Agent Mediated Knowledge Management AMKM 2003

Lecture Notes in Artificial Intelligence Edited by J. G. Carbonell and J. Siekmann Subseries of Lecture Notes in Comput...

Author: Ludger van Elst | Andreas Abecker | Virginia Dignum

21 downloads 816 Views 6MB 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

Lecture Notes in Artificial Intelligence Edited by J. G. Carbonell and J. Siekmann

Subseries of Lecture Notes in Computer Science

2926

3

Berlin Heidelberg New York Hong Kong London Milan Paris Tokyo

Ludger van Elst Virginia Dignum Andreas Abecker (Eds.)

Agent-Mediated Knowledge Management International Symposium AMKM 2003 Stanford, CA, USA, March 24-26, 2003 Revised and Invited Papers

13

Series Editors Jaime G. Carbonell, Carnegie Mellon University, Pittsburgh, PA, USA J¨org Siekmann, University of Saarland, Saarbr¨ucken, Germany Volume Editors Ludger van Elst Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz (DFKI) GmbH Knowledge Management Department Erwin-Schroedinger-Str., 67608 Kaiserlautern, Germany E-mail: [email protected] Virginia Dignum University of Utrecht Institute of Information and Computing Sciences 3508 TB Utrecht, The Netherlands E-mail: [email protected] Andreas Abecker Forschungszentrum Informatik, WIM Haid-und-Neu-Str. 10-14, 76131 Karlsruhe, Germany E-mail: [email protected]

Cataloging-in-Publication Data applied for A catalog record for this book is available from the Library of Congress. Bibliographic information published by Die Deutsche Bibliothek Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the Internet at . CR Subject Classification (1998): I.2.11, I.2.4, I.2, H.5.3, H.4, H.3, C.2.4, K.4.4, J.1 ISSN 0302-9743 ISBN 3-540-20868-2 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. Springer-Verlag is a part of Springer Science+Business Media springeronline.com c Springer-Verlag Berlin Heidelberg 2004 Printed in Germany Typesetting: Camera-ready by author, data conversion by Boller Mediendesign Printed on acid-free paper SPIN: 10976241 06/3142 543210

Preface

In this book, we present a collection of papers around the topic of AgentMediated Knowledge Management. Most of the papers are extended and improved versions of work presented at the symposium on Agent-Mediated Knowledge Management held during the AAAI Spring Symposia Series in March 2003 at Stanford University. The aim of the Agent-Mediated Knowledge Management symposium was to bring together researchers and practitioners of the ﬁelds of KM and agent technologies to discuss the beneﬁts, possibilities and added-value of cross-fertilization. Knowledge Management (KM) has been a predominant trend in business in recent years. Not only is Knowledge Management an important ﬁeld of application for AI and related techniques, such as CBR technology for intelligent lessons-learned systems, it also provides new challenges to the AI community, like, for example, context-aware knowledge delivery. Scaling up research prototypes to real-world solutions usually requires an application-driven integration of several basic technologies, e.g., ontologies for knowledge sharing and reuse, collaboration support like CSCW systems, and personalized information services. Typical characteristics to be dealt with in such an integration are: – – – –

manifold, logically and physically dispersed actors and knowledge sources, diﬀerent degrees of formalization of knowledge, diﬀerent kinds of (Web-based) services and (legacy) systems, conﬂicts between local (individual) and global (group or organizational) goals.

Agent approaches have already been successfully employed in KM for many partial solutions within the overall picture: agent-based workﬂow, cooperative information gathering, intelligent information integration, and personal information agents are established techniques in this area. In order to cope with the inherent complexity of a more comprehensive solution, Agent-Mediated Knowledge Management (AMKM) deals with collective aspects of the domain in an attempt to cope with the conﬂict between the desired order and the actual behavior in dynamic environments. AMKM introduces a social layer that structures the society of agents by deﬁning speciﬁc roles and possible interactions between them. This workshop set the scene for the assessment of the challenges that AgentMediated Knowledge Management faces as well as the opportunities it creates. By focusing on agent-mediated interactions, specialists from diﬀerent disciplines were brought together in a lively and inquisitive environment that provided nice interactions and debates. The main topics for the workshop were: – collaboration and P2P support, – agent-based community support, – agent models for knowledge and organizations,

VI

Preface

– context and personalization, – ontologies and the Semantic Web, – agents and knowledge engineering. Besides extended versions of workshop presentations, this volume includes an introductory chapter, and papers originating from the invited talk and from discussion sessions at the symposium. The result is that this volume contains high-quality papers that really can be called representative of the ﬁeld at this moment. This volume starts with an introduction to the Agent-Mediated Knowledge Management topic. The paper provides an extended motivation and an overview of research and current developments in the ﬁeld. The remainder of the volume has been arranged according to the topics listed above. The ﬁrst section contains four papers on collaboration and peer-to-peer support. The ﬁrst paper in this section by Bonifacio et al. proposes a P2P architecture for distributed KM. Graesser et al. discuss the results of a study on the beneﬁts for KM from intelligent interfaces, namely animated conversional agents. The third paper by Guizzardi et al. presents Help&Learn, an agent-based peerto-peer helpdesk system to support extra-class interactions among students and teachers. The section ends with a paper by Ehrig et al. suggesting a concise framework for evaluation of P2P-based Knowledge Management systems. The second section contains three papers on agent-based community support. The ﬁrst paper by Schulz et al. presents a conceptual framework for trustbased agent-mediated knowledge exchange in mobile communities. Kayama and Okamoto examine knowledge management and representation issues for the support of collaborative learning. The last paper in this section, by Moreale and Watt, describes a mailing list tool, based on the concept of a mailing list assistant. The third section is devoted to agent models for knowledge and organizations. Filipe discusses the coordination and representation of social structures based on using the EDA agent model for normative agents, combined with the notion of an information ﬁeld. Lawless looks at the fundamental relations between the generation of information and knowledge, with agent organizations, decisionmaking, trust, cooperation, and competition. The third paper, by Furtado and Machado, describes an AMKM system for knowledge discovery in databases. Hui et al. report on experience using RDF to provide a rich content language for use with FIPA agent toolkits. The paper by Magalhaes and Lucena, describing a multiagent architecture for tool generation for document classiﬁcation, closes this section. The fourth section, on context and personalization, starts with a paper by Lou¸c˜a who presents a multiagent model to support decision-making in organizations. Novak et al. introduce an agent-based approach to semantic exploration and knowledge discovery in large information spaces. The paper by Evans et al. looks at the use of agents to identify and ﬁlter relevant context information in information domains. The section ends with a paper by Blanzieri et al., presenting the concept of implicit culture for personal agents in KM.

Preface

VII

The ﬁfth section contains four papers that focus on ontologies and the Semantic Web. Cao and Gandon discuss the beneﬁts of societies of agents in a corporate semantic web. Krueger et al. look at ways to fully realize the potential of the Semantic Web, by automatically upgrading information sources with semantic markup. Hassan investigates interfaces to harness knowledge from heterogeneous knowledge assets. Cassin et al. present an architecture for extracting structured information from raw Web pages and describe techniques for extracting ontological meaning from structured information. The paper by Toivonen and Helin presents a DAML ontology for describing interaction protocols. The last paper in this section, by Petrie et al., discusses the beneﬁts of agent technology to the development of Web services. The last section of the book contains six papers related to agent and knowledge engineering. The ﬁrst paper, by Furtado et al., studies the relationship between agent technology, knowledge discovery in databases, and knowledge management. The paper by Molani et al. describes an approach to capture strategic dependencies in organizational settings in order to support the elicitation of requirements for KM systems. Bailin and Truszkowski discuss the role of perspective in conﬂicts in agent communities. The paper by Tacla and Barth`es concerns a multiagent system for knowledge management in R&D projects. Pease and Li introduce a system for collaborative open ontology production. Finally, the paper by Dodero et al. describes an agent-based architecture to support knowledge production and sharing. We want to conclude this preface by extending our thanks to the members of the program committee of the AMKM workshop and to the additional reviewers who carefully read all submissions and provided extensive feedback on all submissions. We also want to thank all authors who were not only willing to submit their papers to our workshop and rework them for this book, but in addition contributed by their lively participation in a spontaneously organized peer review process.

