point of view, these systems, although voluminous, were software defined receivers. Also at the beginning 1980s the development of digital transceivers for base stations and mobile terminals of second generation cellular networks (e.g., GSM, IS-136, IS-95) started. The main goal of these devices was not flexibility but low power consumption for mobile devices. Soon after GSM had been introduced in the 900 MHz band in Europe this standard was also used in the 1800 MHz band in Europe and in the 1900 MHz band in North America. Therefore, the demand for transceivers, that operate in all three GSM bands and that are now called multi band radios, arose. With the deployment of second generation cellular systems in the beginning 1990s commercial systems took over the lead in development of digital transceivers from more military related systems. The definition of UMTS during

33 Evolution of Digital Radios

643

the 1990s included another feature of SDRs: UMTS transceivers have to be able to support GSM, too, and therefore are multi standard transceivers.

33.5 Software Defined Radio (SDR) The functions of a software radio (SR) are realized as programs running on a suitable processor. The transmitter algorithms produce radio signals compliant with the standards that are supported by the SR. The receiver algorithms are developed to optimally recover the information sent by the transmitter. An ideal SR directly samples the antenna output. A SDR is the realizable version of a SR: The received signals are sampled after a suitable band selection filter. According to its operational area, a SDR can be: • •

• •

A multi band system that is supporting more than one dedicated frequency band used for a wireless standard (e.g., GSM 900, GSM 1800, GSM 1900). A multi standard system that is supporting more than one air interface. Multi standard systems can work within one standard family (e.g., UTRA-FDD, UTRA-TDD for UMTS) or across different networks (e.g., DECT, GSM, UMTS, WLAN). A multi service system that provides different services (e.g., telephony, data, video streaming). A multi channel system that supports two or more independent transmission and receiving channels simultaneously.

In this section we focus the discussion on multi mode systems which are combinations of multi band and multi standard systems. The SDR approach allows different levels of reconfiguration within a transceiver: •

•

•

•

Commissioning: The configuration of the system is done once at the time of product shipping when the customer has asked for a dedicated mode (standard or band). This is not a true reconfiguration. Reconfiguration with downtime: Reconfiguration is done only a few times during the product lifetime, e.g., when the network infrastructure changes. The reconfiguration will take some time where the system is switched off. This may include the exchange of (software or hardware) components. Reconfiguration on per-call basis: Here, reconfiguration is a highly dynamic process that works on a per-call based decision. That means no downtime is acceptable. Only parts of the whole system (e.g., front end, digital baseband processing) can be rebooted. Reconfiguration per timeslot: Reconfiguration can even be done during a call.

33.5.1 SDR Processing Figure 33.4 presents the coarse structure of a SDR transceiver that differs from a conventional transceiver only by the fact that it can be reconfigured via a control bus supplying the processing units with the parameters that describe the desired standard (c.f. Table 33.1). Such a structure guarantees that the transmission standard can be changed instantaneously if necessary (e.g., for inter standard handover).

644

Friedrich K. Jondral and Volker Blaschke

Figure 33.4. SDR transceiver.

It is very likely that the principle of choice for a SDR receiver structure will be direct mixing as depicted in Figure 33.2(b). At the output of the A/D converters we get the received signal in its complex digital baseband representation. The next step is the adaptation of the sampling rate to the signal’s standard. The reason for the sampling rate adaptation is that the signal processor should work at the minimum possible rate. For a given standard this minimum rate depends on fc = 1/Tc , the symbol or chip rate, respectively. Usually, a sampling rate of fs = 4fc is sufficient for the subsequent signal processing where, after precise synchronization, the sampling rate may be reduced by another factor of 4. If the fraction of the sampling rates at the resampler’s output and input is rational, the sampling rate conversion can be implemented by an increasing of the sampling rate followed by an interpolation filter and a sampling rate decreasing. The whole procedure is a cascade of interpolation and decimation. If the interpolation lowpass is realized by an FIR filter, the impulse response usually becomes quite long. The solution to this problem is to take the up and down sampling into account within the filter process: Since the upsampled signal is generated by the insertion of zeros, the processing of these zeros can be omitted within the filter. This leads to a polyphase structure of the sampling rate adaptation FIR filter. After sampling rate adaptation the signal is processed within the complex baseband unit (demodulation and decoding). The baseband processing is discussed in some detail in [18]. The baseband unit hands over the decoded bits to the data processing unit. The transmission branch consists of the procedures inverse to that of the receiving branch. I.e., the signal to be transmitted is generated as a complex baseband signal, from which the real part is taken to be shifted to the (transmission) RF.

33.5.2 Parameterization For SDRs adaptability means that the radio is able to process signals of different standards. One method to achieve adaptability is parameterization of standards. As a communications standard we define a set of documents that describe the functions of a system in such a way that a manufacturer can develop terminals or infrastructure equipment on this basis. Standardization is one necessary condition to make a communication system successful on the market as exemplified by GSM. Standardization grasps all kinds of communication networks.

33 Evolution of Digital Radios

645

Of course a standard has to contain all the functions of a system. Especially for mobile systems at least the air interface including the protocol stack has to be specified. Parameterization means that every standard is looked upon as one member of a family of standards [18]. The signal processing and the protocol structures are then developed in such a way that these structures may be switched by parameters to realize the different air interfaces. For implementation there are substantial differences between second generation TDMA standards like GSM and IS-136 on the one hand and third generation standards like UMTS on the other. Within UMTS spreading at the transmitter and despreading at the receiver have to be realized additionally. Looking at the signal processing chains we remark that the error correcting codes of all the second generation standards are quite similar: A combination of block codes for the most important bits and a convolutional code for the larger part of the voice bits is applied. Channel coding for data transmission is done by a powerful convolutional code. UTRA-FDD as a third generation air interface offers net data rates of up to 2 Mbit/s and guarantees bit error rates (BERs) of up to 10−6 for specific applications. To reach these BERs turbo codes are employed. Of course within a SDR all coding/decoding procedures have to be integrated into a general structure. The development of a general coding/decoding structure is essentially a matter of diligence since all procedures rely on linear (recursive) shift registers. General modulator/demodulator structures are presented in [18].

33.5.3 Military SDR - The Software Communications Architecture (SCA) The Joint Tactical Radio System (JTRS) represents the future (mobile) communications infrastructure of the US joint forces. Introducing JTRS stands for an essential step towards the unification of radio communication systems, the transparency of services, and the exchangeability of components, i.e., the portability of waveforms as well as the reconfiguration of hardware (c.f. Figure 33.5, [7]). The development of the JTRS is accompanied and supervised by the US forces’ Joint Program Office (JPO). Development, production, and delivery continue to be the tasks of competing industrial communications software and hardware suppliers. An important new aspect added by the JTRS setup is that the suppliers are guided to aim for a most perfect interchangeability of components due to the supervision function of the JPO. The tool used by the JPO is the software communications architecture (SCA) [17], an open framework that prescribes the developing engineers how the hardware or software blocks have to act together within the JTRS. The communication devices emerging from this philosophy are clearly SDRs. A major group of suppliers and developers of communication software and hardware founded the SDR Forum to promote their interests. The importance of the SDR Forum, however, reaches well beyond the application of SDRs in the JTRS. This is underlined by the SDR Forum membership of European and Asian industrial and research institutions that usually are mainly interested in the evolution of commercial mobile communication networks. The SCA describes how waveforms are to be implemented onto appropriate hardware devices. A waveform is defined by the determination of the lower three layers

646

Friedrich K. Jondral and Volker Blaschke

Figure 33.5. Reconfigurability of a platform and portability of a waveform.

(network, data link, physical) of the ISO/OSI model. Therefore, waveform is another synonym of standard or air interface. Based on the waveform definition, a transmission method is completely determined. The definition of a waveform, therefore, lays down the modulation, coding, access, and duplex modes as well as the protocol structure of the transmission method. The SCA defines the software structure of an SDR that may be usable within the JTRS. The underlying hardware as well as the software is described in object-oriented terms. Moreover, the structures of application program interfaces (APIs) and of the security environment are described. Each component has to be documented in a generally accessible form. The JTRS operating environment (OE) defined in the SCA consists of three main components: • • •

A real-time operating system A real-time request broker The SCA core framework

When developing a SCA compliant radio device the supplier gets the operating system and the CORBA middleware from the commercial market. The core framework as well as the waveform is developed by him or he also gets it from the market or (in future) it may be contributed by the JPO. The SCA is the description of an open architecture with distributed components. It strictly separates applications (waveforms) from the processing platform (hardware, operating system, object request broker, core framework). It segments the application functions and defines common interfaces for the management and the employment of software components. It defines common services and makes use of APIs to support the portability of hardware and software components and of applications. The connections between the applications and the core framework within the SCA are given by the APIs. Standardized APIs are essential in assuring the portability of applications as well as for the exchangeability of devices. APIs guarantee that application and service programs may communicate with one another, independent of the operating system and the programming language used. APIs are waveform specific since uniform APIs for all waveforms would be inefficient for implementations with bounded resources. Therefore, the goal is to have a standard set of APIs for each waveform. The single APIs are essentially given by the layers of the ISO/OSI model:

33 Evolution of Digital Radios

647

(i) A PHY API supports initialization and configuration of the system in non-realtime. In real-time it takes care of the transformation of symbols (or bits) to RF in the transmitter branch. In the receiver branch it transforms RF signals to symbols (bits). (ii) A MAC API supports all the MAC functions of the ISO/OSI layer model (e.g., timeslot control in TDMA or FEC control). (iii) An LLC API makes an interface for the waveform’s link layer performance available (according to the ISO/OSI layer model: data link services) on component level. (iv) A network API makes an interface for the waveform’s network performance available on component level. (v) A security API serves for the integration of data security procedures (INFOSEC, TRANSEC). (vi) An input/output API supports the input and output of audio, video, or other data. The security relevant SCA aspects are described in the SCA security supplement [2]. The SCA security functions and algorithms are of course defined with respect to the military security requirements of JTRS. Currently, many initiatives are undertaken in different countries to adopt the SCA philosophy of JTRS to other national or multinational systems. One example for such an investigation into non-military but security applications is the European Wireless Interoperability for Security (WINTSEC) project.

33.6 Cognitive Radio The increased flexibility of SDR leads to a number of continuative developments. One important advantage of these terminals is the ability for fast adaptation to different operation conditions using the same hardware. This offers the possibility for changing the terminal’s communication features without significant effort in redesigning the terminal’s hardware. Especially under economic aspects this enhancement becomes an important point. Furthermore, the flexibility of SDRs enables the simultaneous adaptation of the terminal to the current working conditions. Basically, this includes all influences to the wireless communication like signal distortion due to the communication channel in general as well as QoS aspects due to user interaction. The system parameters of the 2G communication systems, like GSM or IS-95, are static (c.f. Table 33.1) and defined assuming the worst-case scenario these systems should support. This means that also the supported services should work under bad conditions but cannot profit from better channel conditions in normal situations. Adapting the signal processing to the spectral environment offers additional benefit in efficient wireless data transmission. So, the technological advantages of SDR can build a well-suited basis for enhanced communication technologies using dynamic adaptation for an efficient utilization of radio spectrum. Already at the end of the 1990’s dynamic allocation of radio spectrum went into the focus of research activities [22, 25]. The results presented in [24] show that a capable deregulation of the current static spectrum allocation gives the opportunity to solve the observed spectrum scarcity. The first approaches for flexible spectrum allocation considering the network traffic conditions has been studied during the last years. Today they are summarized by the term Dynamic Spectrum Allocation [23].

648

Friedrich K. Jondral and Volker Blaschke

In addition to an adaptive allocation of dedicated frequency channels, a number of parameters can be considered for optimizing the current transmission. So, a flexible hardware as well as a number of additional information are necessary in order to evaluate the working conditions and to realize a well suitable adaptation of the terminal to the detected situation. The Cognitive Radio (CR) concept published by J. Mitola describes a terminal which is additionally characterized by learning and reasoning algorithms in order to get advance of previous configurations of the terminal and the characteristic user’s behavior. The basic relations of such a self-learning, self-adaptive radio are described in the so-called Cognition Cycle [20]. Starting with the recognition of the current environment of the radio including also time, position and spectral utilization as well as information about the user-specific data input, the Cognition Cycle depicts the different states of cognition, evaluation, action and learning. One important phase of this cycle is the generation and evaluation of alternative terminal configurations in order to optimize the mobile transmission considering the influences of the mobile’s environment. Besides the cognition of the terminal’s outside world also the hardware and software skills have to be known and considered for the decision. Finally, the resulting outcome builds the base for future decisions in the sense of learning from experiences. Based on the fact that most of the daily life’s activities are periodic within each day or week this approach seems to be reasonable. Besides the conceptual approach of Mitola the term Cognitive Radio is also defined in [13]: “Cognitive Radio is an intelligent wireless communication system that is aware of its surrounding environment (i.e., outside world), and uses the methodology of understanding-by-building to learn from the environment and adapt its internal states to statistical variations in the incoming RF stimuli by making corresponding changes in certain operation parameters (e.g., transmit-power, carrier frequency, and modulation strategy) in real-time, with two primary objectives in mind: • Highly reliable communications whenever and wherever needed; • Efficient utilization of the radio spectrum.” Based on Mitola’s concept and the definition given in [13] CR terminals can be characterized by the following attributes: • Awareness: The terminal is able to understand the RF environment as well as its spatial surrounding and position. • Adaptivity: The terminal is able to understand the user’s behavior and can follow and learn from the user’s action. • Autonomy: The terminal is able to act and react in order to adjust itself to a situation without any user interaction. The knowledge about local policy restrictions and the own hardware and software restrictions is required. • Adjustability: The terminal is able to respond to its observed environment including mechanisms for adapting power emission, allocated bandwidth and modulation type in real time. This also requires knowledge about the interference situation at the transmitter and receiver in order to be in line with the local policy restrictions.

33.6.1 Cognitive Radio Structure Based on the definition and features described above, a wide range of applications for CRs can be found. Some of them are discussed in [16] which need a very complex

33 Evolution of Digital Radios

649

data processing compared with todays mobile terminal performance. Nevertheless, the possible applications can be classified into two main groups: •

•

Technology-centric (TC) characteristics like positioning, observation of the spectral environment, learning and reasoning, adaptation of the physical layer configuration, etc. User-oriented characteristics which support the user’s needs for information services like local events, nearest train station, etc. These could also be realized using already existing data base services which are located at the core network side.

The first statement is necessary for the realization of CR features on radio network or mobile device level and builds the precondition for the user-oriented CR features in the second statement [16]. In Figure 33.6 the general structure of a technologycentric CR is depicted which is built by enhancing SDRs with additional functional units providing CR features.

Figure 33.6. General structure of a technology-centric Cognitive Radio.

33.6.2 Functional Enhancement of Cognitive Radios While a main goal of SDR development is an efficient utilization of flexible hardware, the design of CR has been driven by increasing the efficiency of the spectral utilization. This objective is also motivated by the fact that the total utilization within the frequency range of 30 MHz up to 6 GHz is only about 16 percent [10,29]. Furthermore, the allocation of the observed frequency bands is characterized by a significant spatial and temporal variation. On the one side hot-spot scenarios can be

650

Friedrich K. Jondral and Volker Blaschke

pointed out where the request for transmission capacity is higher than the available radio resources (e.g., mega-events). High blocking rates and low QoS performance are some of the consequences. On the other side so called desert areas can be found in the sense that the request for wireless data transmission is significantly smaller than the available resources. Within such frequency bands, the spectral utilization is nearly zero over long periods of time. Furthermore, the different frequency bands can be beneficial or adverse for the quality of the allocated mobile services due to the usable channel bandwidth, current Doppler spread, user velocity or channel attenuation. So, the aim of good QoS performance and efficient and cost-effective spectrum allocation motivates the adaptation of the transmission to the physical conditions. Dynamic Spectrum Management provides a framework for an adaptation of single terminal components up to the complete transmission chain from the transmitter to the receiver [21, 23]. In [12] three general principles for a flexible spectrum management are proposed: • • •

Spectrum allocation Spectrum leases Spectrum sharing

These three principles have in common the adaptive allocation of frequency bands for optimizing the spectral utilization. Depending on the special working scenario, a short term adaptation (spectrum sharing) or a long term adaptation (spectrum allocation) will fulfill the user’s as well as the provider’s requirements. Some of the described principles can already be found in currently established systems even if the system is not explicitly specified as a Cognitive Radio system. The following subsection will give a brief overview about cognitive features in current and future systems.

33.6.3 Cognitive Radio Features in Current and Future Communication Systems The CR structure presented in Figure 33.6 includes several elements a terminal can provide in order to be a CR. Some of them can be found in wireless communication systems which are already on the market without an explicit relation to CR. Also some of them are part of system concepts which are still subject of research. In the following four wireless systems or system concepts using some kind of cognitive features will be briefly described:

WLAN The physical layer specification of the two subgroups, IEEE 802.11a/g and 802.11b, includes several modulation schemes. Regarding to the channel attenuation and the signal to noise ratio (SNR) the number of modulated symbols is adapted. In IEEE 802.11b the data rate is modified by changing the modulation and the channel coding. For a robust transmission using low data rates (1-2 Mbps), the modulation is set to BPSK or QPSK. For higher data rates (5.5-11 Mbps) Complemetary Code Keying (CCK) is used additionally (c.f. Figure 33.7). In IEEE 802.11a/g higher modulation schemes and code rates of R= 21 , 23 , 34 are used for data rate adaptation (c.f. Figure 33.7) [3]. Furthermore, an adaptive resource allocation between bordering

33 Evolution of Digital Radios

651

access points is under specification in the TaskGroup 802.11k [9]. Both mechanisms, the data rate adaptation as well as the adaptive resource allocation, can be described as cognitive features. In contrast to the CR concept described by Mitola [20] the adaptation of the data rate depends only on the present channel conditions. The adaptation is completely independent from previous channel and terminal states.

Figure 33.7. IEEE 802.11 - specified data rates.

Spectrum Pooling The system proposed in [30] enhancing the general thought of [12] for secondary use of unallocated frequency bands describes a potential solution for increasing the spectral utilization. A local area network is installed over a licensed cellular TDMA/FDMA communication system as an overlay system. Assuming a nonpermanent channel allocation of the primary cellular system the overlay network can temporarily allocate the unused transmission resources. In order to minimize the interferences between both systems several preconditions have to be fulfilled. First, the overlay system uses OFDM for a simplified adaptive sub-channel allocation. Moreover, a robust detection of the primary system’s users is required. A special protocol was proposed in order to boost the allocation information to all subscriber terminals within the coverage area of the overlay system’s access point [30]. This approach describes a direct and fast adaptation of the secondary system’s terminals to the current spectral allocation.