September 2003

Ludger van Elst Virginia Dignum Andreas Abecker

VIII

Workshop Organization

Workshop Organization

Workshop Chairs Ludger van Elst Virginia Dignum Andreas Abecker

DFKI, Kaiserslautern, Germany Utrecht University, and Achmea, The Netherlands FZI, Karlsruhe, Germany

Program Committee Paolo Bouquet Rose Dieng Michael N. Huhns Daniel O’Leary Pietro Panzarasa Amit P. Sheth Walt Truszkowski Gerd Wagner

Additional Reviewers Nicola Henze Markus Junker Alexander M¨ adche Heiko Maus Michael Sintek Franz Schmalhofer

University of Trento, Italy INRIA, Sophia-Antipolis, France University of South Carolina, Columbia, USA University of Southern California, Los Angeles, USA University of Southampton, UK University of Georgia, Athens, USA NASA Goddard Space Flight Center, USA Eindhoven University of Technology, The Netherlands

Table of Contents

Towards Agent-Mediated Knowledge Management . . . . . . . . . . . . . . . . . . . . L. van Elst, V. Dignum, A. Abecker

1

Section I: Collaboration and Peer-to-Peer Support Peer-Mediated Distributed Knowledge Management . . . . . . . . . . . . . . . . . . . M. Bonifacio, P. Bouquet, G. Mameli, M. Nori

31

The Impact of Conversational Navigational Guides on the Learning, Use, and Perceptions of Users of a Web Site . . . . . . . . . . . . . . . . . . . . . . . . . . A. Graesser, G.T. Jackson, M. Ventura, J. Mueller, X. Hu, N. Person

48

Agent-Oriented Knowledge Management in Learning Environments: A Peer-to-Peer Helpdesk Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R.S.S. Guizzardi, L. Aroyo, G. Wagner

57

Towards Evaluation of Peer-to-Peer-Based Distributed Knowledge Management Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. Ehrig, C. Schmitz, S. Staab, J. Tane, C. Tempich

73

Section II: Agent Based Community Support TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange for Ubiquitous Peer Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S. Schulz, K. Herrmann, R. Kalckl¨ osch, T. Schwotzer

89

Knowledge Management Framework for Collaborative Learning Support . 107 M. Kayama, T. Okamoto An Agent-Based Approach to Mailing List Knowledge Management . . . . . 118 E. Moreale, S. Watt

Section III: Agent Models for Knowledge and Organizations Information Fields in Organization Modeling Using an EDA Multi-agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 J. Filipe A Quantum Perturbation Model (QPM) of Knowledge Fusion and Organizational Mergers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 W.F. Lawless, J.M. Grayson

X

Table of Contents

Improving Organizational Memory through Agents for Knowledge Discovery in Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 J.J. Vasco Furtado, V. Ponte Machado Experience in Using RDF in Agent-Mediated Knowledge Architectures . . 177 K.-y. Hui, S. Chalmers, P.M.D. Gray, A.D. Preece Using an Agent-Based Framework and Separation of Concerns for the Generation of Document Classiﬁcation Tools . . . . . . . . . . . . . . . . . . . . . . . . . 193 J.A. Pinto de Magalh˜ aes, C.J. Pereira de Lucena

Section IV: Context and Personalization Modelling Context-Aware Distributed Knowledge . . . . . . . . . . . . . . . . . . . . . 201 J. Lou¸c˜ a Discovering, Visualizing, and Sharing Knowledge through Personalized Learning Knowledge Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 J. Novak, M. Wurst, M. Fleischmann, W. Strauss Agentized, Contextualized Filters for Information Management . . . . . . . . . 229 D.A. Evans, G. Grefenstette, Y. Qu, J.G. Shanahan, V.M. Sheftel Implicit Culture-Based Personal Agents for Knowledge Management . . . . . 245 E. Blanzieri, P. Giorgini, F. Giunchiglia, C. Zanoni

Section V: Ontologies and Semantic Web Integrating External Sources in a Corporate Semantic Web Managed by a Multi-agent System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 T.-D. Cao, F. Gandon Automatically Generated DAML Markup for Semistructured Documents . 276 W. Krueger, J. Nilsson, T. Oates, T. Finin A Spreading Activation Framework for Ontology-Enhanced Adaptive Information Access within Organisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288 M.M. Hasan Ontology Extraction for Educational Knowledge Bases . . . . . . . . . . . . . . . . . 297 P. Cassin, C. Eliot, V. Lesser, K. Rawlins, B. Woolf Representing Interaction Protocols in DAML . . . . . . . . . . . . . . . . . . . . . . . . . 310 S. Toivonen, H. Helin Adding AI to Web Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 C. Petrie, M. Genesereth, H. Bjornsson, R. Chirkova, M. Ekstrom, H. Gomi, T. Hinrichs, R. Hoskins, M. Kassoﬀ, D. Kato, K. Kawazoe, J.U. Min, W. Mohsin

Table of Contents

XI

Section VI: Agents and Knowledge Engineering Knowledge Discovery in Databases and Agent-Mediated Knowledge Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 J.J. Vasco Furtado, I. Haimowitz, M. Wurst Intentional Analysis for Distributed Knowledge Management . . . . . . . . . . . 351 A. Perini, P. Bresciani, E. Yu, A. Molani Perspectives: An Analysis of Multiple Viewpoints in Agent-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 S.C. Bailin, W. Truszkowski A Multi-agent Architecture for Evolving Memories . . . . . . . . . . . . . . . . . . . . 388 C. Tacla, J.-P. Barth`es Agent-Mediated Knowledge Engineering Collaboration . . . . . . . . . . . . . . . . . 405 A. Pease, J. Li Dynamic Generation of Agent Communities from Distributed Production and Content-Driven Delivery of Knowledge . . . . . . . . . . . . . . . . 416 J.M. Dodero, S. Arroyo, V.R. Benjamins

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427

Towards Agent-Mediated Knowledge Management Ludger van Elst1 , Virginia Dignum2, and Andreas Abecker3 1

German Research Center for Artificial Intelligence – Knowledge Management Department – [email protected] 2 Achmea & University Utrecht [email protected] 3 Forschungszentrum Informatik – Forschungsbereich Wissensmanagement – [email protected]

Abstract. In this paper, we outline the relation between Knowledge Management (KM) as an application area on the one hand, and software agents as a basic technology for supporting KM on the other. We start by presenting characteristics of KM which account for some drawbacks of today’s – typically centralized – technological approaches for KM. We argue that the basic features of agents (social ability, autonomy, re- and proactiveness) can alleviate several of these drawbacks. A classification schema for the description of agent-based KM systems is established, and a couple of example systems are depicted in terms of this schema. The paper concludes with questions which we think research in Agentmediated Knowledge Management (AMKM) should deal with.

1 Agents and KM Knowledge Management (KM) is defined as a systematic, holistic approach for sustainably improving the handling of knowledge on all levels of an organization (individual, group, organizational, and inter-organizational level) in order to support the organization’s business goals, such as innovation, quality, cost effectiveness etc. (cp. [33]). KM is primarily a management discipline combining methods from human resource management, strategic planning, change management, and organizational behavior. However, the role of information technology as an enabling factor is also widely recognized, and – after a first phase where merely general purpose technology like Internet/Intranets or e-mail1 were found to be useful for facilitating KM – a variety of proposals exist showing how to support KM with specialized information systems (see, e.g., [4]). One class of such systems assumes that a huge amount of organizational knowledge is explicitly formalized (or, “buried”) in documents, and therefore tries to “connect” knowledge workers with useful information items. Typical systems in this category are Organizational Memory Information Systems (OMs, cp. [1, 24]) which acquire and 1

Especially large companies often report that these technologies were the first ways to communicate and distribute knowledge across boundaries of hierarchies.