E2 R Within this European Project the inter working of several networks like cellular, local area or broadcast networks is discussed. This project investigates the enhancements achievable by adapting the complete transmission chain from one end to the other of a connection. Also cognitive features like traffic load and user demands are considered. Due to the cooperation between different network providers extreme

652

Friedrich K. Jondral and Volker Blaschke

traffic situations can be handled more efficiently. Also concepts for Advanced Spectrum Management (ASM), Joint Radio Resource Management (JRRM) and Dynamic Network Planing and Management (DNPM) in the sense of an end-to-end adaptation are investigated [19]. In E2 R ASM enables adaptive allocation of dedicated frequency bands by different Radio Access Technologies (RATs). The JRRM optimizes the throughput of the available RAT which can also be influenced by a dynamic radio cell behavior. This is controlled by the DNPM [19]. Besides the three main functionalities several methods and algorithms for real time spectrum auctioning [5, 6], adaptive resource allocation and vertical system handover are investigated [31].

IEEE 802.22 - Wireless Standard Based on Cognitive Radios In the IEEE 802.22 specification [15] a Cognitive Radio system is described providing frequency agility, dynamic spectrum allocation, adaptive modulation, transmit power control, location awareness and negotiated use. This concept of a Wireless Regional Area Network (WRAN) allows the usage of the VHF/UHF frequency bands from 41 MHz to 910 MHz considering the national regulations of these bands. Within this frequency range, TV and radio broadcast services, wireless microphones and telephones are allocated license users. Therefore, the CR system operates as an overlay network on the condition of minimum interference caused to the licensed terminals. The system specification allows a PHY data rate up to 24 Mbps using adaptable OFDM-modulation of BPSK to 256QAM. Four different QoS classes are supported: constant bit rate, real and non-real time transmission with variable bit rate, and best effort. In order to adjust to the licensed system, the CR terminals have to map to the channel bandwidth of 6, 7 or 8 MHz [4].

33.7 Conclusion Since the commercialization of radio transmission in the very beginning of the 20th century the development of efficient radio communication technology has played an important role. Driven by the request for an efficient wireless data transmission, several major levels of radio technology development can be figured out. All of these processes are limited by the physics of wave propagation theoretically described by the Maxwell equations. Furthermore, the specific influences of non line of sight transmissions and mobile transmitters and receivers have to be taken into account for modeling realistic wave propagation. The wide frequency range from some kHz up to several GHz which is useful for radio communication also leads to a number of aspects that have to be considered designing a modern communication system. Therefore, today’s radio communication networks can be classified according to the provided applications as it was described in Section 33.2. Of course, the different radio applications affect the design of the transmitter and receiver. But in general a transmitter is used for the generation of the radio waves including the modulation of the signal carrier and the shifting to the radio frequency and the receiver does the inverse processing (c.f. Section 33.3). The digitization of the radio technology after the World War II brought one of the important enhancements. Due to digital signal

33 Evolution of Digital Radios

653

processing, the size of transceivers could significantly decrease. The first digital receivers used the superhet principle for down-shifting the incoming radio signal to the baseband as it is described in Section 33.4. The alternative structure is the ZeroIF receiver also described in Section 33.4. Set up to these principles the development of SDRs, described in Section 33.5, depicts the next level of evolution. Due to the moving of the A/D conversion towards the antenna, most of the signal processing can be realized as programs running on a suitable processor. The increased flexibility of the radio hardware results in lower design and production costs. Also the implementation of additional radio applications is simplified as long as the hardware can provide the additional requirements. The basic principles of SDR processing and the advantages of the parameterization were illustrated in the Subsections 33.5.1 and 33.5.2. Also the specific aspects of military SDRs taking into account the interoperability of different software and hardware platforms were explained in Subsection 33.5.3. The Cognitive Radios described in Section 33.6 represent the current level of evolution. Including additional information about the terminal’s spectral and spatial environment and the user’s behavior, the data transmission can be specifically adapted. This includes also aspects of dynamic spectrum management. In Section 33.6 also a brief overview of current systems and concepts using Cognitive Radio features was given. Similar to the previous evolution steps, the implementation of the additional cognitive features and enhancements will also be driven by economic as well as by pragmatic requirements. Nevertheless, the high research effort necessary for each advance in the development of radio technology has built the basis for another step in technological evolution.

References 1. P. Bello. Characterization of Randomly Time-Variant Linear Channels. IEEE Trans. on Communications Systems, pages 4–37, Dec 1963. 2. Bin Lee, T.W. Rondeau, J.H. Reed, C.W. Bostian. Analog-to-digital converters. IEEE Signal Processing Magazine, 22:69–77, Nov 2005. 3. B. Bing, editor. Wireless Local Area Networks. Wiley & Sons, New York, USA, 2002. 4. C. Cordeiro, K. Challapali, D. Birru, Sai Shankar. IEEE 802.22: the first worldwide wireless standard based on cognitive radios. In Proceedings of IEEE 1st International Symposium on New Frontiers in Dynamic Spectrum Access Networks (DySPAN ’05), pages 328–337, Baltimore (MD), USA, 8-11 Nov. 2005. 5. C. Kloeck, H. Jaekel, F. Jondral. Auction sequence as a new resource allocation mechanism. In Proceedings of IEEE 62nd Vehicular Technology Conference, VTC Fall 2005, volume 1, pages 240–244, Dallas (TX), USA, 25-28 Sept. 2005. 6. C. Kloeck, H. Jaekel, F. Jondral. Dynamic and local combined pricing, allocation and billing system with cognitive radios. In Proceedings of IEEE 1st International Symposium on New Frontiers in Dynamic Spectrum Access Networks (DySPAN ’05), pages 73–81, Baltimore (MD), USA, 8-11 Nov. 2005. 7. C. Serra, S. Martin, E. Nicollet. Impact of the SCA on HDR waveform modeling and design. In Proceedings of the 2004 Software Defined Radio Technical Conference, volume A, pages A153–A157, Scottsdale, AZ (USA), Nov 2004.

654

Friedrich K. Jondral and Volker Blaschke

8. R. Clarke. A statistical theory of mobile-radio reception. The Bell System Technical Journal, 47:957–1000, July/August 1968. 9. E. Garcia Villegas et al. Load balancing in WLANs through IEEE 802.11k mechanisms. In Proceedings of 11th IEEE Symposium on Computers and Communications, ISCC ’06, pages 844–850, Cagliari, Sardinia, Italy, 26-29 June 2006. 10. FCC. Spectrum policy task force report, ET Docket No. 02-155. Technical report, FCC, Nov. 2002. 11. B. Fleury. Charakterisierung von Mobil- und Richtfunkkan¨ alen mit schwach station¨ aren Fluktuationen und unkorrelierter Streuung (WSSUS). PhD thesis, ETH Z¨ urich, Z¨ urich, Switzerland, 1990. 12. G. Staple, K. Werbach. The end of spectrum scarcity [spectrum allocation and utilization]. IEEE Spectrum, 41:48–52, Mar. 2004. 13. S. Haykin. Cognitive radio: brain-empowered wireless communications. Selected Areas in Communications, IEEE Journal on, 23(2):201–220, Feb. 2005. 14. C. Heegard and S. Wicker. Turbo Coding. Kluwer Academic Publishers, Boston, MA (USA), 1999. 15. IEEE. P802.22: Cognitive radio, wide regional area network. Technical Specifications, May 2005. 16. J. Mitola, G.Q. Maguire. Cognitive Radio: Making software radios more personal. IEEE Personal Communications, 6:13–18, Aug. 1999. 17. Joint Tactical Radio System (JTRS) Joint Program Office. Software communications architecture specification, jtrs-5000sca v3.0. http://jtrs.army.mil, August 2004. 18. F. Jondral. Software Defined Radio - Enabling Technology, chapter Parametrization - a Technique for SDR Implementation, pages 232–256. John Wiley and Sons, W. Tuttlebee, Ed., London, 2002. 19. K. Moessner et al. Functional architecture of end-to-end reconfigurable systems. In Proceedings of IEEE 63rd Vehicular Technology Conference, VTC Spring 2006, volume 1, pages 196–200, Melbourne, Australia, 7 - 9 May 2006. 20. J. Mitola. Cognitive Radio - An Integrated Agent Architecture for Software Defined Radio. PhD thesis, Royal Institut of Technology (KTH), Kista, Sweden, 2000. 21. M. Nekovee. Dynamic spectrum access with cognitive radios: Future architecures and research challenges. In Proceedings of 1st International Conference on Cognitive Radio Oriented Networks (CrownCom 2006), Mykonos, Greece, 8-10 June 2006. 22. P. Leaves, et al. Dynamic spectrum allocation in a multi-radio environment: Concept and algorithm. In IEE Second International Conference on 3G Mobile Communication Technologies, pages 53–57, London, UK, 26-28 March 2001. 23. P. Leaves, et.al. Dynamic spectrum allocation in composite reconfigurable wireless networks. Communications Magazine, IEEE, 42(5):72–81, March 2004. 24. P. Leaves, J. Huschke, R. Tafazolli. A summary of dynamic spectrum allocation results from DRiVE. In IST Mobile and Wireless Telecommunications Summit, pages 245–250, Thessaloniki, Greece, 16-19 June 2002. 25. R. Toenjes, et.al. Architecture for future generation multi-access wireless system with dynamic spectrum allocation. In Proceedings of IST Mobile Communications Summit, Galway, Ireland, Oct. 2000. 26. L.R. Rabiner and B. Gold. Theory and Application of Digital Signal Processing. Prentice Hall, Englewood Cliffs, NJ (USA), 1975.

33 Evolution of Digital Radios

655

27. C.E. Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27:379–423, 623–656, 1948. 28. C.E. Shannon. Communications in the presence of noise. Proc. IRE, 37:10–21, 1949. 29. Shared Spectrum Company. Comprehensive spectrum occupancy measurements over six different locations. www.sharedspectrum.com, August 2005. 30. T.A. Weiss, F. Jondral. Spectrum pooling: an innovative strategy for the enhancement of spectrum efficiency. Communications Magazine, IEEE, 42(3):8– 14, March 2004. 31. Z. Boufidis, et al. End-to-end architechture for adaptive communication systems. In Proceedings of IEEE 64th Vehicular Technology Conference, VTC Fall 2006, Montreal, Canada, 25-28 Sept. 2006.

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network Tao Chen, Honggang Zhang, Xiaofei Zhou, Gian Mario Maggio, and Imrich Chlamtac CREATE-NET [tao.chen|honggang.zhang|xiaofei.zhou|gian-mario.maggio| imrich.chlamtac]@create-net.org Summary. As the radio spectrum usage paradigm shifting from the traditional command and control allocation scheme to the open spectrum allocation scheme, wireless ad-hoc networks meet new opportunities and challenges. The open spectrum allocation scheme has potential to provide those networks more flexibility, reliability, availability and capacity. However, the freedom brought by the new spectrum allocation scheme introduces spectrum management and network coordination challenges. For instance, wireless ad-hoc networks usually rely on a global common control channel for coordination. Such a control channel may, however, not always available in an open spectrum allocation scheme due to the interference and the need for coexistence with primary users of the spectrum. In this chapter, we propose a cluster-based framework to form a wireless mesh network in the context of open spectrum sharing. Clusters are constructed by neighbor nodes sharing common channels, and the network is formed by interconnecting the clusters gradually. We identify issues in such a network and provide mechanisms for neighbor discovery, cluster construction, network formation, and network topology management. The unique feature of this type of networks is its capability to intelligently adapt to the network and radio environment change.

34.1 Introduction The radio spectrum usage is undergoing a paradigm shift from the traditional command and control allocation scheme to the open spectrum allocation scheme. The report published by the Spectrum-policy Task Force of Federal Communications Commission (FCC) in 2002 [8], which aims at improving the way of utilizing the spectrum resource, catalyzed intensive research activities in this new field of open spectrum sharing. The cognitive radio (CR), which was first coined by Mitola in 1999 [13], is a promising approach to achieve open spectrum sharing flexibly and efficiently [1,9]. It is an intelligent wireless communication system that is aware of its radio environment and is capable of adapting its operation to statistical variations of the incoming radio frequency (RF) stimuli [9]. The research on the CR has already penetrated into a variety of wireless networks and each layer in the network protocol stack [1]. IEEE 802.22 [7] is the first standard based on the CR. IEEE 802.16h [16]

657 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 657–678. c 2007 Springer.

658

Tao Chen et al.

is going to bring CR functions into Worldwide Interoperability for Microwave Access (WiMAX) networks for homogenous and heterogenous network coexistence. A number of cognitive radio testbeds have been developed based on different architectures and radio technologies [1,4,17]. The research on the CR covers a wide range of areas, including spectrum analysis, channel estimation, spectrum sharing, medium access control (MAC), and routing. In this paper, we focus our study on the network formation issue of a decentralized CR based mesh network, which is formed by secondary users of spectrums opportunistically utilizing the detected spectrum holes in an ad-hoc way. We name this kind of cognitive radio based mesh network as CogMesh. Basically, the CogMesh can be regarded as a multi-channel multi-access network, in which the available channels of a node undergoes dynamic changes during the life time of the node. It is different from the traditional mesh or ad hoc networks in the sense that it can opportunistically utilize various spectral holes for smooth peer-topeer communications by virtue of the unique CR functionalities. Correspondingly, the topology management of the CogMesh is affected by two main factors: first, a common control channel may not always available for the whole network; and second, the topology of the network changes over time according to the presence of primary users and secondary users. Therefore, a distributed control plane becomes necessary for the network. However, until now most proposed spectrum control protocols for ad hoc open spectrum sharing networks assume the availability of a common control channel [1]. For instance, an open spectrum sharing protocol proposed in [14] extended the ideas of Request To Send (RTS) - Clear To Send (CTS) exchange and Network Allocation Vector (NAV) in the IEEE 802.11 MAC protocol for the open spectrum access, where the access operation relies on a common control channel. Ma proposed a dynamic open spectrum sharing MAC (DOSS-MAC) protocol for similar networks [12], in which a common channel is used for signaling. Zhao et al observed that although a very limited number of global common channels exist in a network, neighbor nodes may locally share numerous channels with others [19]. A distributed grouping scheme was proposed in [19] to solve the common control channel problem. However, an efficient neighbor discovery process, which is important for an open spectrum access network, is absence. The challenges remaining for an ad-hoc open spectrum sharing network include the neighbor discovery, distributed control, and multi-hop communication. Considering the nature of the CogMesh, we propose a cluster based approach to solve those problems. Lots of cluster formation algorithms have been proposed for ad hoc networks so far [11]- [2]. They are different on the criteria to select clusterheads. However, there are some critical problems to utilize those approaches in the CogMesh: 1. They usually assume a single channel radio on each node; 2. They are designed for fixed network topology, and lack the adequate capability to adapt to dynamic physical topology changes; 3. Most of them only guarantee the network connectivity. The cluster configuration may not be optimized; 4. Some approaches need to know the full topology of the network. Therefore, a unique approach is demanded for a CogMesh network. In this chapter, we introduce the concept of CogMesh, and propose a clusterbased decentralized scheme to solve the network control problem. The users of the

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

659

network are organized into clusters according to their location and available spectrum holes. The network is formed by adaptively interconnecting clusters as illustrated in Figure 34.1. The chapter contains the following contents:

Figure 34.1. An example of CogMesh network.

• • • •

It investigates the issues to setup an ad-hoc open spectrum sharing network which coexists with primary users of the spectrums. It proposes a decentralized cluster-based architecture to form a large scale mesh network. It develops a distributed control scheme at the link layer. The neighbor information is used wisely for network topology management. It provides mechanisms to adapt the network topology to network and radio environment changes.

The remaining of the chapter are organized as follows. The network model and assumptions are provided in Section 34.2. The network architecture are introduced in Section 34.3. The motivation to use the cluster-based structure is explained here. Next, based on the proposed network architecture, the cluster formation, network formation, and network management issues are identified and discussed individually: Section 34.4 introduces the MAC protocol; Section 34.5 describes the spectrum hole detection procedure; Section 34.6 discusses the neighbor discovery and cluster formation procedure; Section 34.7 describes the inter-cluster connection; Section 34.8 addresses network management issues and provides solutions. In Section 34.9 we provide a proof on the network connectivity, which is based on the random graph theory. Then in Section 34.10 the simulation results of the cluster merging algorithm proposed in Section 34.8 are presented. The conclusion is drawn in Section 34.11.

34.2 Network Model In a typical cognitive radio scenario, users of a given frequency band are classified into primary users and secondary users [1]. Primary users are licensed users

660

Tao Chen et al.

of that frequency band. Secondary users, on the other hand, are unlicensed users that opportunistically access the spectrum when no primary users operating on that frequency band. The CogMesh network is formed by secondary users which utilize the spectrum holes for communications. The spectrum holes, as illustrated in Figure 34.2, are white or gray spaces which are free of primary users or partially occupied by low-power interferers [17]. The nodes in the CogMesh are equipped with cognitive radio modules, which are capable of detecting and utilizing spectrum holes efficiently in a distributed way. A given number of spectrum holes are available for the whole network and are identified by their unique channel IDs after effective spectrum sensing and channel state estimation. For each node its available spectrum holes depend on its location. For simplicity, we use the terms channel and spectrum hole interchangeably. We assume the spectrum holes detected by a node change in time but at a relatively slow rate, and the CR nodes move only at a slow speed. The network topology, therefore, is relatively dynamic with stable status. Symmetric links are assumed in the network. A simplified interference avoidance model, i.e.,

Figure 34.2. Spectrum holes. the overlay spectrum sharing model is employed in our network [1], where secondary users only use the spectrums that have not been occupied by primary users. Once detecting the presence of a primary user on a given frequency band, the secondary user simply vacates that band. It is the case in IEEE 802.22 networks [7]. Another interference avoidance model is the interference temperature model [9], which allows the coexistence of primary and secondary users. In summary, the network we study in this paper has following characteristics: • All secondary users are cognitive radio enabled. • Spectrum holes are location dependent and time varying, which means secondary users may own different channel set. • There is no global common channel for the network. However, local common channels exist among adjacent users. • Secondary users form an ad-hoc mesh network, which means distributed control scheme is deployed.