L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 1–30, 2003. c Springer-Verlag Berlin Heidelberg 2003

2

Ludger van Elst, Virginia Dignum, and Andreas Abecker

structure explicit knowledge and aim at high–precision information delivery services (“provide the right people with the right information at the right time”). On the other hand, expert finder systems or community of practice support don’t rely so much on explicitly represented knowledge, but rather bring people together, for instance, to solve a given knowledge–intensive problem (see, for instance [7, 28]). Although such systems also use some explicit knowledge, with respect to the actual knowledge–intensive task, this is more meta than problem–solving knowledge. Often, Information Technology (IT) research for KM focused on the comprehensive use of an organization’s knowledge, thus aiming at the completeness of distribution of relevant information. Technically, this is typically supported by centralized approaches: Knowledge about people, knowledge about processes, and domain knowledge is represented and maintained as information in global repositories which serve as sources to meet a knowledge worker’s (potentially complex) information needs. Such repositories may be structured by global ontologies and made accessible, e.g., through knowledge portals [75, 52]). Or they may be rather “flat” and accessed via shallow (i.e., not knowledge–based) methods like statistics–based information retrieval or collaborative filtering (this is the typical approach of today’s commercial KM tools). In the following, we present some KM characteristics which – in our opinion – account for serious drawbacks of such centralized IT approaches to KM, and which can immediately be coined into requirements for a powerful KM system design: R1 KM has to respect the distributed nature of knowledge in organizations: The division of labor in modern companies leads to a distribution of expertise, problem solving capabilities, and responsibilities. While specialization is certainly a main factor for the productivity of today’s companies, its consequence is that both generation and use of knowledge are not evenly spread within the organization. This leads to high demands on KM: – Departments, groups, and individual experts develop their particular views on given subjects. These views are motivated and justified by the particularities of the actual work, goals, and situation. Obtaining a single, globally agreed–upon vocabulary (or ontologies) within a level of detail which is sufficient for all participants, may incur high costs (e.g., for negotiation). A KM system should therefore allow to balance between (a) global knowledge which might have or might constitute a shared context, but may also be relatively expensive; and (b) local expertise which might represent knowledge that is not easily shareable or is not worth sharing. – As global views cannot always be reached, a KM system has to be able to handle context switches of knowledge assets, e.g., by providing explicit procedures for capturing the context during knowledge acquisition and for recontextualizing during knowledge support. An example for context capturing is a lessons-learned system which is fed by debriefings after a project is finished [43, 42]. Here, a typical question pair is: “What was the most crucial point of the project’s success? What are the characteristics of projects where this point may also occur?” Altogether, we see that distributedness of knowledge in an organizational memory is not a “bug”, but rather a “feature”, which is by far not only a matter of physical

Towards Agent-Mediated Knowledge Management

3

or technical location of some file. It has also manifold logical and content-oriented aspects that in turn lead to derived aspects such as—in an ideal system—the need to deal with matters of – trust (Do I believe in my neighbor’s knowledge?), – responsibility (Is my neighbor obliged to maintain his knowledge base because I might use it? And am I obliged to point out errors that I find in his knowledge base?), – acknowledgement (Who gets the reward if I succeed with my neighbor’s knowledge?), – contextuality of knowledge (Is my neighbor’s knowledge still valid and applicable in my house and my family?), – ... and many others. R2 There is an inherent goal dichotomy between business processes and KM processes: For companies as a whole as well as for the individual knowledge worker KM processes do not directly serve the operational business goals, but are second order processes2 . Within an environment of bounded resources, knowledge workers will always concentrate on their first order business processes. This means they optimize their operational goals locally and only invest very little to fulfil strategic, global KM goals.3 It is clear and pretty well accepted that having and using knowledge is important for optimally fulfilling first-order tasks, but the workload and time pressure is nevertheless usually so high that the effort invested for preparing this, time for knowledge conservation, evolution, organization, etc., is considered a second-order process often neglected in practice. Even cumbersome activities for knowledge search and reuse are often considered to be unacceptable. Therefore, the KM processes should be embedded in the worker’s first-order processes, and proactive tools should minimze the cognitive load for KM tasks. R3 Knowledge work as well as KM in general, is “wicked problem solving” (cf. [15, 21, 22]): This means that a precise a-priori description of how to execute a task or solve the problem doesn’t exist, and consequently, it cannot be said in advance when knowledge should be captured, distributed, or used optimally. An optimal solution for KM problems and the respective knowledge and information flows cannot be prescribed entirely from start to finish, because goals may change or be adapted with each step of working on a task. Therefore knowledge workers and KM systems must be flexible enough to adapt to additional insights and to proactively take opportunities when they arise during work. Solving “wicked problems” is typically a fundamental social process. A KM system should therefore support the necessary complex interactions and underlying, relatively sophisticated processes like planning, coordination and negotiation of knowledge activities. A phenomenon closely related to this is that KM is very much about personal relationships. People want to be recognized as experts, and they are much more willing 2

3

There are a couple of exceptions to this, like R&D departments which have knowledge generation as first order goal. For a discussion of operational processes vs. knowledge processes, see, for instance, [68, 78]. In other words, employees will mostly find a way to get their business done, even if processes and tool support are bad, whereas KM tasks will simply be omitted. This has been our experience in KM systems building from our very first requirements gathering on [47].

4

Ludger van Elst, Virginia Dignum, and Andreas Abecker

to share knowledge face-to-face in collaborative problem-solving and expert chats than putting it anonymously into a central knowledge store. Hence flexible pointto-point connections for powerful online communication and collaboration, as well as individual solutions for knowledge storage, identification, and communication must be allowed. R4 KM has to deal with changing environments: In addition to the intrinsic problems described above, KM systems typically reside in environments which are subject to frequent changes, be it in the organizational structure, in business processes, or in IT infrastructure. Centralized solutions are often ill–suited to deal with continuous modifications in the enterprise, e.g., because the maintenance costs for detailed models and ontologies simply get too high. Furthermore, the implementation of KM systems often follows a more evolutionary approach where functionalities are not implemented “in one step” for a whole company, but partial solutions are deployed to clearly separated sub-structures. In order to obtain a comprehensive system, these elements then have to be integrated under a common ceiling without disturbing their individual value.4 Keeping these requirements in mind, let’s have a look at scenarios which are considered to be rewarding tasks for agent-based software solutions. We quote a number of characteristics from [60] (but similar arguments can be found in many books about multi-agent systems) typically indicating that a scenario could be a good application area for agent technology: agents are best suited to applications that are modular, decentralized, changeable, ill-structured, and complex. Although the match between these five salient features and the KM requirements R1 – R4 listed above is already obvious, we want to elaborate a bit more explicitly on this match. Let us start with the weak definition of agents [83] (with the definitional features autonomy, social ability, reactive behavior, and proactive behavior). Now we will see why agent-based approaches are especially well–suited to support KM with information technology: In the first place, the notion of agents can be seen as a natural metaphor to model KM environments which can be conceived as consisting of a number of interacting entities (individuals, groups, IT, etc.) that constitute a potentially complex organizational structure (see R1, but also R4). Reflecting this in an agent-based architecture may help to maintain integrity of the existing organizational structure and the autonomy of its subparts. Autonomy and social ability of the single agents are the basic means to achieve this. Reactivity and proactivity of agents help to cope with the flexibility needed to deal with the “wicked” nature of KM tasks (see R3). The resulting complex interactions with the related actors in the KM landscape and the environment can be supported and modeled by the complex social skills with which agents can be endowed. Proactiveness as well as autonomy help accomodating to the reality that knowledge workers typically do not adopt KM goals with a high priority (see R2). 4

This requirement of connecting several smaller existing KM islands to create a bigger picture, also fits very well with the frequently suggested KM introduction strategy of looking for “quick wins” (cp. [81]).