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

661

34.3 Network Architecture Traditional multi-channel wireless systems usually use a global control channel for neighbor discovery and access control [15], [18]. However, it is not the case in the CogMesh. The CogMesh uses local control channels for channel access control, where a distributed control scheme is applied. However, the dynamic changing of spectrum holes makes the channel control extremely complex. A pure distributed control scheme like Carrier Sensing Multiple Access/Collision Avoidance (CSMA/CA) may not work. Accordingly, we introduce the concept of cluster into the network. A node forms a cluster on a channel and invites adjacent nodes sharing the same channel to join its cluster. The analysis in Section 34.6 shows the cluster based approach has the advantage on neighbor discovery as compared to the non-cluster based approach. For convenience, the control channel of a cluster is called the master channel of that cluster. The node forming the cluster becomes the clusterhead, which is responsible for intra-cluster channel access control and inter-cluster communications. The channel access scheme is described in Section 34.4, and the cluster formation process is detailed in Section 34.6. By negotiating gateway nodes between clusters, clusters are interconnected into a large network. A gateway node is a member of one cluster that is able to reach the member of other cluster. The cluster interconnection is illustrated in Figure 34.3, and described in Section 34.7. From Figure 34.3, we can see clusters are interconnected in two cases: two clusterheads are connected by one gateway node, or connected by two gateway nodes when no node is 1-hop neighbor of two clusterheads. There are three types of members in a cluster: the clusterhead, ordinary node, and gateway node. For a network to be properly constructed, protocols and mechanisms are necessary

Clusterhead Cluster B

Gateway node Channel 2 Ordinary node

Channel 1 Clusterhead Gateway node Cluster A

Cluster C

Ordinary node

Figure 34.3. Clusters interconnected by gateway nodes.

to specify the behaviors of nodes under different network conditions. We divide those protocols and mechanisms into five parts and discuss them in the following sections: a) MAC protocol to support cognitive radio based multiple channel access b) Spectrum hole detection c) Neighbor discovery and cluster formation. d) Intercluster connection. e) Topology management.

662

Tao Chen et al.

34.4 MAC Protocol The cluster formation and inter-cluster connection are performed distributively based on nodes’ neighbor information. We provide mechanisms to enable nodes exchange their 1-hop and 2-hop neighbors information, which includes neighbors’s identity and their channel list. In the CogMesh, a node may only know partial of its neighbors at the initial stage. The clusters are formed based on the partial neighbor information. As nodes gradually collect more neighbor information based on the proposed neighbor discovery algorithm, clusters are reconstructed and interconnected to a more reliable network structure. All these are done by the MAC protocol. The MAC protocol proposed here is a hybrid MAC protocol that consists of guaranteed access and random access periods, where the guaranteed access period is used for data transmission in and between clusters, and the random access period is used for control message exchange. For each cluster, channel access time is divided into a sequence of superframes. Each superframe consists of five main periods as shown in Figure 34.4. The beacon period is issued by the clusterhead. It contains the time synchronization, control and resource allocation information of the cluster. The following period is the neighborhood broadcasting period (NBP). It is divided into a number of fixed length minislots. Each member of a cluster occupies one mini-slot and uses it to broadcast its identity and 1-hop neighbor list. An entry in the neighbor list includes the identity of the neighbor and its channel list. The master channel of a neighbor is indicated in the cluster list, through which a node knows how to reach the neighbor cluster. A preamble is put at the beginning of each mini slot for other nodes identifying the broadcasting message if they miss the beacon. Moreover, the time and duration of random access periods in this superframe is broadcast in the Frame Map period of its mini-slot. A neighbor of this node, once receiving its neighborhood broadcasting message, has the chance to exchange its neighbor information with the node in the following random access period. The location of a member’s mini-slot is announced by the clusterhead in the beacon period. The number of mini-slots in a superframe is limited by a system parameter in order to avoid too many nodes crowding in one cluster. Next comes the data period. Parallel transmissions are permitted in this period if the transmission sessions use different channels. Time division multiple access (TDMA) is used in each channel. Following the data period, an intra-cluster random access period (RAP) is used for cluster members exchanging control messages. The superframe is ended by a public RAP. Its length is determined by the clusterhead and announced in the beacon. This period has multiple purposes. It uses for a node joining the cluster, nodes exchanging neighbor information, or clusters exchanging control information. The slotted ALOHA is used to resolve collisions in the RAPs. Besides five main periods, there are one or several spectrum detection periods scheduled in a superframe. During these periods, all members of a cluster keep silence and detect spectrum holes. It is desirable to synchronize the spectrum detection periods of adjacent clusters so as to reduce the false alarm of primary users. The false alarm is an event that a secondary user incorrectly reports the presence of primary users due to the interference from other sources. Since the superframes of different clusters are not required to be synchronized, the location of the spectrum detection periods varies from cluster to cluster. Even in a cluster, their location varies from superframe to superframe. A distributed algorithm is demanded to synchronize the

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network Data Period

Superframe

channel n channel 1

Member 4 Channel list

Neighbor ID

Channel list

Member 3

Control

Private Public Random Acess Random Access

Grant List

Neighbor ID

Frame map

Cluster info

Preamble Member 1

Member 2 Node ID

Superframe Map

Preamble

Spectrum Detection Period

Neighborhood Broadcasting Period

Beacon

663

Slotted

Data Frame

Slotted

Figure 34.4. Superframe structure.

spectrum detection periods of adjacent clusters, and determine the location of those periods in superframes on a superframe by superframe basis. Note that the design of the MAC protocol can benefit from the spread spectrum technique for the coexistence of multiple clusters, in which two kinds of spreading codes are used in the periods of a superframe. One is the public spreading code, which is globally known to every node and used by the beacon, NBP, and public RAP for broadcasting control messages. The other is the private spreading code used for data transmission. A clusterhead chooses the private spreading code in a way so that adjacent clusters use different codes. The code distribution algorithm is required here.

34.5 Spectrum Hole Detection The spectrum detection is performed by each node periodically. The techniques used for the spectrum hole detection include matched filter detection, energy detection, cyclostationary feature detection [1], and so on. In [9], multitaper spectral estimation plus singular value decomposition (MTM-SVD) is suggested for spectrum hole detection. We assume nodes in the network are equipped with one or combination of those detection techniques, and are able to detect all spectrum holes accurately and efficiently. Synchronization in an ad-hoc network is always a challenge due to the absence of centralized coordination in nature. However, there are solutions to partially solve this problem [10]. Cordeiro proposed an algorithm in 802.22 networks to synchronize the superframe of different base stations [7]. The network scenario of 802.22 is similar to the CogMesh in the sense that they both use the spectrum in an opportunistic way. However, 802.22 uses fixed base stations for the channel access control, while in the CogMesh, the clusterhead of a cluster is not fixed. Therefore a more sophisticated mechanisms is demanded in the CogMesh. The initial idea is that we only synchronize the detection periods instead of superframes. A time stamp algorithm can be employed to solve this problem. Each new cluster is stamped by its created time. If it detects a neighbor cluster has a time stamp earlier than its own,

664

Tao Chen et al.

the cluster synchronizes its detection period to that neighbor cluster and replaces its time stamp with the neighbor cluster’s one. At the end, all clusters use the detection period of the earliest formed cluster. For a large scale network, it may take long time to achieve the synchronization.

34.6 Neighbor Discovery and Cluster Formation The neighbor discovery and cluster formation process are introduced together since they are highly coupled. For convenience, we give the following definitions: the host cluster of a node is the cluster that the node belongs to; the neighbor cluster of a node is the cluster that the node does not belong to but has 1-hop neighbors as its members; the total neighbor clusters of all members of a cluster are called the cluster’s neighbor clusters. The neighbor discovery is performed during clusters’ NBPs. When a node wants to join the network, it first detects the available channels. Then it scans one of its channels for a given period of time, waiting for beacons on that channel. The node starts the scanning process from the lowest frequency band channel, which is called the lowest channel. The scanning time on a channel is chosen so that it exceeds the period of the longest superframe. We call a scanning period as scanning interval, and the first scanning interval a new node starts as the first scanning interval. If there is a neighbor cluster on the frequency band a node listens on, it is able to capture its beacon during a scanning interval. We divide the first scanning interval into three cases: no message comes; a beacon comes; or neighbor messages come but no beacon comes. In the first case, the node forms a cluster on the scanning channel and becomes the clusterhead. In the second case, the node requests to join the cluster through the public RAP of the cluster. If the clusterhead accepts the request, it assigns a mini-slot to the requesting node. Starting from next superframe, the new joining node broadcasts its neighbor list in that mini-slot. However, if there is no empty mini-slot in a cluster, the clusterhead will reject the request. The requesting node then chooses the second lowest channel to start a new scanning process, or form its own cluster if finding the detected clusters are all full after iterating all channels. The third case means the node has neighbor clusters but it is 2-hop away from clusterheads. The node then records neighbor information, and tries to exchange neighbor information with that neighbor through the public RAP of the corresponding neighbor cluster. After that, it continues its scanning process on the next available channel. If the node can not find a channel satisfying the case one and two after iterating all channels, it starts its own cluster on a randomly chosen channel. After a node joining a cluster, it periodically chooses from its channel list a nonmaster channel to scan so as to discover other neighbor nodes. An algorithm can be developed to intelligently choose the non-master channel according to the neighbor information the node detects. For instance, if it discovers new 2-hop neighbors on a non-master channel, it listens on that channel first. Let us explain the neighbor discovery and cluster formation by an example illustrated in Figure 34.5. The numbers in the bracket close to each node are available channels of that node. The smaller number represents the lower spectrum hole. We assume the spectrum holes do not change during the cluster formation procedure. The edge between two nodes indicates they can hear each other. Assume the node

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

665

A is the first node forming the cluster on the channel 1. The cluster is labeled as the cluster A. Its 1-hop neighbors B, C, D listen on their lowest frequency band, i.e., the channel 1, detect the beacon issued by the cluster A. They join the cluster A through corresponding association processes. From the neighbor discovery process, the node B knows the node C is its 1-hop neighbor, and the node D is its 2-hop neighbor. Next, the node E, F, G form a cluster on the channel 2. Assume the node E forms the cluster, labeled as the cluster E. The node F, G join the cluster E right after. The node B listens on the non-master channel 2. It discovers E, F as its 1-hop neighbors, and G as its 2-hop neighbor. The cluster A and E therefore are interconnected by the node B. Then, assume the node I forms the cluster I on the channel 3. The node H receives B’s broadcasting message and detects B as its 1-hop neighbor. However, H can not receive beacons from the cluster A. It starts a new listen process on the channel 3 and finally joins the cluster I. The node H informs B that its new neighbor list through the public RAP of the cluster A. The node B knows from H there is a cluster on the channel 3. It knows the neighbors H, I on the channel 3 through a scanning process on that channel. Furthermore, the node C will know B has new neighbors H and I from the NBP of the cluster A and finally know its neighbor I on the channel 3. At this stage, three clusters are formed, and the clusterheads has enough neighbor information for inter-cluster connection. The clusters then negotiate with each other to form a large network through the public RAP of each other.

Figure 34.5. Cluster formation process.

666

Tao Chen et al.

34.6.1 Analysis of Neighbor Discovery Approaches The analysis shows that the proposed cluster based approach significantly reduces the neighbor discovery time as compared to non-cluster based approaches. We use the following non-cluster based approach as the reference model: assuming channel time is divided into superframes and nodes in the network synchronize on the superframe basis. In each superframe a node randomly picks up a channel to stay and is able to discover all 1-hop neighbors in the same channel. Given total N channels and M 1-hop neighbors, the question is how much time as the unit of the superframe a node needs to take in order to discover all its neighbors. Assuming the probability that a node chooses a channel to perform neighbor discovery is p = 1/N , and the node has r remaining undiscovered 1-hop neighbors, the probability that i neighbors stay in the same channel of the node is: r P {N = i} = i pi (1 − p)r−i which is a binomial distribution. The average number of neighbors being discovered is thus rp. Assume in the round k, there is rk neighbors remains undiscovered. In round k + 1, the number rk+1 becomes rk − rk p. Considering in the first round r1 = M , we get: rk+1 = (1 − p)rk = (1 − p)k M Assuming all neighbors are discovered by the node when rk+1 < 1, it is easy to get the average number of rounds the discovery process takes : k > − ln M/ ln(1 − p) So we get k = O(lnM ). For the cluster based approach, we assume the clusters are constructed evenly over the channels. The number of superframes used to form a cluster is two. A node then takes maximum N − 1 superframes to detects neighbors distributed on all non-master channels. The number of rounds for 1-hop neighbor discovery is therefore k = N − 1 + 2 = N + 1. So we get k = O(N ). Note that usually the condition M >> N holds. We can thus conclude that the cluster based approach outperforms the non-cluster based approach in most cases. The Figure 34.6 shows the average number of rounds it will take for each approach under different neighbors M and channels N .

34.7 Inter-Cluster Connection The network is formed by interconnecting clusters through gateway nodes. A gateway node is the intermediate node of two clusters, through which two clusters can exchange control messages and data. They are chosen by clusters through a certain algorithm. There are two cases for the inter-cluster connection: two clusters are overlapping, or non-overlapping. In the first case at least one node belonging to any of the clusters is 1-hop neighbor of two corresponding clusterheads. This node is called the intermediate node. One of the intermediate nodes is chosen by its clusterhead as the gateway node to reach the peer cluster. The gateway node in this case is called 1-hop gateway node. In the second case, if two clusters are non-overlapping

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

667

Figure 34.6. Comparison of two neighbor discovery approaches.

but there are nodes belonging to two clusters can hear each other, they are chosen to interconnect two clusters. We call those nodes 2-hop gateway nodes. In the following, we describe the procedure of the cluster interconnection in detail. The 1-hop gateway interconnection is illustrated in Figure 34.7. The node A and C are clusterheads of two clusters, named the cluster A and C, respectively. The node B1 and B2 are a member of the cluster A and C, respectively. The clusterhead A knows it can reach the cluster C through the node B1. It chooses the node B1 as the gateway node to the cluster C and commands the node B1 to inform the clusterhead C its choice. The node B1 listens on the beacon of the cluster C and sends the gateway choice message to the clusterhead C through the private random access period of the cluster C. When the clusterhead A has control messages send to the clusterhead C, it firstly sends to the node B1. Then the node B1 sends to the clusterhead C through C’s private random access period. In the reverse path, the clusterhead C sends control messages to the node B1 in the cluster A’s private random access period. The node B1 relays the messages to the clusterhead A. The clusterhead C can select the node B2 as its gateway node to the cluster A, and uses the path C to B2 to A when it has control messages send to the clusterhead A. Note that it does not matter whether the cluster A or C operates on the same frequency band since the gateway node works in a store and forward mode. The 2-hop gateway interconnection is more complex since it involves the coordination of two gateway nodes. The 2-hop gateway interconnection is illustrated in Figure 34.7. Similar to the 1-hop gateway scenario, the node A, C are clusterheads of cluster A and C, and the node B1 and B2 belongs to the cluster A and C, respectively. The cluster A and C are not overlapping. But the node B1 and B2 can hear each other. The clusterhead A knows the cluster C can be reached through the node B1. It initializes the gateway setup process by sending to the node B1 a request message. The node B1 relays the request to the node B2 in the cluster C’s public RAP. The node B2 receives the request and relays it to the clusterhead C.

668

Tao Chen et al.

The acknowledgement flows from the clusterhead C to A through the node B2 and B1 following the similar procedure. As nodes discovering more neighbors through Ordinary node

Clusterhead

Clusterhead

Cluster C

Cluster C C

Channel 2 B1

Gateway node

Gateway node C B2

Clusterhead

Cluster A

Ordinary node

B1

Channel 1 A

Channel 1 B2

Gateway node Ordinary node

1-Hop Gateway

A Cluster A

2-Hop Gateways

Figure 34.7. Inter-Cluster connection.

the NBPs, the network connectivity can be gradually improved. Based on link layer connections, the upper layer protocols are able to deliver end-to-end services.

34.8 Topology Management There are several motivations for efficient topology management in the CogMesh. First of all, the random nature of the ICC phase makes the formed clusters hardly being optimized results in line with the physical topology. The number of clusters can be reduced while maintaining the network connectivity. As a result, the control overhead is reduced, and the spectrum efficiency is improved. Secondly, in the cognitive radio scenario, the available channels for each node fluctuate with regard to the radio environment. Consequently, the topologies of clusters are not static, and the reconfigurations are required over time. For this matter, the topologies of clusters must be optimized time by time so as to adapt to the radio environment. In the following section, we examine network maintenance issues at the link layer, and provide corresponding solutions. The issues include: 1. 2. 3. 4. 5.

Nodes join the network; Nodes leave the network; Spectrum holes of a node change; Clusterhead shifts the master channel; Cluster merging.

34.8.1 Nodes Join Network When a new node joins the network, it first detects the spectrum holes. Then it starts the aforementioned process in section 34.6 to join a cluster. The problem here is how the node knows its neighbors and how its neighbors on different master channels know this node. The node and its neighbors in the same cluster know each other through the neighbor discovery process. From the neighbor information it obtains,

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

669

the node can identify potential neighbor clusters and their master channels. It can, thereafter, actively shift to the master channels of other clusters for the neighbor discovery. On the other hand, when its 1-hop neighbors detect a new neighbor, they will broadcast their updated neighbor list through neighbor discovery process. Other nodes in neighbor clusters, when discovering this node as their 2-hop neighbor, may schedule scanning processes to check if it is their 1-hop neighbor. After several rounds of neighbor information exchange, the new node and its neighbors will obtain accurate neighbor information accordingly.

34.8.2 Nodes Leave Network When a node leaves the network, it is important to inform its neighbors the leaving event timely. Note that a node may have different roles in a cluster. The cluster and neighbors of the leaving node should detect and handle the leaving event properly according to the node’s role. There are two kinds of leaving events: a node may leave the network following a disassociation process or disappear suddenly due to node malfunction. We call the former the disassociation process and the latter the absent process, which are described in the following.

Disassociation Process Nodes with different roles have different disassociation processes. For an ordinary node, it informs its clusterhead through a specific message sent in the private RAP of the host cluster. The clusterhead announces the leaving of the node by broadcasting a special message in its mini-slot. The members of the host cluster detect the leaving event through the special message, updating their neighbor list accordingly. Members of the host cluster will inform their neighbor clusters the leaving event. Those neighbor clusters in turn inform their members who are neighbors of the leaving node accordingly. When a gateway node disassociates from the network, it informs the clusterhead its leaving. The clusterhead will start a process to negotiate a new gateway node. The remaining leaving procedure follows the ordinary node disassociation process. The disassociation process of a clusterhead is little more complex. For a cluster, a secondary clusterhead can be optionally chosen by the primary clusterhead during the lifetime of the cluster, which is the member of the cluster with maximum 1-hop neighborhood to other cluster members. When a cluster has a secondary clusterhead, the primary clusterhead hands over the clusterhead role to the secondary clusterhead before it leaves the network. The members of the old cluster who are 1-hop neighbors of the new clusterhead remain in the cluster. Those who are 2-hop neighbors of the new clusterhead leave the cluster and start an aforementioned node joining process. The procedure to inform the neighbors the leaving event of the clusterhead is similar to that in the ordinary node disassociation process. If a cluster has no secondary clusterhead, the clusterhead broadcasts a cluster dismissed message to its members. The members then start node joining processes accordingly. The members may inform the neighbor clusters the disassociation of the clusterhead through public RAP of those clusters. The neighbor clusters dismiss their gateway nodes to this cluster and update the neighbor lists of their members accordingly.

670

Tao Chen et al.