Towards Agent-Mediated Knowledge Management

5

Regarding primarily the software-technology aspects of agents, they represent a way of incorporating legacy systems into modern distributed information systems; wrapping a legacy system with an agent will enable the legacy system to interact with other systems much more easily. Furthermore, agent approaches allow for extensibility and openness in situations when it is impossible to know at design time exactly which components and uses the system will have. Both arguments reflect pretty well the technical consequences of abstract requirements such as R4 and R3 (changing environments demand continuous reconfiguration, the unpredictable nature of wicked-problem solving require flexible approaches), R2 (competition between operational work and KM meta work call for stepwise deployment and highly integrated KM solutions), or R1 (already existing local solutions must be confederated). There have been a number of more or less theoretical analyses of requirements and ambitious approaches to agent-based solutions for KM (see, e.g., [56, 72]), as well as experimental systems exploring the use of agents for investigating the one or other aspect (such as weakly-structured workflow, ontology mediation, metadata for knowledge retrieval, or contextuality) of comprehensive agent-based KM frameworks (like FRODO, CoMMA, Edamok [3, 31, 11, 36], some of them are included in this book). We are well aware that nowadays we are far from reaching a state where we can oversee all methodological, technological, and practical benefits and prospects, problems and pitfalls, and challenges and achievements of Agent-Mediated Knowledge Management. But we hope and we are pretty sure that this paper as well as this volume gives a good idea of the AMKM landscape, opens up some new ways for interesting future work and shows how far we have already come.

2 A Description Schema for Agent-Based KM Approaches In research as well as in first generation “real-world applications” several agent–based systems exist to support various aspects of Knowledge Management, from personal information agents for knowledge retrieval to agent–based workflows for business process–oriented KM. In order to be able to compare different agent approaches to KM, we need to describe agent and multi–agent architectures in a way that abstracts from the particularities of individual implementations, but still captures their relevant characteristics. A couple of helpful classification schemas for single agents and multi–agents systems have already been proposed (e.g., Franklin and Graesser’s taxonomy of agents [35]), discriminating agents for example by their tasks (information filtering, interface agents etc.), their abstract architecture (e.g., purely reactive vs. agents with state) or concrete architecture (e.g., belief-desire-intention vs. layered) architectures (cf. [82]), or other specific features (mobility, adaptivity, cooperativeness, etc.). For instance, [61] presented an interesting top-level characterization of agent applications, basically distinguishing three kinds of domains: 1. Digital domains where the whole environment of the agents is constituted from digital entities, as is the case, e.g., in telecommunications or static optimization problems. 2. Social environments where software agents interact with human beings.

6

Ludger van Elst, Virginia Dignum, and Andreas Abecker

3. Electromechanical environments where agents manipulate and experience the nonhuman physical world via sensors and actuators, as is the case, e.g., in robotics, factories, etc. A further classification dimension can be added directly because besides the domain to be handled by the agents we also have to consider the kinds of interfaces to be provided by an agent-based application. Here we have the same options as above: we need social interfaces to integrate people, digital interfaces to interact with other agents, and electromechanical interfaces to link to the physical world. In the case of KM applications we normally have to consider a (highly) social environment with both social and (usually a number of different) digital interfaces. For the purpose of this paper, we propose a description schema that is on the one hand more specific than these classifications and on the other hand also captures the whole life cycle of agent–oriented system development. To get an overview of agent approaches for KM, we think that a categorization along three dimensions is especially beneficial: 1. the stage in a system’s development process where agents are used (analysis, conceptual design, or implementation); 2. the architecture / topology of the agent system; and 3. the KM functionality / application focused on. We discuss these dimensions in the following three subsections. 2.1 System Development Level Agent–oriented Software Engineering emphasizes the adequacy of the agent metaphor for design and implementation of complex information systems with multiple distinct and independent components. Agents also enable the aggregation of different functionalities (such as planning, learning, coordination, etc.) in a conceptually embodied and situated whole [51]; agents also provide ways to relate directly to these abstractions in the design and development of large systems. In Knowledge Management, not only are the IT systems highly complex and distributed, but also the organizational environment in which these systems are situated. Especially in more comprehensive KM approaches, the complexity of the organization has to be reflected in the IT architecture. Often, “real world entities” of the organization have a relatively direct counterpart in the computer system, leading to a rather tight coupling between the real and the virtual worlds. Therefore, an organizational analysis is commonly an integral part of methodologies for the development of Knowledge Management IT (see, e.g., the CommonKADS [74], or the DECOR [59] methods). Originating in the realm of human collaboration, the notion of agents can be an epistemologically adequate abstraction to capture and model relevant people, roles, tasks, and social interactions. These models can be valuable input for the requirements analysis phase for the development of the KM system. So, due to the fundamentally social nature of KM applications, the agent paradigm can be — and actually has been — applied at different development levels, such as analysis, modeling and design, and not just to represent technological components of

Towards Agent-Mediated Knowledge Management

7

Fig. 1. Notion of Agents at Different Stages in the Development Cycle of an Agent– Based KM System

implemented systems. Figure 1gives an overview of the use of agents on different levels in the system engineering cycle. Of course, on each level we can have different specific agent theories (that is, how agents are conceptualized, what basic properties they have, etc. [83]) and respective representation languages (which on the implementation level may be operational programming languages) for defining concrete agents and their relations. Methodologies for agent–oriented software engineering like Tropos [40] and Gaia [85] not only define these representation languages for different levels, they are also the glue between them by providing mappings and processes for the transition from one level to another. The hope is, of course, that on the basis of a high correspondence of the primitives on each level these transitions will be smooth and less error–prone. Even though such methodologies provide a powerful tool to design multi-agent systems, and are currently widely used, they are not always suitable to deal with the complexity of fully fledged KM environments, including openness and heterogeneity. In [27] overall design requirements for KM environments were identified, which include the need to separate the specification of the organizational structure for the internal architecture of its component entities, and the need for explicit representation of normative issues. A recent proposal for a methodology for agent societies that meets these requirements, is presented in [25]. However, even when it seems likely that the entire development life cycle for KM applications can benefit from the concept of agents, we are well aware that in concrete,

8

Ludger van Elst, Virginia Dignum, and Andreas Abecker

real–life situations often pragmatic reasons5 may lead to the use of agents at just one or two development levels. On the other hand, having to implement a KM system on the basis of “conventional software” (like relational databases or client/server–based groupware solutions) or on the basis of modern, strongly related technologies like peerto-peer networks or web services should not necessarily hinder an agent–oriented analysis and system design6 . 2.2 Macro-level Structure of the Agent System Agent theories, abstract agent architectures, and agent languages as defined in [83] mainly take a micro–level view, i.e., they focus on the concept of one agent: What properties does an agent have, how can these properties be realized in a computer system, what are the appropriate programming languages for that? For Knowledge Management—which typically employs a strong organizational perspective—the macro–level structure is also of special interest. How many agents do we have? What types of agents? What is the topology with respect to the flow of information, or with respect to the co-ordination of decisions? One possible dimension to characterize the macro– level of an agent–based KM system is the degree of sociability as depicted in Figure 2:

Single A gent •Personal Inform ation A gent

H om ogeneous M A S •C ooperative R etrieval A gents

•A gent-based O M A rchitecture

(H eterogeneous) A gentSocieties •A gent-based D istributed O M A rchitecture

Fig. 2. Degree of Sociability

– Single–agent architectures are at one end of the spectrum. Typical examples come from the area of user interface or personal information agents which build a model of a user’s interest and behavior, and exploit this knowledge to support him or her by providing relevant information, e.g., from the Web. These agents can perceive their environment and access some objects like web resources, but they normally have no elaborated interaction (like collaboration or negotiation) with other agents (except for the human user). 5

6

In [84], Wooldridge and Jennings nicely describe classes of pitfalls for the development of agent–based systems, including the “overselling”, “being dogmatic”, and “agents as silver bullet” pitfalls. Parunak [60, 61] also discusses the pragmatics of agent–based software development in real–world settings. Actually we observe that technologies like P2P and web services incorporate many aspects of the notion of agents when encountering application domains that have characteristics like those described in Section 1 for KM applications.

Towards Agent-Mediated Knowledge Management

9

– Homogeneous multi–agent architectures already have a higher degree of sociability. Agents can co-operate with other agents in order to solve their tasks. Homogeneity means that the system consists mainly of one type or class of agents. These agents do not necessarily have to have exactly the same goals, but their tasks and capabilities are comparable. Agent–based collaborative filtering is a typical example for this class of MAS: All agents are seen as peers which can provide information on what entities they use or like, and each agent can collect this information to provide the user with valuable hints about interesting new information. Nevertheless, all agents may have individual information collection and integration strategies. – Heterogeneous multi–agent architectures contain multiple agent classes which may have completely different purposes, knowledge and capabilities. Various information integration architectures (e.g., Knowledge Rovers [44], MOMIS/MIKS [8]) are described as heterogeneous MAS: Specialists exist for wrapping information sources, agents for integrating different description schemas, and for adequately presenting information to the users. All these different agent types have to cooperate and bring in their complementary expertise in order to accomplish the overall goal of the system. A characterization of the macro–level structure of an agent–based KM system may, in addition to the description of the number of agents and the system’s heterogeneity, also include facets like – co-ordination form: How are decisions and information flow coordinated? On the basis of a market model? As a fully connected network? Or in a hierarchical manner? – open vs. closed system: Can new agents enter the system? If yes, does their agent class (competencies, purpose, etc.) have to be known in advance? Or can new types of agents be integrated easily (even at runtime)? – implicit vs. explicit social structure: Do the agents have an explicit representation of their role in the system which allows for a certain assurance of the system’s global behavior? Do they even have a machinery for reasoning about their rights and obligations? Are roles globally defined or negotiated? Or is the agent’s social behavior only locally controlled and the system’s behavior completely emergent? Electronic institutions are a typical example of a complex society architecture. Electronic institutions provide a computational analogue of human organizations in which agents interact through roles that are defined as specified patterns of behavior [79]. Similarly, virtual organizations can potentially take advantage of the new electronic environments through coalition formation among disparate partners to form aggregate entities capable of offering new, different or better services than might otherwise be available. To design such systems requires a theory of organization design, and knowledge of how organizations may change and evolve over time. Sociological organization theory and social psychology are clearly important inputs to the design. Moreover, for the design of open societies, political theory may be necessary. Open systems permit the involvement of agents from diverse design teams, with diverse objectives, which may all be unknown at the time of design of the system itself. How the system as a whole makes decisions or agrees on joint goals will require the adoption of specific

10

Ludger van Elst, Virginia Dignum, and Andreas Abecker

political philosophies, for example whether issues are subject to simple majority voting or transferable preference voting, etc. (cp. [51]). Of course, the above examples for different degrees of sociability — single–agent, homgeneous/heterogeneous MAS — do not form a discrete, categorical discrimination. On the contrary they are exemplary operating points on a continuous scale. Heterogeneous MAS, e.g., may have sub–societies that are homogeneous themselves. Or, a system may be mainly homogeneous, but has one specialist agent for a certain task. And even between several aspects of communication the structure of the system may differ. So, the topology for making decisions may be a hierarchy, while information may be spread based on a market or fully connected network model [29]. It is also clear that there are dependencies between the three facets of system description. Not all possible combinations fit equally well together, not all of them are equally useful. For instance, if we have a highly structured agent society (like the electronic institutions outlined above) we can normally profit from known social structures when designing effective co-ordination, communication, and decision-making mechanisms, and do not have to use such a general, but “expensive” mechanism as a fully connected network. On the other hand, the more social structure is explicitly implemented into an agent society, the more “closed” this society might be in the sense that entering it will probably be based on a well-specified procedure, depending both on the current status of the society and on the capabilities and goals of a new agent that wants to enter it. On the other hand, if we have a relatively “democratic” way of co-ordination, like a market model, and a completely implicit social structure, it might be pretty easy for such an agent society to act as a pretty “open” system. 2.3 KM Application Area The two classification dimensions for multi agent systems described in the previous subsections are not directly related to applications in the KM domain. Up to now, we looked at the level in system development where the notion of agents is used, and at the macro–level structure of the agent system7 . The third dimension for characterizing agent–based KM applications, described in this subsection, deals with the specific knowledge management functionality of the system: What is the scope of the systems? Which Knowledge Management processes or tasks are supported? In this paper, we do not want to prescribe a detailed framework for this dimension, but only want to gather and offer some possibilities and general directions. Principally, all high–level Knowledge Management models can be seen as a starting point to form the vocabulary for this dimension, and there are many such KM models. We will start with the famous KM cycle by Probst et al. [64] which — in addition to the management–oriented tasks of defining knowledge goals and assessing the organization’s knowledge — e.g., identifies six building blocks: – Identification processes analyze what knowledge exists in an organization, what the knowledge containers are, who the stakeholders are, etc. 7

Though it should be noted that the emphasis on the degree of sociability as an important dimension of characterization is strongly biased by our theoretical analysis of KM in Section 1.

Towards Agent-Mediated Knowledge Management

11

Acquisition is the process of integrating external knowledge into an organization. Development processes generate new knowledge in the organization. Distribution processes connect knowledge containers with potential users. Preservation aims at the sustainability of knowledge, i.e., that is accessible and understandable over a period time. – Utilization means to operationalize available knowledge in order to solve actual business tasks (better). – – – –

Originating in the management sciences, Probst et al.’s view has been widely adopted and adapted in technology–oriented KM literature (e.g., [1, 76]). Likewise, the classical model of Nonaka and Takeuchi [58]—which focuses on knowledge generation—can be used to describe the KM application area of a system. These authors claim that new knowledge is created by four types of transformation processes between implicit / internal knowledge (e.g., competencies, experiences, skills) and explicit / external knowledge (e.g., facts, coded rules, formal business processes): – With socialization, knowledge that is implicit to a person is transferred to another person by sharing experiences. Apprenticeship learning, for example, makes heavy use of socialization. – Externalization is the process of making implicit knowledge explicit, e.g., by talking about it, writing it down informally or by formalizing it. Knowledge acquisition techniques developed in expert system research mainly aim at externalization. – Combination is the basis for generating new knowledge from external knowledge by relating knowledge pieces with other knowledge pieces. Data mining and machine learning are technical approaches of this type of knowledge creation process. – Internalization is the transformation of explicit knowledge into implicit knowledge and thereby making it applicable. From these classical models, several further distinctions have been developed in Knowledge Management research that can be utilized to describe the application area. For example, systems can take a more process–oriented or a more product–oriented view [47, 54]. The latter emphasizes the management of explicit knowledge contained in ”kowledge products” such as databases, documents, formal knowledge bases etc.; the former focuses on human beings and their internal knowledge, i.e., the ”process of knowing” and the ”process of knowledge exchange” between people. Typical systems with a product–oriented view are document retrieval agents. Expert finder systems, on the other hand, take a more process–oriented view. Furthermore, a KM system can support individuals and their tasks at hand, it can support teams and groups, or it may take a more global, organizational perspective. The theoretical analysis of Knowledge Management characteristics in Section 1 may be the source of further possible application areas for information technology, e.g., facilitating trust, motivating users to share knowledge, or establishing group awareness. Concrete agent–based KM applications may deal with one or a few of these aspects, or they may be more comprehensive frameworks that try to cover large parts of the KM cycle. In the following section we will analyze existing agent-based KM applications, illustrative for the different approaches.

12

Ludger van Elst, Virginia Dignum, and Andreas Abecker

3 Exemplary Agent-Based KM Applications In the previous section, we proposed three dimensions to describe agent–based Knowledge Management systems: i) the system development level (analysis, design, implementation), ii) the macro–level structure of the system (single agent, heterogeneous, or homogeneous MAS), and iii) the KM application area (knowledge distribution, generation, use, etc.). In this section, we will present some examples of agent-based systems developed to support and/or model Knowledge Management domains. We group these systems by the second dimension (macro–level structure), because this also largely reflects and matches the historical evolution of research in this area. Since compiling a complete overview of the systems in all three dimensions is well beyond the scope of this paper, we briefly sketch some systems which we consider typical for the specific approach. Our aim is to present current developments in Agent-Mediated Knowledge Management, indicate their differences to conventional approaches, expose their benefits, and suggest areas for further work. 3.1 Predominantly Single Agent Approaches Most KM support systems that take a single agent approach are User Interface Agents or Information Agents. A User Interface Agent embodies the metaphor of “a personal assistant who is collaborating with the user in the same work environment” [53]. Though this rather general definition would comprise agent support for all kinds of KM activities that a knowledge worker can perform (e.g., distribute knowledge, generate new knowledge), virtually all systems in this class are information agents8 . These agents typically – have access to a variety of information sources, – handle a model of the user’s information needs and preferences, and – try to provide relevant information to the user in an adequate way, either by filtering incoming information from the sources or by actively retrieving it. Prototypical systems in this category use e-mail in-boxes, news forums, dedicated KM databases within the company, intranet documents, or internet search engines as information sources. A representative architecture for an intelligent information agent that assists the user in accessing a (not agent–based) Organizational Memory, in this case the OntoBroker system, is described in [77, 73]. The agent relies on an explicit model of the business process the user is engaged with and uses this knowledge of the work context to determine when information support may be appropriate and what information may be useful in that context. Two variants of the system are available, a reactive and a proactive one. In the reactive case, the user triggers the agent by selecting a specific (pre–modelled) query in a specific application context. The agent then tries to retrieve relevant knowledge from the Organizational Memory and passes it on directly to the respective application 8

For an overview of personal information agents, also for other tasks like expert finding and information visualization see [49].