Absent Process In an absent process, a node leaves the network without informing its cluster. So it is important to detect the leaving event timely. For an ordinary node, the clusterhead detects the leaving event through the neighbor discovery process. A node has to broadcast its neighbor list in the assigned mini-slot at least once during a given number of superframe periods. The clusterhead and its members infer the leaving of the node through the absence of corresponding neighborhood broadcasting messages during that period. Once the leaving event is detected, the remaining procedure is similar to that in the disassociation process. If a clusterhead disappears, the members of the cluster will fail to received beacon. They can therefore infer the leaving of the clusterhead. If there is a secondary clusterhead, it takes over the clusterhead role. Otherwise, the cluster is dismissed. In addition to actively detecting the leaving events, each node maintains an expired timer for each entry in its neighbor list. Without receiving a neighborhood broadcasting message from its neighbor in the given expired time, the node deletes the entry of that neighbor from its neighbor list. Different expired time are used for different types of neighbors. The 1-hop neighbors have shorter expired time than 2-hop neighbors. The neighbors belonging to the same cluster have shorter expired time than those belonging to different clusters. The expired timer of 1-hop neighbor entry is renewed if the node receives neighborhood broadcasting message from the corresponding neighbor during the expired time. For a 2-hop neighbors entry, its expired timer is updated if the node finds that neighbor listed in received neighborhood broadcasting messages. Once an 1-hop neighbor entry is expired, the node inform its host and neighbor clusters accordingly.

34.8.3 Spectrum Holes Change Each node senses spectrum holes periodically. If it detects new spectrum holes, it simply informs its cluster and neighbor clusters accordingly. The new spectrum holes will not immediately affect the operation of the host and neighbor clusters. However, if some spectrum holes of a node are not available any more, the host and neighbor clusters may malfunction if their control functions rely on those spectrum holes. The network needs rapidly adapt to spectrum hole occupied events. If a node detects its master channel is not available, it hand over to one of its neighbor cluster, and inform its old host cluster and other neighbor clusters the change. If no neighbor cluster is available, it forms a cluster by itself. If a gateway node detects the channel used to connect the peer cluster become unavailable, it informs its clusterhead to adapt to the change. If a gateway node shifts to another cluster due to the spectrum hole change event, its previous host cluster is able to detect the handover event through the neighbor discovery process in NBP and deal with this event accordingly. In case the master channel of a clusterhead is not available, the cluster will be dismissed.

34.8.4 Cluster Shift Master Channel A clusterhead has the need to change it master channel for several reasons: •

The current master channel is too crowd because several neighbor clusters share the same channel;

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network • • •

671

The channel quality deteriorates due to the increasing interference; There are lots of 1-hops neighbor nodes on other channels that can be merged into this cluster; From the historical statistic information, the current master channel needs to be vacated for primary users in some time periods.

It is the clusterhead’s duty to determine the channel shift. It inform its members the shift operation. After all member nodes confirm the change, the clusterhead starts the beacon in the new master channel from the next superframe. The neighbor clusters of this cluster is informed after the shifting. The gateway nodes to neighbor clusters are reselected. In case a cluster can benefit from the channel shift while few of its members have no spectrum hole on the new master channel, it requests those members to leave the cluster. Algorithms can be customized to meet different performance goals.

34.8.5 Merge Clusters The clusters of the network are formed spontaneously according to the initial network conditions. The resultant topology may not be optimized. Moreover, due to the nature of open spectrum sharing, the network topology may undergo frequent changes during the lifetime of the network. The cluster structures, therefore, need to be adjusted in accordance with the network conditions. The cluster merging process provides a method to adapt cluster structures to the radio environment automatically and distributively. The clusters that share some properties are merged in a way that the overall number of clusters in the network is reduced. Consequently, the communication overhead is reduced. According to different criteria, different merging algorithm can be developed. For instance, statistic learning algorithms can apply here to well adapt to radio environment changes. In the following, we propose a distributed merging algorithm based on the dominating set (DS) theory in the graph theory.

Cluster Merging Algorithm The cluster optimization problem can be considered as a DS problem in graph theory. The DS problem consists of finding a subset of nodes with the following properties: each node is either in the DS, or is adjacent to a node in the DS [3]. In our network, the DS is the collection of clusterheads. The cluster optimization problem is to find a minimal dominating set (MDS) of the CogMesh network according to its physical topology, which is presented by a graph G = (V, E), where V is the set of network nodes, and E is the set of radio links between nodes. The MDS problem is proven to be a NP-hard problem even when the complete network topology is available [2]. However, a sub-optimum DS can be obtained through a local minimum election of the dominators by a heuristic algorithm. The algorithm is run periodically and distributively on each node and only relies on the discovered neighbor information to determine the locally optimized cluster configuration. As a result, the collection of clusterheads is gradually converged to a sub-optimum DS. When the physical topology changes due to the events such as new nodes joining the network, nodes leaving the network, or radio environment changing, the affected

672

Tao Chen et al.

nodes or clusters are reconfigured to immediately absorb the changes. The optimization algorithm is performed thereafter to optimize the changed physical topology. The basic rule for the reconfiguration is when an affected node currently belongs to no cluster, it takes action to associate with one cluster or start a new cluster. In other cases, the clusterheads coordinate the changes. Note that after reconfiguration, gateway nodes of affected clusters may need renegotiation.

Table 34.1. Notation. CHi MCHi k NA C(A) k (c) NA CH(k) NA |.|

clusterhead i all members of CHi k-hop neighbors of node A available channel list of node A k-hop neighbors of node A on channel c k-hop neighbor clusterheads of node A the number of members in a set

The algorithm is shown in Figure 34.8 and works as follows. For convenience, we use the notation as shown in Table 34.1. From the neighbor list, a node, denoted as node A, obtains the node set VA , which includes all members of its 1-hop neighbor clusters and its host cluster. It is the target node set to be optimized. The objective is to construct clusters based on a MDS of the graph GA = (VA , EA ) so that the number of clusters in VA can be minimized. GA is the physical connection graph of node set VA , in which EA is the set of links between node pair of VA . Each channel between two nodes has a link in EA . The MDS is obtained by a heuristic algorithm [6] as shown in Figure 34.9. The algorithm takes the multiple channels of a node into account. First, a cluster is formed by taking node A as the clusterhead and the channel with maximal degree as the master channel. The 1-hop neighbors of node A in VA are assigned to the cluster if they share the master channel with node A. The members of the formed cluster are eliminated from VA . The remaining nodes is processed as following. As in a Max Degree algorithm [3], a node with max degree on a channel is chosen to form a cluster with corresponding neighbors in order until all nodes join the network. Finally, the new cluster configuration comes out with clusterheads list, the master channels and assigned members of each cluster. If the number of resultant clusters is smaller than current one, node A starts a negotiation process to reconfigure its surrounding clusters. To start the negotiation process, node A sends rearrangement requests to the clusterheads it wants to reconfigure, indicating the gain that can be obtained from the rearrangement, and the reconfiguration instruction. The gain is the total number of clusters being reduced if the rearrangement is taken. A clusterhead, once accepts the request, sends an acknowledge to node A. Node A negotiates with the target clusterheads to complete the remaining configuration process only after receiving all acknowledges back. Otherwise, it cancels the process to avoid increasing the cluster number by an incomplete reconfiguration.

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

673

Input: Neighbor information of node A CH(1) curCHSet = NA ; FormSG = (V, E) where S V = CH ∈N CH(1) MCHi A, and {U, V } ∈ E i

A

if U ∈ NV1 ; [CHSet, ClusterSet] = GetClustersFromDS(G, A); if |CHSet| < curCHSet then ReconfigureClusters(ClusterSet); end Figure 34.8. Main function for Cluster Optimization. Function [CHSet,ClusterSet] = GetClustersFromDS(G, myID) T 1 c0 = arg maxc∈C(myID) (|NmyID (c) V |); CHSet ← myID; 0 ClusterSet ← {CH: myID, S Master Channel: c , 1 0 T Members: NmyID (cS) V myID}; 1 S = V \NmyID (c0 ) myID; while S 6= ∅ do foreach ni ∈ S do T N bri = maxc∈C(ni ) (|Nn1i (c) TS|); 1 ci = arg maxc∈C(ni ) (|Nni (c) S|); end i = arg max(i|ni ∈S) (N bri ); CHSet ← ni ; ClusterSet ← {CH: Tni ,SMaster Channel: ci , Members: Nn1i (cTi ) SS ni }; S = S\Nn1i (ci ) S ni ; end return; Figure 34.9. Function for getting clusters from DS.

34.9 Correctness of Network Connectivity We prove the correctness of network connectivity under the assumption that the neighbor discovery is perfect and the network is connected in the physical topology graph. Propostion 1 After the ICC phase, the collection of clusterheads forms a DS. Proof. Assuming the neighbor discovery process is perfect, a node knows all its 1-hop and 2-hop neighbors in finite time. Following the given ICC process, a node

674

Tao Chen et al.

will either join a cluster or form its own cluster in finite time. Therefore the collection of clusterheads forms a DS. Lemma 1 The maximum distance a clusterhead to another closest clusterhead is 3. Proof. Assuming the distance is 4, according to our cluster formation algorithm, the middle node between those two clusterheads must belong to a clusterhead. Therefore this clusterhead is exactly 2-hop away from those two clusterheads, which leads to a contradiction. The proof that two clusterheads can not be 5 or more hops away can be obtained from the result that two clusterheads can not be 4-hop away iteratively. As a conclusion, any clusterhead is at most 3-hop away from its closest clusterheads. Theorem 1 If the clusterheads form a dominating independent set in the current network graph, then a connected backbone is guaranteed to arise if each clusterhead establishes connections to all other clusterheads that are at most at a distance of 3 [5]. Corollary 1 After the ICC phase, the network is connected by clusters at the link layer. Proof. This conclusion can be easily obtained from Proposition 1, Lemma 1 and Theorem 1. Propostion 2 The cluster optimization algorithm does not change the network connectivity. Proof. The cluster optimization algorithm does not exclude any node from new formed clusters or original clusters. Therefore the collection of clusterheads after the algorithm is still a DS of the network graph. According to Lemma 1, and Theorem 1, the algorithm does not change the network connectivity. Propostion 3 If the events such as nodes leaving the network or radio environment changes do not change the physical topology graph, then after reconfiguration, the link layer connectivity does not change. Proof. After reconfiguration, the affected nodes will either join other clusters or form their own clusters. Therefore the clusterheads still form a DS. As in Proposition 2, the connectivity at the link layer does not change.

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

675

34.10 Simulation Results The setup of the simulation is following. As shown in Figure 34.1, multiple nodes of the CogMesh are randomly placed in a 600m × 600m 2-dimension square according to the Poisson distribution. The maximum transmission range of a node is set to 100m. The available channels for a nodes are determined by its location in the square. The square is divided into 16 equal size sub-squares. Secondary users in the same sub-square share identical available channels. The available channels for secondary users in a sub-square is randomly picked from a channel pool (CP). We specify that each sub-square has at least one available channel. The simulation consists of two scenarios in line with two channel conditions. In the stationary channel scenario, the available channels of each node are fixed during the life time of the node. In this scenario, we employ two reference algorithms for performance comparison. The first is the lowest ID algorithm (Lowest ID) [3], in which the node with the lowest ID among its neighbors has the highest priority to form a cluster. The second is the max degree algorithm (Max Degree) [3], in which the node with max degree among its neighbors forms a cluster first. We name our algorithm as local minimal dominating set (LMDS) algorithm. In dynamic channel scenario, the available channels of each sub-square change periodically in a way that during two consecutive changes, the channels for each node are frozen for affected nodes rejoining the network and being optimized by the proposed algorithm. Figure 34.10 shows

Figure 34.10. Cluster statistic in stationary channel scenario.

the number of clusters and average cluster size obtained by different algorithms in the first scenario. The number of channels in the CP is set to two. It is seen from Figure 34.10 that after the ICC phase, the number of clusters is high. However, after the optimization, the number of clusters is reduced significantly and kept below 30 even when the total number of nodes becomes 210. It shows the capability of the proposed algorithm to merge small clusters. Moreover, the proposed algorithm exhibits similar efficiency as Max Degree, and better performance than Lowest ID. The efficiency of three algorithms under different CP is shown in Figure 34.11. When the size of the CP is one, the network reduces to a single channel network. In this case, a node can connect to any other node in its transmission range. That explains in Figure 34.11 why the maximum channel 1 case has smaller cluster size than the

676

Tao Chen et al.

Figure 34.11. Cluster number in stationary channel scenario, with various spectrum holes. other two cases. For size of the CP equal to 2 and 5, similar performance is shown under each algorithm. Figure 34.11 shows that in ICC phase, the size of the CP has small impact on the number of formed clusters. Figure 34.12 describes the number

Figure 34.12. Cluster statistic in dynamic channel scenario.

of clusters and the average cluster size in the dynamic channel scenario. Note the case that the size of CP is one provides a reference to compare the performance of the other two cases. As expected, when the network has multiple channels, the channel changes significantly affects the cluster configuration. As seen from the right

34 CogMesh: A Cluster Based Cognitive Radio Mesh Network

677

plot of Figure 34.12, the number of clusters dramatically increases after the radio environment change. However, after optimization, the number of clusters is reduced to the same level before the radio environment change. It verifies that the proposed algorithm is able to adapt to the radio environment change under different channel conditions and maintain the cluster configuration to a relatively optimal level.

34.11 Conclusion In this chapter, we propose a clustered based cognitive radio based mesh network which utilizes open spectrums opportunistically. The network is constructed and controlled in a distributed way and provides coexistence with primary users of the spectrums. The basic unit of the network is the cluster, which is a sub-network formed by a group of neighbor nodes sharing common channels, and coordinated by a selected node in the cluster called clusterhead. The network is constructed by interconnecting clusters after they learn each other through neighbor discovery process. In the network, the neighbor information plays a crucial role in cluster formation, network construction, and network management. The accuracy of the neighbor information determines the network performance. The chapter provides an MAC protocol nodes efficiently exchanging neighbor information over multiple channels. Moreover, the issues in cluster formation, network formation, and network topology management are addressed and corresponding solutions are provided.

References 1. I.F. Akyildiz, W.Y. Lee, M.C. Vuran, and S. Mohanty. Next Generation/ Dynamic Spectrum Access/Cognitive Radio Wireless Networks: A Survey. Computer Networks, 50(13):2127–2159, 2006. 2. AD Amis, R. Prakash, THP Vuong, and DT Huynh. Max-Min d-Cluster Formation In Wireless Ad Hoc Networks. INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, 1, 2000. 3. L. Bao and JJ Garcia-Luna-Aceves. Topology Management In Ad Hoc Networks. Proceedings of the 4th ACM international symposium on Mobile ad hoc networking & computing, pages 129–140, 2003. 4. R.W. Brodersen, A. Wolisz, D. Cabric, S.M. Mishra, and D. Willkomm. CORVUS: A Cognitive Radio Approach For Usage Of Virtual Unlicensed Spectrum. White paper, Available for download from http://bwrc.eecs.berkeley.edu/ Research/MCMA, 2004. 5. I. Chlamtac and A. Farag´ o. A New Approach To The Design And Analysis Of Peer-To-Peer Mobile Networks. Wireless Networks, 5(3):149–156, 1999. 6. V. Chvatal. A Greedy Heuristic For The Set-Covering Problem. Mathematics of Operations Research, 4(3):233–235, 1979. 7. C. Cordeiro, K. Challapali, and M. Ghosh. Cognitive PHY and MAC layers for Dynamic Spectrum Access and Sharing of TV Bands. Wireless Internet Conference (WICON), 2006. 8. Federal Communications Commission. Spectrum Policy Taks Force. Rep. ET Docket No. 02-135, Nov., 2002.

678

Tao Chen et al.

9. S. Haykin. Cognitive Radio: Brain-Empowered Wireless Communications. Selected Areas in Communications, IEEE Journal on, 23(2):201, 2005. 10. Q. Li and D. Rus. Global Clock Synchronization in Sensor Networks. Computers, IEEE Transactions on, 55(2):214–226, 2006. 11. CR Lin and M. Gerla. Adaptive Clustering For Mobile Wireless Networks. Selected Areas in Communications, IEEE Journal on, 15(7):1265–1275, 1997. 12. L. Ma, X. Han, and C. Shen. Dynamic Open Spectrum Sharing MAC Protocol for Wireless Ad Hoc Networks. Dyspan 2005, Baltimore, Nov, 2005. 13. J. Mitola III and GQ Maguire Jr. Cognitive Radio: Making Software Radios More Personal. Personal Communications, IEEE [see also IEEE Wireless Communications], 6(4):13–18, 1999. 14. S. Sankaranarayanan, P. Papadimitratos, A. Mishra, and S. Hershey. A Bandwith Sharing Approach to Improve Licensed Spectrum Utilzation. Dyspan 2005, Baltimore, Nov, 2005. 15. J. So and N.H. Vaidya. Multi-Channel MAC For Ad Hoc Networks: Handling Multi-Channel Hidden Terminals Using A Single Transceiver. Proceedings of the 5th ACM international symposium on Mobile ad hoc networking and computing, pages 222–233, 2004. 16. J. Sydor. Messaging And Spectrum Sharing Between Ad-Hoc Cognitive Radio Networks. IEEE International Symposium on Circuits and Systems, 2006. 17. TA Weiss and FK Jondral. Spectrum Pooling: An Innovative Strategy For The Enhancement Of Spectrum Efficiency. Communications Magazine, IEEE, 42(3):S8–14, 2004. 18. S.L. Wu, C.Y. Lin, Y.C. Tseng, and J.P. Sheu. A New Multi-Channel MAC Protocol with On-Demand Channel Assignment for Multi-Hop Mobile Ad Hoc Networks. International Symposium on Parallel Architectures, Algorithms, and Networks, I-SPAN, pages 232–237, 2000. 19. J. Zhao, H. Zheng, and G. Yang. Distributed Coordination In Dynamic Spectrum Allocation Networks. Dyspan 2005, Baltimore, Nov, 2005.

35 Coordinating User and Device Behavior in Wireless Grids Lee W. McKnight1 , William Lehr2 , and James Howison1 1

2

School of Information Studies, Syracuse University [lmcknigh|jhowison]@syr.edu Massachusetts Institute of Technology [email protected]

Summary. The evolution of computing is characterized by decentralization and decreasing institutional control over resources. Wireless Grids, that is, fixed and mobile ad-hoc resource sharing networks, are challenging environments in which users strategic behaviors are crucial to system performance. We review the mechanisms employed to regulate strategic behavior online–technical, social, legal and economic–and discuss trends in their operation and application in distributed wireless grid computing.