Towards Agent-Mediated Knowledge Management

13

that triggered the information need and thereby to the user. The reactive agent must have complete knowledge about the process context and about the information needs. The proactive agent, on the other hand, relaxes these two requirements, i.e., the application context and the relevant queries may be only partially defined when the agent becomes active. Instead, the agent has a proactive inferencing mechanism which employs heuristics to retrieve relevant information based on uncomplete context and query specifications. In order to cope with the potentially huge number of possible results and related problems (e.g. storage, processing time) from inferencing with underspecified context, the proactive agent is equipped with a mechanism for bounded resource consumption. For their actual knowledge retrieval step, OntoBroker agents, both reactive and proactive, exploit the ontology-based structure of the Organizational Memory . However, many personal information agents are designed for an environment where such an ontological structure of the information sources cannot be assumed, e.g., the World Wide Web. In this case, agents often rely on standard information retrieval techniques for searching. Rhodes and Maes [70] present three just-in-time information retrieval (JITIR) agents: The Remembrance Agent continually presents a list of documents that are related to a document that is currently being written or read in the Emacs editor, Margin Notes uses documents loaded in a Web Browser as context, and Jimminy uses the physical environment (location, people in the room, etc.) to determine what information may be relevant. All three agents use the same back–end system Savant [69] for the actual information retrieval step. Nevertheless, the primary contributions of research in personal information agents are not so much the various core retrieval techniques (from statistics–based similarities of text documents up to ontology–based access to formalized knowledge items), but the development of adequate sensors and effectors for personal information agents. Sensors define the way the agents can assess the context of their services, i.e., when to perform a service proactively and what the user’s actual information need is. Here, a wide range of approaches are covered in literature, from the pre–modelled business processes described above, to observing knowledge workers in their usage of standard office applications like text processors, web browsers or mailing tools (cf. Watson [17] or Letizia [48]). The effectors of user interface agents, on the other hand, determine the way information can be presented to the user. The JITIR agent Margin Notes [70], for example, automatically rewrites Web pages as they are loaded, and places links to personal information items in a dedicated area of the page. Watson presents suggestions in a dedicated window, and in KnowMore [2], information from the Organizational Memory can be directly handed over to specific fields in a form–based application. We now discuss the characteristics of personal assistants along the other two characterization dimensions for AMKM applications described in section 2. Concerning the level of system development, personal assistant approaches are mostly deployed at the modelling level. The most relevant aspect used from the agent metaphor is that an agent acts on behalf of a user who has specific goals and interests. Regarding the implementation level, personal assistants are currently mostly implemented using conventional programming techniques, i.e., without using a more general “agent development kit for personal information agents”. A well–known exception is Letizia, developed at MIT

14

Ludger van Elst, Virginia Dignum, and Andreas Abecker

[48]. With respect to the KM application area, personal assistants, as user–directed approaches, are mainly related to the dissemination of knowledge to be used by knowledge workers, in a just–in–time, just–enough fashion. Applications such as OntoBroker take a product–oriented view on knowledge, as they emphasize the management of explicit knowledge sources. To sum up, we can say that many of the presented ideas are already well-developed in the technological sense, and some of them have even found their way into commercial software products of advanced vendors. In those applications, the agent term is often not used in the narrower technical sense, but merely as a communication or as a design metaphor, but not built upon dedicated agent software platforms. There is a clear, but indirect, link between the functionalities achieved by such systems and our KM software requirements R1 – R4 defined above. Usually one can see that the software functionalities provided here are useful, because they address issues caused by our items R1 – R4 (e.g., in frequently changing environments, push services achieved by personal information agents are much more important than in stable environments, since an agent can continuously monitor whether some relevant change has happened). Altogether, though the software functionalities are stable to some extent and apparently useful, the logical next step for research and application has seldom been done, namely a rigorous assessment of usability and usefulness issues. There are a few specific experiments about evaluation of Personal Information Agents and the influence of process-aware, proactive information delivery, respectively (see [16, 18, 32, 70]), but in our opinion there is still a need for broad and long-term experiments about usability issues, user acceptance, and influence on working behavior and working efficiency / effectiveness by KM tools. 3.2 Homogeneous Multi-agent Approaches As described in Section 2.2, homogeneous multi–agent systems are formed by several agents belonging mainly to a common “agent class”, i.e., on an abstract level they have comparable competencies and goals (albeit they might act on behalf of different users)9 . Pure homogeneous multi–agent systems are rarely found in literature. Typically, facilitation functions (e.g., matchmaking and management of collaboration) are encapsulated as (centralized) service agents, different from the other agents, which might be homogeneous. Examples of such “weakly homogeneous” systems, mostly specialized on one KM task, are presented later in this section. An obvious extension to the personal information agents described in the previous section is to see each user not only as an information consumer, but also as a provider. In this case, besides retrieval and presentation support, the personal agent should assist the user in serving as a source of information. A very simple example for such agents are the clients for peer-to-peer file sharing support like Kazaa, ED2K, or – in the domain of learning resources – Edutella [57]. These agents have specialized interfaces for expressing queries, passing them on to other agents and displaying the results. But they 9

This definition identifies homogeneous multi–agent systems as close conceptual relatives of peer-to-peer (P2P) systems, even though their implementational basis can be quite different (cf. Section 2.1).

Towards Agent-Mediated Knowledge Management

15

are also able to receive queries and process them by answering with result documents or by passing a query to other agents. Such interaction between different personal assistants can be considered as a multi–agent system. In the following, more elaborate approaches are also described. MARS, an adaptive social network for information access described and evaluated in [86] has a purely homogeneous structure that is based on the idea described in the previous paragraph. Each agent basically has two competencies: i) to deliver some domain information with respect to a query, and ii) to refer to other agents that may fulfill a specific information need. Additionally, the agents learn assessments of the other agents in the network with respect to the two aspects. This means they assess the other agents’ expertise (ability to produce correct domain answers) as well as their ability to produce accurate referrals. DIAMS [20] is a system of distributed, collaborative information agents that help users access, collect, organize and exchange information on the World Wide Web. DIAMS aims at encouraging collaboration among users. Personal agents provide their owners with dynamic views on well–organized information collections, as well as with user–friendly information management utilities. These agents work closely together with each other and with other types of information agents such as matchmakers and knowledge experts to facilitate collaboration and communication. In order to promote easy information sharing and exchange, an object–based structure is used for the information repositories. DIAMS furthermore uses a flexible hierarchical presentation of information integrated with indexed query functionalities to ensure effective information access. Automatic indexing methods are employed to support translation between user queries and communication between agents. Collaboration between users is aided by the easy sharing of information and is facilitated by automated information exchange. Connections between users with similar interests can be established with the help of matchmaker agents. The focus of the research described in [62] is to add context–awareness to personal information agents that are (homogeneous) peers in a larger society of agents. The so-called CAPIAs (Context–Aware Personal Information Agents) have a model of their social and potential process context (e.g., the user’s schedule) as well as of their physical context (time and location). In the COMRIS Conference Center system the CAPIAs are employed for context–sensitive presentation of relevant information, e.g., whether “interesting” conference attendees or events (sessions, exhibition booths) are to be found nearby. Homogeneous multi–agent approaches in Knowledge Management seem to be a good way for leveraging single–agent approaches by taking advantage of the knowledge of other users in the organization. In the GroupLens project these leveraging effects are systematically investigated [41]. However, such systems are often not designed as agent systems. Due to their focus on one KM task (e.g., recommendation of one specific type of information objects) and a relatively controlled environment, centralized implementations are common. For example Let’s Browse [50], the successor of the personal information agent Letizia [48], does not model its collaborative web browsing as a cooperation between independent agents, but as one central agent that comprises the profiles of several users. An interesting but open question is to what extent multi–agent