35.1 Introduction Computing and communication networks have evolved from centralized, hierarchical systems under the management of a single entity toward decentralized, distributed systems under the collective management of many entities. Intelligence has shifted to edge-nodes, which increasingly are capable of acting as autonomous agents making complex decisions to create, deliver, or receive services [28, 46, 51]. Grid computing historically focused on the large-scale sharing of computing resources such as software, hardware, databases and data sources [14, 15]. Wireless grids organized as ad hoc networks of hardware, software, and content resources represent the epitome of this evolution from centralized systems toward ad hoc cognitive and cooperative networks–that is, what we call wireless grids–at the edge [18, 31, 32]. This article discusses the implications of this change for system and service design for distributed network applications including wireless grid applications. We identify some of the academic literatures that are likely to be increasingly relevant for adapting to these new challenges. In Section 35.2 of this article, we provide a stylized overview of the evolution of computing networks to wireless grids, to explain why the need to design for strategic behavior is becoming increasingly critical. We then briefly summarize the critical characteristics of wireless grids, as identified by our preliminary research on this issue [18, 32, 37]. Section 35.3 reviews the four principal mechanisms–technical, social, legal and economic–that are relevant for coordinating behavior in wireless grids and other distributed computing networks. Section 35.4 argues that these mechanisms evolve through the life-cycle of a technology and describes current trends in this evolution. The chapter concludes by sketching our

679 F.H.P. Fitzek and M.D. Katz (eds.), Cognitive Wireless Networks, 679–697. c 2007 Springer.

680

Lee W. McKnight et al.

future work on considering the implications of our analysis for the design of wireless grids.

35.2 From Systems Management to Grid Coordination From a systems management perspective, a change is underway which is akin to the transition in the Soviet Union in the 1990s from a centrally-planned socialist system to a decentralized capitalist economy in Russia. Centralized command and control as modes of coordination are giving way to new mechanisms for allocating resources and moderating behavior [29]. Distributed ownership and decentralized control are raising new challenges for assuring system security and reliability. New network management mechanisms need to draw increasingly from the social, political/legal, and economic models of coordination used elsewhere in society. As with any significant change, there are both risks and new opportunities that must be better understood. Traditional communication networks were designed on the basis of centralized, hierarchical control. In the 1960s, users connected to mainframe computers using dumb terminals. In such an environment, controlling and coordinating the behavior of edge-nodes was relatively simple and security protection could be handled largely by admission control. In the early days, computing resources were firmly under the control of a select cadre of IT professionals. With the emergence of distributed processing and smart terminals in the 1970s, the problem of allocating resources and controlling the behavior of edge nodes became more complicated. However, most computing networks were still under the control of centralized network management supported by the power of management over employees. In the 1980s, with the emergence of personal computing, Local Area Networks (LANs) and Wide Area Networks (WANs), computing and communications became increasingly integrated and distributed. A greater share of network intelligence was located in a continuously growing set of edge nodes. The heterogeneity of behavior that needed to be managed became even greater. Additionally, IT resources were increasingly under the direct control of end-users with much more diverse IT expertise. Corporate data managers now had to contend with non-IT specialists moving PCs among offices and loading or modifying application software in ways that were hard to monitor and manage. The resource allocation and coordination problem continued to grow more complex. In the 1990s, the commercial emergence of the Internet expanded data communications and computing to a mass market, and increasingly provided a platform for interconnecting networks around the globe. The Internet’s end-to-end architecture which facilitated peer communications among nodes stood in marked contrast to the traditional telecommunications networks which were based on hierarchical, centralized network management [28, 46, 51]. In the Internet, control is distributed to edge-nodes. However, the potential chaos that such a transition risked was moderated because key resources (e.g., DNS and routing infrastructure) were largely under the control of corporate data managers and carriers descended from the traditional telecommunications networks. This technical architecture was mirrored by changes in industry structure and the policy environment. Traditional telecommunication and computing networks tended

35 Coordinating User and Device Behavior in Wireless Grids

681

to be owned and managed end-to-end by a single entity (e.g., a carrier network or a corporate enterprise network). When these networks interconnected, these occurred at well-defined locations under bilateral (or multilateral) peering points. In the case of telecommunication carriers, the operation of these networks was also subject to substantial government regulation which constrained both the pricing and technical terms under which services were offered and interconnected. Network management and ownership in the Internet, by contrast, is distributed among a global collection of heterogeneous end nodes, some of which are single computers or devices, while others are large networks in their own right. The diversity in ownership and computing/communication technology reflected in these edge-nodes raises the coordination problem to a new level of complexity. The Internet is also much more open than traditional network environments. The open, distributed nature of the Internet has facilitated the proliferation of computing in business and society, and contributed to dramatic growth in the ICT sectors and the global economy as a whole, but it has also raised problems for system designers. Computer and network designers can no longer assume that systems will be owned and managed by a single entity with a single, coordinated set of goals. Increasingly, nodes are capable of self-interested behavior that can impact overall system performance in unpredictable and potentially adverse ways. The diversity of ownership in networking resources gives rise to diversity in strategic interests. Coordinating behavior among nodes in a distributed network where all participants share common strategic interests is a difficult but well-defined problem for decision science. However, in an Internet-style environment, network management requires coordination among agents that are likely to have divergent capabilities and strategic interests. Resource allocation and control becomes a ’microeconomic’ coordination problem. That is, whereas decision science provides a toolset for determining the optimal solution to single agent (common objective) problems, microeconomics provides a language/framework for studying the interactions of selfinterested, strategically-independent agents. Its tools include the study of market behavior and game theory. Of course, many other academic disciplines also offer insights that are helpful in understanding how to design for strategic behavior, including computer science (parallel processing, ad hoc networks, and artificial intelligence), sociology and psychology, political science (including understanding interest group behavior and motivation), legal theory, and biology (especially evolutionary systems). Moreover, computing/communication networks are becoming ever more important parts of our social (entertainment, cyber communities), economic (eCommerce), and political lives (eGovernment). In this environment, network design and management cannot be separated from the legal, political, social, and economic institutions governing human interactions in other spheres. Unsurprisingly, as computer networks become more central to our lives, the modes in which we regulate our lives in other spheres will become more relevant for how computer networks operate. The openness of Internet-type networks allows businesses, their suppliers, and consumers to communicate and interact freely. The distributed and flexible architecture allows resources to be combined and used in novel ways, encouraging innovation and enhancing capabilities [9, 34]. However, this also increases the problem of protecting systems from myriad challenges ranging from viruses, denial of service attacks, intellectual property infringement (including protecting copyright in an era of resource sharing systems), and the abuse of privacy. Fraud and theft are also more

682

Lee W. McKnight et al.

common as criminals follow the money onto the information grid. Assuring system reliability and managing quality of service for diverse applications (delay-sensitive voice and file transfers on a shared network) is also more complicated when the identity, capabilities, and goals/incentives of end-nodes are not pre-configured and controllable. The emergence of Peer-to-Peer (P2P) networks, such as Napster, Gnutella, Freenet and BitTorrent to name a few as well as computational networks such as Seti@Home and distributed.net reflects a reassertion of the end-to-end architectural model of the Internet and illustrates the importance of user behavior to system performance. In P2P networks the resources making up the network, storage space, routing and computational cycles are voluntarily provided by individual end-users with little or no institutional connections or trust. Shneidman and Parkes argue that, “perhaps the key defining characteristic of a peer to peer network is that one cannot distinguish between strategic nodes and the network infrastructure” [54]. Yet this risks overstatement as P2P networks are properly called overlay networks to emphasize that they run over the existing institutionally owned and managed infrastructure. This overlay nature gives leverage to attempts to centrally control peer to peer activities which we describe below as significant for the emergence of wireless grids. The growth of wireless accelerates these trends because it increases opportunities for computing to become ubiquitous (always available, always connected), expands the heterogeneity of networking resources that need to be managed (mobility management and wireless/wireline interconnection), and the shear number of sensor network end nodes that need to be managed (connected computers in everything from our bodies to clothes, appliances, cars, and walls). Wireless grids represent the epitome of this transition. In a wireless grid, even more than in overlay P2P networks, the edge nodes are the network. Designers and network managers of an ad hoc wireless grid will need to anticipate the strategic behavior of the end-nodes that will comprise the network. The challenge will be two fold: first, end-nodes will have to be induced to contribute resources to the network; and, second, to behave while part of the network in a way that helps maximize the total net benefits realized by the network. For example, in a wireless grid network, edge devices will likely need to be induced to contribute computing/communication resources to process traffic from other edge nodes, while at the same time refraining from behavior that deteriorates the service offered to other users (e.g., excessive use of shared resources) [6, 55]. Wireless grids are emerging from the coalescence of a number of independent research efforts and industry trends (see Figure 35.1). There are important developments associated in each of these areas that are critical to the evolution of wireless grids, but a concrete overall view is yet to emerge. Wireless grids will not be a computing network separate from the social/economic framework in which they operate. Continued multidisciplinary research is needed to properly design wireless grid networks1 . We are engaged in a collaborative project to design infrastructure for wireless grids and to understand the virtual markets whose emergence we anticipate. This 1

Several conference papers were our first efforts to explicate and define the critical features of wireless grids. These may be found on the website www.wirelessgrids.net. Note especially [3, 8, 18, 20, 32, 44, 57].

35 Coordinating User and Device Behavior in Wireless Grids

683

Figure 35.1. Research Issues and Industry Trends leading to wireless grids. work will require the development of appropriate coordination mechanisms and will need to draw on and be compatible with coordination structures/institutions used elsewhere in networks and in society. In the next section we present a taxonomy of coordination frameworks, discuss their current realization on the Internet and their relevance for organizing behavior in distributed networks including wireless grids.

35.3 Coordinating Strategic Behavior in Distributed Networks There are four prototypical ways in which to coordinate and allocate resources in distributed networks: (1) Technical; (2) Social (3) Legal and (4) Economic. Each of these is discussed further below, along with examples of their use in network system design and operation.

35.3.1 Technical The traditional and still most common approach to network management is to use technical means to regulate behavior. Appropriate behavior can be ’hard-wired’ into the network through hardware and software design. In biological systems, genetic coding may hard-wire in behavior and evolution can encourage and re-enforce behaviors that enhance a species prospects for survival. In computer and communication networks, standards and communication protocols limit the range of allowed designs and behaviors that may be encountered,

684

Lee W. McKnight et al.

thereby rendering system performance more predictable and controllable2 . Of course, designing suitable protocols that do not unduly constrain flexibility is quite difficult, especially in a distributed environment. The key is to define open interfaces that provide sufficient assurance as to the functionality that will be supported to allow interoperability without dictating detailed implementation rules that might limit innovation. In many cases, the determination of which behavior is consistent with optimal network performance will depend on local and system-wide conditions that may be changing dynamically. Allowing nodes autonomy to moderate behavior in response to local needs and conditions can enhance overall performance, but this local autonomy than creates the potential for strategic manipulation [26]. Over time, network design has moved to a layered architecture with well-defined interfaces supporting communication across layers. The trend towards technical standardization based on open interfaces has resulted in a number of important developments for the industry. For example, open interfaces can allow end-users to ’mix and match’ components (e.g., like when consumers mix-and-match stereo components or software applications on personal computers) to create customized systems. The open interfaces can also enhance industry competition by supporting both system-level and component-level competition. Industry standardization can also give rise to positive network externalities that expand demand and scale/scope economies that can lower industry costs. Because the choice of where to define interfaces and what technologies to accommodate has such important implications for industry economics, industry standardization is inherently strategic. Getting the industry to agree on what standard to adopt is often quite difficult. The process can be contentious, expensive, and slow. Indeed, the process may be slow precisely because the standard development organizations have adopted bureaucratic rules in order to protect standardization from strategic manipulation. Even after a standard has been defined, assuring compliance can be quite difficult. The standard which allows a lot of implementation flexibility may not assure adequate interoperability. When the networks are owned and managed by a relatively few number of players “as was more often the case in traditional telephone networks” enforcing interoperability was relatively easier. Adoption and implementation of the standard can be managed centrally. On the Internet the adoption and implementation of standards is focused on the IETF/IESG Request For Comment process3 which, as discussed below, is implemented through a voluntary process supported by an informal reputation system. In a wireless grid network, the proliferation of edge-nodes under autonomous control makes technical coordination much more difficult. One approach that has been used to manage interoperability in the distributed control environment that characterizes unlicensed spectrum is to require equipment certification. This ex ante testing is used to certify that equipment will comply with the communication protocols that have been adopted. In the case of unlicensed spectrum use, the principle concern is that a transmitter will comply with limits on radiated power. Most other details concerning how the transmitter will behave

2 3

Lessig makes a related argument in [27]. See RFC 2555 for a summary of the development of the RFCs and their process.

35 Coordinating User and Device Behavior in Wireless Grids

685

are left unspecified and it is left up to users of the spectrum to adopt suitable communication protocols and strategies for contending with congestion. The certification approach facilitates distributed and asynchronous deployment of network equipment, but it limits flexibility and becomes less tractable as radio transmitters and receivers become more adaptive and software-controlled. There are a number of reasons for this. First, certifying the behavior of software is inherently more difficult than for hardware. Second, power modulation represents an important option for managing spectrum use efficiently, and a priori power limits are overly restrictive. Third, the certification approach may tilt the industry playing field in favor of incumbents (e.g., established equipment makers), potentially harming innovative approaches. In the Internet, the TCP/IP protocols manage congestion via statistical back-off: when nodes experience congestion, they slow down their transmissions randomly. This works quite well when networks are lightly loaded and its simplicity makes it easy (low cost) to implement in a distributed network. Nodes only need local information to self-regulate their behavior. The downside of this approach is that it does not support quality-differentiated services which are important once the network starts having to contend with traffic that has heterogeneous requirements (e.g., delay-sensitive telephony and delay-insensitive email) and intrinsic values (e.g., network control messages and music downloads). While technical approaches to coordinating behavior based on standards and communication protocols or network etiquette will remain important, they are unlikely to be sufficient by themselves. For example, it is possible to tweak TCP/IP parameters to capture an excessive share of network resources. This was not a significant problem in the early days of the Internet when it was a government-subsidized network used mostly by academics. With the Internet’s growing social and commercial relevance, the control of quality of service has moved beyond purely technical approaches.

35.3.2 Social The second common mechanism employed to regulate strategic behavior in networks operates through the social networks in which actors are embedded. Professional and cultural ties provide leverage by which network managers, and participants, can punish undesirable behavior and reward behaviors supportive of the goals of the system. Social mechanisms often support and provide the leverage to enforce the behaviors encoded in the technical protocols and standards discussed above. The social mechanisms of greatest interest are those that operate in two ways: through cognitive factors such as conscience (or morality) and social influences, especially reputation in the context of group membership [11]. While reputation concerns the opinions of others about an actor, conscience concerns the opinions of an actor about their own actions. Both mechanisms act to regulate behavior however from a network application designers perspective reputation is the most useful mechanism. The basic proposition of behavior regulation through reputation is that because people care about their reputation they will not act in ways that damage it and

686

Lee W. McKnight et al.

will act in ways that enhance it4 . Yet reputation, as a strongly socially contextual concept, has varying mechanisms, impacts on behavior and scaling properties5 . One of the best known and most studied reputation system is the “feedback” mechanism employed by Ebay which allows buyers and sellers to exchange public information regarding their satisfaction with the transaction [48, 49]. This system aims to regulate the potentially selfish behaviors such as fraud or a bait and switch tactics. This system has been credited with the growth and rapid acceptance of the Ebay system and the ability for Ebay to avoid the need to provide costly dispute resolution systems or guarantees. The system also seems to provide desirable outcomes for sellers: Resnick et al. in [50] concluded that sellers with high reputation earned approximately seven percent more than low reputation sellers. Reputation is also employed as a tool for combating email Spam through blacklists of mailservers known (or believed to) send spam. The best known of these is the Realtime Blackhole List (RBL)6 . While far from perfect7 , these systems have helped reduce spam from operators of open mail relays. Open mail relays are tempting to self-interested systems administrators because they offer convenience in configuration and for their intended users who do not have to deal with authentication or changing outgoing mailservers when moving between networks and IP addresses. However open relays provide conduits for the senders of spam into the Internet mail infrastructure, an activity which causes significant inconvenience to end users and consumes significant amounts of network bandwidth. Listing a mail server in a blacklist is a statement that the server has a bad reputation and means that servers which subscribe to that list will not accept mail from the legitimate users of the server and will ’bounce’ the messages with a statement that the users mail server is suspected of spamming. The operator of an open relay is therefore encouraged to adopt more system-friendly behavior through a combination of technical (blocking) and reputation (reports made to users of the server and other systems administrators that are embarrassing to admin of an open relay). Reputation has been employed also to coordinate behavior in P2P files sharing systems. Here the system designers goal is to increase the quantity and quality of content available on the network. Accordingly, Gnutella and Kazaa both provide mechanisms to prioritize the downloads of clients that have established a good reputation for providing uploads. While these mechanisms are currently quite basic they are developing rapidly, for example, BitTorrent, which provides swarming downloads by re-using clients currently downloading from a server as parallel servers for other clients employs a version of the Tit for Tat strategy developed in formal analyses of the Prisoners Dilemma game [10]. 4

5

6

7

In [39], Moreton and Twigg discuss the similarities between reputation systems and markets in which actors are motivated by money. Economic mechanisms are considered in Section 35.3.4 below. A useful taxonomy of types of reputation and their characteristics is provided by [40]. In early 2003 an NSF funded workshop was organized to support and develop this field. Resnick and Dellarocas’ summary of the workshop provides an excellent introduction [47]. http://www.declude.com/junkmail/support/ip4r.htm lists over 90 known blacklist services. At the time of writing spam fighting blacklists where under sustained denial of service attacks believed to be launched by spammers.

35 Coordinating User and Device Behavior in Wireless Grids

687

Scholars are seeking to formalize reputation systems to support the development of P2P and distributed computing applications. For example, [1] and [24] propose to utilize a distributed data structure to store complaints about nodes in a P2P network and in an electronic market. These reputation mechanisms share aspects of community-public goods. Because providing reputation reports is costly in time and resources, participants may have a natural preference for free-ride on the information provided by others. Thus a major challenge in building systems to formalize and extend social regulation through reputation is addressing the issues of incentive compatibility that arise. This challenge is addressed in [23] and [38]. Reputation has been utilized in promoting desired end-node behavior in distributed computing projects which are pre-cursors of Grid applications. The Seti@Home project leverages both reputation and conscience by compiling and making available statistics on the number of units that users have processed. This information is made prominent on the local client, leveraging conscience, and through league tables, periodically released on the Seti@Home site and lists which leverage reputation. Furthermore when interesting results are discovered the user or team who undertook the processing is highlighted despite the random distribution of work units. It is not clear what use this type of reputation is to the actors but the emergence of highly competitive teams (containing thousands of members) aiming to process the largest number of work units suggests that it is an effective motivator of desirable behaviors8 . However, this motivator is far from unproblematic–cheating through altered software has been discovered within the Seti@Home system [54]. Social mechanisms rely on the strength of social ties or group identification to regulate behavior in networks. This mechanism is clearly limited by the growth and expansion of actors interconnecting through networks, which, by the sheer increase in numbers, reduces the effectiveness of both informal reputation systems and morality derived from group membership. In addition the rising financial rewards available through network misbehavior, such as Spamming, motivate actors to compare these rewards to the often less quantifiable reputation rewards. Nevertheless social regulation remains an important mechanism particularly in situations characterized by high levels of repeat interactions.