16

Ludger van Elst, Virginia Dignum, and Andreas Abecker

modelling has an “added value” (e.g., wrt. user trust, privacy concerns, willingness to disclose information, ...) compared to a “functionally” (e.g., with respect to the quality of the recommendations) equivalent monolithic system. As with the single agent approaches presented above, homogeneous multi–agent systems applications to KM are mainly seen at the modelling level of development. On the other hand, in relation to the KM dimension, multi–agent approaches are mostly directed to the modelling of collaboration and interaction between users and systems, that is, with socialization issues. While most systems still lean considerably towards a product–oriented view of knowledge, these systems take a more process–oriented view on the management of knowledge than single agent approaches do, and can support teams and groups, as well as individual users. Homogeneous multi–agent approaches mostly provide a multiplication of a single–agent, and as such may not be able to support enough depth needed at the analysis and design level for comprehensive KM. Complex KM domains often require the combination of global and individual perspectives, and activities to follow desired structures, while enabling autonomous decisions on how to accomplish results. In order to cope with these requirements, heterogeneous approaches may be more appropriate, such as those described in the next subsection. 3.3 Heterogeneous Multi-agent and Society-Oriented Approaches Heterogeneous multi–agent systems not only consist of a potentially high number of agents, but these agents also belong to different classes. This means the agents have diverse competencies and types of goals. The heterogeneity can be due to the large number of “real–world” entities of the organization that are reflected in the system, or due to a purely functional decomposition from a software engineering point of view. Also, the more Knowledge Management functions a systems covers, the more heterogeneous the system will be. The systems we present in this section comprise both types of heterogeneity. Some of them only have a limited scope in terms of KM functionality (e.g., storing and retrieving knowledge objects), but encapsulate various service functions in separate specialized agents. Others are meant to be more comprehensive KM backbones and therefore employ agents for more diverse aspects like process support, retrieval support, and personalization. The society–oriented approaches we sketch at the end of this section demonstrate a potential way to cope with this heterogeneity and the complexity of such systems. The design of many agent–based Knowledge Management systems emerges from the “standard” three–tier enterprise information architectures that are often the basis for business applications (e.g., [55, 34, 45] and others): – The data layer manages repositories with knowledge objects such as documents, e-mail, etc. – The application layer realizes the business logic of the system. – The presentation layer organizes the interaction of the system with its users. KAoS [14, 19], a generic agent architecture for aerospace applications, is quite an early agent–based system for the management of technical information contained in documents, that is based on such a layer model. Aiming mainly at flexible information

Towards Agent-Mediated Knowledge Management

17

delivery from heterogeneous information sources in a distributed environment, KAoS employs agents on all three layers. In addition, a layer with generic service agents provides the middleware functionality of an agent platform (whitepage and matchmaking services for agents, proxies for connections to other agent domains, agent context management). The data services wrap the information sources by encapsulating indexing, search and retrieval functions, but also monitor them to allow for proactive information push. The prototype system Gaudi uses the KAoS platform for situation–specific, adaptive information delivery in the context of training and customer support in the airplane industry [13]. Recent versions of KAoS also incorporate social aspects in agent communities [34]. However, the relevance of this approach for Knowledge Management applications has not yet been discussed.

Fig. 3. Three-layer KM Architecture [45] (reprinted with kind permission)

The focus of KM systems based on a layer architecture like the one presented above is mostly the reuse of information contained in the information sources. Consequently, the knowledge flow is mainly from the data layer to the presentation layer. The conceptual model for Knowledge Management that Kerschberg presents with his Knowledge Rover architecture [44] does not have this principal restriction. He broadens the presentation layer to a Knowledge Presentation and Creation Layer, which also comprises discussion groups and other types of potential knowledge creating services [45] (cf. Figure 3). Hence, knowledge flow from the presentation to the data layer is also taken into account. Consequently, the application layer comprehensively embraces all basic KM

18

Ludger van Elst, Virginia Dignum, and Andreas Abecker

processes — acquisition, refinement, storage/retrieval, distribution, and presentation of knowledge (cf. Section 2.3). For a knowledge reuse–oriented view, the integration of information from various sources (cf. [80]) is essential. One project that deals explicitly with the fusion of knowledge from multiple, distributed and heterogeneous sources is KRAFT [63]. KRAFT has an agent–based architecture, in which all knowledge processing components are realized as software agents. The architecture uses constraints as a common knowledge interchange format, expressed in terms of a common ontology. Knowledge held in local sources can be translated into the common constraint language, fused with knowledge from other sources, and is then used to solve a specific problem, or to deliver some information to a user. The generic framework of the architecture can be reused across a wide range of knowledge domains and has been used in a network data services application as well as in prototype systems for advising students on university transfers, and for advising health care practitioners on drug therapies. The implementation of KRAFT is based on the FIPA standard with RDF as a content language. Sharing knowledge between people can take place directly, e.g., in face–to–face collaborations or with synchronous media like video conferencing, or indirectly, e.g., via information objects that are exchanged. Even hybrid approaches are possible, for example by analyzing the use of information objects and establishing direct links between people using the same objects. This direction was investigated in the Campiello project [46]. Campiello aims at using innovative information and communication technology to develop new links between local communities and visitors of historical cities of art and culture. The objectives of the project are to connect local inhabitants of historical places better, to make them active participants in the construction of cultural information and to support new and improved connections with cultural managers and tourists. The system includes a recommender module, a search module, and a shared data space. In order to facilitate the integration, tailoring and extensibility of these components, an agent model was chosen for the services in Campiello. The architecture supports interaction between distributed, heterogeneous agents and is built on top of the Voyager platform10 which was extended towards an agent platform by adding directory and broker services, administration tools and agent classes. In an organizational environment, one of the main context aspects is the business process a knowledge worker is involved in. Business process–oriented Knowledge Management (BPOKM, cf. [5]) considers these processes i) as knowledge objects themselves, ii) as knowledge creation context, iii) as trigger, when some knowledge objects may be relevant, and iv) as context what knowledge may be relevant. The EULE system [67] shows an integration of business process modeling and knowledge management. The system takes a micro–level view on business processes by modeling and supporting “office tasks” of a single worker by just–in–time information delivery, but does not coordinate complete workflows performed by groups of people. While EULE is not an explicitly agent–based system, in the FRODO framework for Distributed Organizational Memories [3] workflows themselves are first–order citizens in an agent–society for KM in distributed environments. An Organizational Memory in FRODO can be seen as a meta-information system with tight integration into enterprise business processes, 10