35.3.3 Legal Legal and political systems are designed to regulate and enforce a wide variety of behavioral prescriptions and prohibitions in the interests of promoting the well-being of the broader community. Roman law (unitary law) and common law (Anglo-Saxon) legal traditions share many elements of commonality, while differing in their approaches to legal change and adaptation [27]. For the United States, the Communications Act of 1934 (as most recently amended in the U.S. Telecommunications Act of 1996) defines the legal framework for media and telecommunications systems and services. The Act includes detailed specification and regulatory guidelines for interconnection of networks. Behaviors affecting the use of radio spectrum have been addressed primarily through legal means, including provisions for licensed and unlicensed (Part 15) 8

A sample league table can be seen here http://www.muskratgroup.com/kwsn/ teams.html

688

Lee W. McKnight et al.

frequency use. While the recognition and treatment of property rights is fundamental to capitalist economies, including the unlimited right to exclude, this centralized legal framework for spectrum management has been undermined by the development of new spectrum sharing technologies. In other work we have critiqued the lack of foresight exhibited by the legislators who enacted that law [41]. Subsequent events, including the emergence of wireless grids, prove our point. Ubiquitous wireless grid environments will pose challenges to many areas of law and law enforcement as diverse heterogeneous market, policy, and user requirements must be simultaneously resolved in a shared resource environment. The Digital Millennium Copyright Act has been employed to limit the behavior of network users. Section 1201 of this act was the basis for US vs ElcomSoft and Sklyarov in which a Russian programmer, Dimitry Sklyarov, was arrested on a visit to the US for providing a circumvention device able to remove the encryption from Adobe PDF files. ElcomSoft, his employer, also faced charges. The DCMA also forms the legal basis for actions designed to obtain evidence of copyright infringement from ISPs. In 2003 the RIAA obtained subpoenas against, amongst others, Verizons ISP, who was required to release the details of subscribers accused of sharing copyrighted music on P2P services. These subpoenas made possible the contributory copyright infringement suits made against over 260 individuals in 2003. The RIAAs stated strategy is to utilize the threat of such lawsuits to reduce the, from their perspective, undesirable behavior of users providing resources to P2P music file sharing networks9 . Contract law has also been employed to regulate behavior online. The Terms and Conditions required of ISP customers usually contain acceptable use provisions which restrict activities considered to be undesirable by the network designers, such as running servers on home access accounts. These contracts also facilitate ISPs cooperating with law enforcement officials or legal subpoenas for evidence.

35.3.4 Economic The market’s Invisible Hand provides another potent mechanism for coordinating behavior in distributed systems. Competitive markets, when they are operable, provide an efficient mechanism for allocating resources that do not presume any common interest among resource producers or consumers. Buyers and sellers, each seeking to maximize their individual welfare, will compete for scarce resources. Excess demand for resources drives market prices up, inducing consumers who value the resource the less than the current price to leave the market and inducing suppliers who can produce at lower cost to increase supply. Excess supply has the opposite effect. In the idealized competitive market, the atomistic buyers and sellers each act independently, ignoring their impact on the market price, yet collectively their distributed behavior drives the market to equilibrium. In the efficient equilibrium, supply and demand are balanced, resources are produced at the lowest possible cost, and allocated to the highest-value demand. Unfortunately, the ideal of perfect competition is seldom realized in the real world; and even the ideal economic model is somewhat sketchy with respect to the dynamics of how a market approaches equilibrium [25]. Indeed, real world markets depend critically on the social, legal, and technical environment that shapes the way 9

See http://www.riaa.com/news/newsletter/090803.asp

35 Coordinating User and Device Behavior in Wireless Grids

689

in which actors exchange information, negotiate for the exchange of goods, complete their transaction, and in the event of disputes, reconcile any difficulties. The development of functional economic systems for computer networks has been studied for over thirty years. One of the key developments in this field is the recognition that the systems cannot simply provide efficient allocations of resources (as per [56, 59]) but must manage incentives and strategies of the participants (as per [12, 13, 53, 58]). Unfortunately this second step has proven to be difficult in both theory and implementation. This is clear from the market controlled approaches to resource allocation in Grids, summarized in [19], which fail to adequately address strategic issues. Buyya et al. in [7] demonstrate both the usefulness of an economic approach to resource allocation within Grid computing environments and the difficulties faced. They developed and implemented a market-making scheme involving the interaction of consumer and producer agents undertaking a wide range of economic interaction models, including auctions and announced prices. This market-making scheme was able to demonstrate efficiency in the allocation of resources on the Grid. However the model suffers from two key difficulties that will serve to illustrate the complexity of difficulties of implementing computational market system: bootstrapping and incentive management. The system suffers from a bootstrapping problem: Grid services provide the underlying services for a market designed to motivate the provision of Grid Services, The Grid computing environments provide necessary infrastructure including security, information, transparent access to remote resources, and information services that enable us to bring these two entities together [7, p. 2]. Without these basic requirements markets do not function effectively. Real world markets are embedded in social relations, not the least of which is the system of contract law and the enforcement mechanisms that support it. Buyya et al.s approach, grounded as it is within traditional resource allocation literature, does not adequately address the strategic challenges of networked computing. Buyya et al. acknowledge this when they present their function for Resource Value, Resource Value = Function (Resource strength, Cost of physical resources, Service overhead, Demand, Value perceived by the user, Preferences) And state, The last three are difficult to capture from consumers unless they see any benefit in disclosing their actual demand, preference, and/or resource value, which varies from one application to another. [7, p. 4] If these parameters have to be truthfully disclosed to reach the desired resource allocation efficiencies then the system must be designed in such a way that it is to the agents benefit to reveal such private information. Otherwise the system is open to systematic under- or over-statement of private valuations and will not achieve the desired (and expected) efficiencies. This is the heart of Shneidman and Parkes recent criticism of the literature on economic analyses of Grid computing where they argue that recent papers on economic models for resource scheduling in scientific Grid computing have not explored issues of rationality [54, p. 6] (referencing [7]). An important challenge for designers of wireless grid technology will be to design for virtual markets. The mechanisms for determining who participates in these markets, how information is exchanged, how participants negotiate for the exchange of

690

Lee W. McKnight et al.

resources, payment/compensation mechanisms, and monitoring/enforcement structures will all be critical elements that must be developed. These mechanisms must be incentive compatible. That is participants have to trust these mechanisms to behave as expected and in such a way that induces them to participate and elicits cooperative behavior that is also self-interested and selfish. For prices to emerge and markets to function appropriately, it must be possible to define common resources using a collective and public language that can allow resources to be “commoditized”. Participants have to know what they are negotiating for when they decide to purchase or supply a unit of computing or communication power. Figuring out what are the right ways to describe and quantify commodities and the terms and time limits for purchase/supply contracts will represent a difficult challenge. Eventually, we will need service level agreements for wireless grids [33]. There are a diverse range of market mechanisms in use. These range from free exchange (e.g., subsidized) to barter systems (exchange of goods without money) to the arm’s length exchange (exchange for money with limited prior contact or on-going contractual relationship) to bilateral or multilateral exchange. All of these have been used in various contexts within modern communication networks. For example, WiFi free nets and the enterprise networks provided to corporate employees or university students are often subsidized. Although they obviously cost the provider, the consumer does not directly pay for access to the resources. Network peering may be considered a form of barter exchange in which interconnecting carriers agree to exchange traffic at no charge. In the Internet, the lack of a more developed economic system has plagued multi-lateral ’free’ peering with consistent congestion problems, leading most backbone carriers to move to bilateral peering. Telephone service markets offer numerous well-developed versions of more advanced economic market systems. Traditionally, these were regulated as common carriers, which protected atomistic consumers from being discriminated against. Atomistic residential and small business consumers purchase service without term commitments according to regulated tariffs. The more competitive markets such as long distance services and cellular services are less heavily regulated. Consumers churn among alternative providers in response to more attractive price/quality offerings. The competition for consumers forces carriers to lower costs and enhance quality. Advertising and marketing help suppliers and consumers learn about available options. Wireless grids are likely to make use of all of these market models as they develop. In anticipation, it would be useful to consider how to design for flexible market models that do not presume a particular industry structure or mode of exchange. A key element will be to design for market interfaces. These are most likely to occur via open interfaces that can be standardized so that the requisite information may be exchanged among parties that may be exchanging resources at arm’s length. If the parties have an on-going relationship and shared common interests, than the market exchange interface may be quite simple (e.g., exchange within a single firm). Alternatively, if the relevant commodity can be provided in a market situation that approaches the competitive ideal, than again the market interface may be quite simple–the Invisible Hand of the market can supply coordination. More typically, the transaction will involve agents with potentially conflicting, self-interested goals and the designer will need to consider the game-theoretic aspects of exchange (e.g., asymmetric and incomplete information, reward/penalty structures, sequencing of actions, player strategy spaces, etc.).

35 Coordinating User and Device Behavior in Wireless Grids

691

Understanding the need for incentive-compatible optimal design is much easier than explicating how this might be accomplished without complexity that will hinder the adoption of the mechanisms10 . No single economic approach will be ideal for all circumstances. The appropriate economic design will depend critically on the other elements of the environment: the technical, social, and legal context in which participants will interact.

35.4 Interactions and Dynamics in Regulation The process of change in the computing environment described in Section 35.3 gives rise to a particular dynamic among the four models described above. Early in a technologies life cycle, technical and social coordination mechanisms are most useful and were clearly emphasized in the development of the Internet. They allow for the greatest level of innovation and utilize the familiarity and shared intentions of the development community as a trusted base to support this innovation. However, a natural byproduct of technology becoming more mainstream is that the range of parties that are interacting become less familiar to each other–there are less repeat interactions, less common expertise/knowledge/experience to induce conformation–so self-interested and potentially harmful behavior increases. The Internet is reaching this second phase. Simultaneously, the stakes of non-cooperation have risen sharply as businesses rely on Internet services to invest and risk real money. This section briefly describes two currently developing responses to this situation. The first is a move to ’harden’ technical regulation and to substitute law for the social regulation that had supported technical regulation. The second is an expansion in legal provisions relating to behavior online and the development of surveillance systems to support their operation.

35.4.1 Hardening Technical Regulation with Legal Enforcement There are a number of current proposals that would strengthen technical regulation through both hardware and software initiatives and through legal means to mandate their use. This reflects a loss of confidence with current voluntary technical regulation. “Trusted Computing” has been proposed as a solution to computer insecurity and viruses and the use of computers for copyright infringement–all areas of ’misbehavior’ online. The Trusted Computing Group, an industry body lead by Intel and Microsoft, propose designing systems which are only able to run code which has been verified through a digital signature. The system would be incapable of running non-signed or altered code and network applications would be able to ascertain that their peers where running particular versions whose behavior could be relied upon [2, 52]. Trusted Computing would thereby create a technical mechanism would could be used to protect against the execution of virus code as well as to prevent the 10

There is significant work underway in the area of Distributed Algorithmic Mechanism Design. See [16, 42, 45, 60]. However [43] reminds us of risks in decision marking complexity in online markets.

692

Lee W. McKnight et al.

execution of infringing digital media behavior. This proposal sidesteps issues of providing incentives for desired behavior by recreating the institutional fence whose breakdown we examined in Section 35.2, by providing the technical hook for external control over the uses of computing devices. In [52] Schecter et al. consider the ability of Trusted Computing to control end node behavior. They introduce this through an ironic demonstration that this capability could be used by P2P music sharing network designers to protect themselves from the attempts of content owners to disrupt the networks. Injecting corrupted content and flooding networks are tactics which have been adopted by the content industries and are, from the network designers point of view, undesirable and detrimental to system performance. Trusted Computing platforms would allow network clients to ascertain that a peer is running application code without these detrimental behaviors and to exclude misbehaving clients from the network. In [52] example clients are able to exclude clients designed to reduce network throughput by flooding bandwidth with extraneous traffic. It is clear, then, that Trusted Computing would merely provide a technical hook for end node control but that market and legal provisions will determine how that hook is used. A similar development can be observed in the TV broadcast industry where the digitization of content is viewed as creating opportunities for violations of copyright that would threaten the viability of the conversion to DTV. The Broadcast Protection Discussion Group (BPDG), an industry body charged with preventing this self-interested behavior has been proposed that there be a ’broadcast flag’ attached to ’protected’ content which would indicate that that content may not be used in certain ways, and that compliant devices be designed to respect this flag. This proposal has been incorporated into the ATSC standards as an optional part11 , however in August 2002 the FCC issued a Notice of Proposed Rulemaking12 in which it stated that that it was inviting discussion on the question, “Should the FCC mandate that consumer electronics devices recognize and give effect to the broadcast flag?”. Legislative proposals such as The Consumer Broadband and Digital Television Promotion Act (CBDPTA), proposed in 2002 by Sen. Ernest Hollings, would mandate the use of copy protection scheme in any device that can “retrieve or access copyrighted works in digital form” and it has been suggested that this implies the legislation of Trusted Computing. These bills have so far not received broad support but reflect the trend of providing legal backing to the use of hardened technical standards for the regulation of online behavior. These are an attempt to return to the systems management paradigm discussed in Section 35.2.

35.4.2 New Legal Provisions and Their Surveillance Implications The second response to declining trust, rising misbehavior and increasing stakes are efforts to utilize the civil and criminal justices systems to enforce desired behavior online. It is the nature of justice proceedings that they occur after infringing behavior in question and that admissible evidence of infringing behavior must be brought before a court. For this reason legislative proposals typically imply an increase in surveillance of online behavior. 11 12

ATSC Standard A65/A. FCC Digital Broadcast Copy Protection MB Docket No. 02-230.

35 Coordinating User and Device Behavior in Wireless Grids

693

The recent high profile investigation, arrest and pending prosecution of the juvenile writers of the SoBig virus reflect the increasing use of law enforcement to combat online misbehavior that threatens the stability and performance of computer networks. These actions are quite distinct from the enforcement of laws relating to pornography or fraud because the social evil targeted is a decline in system performance which had hitherto been considered purely a technical matter. Many US States and European countries have, or are in the process of passing laws against unsolicited email, known as Spam which contain steep financial penalties. The DCMAs copyright provisions, as discussed above, are increasingly being used to target online behavior. Common to all these laws is the need to collect admissible evidence of infringing behavior online. The Internet community has struggled with calls for lawful interception of internet traffic for the purpose of evidence collection [5]. In May 2000, after an internal debate, the IETF issued RFC 2804, IETF Policy on Wiretapping in which it writes, “The Internet Engineering Task Force (IETF) has been asked to take a position on the inclusion into IETF standards-track documents of functionality designed to facilitate wiretapping. This memo explains what the IETF thinks the question means, why its answer is “no”, and what that answer means.” [22, p. 1]. Yet despite this policy the issue has not subsided. Cisco Systems has made available, as an optional router software feature that must be specifically requested, the capacity to give access to data flowing through routers in a form specifically designed to be legally admissible [30]. Their initiative to publish this capability as an Internet Draft [4] indicates that this debate is far from closed and that the pressure to collect evidence to support Legal regulations of online behavior remains strong.

35.5 Conclusion and Implications for Wireless Grids Despite the continuously increasing interests in wireless grid13 , research on wireless grids is scattered. This chapter has examined the evolution of computing from systems management within known institutional contexts to the decentralized and end-user centric model, i.e., from a user perspective [36]. We highlighted the increasing importance of self-interested strategic behavior and the need for network application designers to be able to promote desired behaviors while discouraging undesired behaviors. We argued that wireless grids are the epitome of these developments. We identified and examined four mechanisms for the regulation of strategic behavior–technical, social, legal and economic–and examined dynamics within them. It is clear that designers of wireless grids and their applications will need to draw on all of these mechanisms. We expect a similar dynamic in the development of wireless grids to that which has occurred through the technology life cycle of existing network technologies. It is, therefore, sensible that the initial focus be on traditional technical and informal reputation systems which are better placed to promote innovation and experimentation. It is similarly sensible to design such that the evolution towards economically managed situations is easier and so that this evolution, and future innovation, is not foreclosed by premature hardened technical 13

For instance, IEEE Internet Computing launched a special issue on wireless grids in 2004 guest edited by Lee McKnight, where researchers studied wireless grids via various perspectives [17, 21, 35].

694

Lee W. McKnight et al.

and legal proscription. The best manner in which to accomplish this is a topic for our further research.

Acknowledgement. This article is based upon research supported in part by the National Science Foundation, Partners for Innovation program grant #0227879, involving faculty and students at MIT, Boston, Northeastern, Syracuse, and Tufts Universities, the Swiss Federal Institute of Technology (ETH Zurich), University of Colorado, Boulder, BTexact, Cisco Systems, Hitachi, Nokia, Novell, and other firms, in partnership with TeleCom City, TRITEC, the Museum of Science Boston, and the School of Information Studies and CASE Center at Syracuse University. See www.wirelessgrids.net for more information. The opinions expressed in this article are those of the authors. Any errors of fact or by omission are the responsibility of the authors and not the institutions with which they may be affiliated.

References 1. K. Aberer and Z. Despotovic. Managing trust in a peer-2-peer information system. In the Tenth International Conference on Information and Knowledge Management, 2001. 2. R. Anderson. Trusted computing FAQ. http://www.cl.cam.ac.uk/users/ rja14/tcpa-faq.html, 2003. 3. D. Anius. Enhancing innovation within a regional wireless grid. In The International Conference on Computer, Communication and Control Technologies, Orlando, Florida, 2003. 4. F. Baker, B. Foster, and C. Sharp. Cisco architecture for lawful intercept in IP networks, Internet Draft (Informational). Technical report, IETF, 2003. 5. P. A. Branch. Lawful interception of the internet. Technical report, Centre for Advanced Internet Architectures, 2003. 6. L. Butty´ an and J.-P. Hubaux. Stimulating cooperation in self-organizing mobile ad hoc networks. Mobile Networks and Applications, 83(5):589–592, 2003. 7. R. Buyya, H. Stockinger, J. Giddy, and D. Abramson. Economic models for management of resources in peer-to-peer and grid computing. In the SPIE International Conference on Commercial Applications for High-Performance Computing, Denver, USA, 2001. 8. B. Chen and C. H. Chang. Mobility impact on energy conservation of ad hoc routing protocols. In International Conference on Advances in Infrastructure for Electronic Business, Education, Science, Medicine, and Mobile Technologies on the Internet, L’Aquila, Italy, 2003. 9. D. Clark. The design philosophy of the darpa internet protocols. In SIGCOMM ’88: Symposium proceedings on Communications architectures and protocols, pages 106–114, New York, NY, USA, 1988. ACM Press. 10. B. Cohen. Incentives build robustness in bittorrent. In Workshop on Economics of Peer-to-Peer Systems, 2003. 11. R. Conte and C. Castelfranchi. Cognitive and Social Action. UCL Press, 1995. 12. J. Crowcroft, R. Gibbens, F. Kelly, and S. Ostring. Modelling incentives for collaboration in mobile ad hoc networks. In WiOpt ’03, 2003.

35 Coordinating User and Device Behavior in Wireless Grids

695

13. E. Drexler and M. Miller. Ecology of Computation, chapter Incentive engineering for computational resource management. Elsevier Science Publishers B. V., 1988. 14. Foster, C. Kesselman, J. Nick, and S. Tuecke. The physiology of the grid: An open grid services architecture for distributed systems integration. citeseer. nj.nec.com/foster02physiology.html, 2002. 15. Ian Foster, Carl Kesselman, and Steven Tuecke. The anatomy of the grid: Enabling scalable virtual organizations. Int. J. High Perform. Comput. Appl., 15(3):200–222, 2001. 16. E. Friedman and D. Parkes. Pricing wifi at starbucks - issues in online mechanism design. In Fourth ACM Conf. on Elec. Commerce, 2003. 17. M. Gaynor, S.L. Moulton, M. Welsh, E. LaCombe, A. Rowan, and J Wynne. Integrating wireless sensor networks with the grid. IEEE Internet Computing, 8(4):32–39, 2004. 18. M. Gaynor, L. W.McKnight, J. Hwang, and J. Freedman. Wireless grid networks and virtual markets. In the International Conference on Computer, Communication and Control Technologies, 2003. 19. J. Gomoluch and M. Schroeder. Market-based resource allocation for grid computing: A model and simulation. In 1st International Workshop on Middleware for Grid Computing, Rio de Janeiro, Brazil, 2003. 20. J. Hwang and P. Aravamudham. Proxy-based middleware services for peer-topeer computing in virtually clustered wireless grid networks. In the International Conference on Computer, Communication and Control Technologies, Orlando, Florida, 2003. 21. Junseok Hwang and P. Aravamudham. Middleware services for P2P computing in wireless grid networks. IEEE Internet Computing, 8(4):40–46, 2004. 22. IESG IAB. IETF policy on wiretapping. Technical Report RFC 2804, IETF, 2000. 23. R. Jurca and B. Faltings. An incentive compatible reputation mechanism. In the IEEE Conference on E-Commerce, Newport Beach, CA, USA, 2003. 24. S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina. The eigentrust algorithm for reputation management in p2p networks. In the Twelfth International World Wide Web Conference, 2003. 25. T. F. Kompas. Traditional notions of equilibrium reconsidered. Equilibrium, 11:484–495, 2000. 26. W. Lehr. Standardization: Understanding the process. Journal of the American Society for Information Science, 43(8):550–555, 1992. 27. L. Lessig. Code and Other Laws of Cyberspace. Basic Books, 1999. 28. J. MacKie-Mason, S. Shenker, and H. Varian. Network architecture and content provision: An economic analysis. Telecommunications Policy, 20(3):203–217, 1996. 29. T. W. Malone, J. Yates, and R. I. Benjamin. Electronic markets and electronic hierarchies. Communications of the ACM, 30(6):484–497, 1987. 30. D. McCullagh. Inside Cisco’s eavesdropping apparatus. http://news.com.com/ 2010-1071-997528.html, 2003. 31. L. W. McKnight, D. Anius, and O. Uzuner. Virtual markets in wireless grids: Peering policy obstacles. In Proceedings of the International Conference on Computer, Communication and Control Technologies, 2003.

696

Lee W. McKnight et al.

32. L. W. McKnight and J. Howison. Towards a sharing protocol for wireless grids. In Proceedings of the International Conference on Computer, Communication and Control Technologies, 2003. 33. L. W. McKnight and W. Lehr. Show me the money: Agents and contracts in service level agreement markets. INFO, 40(2):24–36, 2002. 34. L. W. McKnight, P. M. Vaaler, and R. L. Katz. Creative destruction: business survival strategies in the global Internet economy. MIT Press, 2001. 35. Lee McKnight. Wireless Grids–Distributed Resource Sharing by Mobile, Nomadic, and Fixed Devices. IEEE Internet Computing, 8(4):2–10, 2004. 36. L.W. McKnight, R. Sharif, and L. van de Wijngaert. Wireless Grids:Assessing a New Technology from a User Perspective, pages 169–182. Springer, 2005. 37. W. McKnight, D. Anius, and O. Uzuner. Internet Services. Quality of Service in Grids, Networks and Markets. In press. 38. N. Miller, P. Resnick, and R. Zeckhauser. Eliciting honest feedback in electronic markets. In Working Paper for the SITE02 workshop, 2003. 39. T. Moreton and A. Twigg. Trading in trust, tokens and stamps. In 1st Workshop on the Economics of P2P systems, Cambridge University, 2003. 40. L. Mui, M. Mohtashemi, and A. Halberstadt. Notions of reputation in multiagents systems: a review. In the first international joint conference on Autonomous agents and multiagent systems, 2002. 41. W. R. Neuman, L. W. McKnight, and R. J. Solomon. The Gordian Knot: Political Gridlock on the Information Highway. MIT Press, 1997. 42. C. Ng, D. Parkes, and M. Seltzer. Strategyproof computing: Systems infrastructures for self-interested parties. In 1st Workshop on the Economics of P2P systems, 2003. 43. Odlyzko. Internet pricing and the history of communications. Computer Networks, 36(5-6):493–517, 2001. 44. J. S. Park, A. Nanda, and J. Howison. Security challenges and countermeasures in wlans. In the International Conference on Computer, Communication and Control Technologies, Orlando, Florida, 2003. 45. D. C. Parkes. Iterative combinatorial auctions: Achieving economic and computational efficiency. Technical report, University of Pennsylvania, 2001. 46. D. Reed, J. Saltzer, and D. Clark. Active networking and end-to-end arguments. IEEE Network, 12(3):67–71, 1998. 47. P. Resnick and C. Dellarocas. Online Reputation Mechanisms: A Roadmap for Future Research. Technical report, First Interdisciplinary Symposium on Online Reputation Mechanisms, 2006. 48. P. Resnick and R. Zeckhauser. Trust among strangers in internet transactions: Empirical analysis of ebay’s reputation system. Advances in Applied Microeconomics, Elsevier Science, 11, 2003. 49. P. Resnick, R. Zeckhauser, E. Friedman, and K. Kuwabara. Reputation systems: Facilitating trust in internet interactions. Communications of the ACM, 43(12):45–48, 2000. 50. P. Resnick, R. Zeckhauser, J. Swanson, and K. Lockwood. The value of reputation on ebay: A controlled experiment. Technical report, School of Information, University of Michigan, 2002. 51. J. H. Saltzer, D. P. Reed, and D. D. Clark. End-to-end arguments in system design. ACM Transactions on Computer Systems, 2(4):277–288, 1984.

35 Coordinating User and Device Behavior in Wireless Grids

697

52. S. E. Schecter, R. A. Greenstadt, and M. D. Smith. Trusted computing, peerto-peer distribution, and the economics of pirated entertainment. In The Second Workshop on Economics and Information Security, 2003. 53. S. Shenker, D. Clark, D. Estrin, and S. Herzog. Pricing in computer networks: Reshaping the research agenda. ACM Computer Communication Review, 26(6):19–43, 1996. 54. J. Shneidman and D. Parkes. Rationality and self-interest in peer to peer networks. In Proceedings of the 2nd International Workshop on Peer-to-Peer Systems, 2003. 55. B. Strulo, J. Farr, and A. Smith. Securing mobile ad hoc networks – a motivational approach. BT Technology Journal, 21(3):81–89, 2003. 56. Sutherland. A futures market in computer time. Communications of the ACM, 11(6):449–451, 1968. 57. O. Uzuner and R. Davis. Digital fingerprinting for distribution volume tracking:intellectual property protection in wireless grids. In the International Conference on Computer, Communication and Control Technologies, Orlando, Florida, 2003. 58. H. R. Varian. Economic mechanism design for computerized agents. In Workshop on Electronic Commerce, 2000. 59. C. A. Waldspurger, T. Hogg, B. A. Huberman, J. O. Kephart, and W. S. Stornetta. Spawn: A distributed computational economy. Software Engineering, 18(2):103–117, 1992. 60. C. J. Woodard and D. C. Parkes. Strategyproof mechanisms for ad hoc network formation. In 1st Workshop on the Economics of P2P systems, 2003.

Acknowledgements

The present book is the result of the coordinated efforts of many people around the globe, and it would have not been possible without the key contributions of our invited authors. We thank you all for sharing with us your technical expertise, and for the professionalism and endless enthusiasm showed during the writing period. We are deeply thankful for the foreword of Professor Bernhard Walke introducing the book and sharing his view on future wireless communication systems. This idea of editing this book came not only as the result of our own research but also due to numerous discussions and panels that we had with colleagues working in this field. The results of those enlightening discussions are reflected in our work. A special thank to Professor Frank Reichert for the discussion about pricing strategies in cooperative networks and for sharing his view on the book structure. Thanks to Kasper Rodil for another exciting design for the cover. Parts of the book were partially financed by the Danish government on behalf of the FTP activities within the X3MP project and the Center for TeleInFraktur (CTIF) within the EDWIN project. Our thanks go to our co-workers in the EDWIN group such as Morten Holm Larsen, Morten V. Pedersen, Søren Vang Andersen, Andreas Kopesel and our colleagues at Aalborg University for the inspiring discussions about the project results. We would like also to thank Harri Pennanen, Nina Tammelin, and Per Møller from Nokia for the great help in supporting us in any mobile phones related activities. We are pleased by this collaboration and we are more than thankful that three years ago Harri gave us the possibility to join the academic activities of Nokia. VTT, the Technical Research Centre of Finland, also supported financially and logistically part of this book. We are grateful to Technology Director Dr. Jussi Paakkari, Technology Manager Ky¨ osti Rautiola and Research Professor Dr. Aarne M¨ ammel¨ a for their unconditional support during this initiative. We also thank our research colleagues at VTT (Communications Platform Group, and in particular the Cooperative and Cognitive Networks Team) for their technical contributions,

699

700

Acknowledgements

motivating discussions and for creating a truly pleasing working atmosphere. We are particularly thankful to Mark de Jongh and Cindy Zitter, from Springer for their encouragement, patience and flexibility during the whole edition process. We wish sincerely to thank Dr. Tim Brown and Robert Sheahan for their invaluable help in proof reading several chapters of the book. We are grateful to Thomas Arildsen for his wonderful support on the book editing process. Finally, but most importantly, we Editors would like to, once again, thank our wives, Sterica and Paula, and our children, Lilith and Samuel, for their unfailing support and for being so understanding while we worked on this new book.

List of Abbreviations and Symbols 16-QAM 16 Quadrature Amplitude Modulation µC MicroController µP MicroProcessor (I)FFT (Inverse) fast Fourier transform 1G First Generation 256-QAM 256 Quadrature Amplitude Modulation 2D Two-Dimensional 2G Second Generation 3D Three-Dimensional 3GPP 3G Partnership Project 3G Third generation 4G Fourth Generation 5G Fifth Generation 64-QAM 64 Quadrature Amplitude Modulation ACPS Ad-Coop Positioning System A-GPS Assisted-GPS AARF Adaptive Automatic Rate Fallback ACK Acknowledgment ACPS Ad-Coop Positioning System AC Alternating Current aDA advanced Dynamic Algorithm ADC Analog-to-Digital Converter AEP Asymptotic Equipartition Property AFDP Adaptive File Distribution Protocol AF Amplify-and-Forward AHN Ad Hoc Network AIC Additional Information Container AMAP Autonomic Managment of Access Points AMC Adaptive Modulation and Coding AMR Adaptive Multi-Rate AOA Angle of Arrival AODV Ad Hoc On-demand Distance Vector Routing AOMDV Ad Hoc On-demand Multipath Distance Vector Routing API Application Programming Interface APP Application layer AP Access Point ARF Automatic Rate Fallback ARQ Automatic Repeat Request AR ADDR Access Request Address ASIC Application Specific Integrated Circuits ASM Advanced Spectrum Management ASSP Application specific standard product AWGN Additive White Gaussian Noise A&F Amplify-and-Forward AaF Amplify-and-Forward B3G Beyond third generation (mobile communication systems) BAN Body Area Network

701

702

List of Abbreviations and Symbols

BC Broadcast Channel BD ADDR Bluetooth Device Address BER Bit Error Rate BLER Block Error Ratio BOM Bill Of Materials BPF Bandpass Filter BPSK Binary Phase Shift Keying BS Base Station BTC Block based Turbo-Codes C2C Car-to-Car Communications CBR Constant Bit Rate CCK Complementary Code Keying CCP2P Cellular Controlled Peer-to-Peer CCS Cognitive Communication Systems CC Coded Cooperation CC Convolutional Codes CDMA Code Division Multiple Access CEP Circular Error Probability CFAR Constant False Alarm Rate CF Compress-and-Forward CH Cluster Head CIC Cascaded Integrator-Comb CIEP Cognitive Information Exchange Protocol CIF Common Intermediate Format CINR Carrier-to-Interference-and-Noise Ratio CITADE Correlation based Iterative Tap Amplitude and Delayc Estimation CL Cellular Link CM Cluster Member CNR Channel Gain-to-Noise Ratio CNode Conscious Node CO Cognitive Operation CPE Common Phase Error CPLD Complex Programmable Logic Device CPU Central Processing Unit CP Channel Pool CP Content Provider CP Cyclic Prefix CRB Complete Recovery Bit CRC Cyclic Redundancy Check CRTP Compressed Real Time Protocol CRTS Cooperative Request-to-Send CR Cognitive Radio CSD Circuit Switched Data CSI Channel State Information CSMA/CA Carrier Sensing Multiple Access/Collision Avoidance CSMA Carrier Sense Multiple Access CSR Cooperative Spatial Reuse CTH Clear-to-Help CTS Clear-to-Send

List of Abbreviations and Symbols C&F Compress-and-Forward C REMS Cognitive Reconfigurable Equipment Management System CaF Compress-and-Forward CoPS Cognitive Protocol Stack CogBus CogNet Bus CogNet Cognitive Complete Knowledge Network CogPlane Cognitive Plane DAB Digital Audio Broadcasting DAC Digital-to-Analog Converter DCCH Dedicated control channel DCF Distributed Coordination Function DCT Discrete Cosine Transform DC Direct Current DDC Delay Diversity Code DECT Digital Enhanced Cordless Telecommunications DF Decode and Forward DIFS DCF Interframe Space DIV Distortion In Interval DLC Data Link Control layer DNPM Dynamic Network Planning and Management DR Digital Radio DS-CDMA Direct Sequence Code Division Multiple Access DSK DSP Starter Kit DSP Digital Signal Processor DSR Dynamic Source Routing DSSS Direct Sequence Spread Spectrum DS Dominating Set DTCH Dedicated Transport Channel DTC Distributed Turbo Coding DTV Digital Television DVB-H Digital Video Broadcasting on Handheld DVB Digital Video Broadcast DVD Digital Video Disc D&F Decode-and-Forward DaF Decode-and-Forward E-911 Enhanced-911 EDGE Enhanced Data Rates for GSM Evolution EDR Enhanced Data Rate EGPRS Enhanced General Packet Radio Service EIFS Extended Interframe Space ESS Extended Service Set EVM Error Vector Magnitude FALCON Flexible Access Logic for COmmunication Networks FCC Federal Communications Commission FCH Frame Control Header FCS Frame Check Sequence FDD Frequency Division Duplex FDMA Frequency Division Multiple Access FEC Forward Error Correction

703

704

List of Abbreviations and Symbols

FFT Fast Fourier Transformation FIFO First-In-First-Out FOV Field Of View FPGA Field Programmable Gate Array GMSK Gaussian Minimum Shift Keying GPRS General Packet Radio Service GPS Global Positioning System GSM Global System for Mobile Communications GSR GNU Software Radio GUI Graphic User Interface GoP Group of Pictures HAP High Altitude Platforms HARQ Hybrid ARQ HAWK Highly Adaptable Wireless Kit HIPERLAN HIgh PErformance Radio LAN HSCSD High Speed Circuit Switched Data HSDPA High Speed Downlink Packet Access HSPA High Speed Packet Access HSUPA High Speed Uplink Packet Access HTML Hypertext Mark-up Language HTS Helper-Ready-to-Send HTTP Hypertext Transfer Protocol I/Q Inphase / quadrature IC-DMS Interference Channel with Degraded Message Sets ICC Initial Cluster Construction ICI Inter-carrier Interference IEC International Electrotechnical Commission IEEE Institute of Electrical and Electronics Engineers IF Intermediate Frequency iid Independent, identically distributed IM/DD Intensity Modulated Direct Detection IMS IP Multimedia System IMT-2000 International Mobile Telecommunications 2000 IPC Inter-Process Communication IPD Iterated Prisoner’s Dilemma IP Intellectual Property IP Internet Protocol IR Infra-Red ISI Inter-Symbol Interference ISO International Standardization Organisation ITU International Telecommunications Union IrDA Infrared Data Association JPCR Joint Power Control and Routing algorithm JPO Joint Program Office JRRM Joint Radio Resource Management JTAG Joint Test Action Group JTRS Joint Tactical Radio System kbps Kilobit Per Second LAN Local Area Network

List of Abbreviations and Symbols

705

LDPC Low-Density Parity-Check LED Light Emitting Diode LLC Logical Link Control layer LLR Log-Likelihood Ratio LMDS Local Minimal Dominating Set LOS Line-Of-Sight LO Linking Operation LPI Lost Packet Information Vector LPM Lost Packet Matrix LS Least Square M2M Machine-to-Machine MAC Medium Access Control MAGIC Mobile Multimedia; Anytime, Anywhere, Anyone; Global Mobility Support; Integrated Wireless Solution; Customized Personal Service MAN Metropolitan Area Network MAP Maximum A-Posteriori MBMS Multimedia Broadcast and Multicast Service MBWA Mobile Broadband Wireless Access MCM Multi Carrier Modulation MDC Multiple Description Coding MDS Minimal Dominating Set MD Mobile Device MIMO Multiple Input Multiple Output MISO Multiple Input Single Output MMSE Minimum Mean Square Error MOS Mean Opinion Score MPEG Moving Pictures Expert Group MPR Multi-Packet Reception MRC Maximum Ratio Combining MRSS Multi-Resolution Spectrum Sensing MSB Most Significant Bit MSS Maximum Segment Size MS Mobile Station MTM-SVD Multitaper Spectral Estimation Plus Singular Value Decomposition MTP Multicast Transport Protocols MTU Maximum Transmission Unit MUD Multiuser Detector MUSTANG MUlti-STAndard single chip transceiver for the Next Generation Mbps Megabit Per Second NACK Negative Acknowledge NAV Network Allocation Vector NBP Neighborhood Broadcasting Period NC Network Coding NFC Near-Field Communications NLLS Non Linear Least Square NLOS Non-Line-of-Sight NO Network Operator NeMo Network Mobility OFDMA Orthogonal Frequency Division Multiple Access

706

List of Abbreviations and Symbols

OFDM Orthogonal Frequency Division Multiplexing OSIE Open System Interconnection Environment OSI Open Systems Interconnection OSR Oversampling Rate OS Operating System OVSF Orthogonal Variable Spreading Factor OW Optical Wireless P-SCH Primary Synchronization Channel P2P Peer-to-Peer P3 People-to-People-to-geographical-Places PAN Personal Area Networks PCB Printed Circuit Board PCCC Parallel Concatenated Convolutional Code PCF Point Coordination Function PDA Personal Digital Assistant PDU Protocol Data Unit PD Prisoner’s Dilemma PES Parallel Elementary Streams PGM Pragmatic General Multicast Protocol PHS Personal Phone System PHY Physical Layer PIFS PCF Interframe Space PLC Power Line Communications PLL Phase locked loop PLR Packet Loss Rate PMR Professional Mobile Radio PM ADDR Parked Member Address PN Petri Net POR Power On Reset POT Cooperative scheme with potential game implementation PRN Pseudo Random Noise PSNR Peak Signal-to-Noise Ratio PU Primary User QAM Quadrature Amplitude Modulation QPSK Quadrature Phase Shift Keying QRP Query Routing Protocol QoS Quality of Service RAP Random Access Period RAT Radio Access Technology RBAR Receiver-Based AutoRate RCPC Rate-Compatible Punctured Convolutional RERR route error RFC Request for Comments RFID Radio Frequency Identification RF Radio Frequency RMSE Root Mean Square Error RMS Root Mean Square RND Random channel assignment scheme ROHC Robust Header Compression

List of Abbreviations and Symbols RO&P Regular operation & prepare RO Regular operation RRC Root Raised Cosine RREP route reply RREQ route request RSSI Received Signal Strength Indicator RSS Received Signal Strength RS Relay Station RTDX Real time data exchange RTH Request-to-Help RTP Real Time Protocol RTS Request-to-Send RTT Round Trip Time S-SCH Secondary synchronization channel SCA Software Communications Architecture SCF Spectral Correlation Function SCM Single Carrier Modulation SCR Software Controlled Radio SDD Space Division Duplex SDR Software Defined Radio SDU Service Data Unit SEP Symbol Error Probability SES Sequential Elementary Streams SGSN Serving GPRS Support Node SIFS Short Interframe Space SINR Signal to Interference-plus-Noise Ratio SIR Signal to Interference Ratio SMS Short Message Service SNR Signal-to-Noise Ratio SOC System On Chip SPA Sum Product Algorithm SPI Serial Peripheral Interface SPTF Spectrum Policy Task Force SP Service Provider SRL Short Range Link SRM Scalable Reliable Multicast Protocol SR Software Radio SS Subscriber Station STC Space-Time Codes STTC Space-Time Trellis Code SU Secondary User SVC Scalable Video Coding S&F Store-and-Forward SaF Store-and-Forward TADE Tap Amplitude and Delay Estimation TCP Transmission Control Protocol TC Turbo-Codes TDD Time Division Duplex TDMA Time Division Multiple Access

707

708

List of Abbreviations and Symbols

TDM Time Division Multiplex TDOA Time Difference of Arrival TFT Tit-for-Tat TOA Time of Arrival TRX Transceiver TTA Telecommunications Technology Association TTL Time-to-Live ToS Type of Service UART Universal Asynchronous Receiver Transmitter UDP User Datagram Protocol UEP Unequal Error Protection UE User equipment UI User Interface UMTS Universal Mobile Telecommunications System USB Universal Serial Bus USRP Universal Software Radio Peripheral UWB Ultra Wide Band VAA Virtual Antenna Array VBR Variable Bit Rate VCG Vickrey-Clarke-Groves VCO Voltage Controlled Oscillator VLC Visible Light Communications VQM Video Queue Management VoIP Voice over IP WAN Wide area network WAP Wireless Application Protocol WBAN Wireless Body Area Network WCDMA Wide-band Code Division Multiple Access WDM Wavelength Division Multiplexed WIMAX Worldwide Interoperability for Microwave Access WLAN Wireless Local Area Network WMAN Wireless Metropolitan Area Network WMN Wireless Mesh Network WNLLS Weighted Non Linear Least Square WPAN Wireless Personal Area Network WSN Wireless Sensor Network WSSUS Wide Sense Stationary Uncorrelated Scattering WShRN Wireless Short Range Network WTs Wireless Terminals WUSB Wireless universal serial bus WWAN Wireless Wide Area Network WWRF Wireless World Research Forum WWW World-Wide Web Wi-Fi Wireless Fidelity WiBro Wireless broadband WiMAX Worldwide Interoperability for Microwave Access XTAL Crystal oscillator ZF Zero-forcing

Index

µC, microcontroller 309 µP, microprocessor 309 16QAM, 16-ary quadrature amplitude modulation 317 8PSK, octernary phase shift keying 317 A-MPDU 518 A-MSDU 518 acceptance probability 120 active networks 426 active awareness 354 Ad hoc networks 179 ADC, analog-to-digital converter 309 Agentset 587 always best connected 114 ambient network 112 Amplify-and-Forward 193 Amplify-and-Forward (AaF) 159 Application Programmable Interface, API 402 ARGOS 307, 317 ARQ 487 ARQ, automatic repeat request 320 ASSP, application specific standard product 309 Auction Design 559 AWGN, additive white Gaussian noise 315 backoff 517 bandwidth derivatives barter systems 94 barter trade 130

129

base station 309 Bayesian process 242 Bayesian reasoning 402 BB, baseband 313 beacon approach 355 belief vector 412 BER 409 best response correspondence beta distribution 242 Bluetooth 308, 473, 504 BOM, bill of material 309 bottleneck 292 Button 590

125

C-Cube 423 CDMA, code division multiple access 317 centralized scheduling 285, 286 Channel inter-node 158 Channel access delay 517 CIEP 279 classification based on response time 358 classification based on topology 359 CO, cognitive operation 311 Co-evolution 63 coalitional game theory 535 Coded Cooperation 160 performance 163 CogBus 277 CogMesh 658 CogNet 273 CogNet AP 271

709

710

Index

Cognitive Transport Module 280 Cognitive Bus 277 Cognitive Engine, CE 399 Cognitive Information Exchange Protocol 279 cognitive networks 426, 537 Cognitive Plane 277 cognitive radio 310, 354, 397, 533, 537 Cognitive Radio Spectrum Sharing 571 Cognitive Resource Manager, CRM 397, 399 Cognitive Wireless Network, CWN 398, 411 cognitive wireless networks 306 cognitive wireless solution 307, 308 CogPlane 277 CogTCP 280 Coherent systems 630 Common Application Requirements Interface, CAPRI 405 Commons Tragedy 66 competitive wireless access 119 composition agreement 113 Compress-and-Forward 182 Compress-and-Forward (CaF) 159 congestion game 536, 546, 547 connection manager 403 connection profile 240 control plane interworking 112 convex optimization 293 Cooperation destination-initiated 171 diversity 156 implementation 157, 173 level 160 partner 159 partner-initiated 171 sender-initiated 170 cooperation 87 Cooperation between RF and optical wireless systems 627 cooperation range 477 Cooperation scheme adaptive/non-adaptive 160 performance 162 regenerative/non-regenerative 160 Cooperative link 156

Medium Access Control (MAC) 157 networks 156 protocol 159 resource allocation 157 cooperative networks 307 cooperative spectrum sensing 366 Cooperative transceivers 628 cooperative wireless access 115 coopetition 128 Coopetive wireless access 129 CoPS 423 CPLD, complex programmable logic device 313 cross-layer 299 cross-layer optimization 399 crossover 408 CSMA/CA 514 DAC, digital-to-analog converter 314 database approach 356 decentralized access selection 114 Decode-and-Forward 160, 189 decoding 309 DECT, Digital Enhanced Cordless Telecommunications 317 Delay Diversity Code 193 design 306, 309 detection 309 device-centric 294 Diffuse optical wireless systems 624 digitization 309 directed graph 288 disagreement point 125 Discrete Memoryless Relay Channel 182 distributed algorithm 297 Distributed Turbo Coding (DTC) 160 dominant mode 287 DSP, digital signal processor 309 dual-stage spectrum sensing 367 duplexing constraint 300 DVB-H 473 MPED-FEC 474 parallel elementary streams (PESs) 474 sequential elementary streams (SESs) 474 time slicing 473

Index Dynamic Spectrum Access, DSA 410, 414 dynamic spectrum management 397 energy detection 361 ENOB, effective number of bits 323 entry barriers 111 etiquette 534, 550 events 200 EVM, error vector magnitude 323 EVM, evaulation module 313 Evolutionary algorithms 63 evolutionary algorithms, EA 401 expected expected utility 413 expected utility 412 Expected Value of Perfect Information, EVPI 414 fairness 100 FALCON, Flexible Access Logic for COmmunication Networks 307, 317 FDMA, frequency division multiple access 317 feature detection 362 FEC 488 FFT, fast Fourier transform 315 FIFO, first-in first-out 319 filtering 309 firmware 312 FM, frequency modulation 317 frame concatenation 518 frame packing 518 FreezeTimer 406 frequency scanning 309 future proofness 308 game theory 90, 557, 583 genetic algorithms, GA 401, 407 GFSK, Gaussian frequency shift keying 317 GMSK, Gaussian minimum shift keying 317 gnuRadio 399, 402 Gnutella protocol 236 Graph Coloring 560 Graphic interface 580 Group selection 66

711

GSM, Global System for Mobile Communciations 317 Half-duplex constraint 180 hardware 309, 312 HARQ 488 HAWK, Highly Adaptable Wireless Kit 307, 317 HDL, hardware description language 313 Hidden terminals 168 HOM, higher order modulation 317 ICs, integrated chips 313 IDE, integrated design environment 312 IEEE 802.16d/e 308 IF, intermediate frequency 313 IFFT, inverse fast Fourier transform 314 IMD, intermodulation distortion 323 incentives 87 information rate 309 infrastructure components 306 interconnection 118 interference temperature concept 363 IP header compression 499 IPD model 584 ISM-band 399 L-value 315 Level of cooperation 160 Line of sight optical wireless systems 625 LLR, log-likelihood ratio 315 LO, linking operation 311 machine learning 401 matched filter 315 matched filter detection 360 matching 301 max-min fair 287, 291 mechanism design 87 medium access control (MAC) 309 Memes 63 microscopic diversity 117 MIMO-OFDM 163 minimum regret learning 537, 539, 544 mixed signal 309

712

Index

MMI, man machine interface 316 mobility 205 mobility management 200 Mobility scenario 163 symmetric/asymmetric 165 mode activity vector 288 Models Library 581 monopsony 114 Moore’s Law 31 MSK, minimum shift keying 317 multi-objective optimization 409, 412 multicast 485 Multimedia 307 multimode terminal 111 MUSTANG, MUlti-STAndard single chip transceiver for the Next Generation 307, 317 Nash bargaining game 125 Nash equilibrium 536, 537, 544, 545, 547, 558 Nash equilibrium point 125 National Border Spectrum Sharing 562 natural monopolies 115 negotiated spectrum use 355 network advertisement 114 Network coded cooperation 160 network composition 112 network discovery 114 Network Operator Spectrum Sharing 561 network sharing 119 non-cooperative game framework 535 noncooperative game 124 Observer 588 OFDM, Orthogonal Frequency Division Multiplexing 308 OFDMA Spectrum Sharing 574 open systems 119 Opportunistic relaying 170, 174 Optical amplification 630 Optical wireless communications 623 Optical wireless components 625 Optical wireless hotspots 628 optical, unguided waves 309 Orthogonality constraint 180 packet aggregation

518

pareto boundary 294 Pareto efficiency 536 Pareto front 409 Pareto-optimality 558 Partially Observed Markov Decision Process, POMDP 410 Partner 156, 159 Partner selection 169 passive awareness 354 Patches 589 Patterns 170 payment systems 93 PCB, printed circuit board 310 Peer-to-Peer (P2P) 235 Peer-to-Peer overlay network 64 persistent agent strategy 122 Petri net 310 platform 306 platform approach 309 PLL, phase locked loop 319 policy based approach 356 potential game 536, 544, 545, 547 potential offered load 122 power class 309 price elasticity of demand 122 Price of Anarchy 558 price skimming 130 pricing 534, 535, 539–541, 544 primary user 354 Prisoner’s Dilemma 67, 584 processor core 309 production cost 110 Properties of optical and RF channels 627 proportional fair 287, 293 proportionally fair divisible auction 120 protocol, higher layer 309 QoS, Quality of Service 316 QPSK, quadrature phase shift keying 317 Quality of Service 500 query profile 243 Query Routing Protocol (QRP) 238 radio 309 Radio Resource Manager, RRM rate region 294

398

Index reconfiguration 309 reinforcement learning 402 Relay-adaptive cooperation 160 Relaying broadcast 158 simple 158 unicast 158 user cooperative 159 remote protocol component 427 reply profile 243 reputation schemes 534, 539 reputation systems 91 reservation price 120 resource management 399 RO, regular operation 311 RO&P, regular operation & prepare 311 RSSI, radio signal strength indication 315 RTS/CTS 514 SD, secure digital (card) 320 secondary user 354 Selection relaying 160 semantic profile 239 SemPeer 244 service class 309 Shapley value 536 sharing agreement 113 Slider 590 software 309, 312 software controlled radio 437 software defined radio 173, 310, 415, 437, 533 Solid-state lighting 629 Space-time cooperation 160 Space-Time Trellis Code 194 specialized access systems 111 spectrum awareness 354 spectrum broker 357 spectrum sensing 357 spectrum server 285–287, 356 Spectrum Sharing 555 spectrum use pattern 354 submodular game 536 supermodular game 536 Symella 237, 247 synset 238

713

Tags 69 taxonomy 239 TCP 399, 406 TDMA, time division multiple access 317 terminal 306 Testbeds 173 theory of comparative advantage 116 throughput 516 time-series analysis 402 Time-to-Live (TTL) 236 toolbox 399, 415 Tournament Selection 64 trade-agent 115 transceiver 309 transmission mode 288 TRG 200, 201 architectural requirements 203 Consumers 200 filters 205 implementation 203, 205 Producers 200 triggers 200 Turtle 580 Typical sequences Strongly typical sequences 185 Weakly typical sequences 185 ULLA Query Language 405 UMTS, Universal Mobile Telecommunications System 308 UMTS, Wideband Code Division Multiple Access (WCDMA) 308 Unified Link-Layer API, ULLA 403 Universal Network Interface, UNI 405 Unlicensed Bands 567 USB, universal serial bus 314 User cooperative relaying 159 utility 411 Value of Perfect Information, VPI 413 Virtual File System, VFS 417 Virtual World 580 Visible light communications 629 Voice over IP 499 VR, virtual reality 307 WiFi Operators

569

714

Index

WiMAX, Worldwide Interoperability for Microwave Access 308 wireless grids

34

Wireless Local Area Network (WLAN) 308, 513

Wireless Metropolitan Area Network (WMAN) 163 wireless terminal 309 Wyner-Ziv coding 182 XML

405

Cognitive Wireless Communication Networks

Read more

Wireless Communications Systems and Networks

Read more

Security in Wireless Mesh Networks (Wireless Networks and Mobile Communications)

Read more

Security in Wireless Mesh Networks (Wireless Networks and Mobile Communications)

Read more

Wireless Communications

Read more

Wireless Communications

Read more

Wireless Communications

Read more

Wireless Communications

Read more

The Future of Wireless Communications

Read more

Wireless Communications

Read more

Wireless Networks

Read more

Wireless Networks

Read more

Wireless networks

Read more

Wireless Sensor Networks Signal Processing and Communications

Read more

Wireless Communications & Networks (2nd Edition)

Read more

Roaming in Wireless Networks (Communications Engineering)

Read more

Optical Wireless Communications: IR for Wireless Connectivity

Read more

Problem Solving for Wireless Sensor Networks (Computer Communications and Networks)

Read more

Guide to Wireless Sensor Networks (Computer Communications and Networks)

Read more

Wireless Quality of Service: Techniques, Standards, and Applications (Wireless Networks and Mobile Communications)

Read more

Optical Wireless Communications: IR for Wireless Connectivity

Read more

Fundamentals of Wireless Sensor Networks: Theory and Practice (Wireless Communications and Mobile Computing)

Read more

Globalization of Mobile and Wireless Communications

Read more

IEEE Wireless Dictionary (IEEE Standards Wireless Networks)

Read more

Wireless Communications Security

Read more

IFIP-Personal Wireless Communications

Read more

4G wireless video communications

Read more

Mobile Wireless Communications

Read more

Modern Wireless communications

Read more

Modern Wireless communications

Read more

Recommend Documents

Cognitive Wireless Communication Networks

Cognitive Wireless Communication Networks Ekram Hossain • Vijay Bhargava Editors Cognitive Wireless Communication Ne...

Wireless Communications Systems and Networks

TeAM YYePG Digitally signed by TeAM YYePG DN: cn=TeAM YYePG, c=US, o=TeAM YYePG, ou=TeAM YYePG, [email protected] Re...

Security in Wireless Mesh Networks (Wireless Networks and Mobile Communications)

P1: Binaya Dash July 31, 2008 12:41 AU8250 AU8250˙C000 P1: Binaya Dash July 31, 2008 12:41 AU8250 AU8250˙C000 ...

Security in Wireless Mesh Networks (Wireless Networks and Mobile Communications)

Security in Wireless Mesh Networks Series Editor: Yan Zhang Millimeter Wave Technology in Wireless PAN, LAN, and MAN ...

Wireless Communications

WIRELESS COMMUNICATIONS Wirelesss technology is a truly revolutionary paradigm shift, enabling multimedia communication...

Wireless Communications

WIRELESS COMMUNICATIONS Andrea Goldsmith Stanford University The possession of knowledge does not kill the sense of w...

Wireless Communications

WIRELESS COMMUNICATIONS Wirelesss technology is a truly revolutionary paradigm shift, enabling multimedia communication...

Wireless Communications

WIRELESS COMMUNICATIONS Andrea Goldsmith Stanford University The possession of knowledge does not kill the sense of w...

The Future of Wireless Communications

The Future of Wireless Communications The Future of Wireless Communications William Webb Artech House Boston London...

Wireless Communications

CHAPTER 1 Introduction to Wireless Communication Systems he ability to communicate with people on the move has evolv...