http://www.recursionsw.com/products/voyager/voyager.asp

Towards Agent-Mediated Knowledge Management

19

which relies on appropriate formal models and ontologies as a basis for common understanding and automatic processing capabilities [1]. Figure 4 shows FRODO’s four layer architecture for each Organizational Memory (OM): i) The application layer manages the process context in form of weakly–structured workflows [32]. ii) The source layer contains information sources with various levels of formalization (process models, text documents, etc.). iii) The knowledge description layer provides uniform access to the sources by means of ontologies. iv) By utilizing these descriptions, the knowledge access layer connects the application with the source layer. Agents in a FRODO OM reside on all four layers: – Workflow–related agents (task agents, workflow model manager, ...) are on the application layer and control the execution of business processes. – Personal User Agents are also on the application layer and provide the interface to the individual knowledge worker. – On the knowledge access layer, Info Agents and Context Providers realize retrieval and other information processing services to support the task and user agents. – The knowledge descriptions are handled by Domain Ontology Agents. Dedicated Distributed Domain Ontology Agents serve as bridges between several OMs. – Wrapper Agents and Document Analysis and Understanding Agents enable access to the sources and informal–formal transitions of information, and are thus located in the knowledge object layer or at the intersection between knowledge objects and knowledge descriptions, respectively. In order to cope with the heterogeneity and complexity, as well as to constrain the overall behavior of the system, agents in FRODO are organized in societies. Therefore, a FRODO agent is not only described by its knowledge, goals and competencies, but also by its rights and obligations. The description of ontology societies in [30] exemplifies FRODO’s concept of socially–enabled agents for KM. The implementation is based on the FIPA–compliant agent platform JADE11 . FRODO’s approach towards Distributed Organizational Memories is strongly driven by the general considerations of KM presented in Section 1. The overall goal is to find a balance between the organizational KM needs and the individual needs of knowledge workers. This is reflected in the way domain ontologies are handled in the distributed environment. Coming from a comparable analysis of KM characteristics [11], the Edamok project12 also aims at enabling autonomous and distributed management of knowledge. Edamok completely abandons centralized approaches, resulting in the peer–to–peer architecture KEx [10]. Each peer in KEx has the competence to create and organize the knowledge that is local to an individual or a group. Social structures between these peers are established that allow for knowledge exchange between them. In addition to the semantic coordination techniques that are required for this approach, the Edamok project also investigates contextual reasoning, natural language processing techniques and methodological aspects of distributed KM. An approach which is closely related to FRODO and Edamok has been developed in the CoMMA project [9]. The CoMMA architecture also employs societies of agents for personalized information delivery [38]: 11 12

http://sharon.cselt.it/projects/jade/ http://edamok.itc.it/

20

Ludger van Elst, Virginia Dignum, and Andreas Abecker

Fig. 4. FRODO Architecture for a Single Organizational Memory

– Agents in the ontology dedicated sub–society are concerned with the management of the ontological aspects of the information retrieval activity. – The annotation dedicated sub–society is in charge of storing and searching document annotations in a local repository and also of distributed query solving and annotation allocation. – The connection dedicated sub–society provides white page and yellow page services to the agents. – The user dedicated sub–society manages user profiles as well as the interface to the knowledge worker. The sub–societies in CoMMA can be organized hierarchically or peer–to–peer [39]. The position of an agent in a society is defined by its role [37]. The system was implemented on top of the JADE agent platform, and special attention was paid to the use of XML and RDF for representing document annotations and queries. As already stated above, business processes play an important role for providing context of knowledge generation and reuse. The utilization of the process context ranges from rather static access structures to knowledge objects (e.g., as a browsing hierarchy in a portal, or as an annotation that can be exploited by a search agent) to workflow–like agent–supported execution and the triggering of proactive information delivery. An interesting system that uses a concrete, domain–specific process model for its information support is K-InCA [71, 6]. In K-InCA, agents are used to guide, monitor and stimulate managers towards the understanding of KM concepts and the adoption of KM practices in organizational contexts, so that the system behaves as a personal

Towards Agent-Mediated Knowledge Management

21

KM coach for its users. The underlying process model of a K-InCA agent describes how changes are adopted by individuals and thereby new knowledge is incorporated into a person’s spectrum of working habits. K-InCA agents can be seen as experts on organizational behavior and change management, assisting users in the transition from their current working habits to new habits that integrate some new behavior (e.g. KM practices, entrepreneurial attitude, etc.). The system allows for different modes of interaction (practice and coaching), aiming at bringing the user to adopt a desired behavior. In order to achieve this goal, agents react to the current user activity on the basis of information stored in a domain model and a user model, as well as through interaction with other agents. With respect to the question of where in the development cycle the notion of agents is used (cf. Section 2.1), most of the systems presented up to now take a kind of middle– out approach: All of them have an agent–based description of the system’s components. This description is partly motivated by a functional decomposition from an IT point of view and partly a result of reflecting real–world entities (users, groups, etc.) in the system. Some of these architectures are then implemented using “conventional” software technology (e.g., most user interface agents), others build upon dedicated platforms for agent systems (e.g., based on the FIPA13 specifications). Only a few of the described systems complement their architectures with an agent–based Knowledge Management methodology for guiding the development of such a system in an organizational context (e.g., Edamok, stemming from the general MAS methodology Tropos and developing it towards KM).

Organizational model

Legend: role agent

Social model

Interaction model

structural interaction actual interaction (contract)

Fig. 5. Relations between the Different Models in OperA

13

http://www.fipa.org/

22

Ludger van Elst, Virginia Dignum, and Andreas Abecker

A recent proposal for a design methodology specifically tailored to agent societies is OperA [25]. This methodology is based on a three–tiered framework for agent societies that distinguishes between the specification of the intended organizational structure and the individual desires and behavior of the participating agents: 1. The organizational structure of the society, as intended by the organizational stakeholders, is described in the Organizational Model (OM). 2. The agent population of an OM is specified in the Social Model (SM) in terms of social contracts that make explicit the commitments which are regulating the enactment of roles by individual agents. 3. Finally, given an agent population for a society, the Interaction Model (IM) describes possible interaction between agents. After all models have been specified, the characteristics and requirements of the society can be incorporated in the implemented software agents themselves. Agents will thus contain enough information and capability to interact with others according to the society specification. Figure 5 depicts the relation between the different models. The OperA methodology supports the specification of an Organizational Model by analyzing a given domain and determining the type and structure of the agent society that best models that domain is described in [29]. The methodology provides generic facilitation and interaction frameworks for agent societies that implement the functionality derived from the co–ordination model applicable to the problem domain. Standard society types such as market, hierarchy and network, can be used as starting points for development and can be extended where needed and determine the basic norms and facilitation roles necessary for the society. These coordination models describe the different types of roles that can be identified in the society and issues such as communication forms, desired social order and co-operation possibilities between partners. The OperA methodology and framework have been applied to the design of Knowledge Market, an agent society to support peer–to–peer knowledge sharing in a Community of Practice; this has been designed in such a way that it preserves and recognizes individual ownership of knowledge and enables the specification and monitoring of reciprocity agreements [26]. 3.4 Description of Example Systems: Concluding Remarks In Section 2, we presented a framework for the description of agent–based Knowledge Management systems with the main dimensions system development level, macro–level structure, and KM application area. The analysis of several KM systems in Sections 3.1–3.3 shows that this space is not fully covered by the research approaches and prototypes presented (see also Table 1). Two factors may contribute to this fact: 1. Though at first glance, only the last dimension — the application area — seems to be KM specific, the dimensions are not really independent. If for example, knowledge use and internalization by specialized presentation techniques is the focus of research, an “agentification” of all knowledge sources may well be technological overkill. Or, the other way around, comprehensive KM frameworks may require more powerful agent architectures to cope with the complexity of various KM tasks.

Towards Agent-Mediated Knowledge Management

Macro–level Structure

Example application

Single–agent System

Homogeneous MAS

OntoBroker [77], MARS [86], DIAMS Jimminy [70], Remem- [20], GroupLens [41], brance Agent [70], CAPIA [62] MarginNotes [70], Watson [17], Letizia [48], Lets Browse [50]

23

Heterogeneous MAS

KAoS [34], Knowledge Rover [45], KRAFT [63], Campiello [46], FRODO [3], CoMMA [38], KEx [10], KInCA [71], OperA [25]

Organizational analysis (seldom)

System– development Level

Design (acting on be- Design (restricted no- Design (more comhalf of –metaphor) tion of agents) prehensive notion of agents: belief-desireintention architectures, speech acts) Implemented mostly Implemented on top Implementation with with conventional of middleware for dedicated agent plattechniques distributed systems forms and/or Semantic (Web, Peer–to–Peer) Web technology

KM Applica- Distribution and utiliza- Distribution, utilization tion Area tion of knowledge and preservation of knowledge Adequate presentation Presentation for interto ease internalization nalization, connecting people for socialization Mostly “knowledge as Product and (rudimenproduct” tal) process view

Often KM frameworks

Aiming at covering large areas of the knowledge cycle Product and/or process view

Table 1. Typical Operation Points within the Design Space of Agent–based KM Systems

2. Sparsely populated areas in the design space spanned by the description framework just may not yet be investigated by current research. While the first case covers operating points that simply make no sense for agent–based Knowledge Management, the second may lead towards new research aspects. We think that some papers in this book are well suited for stimulating thoughts in new directions.

24

Ludger van Elst, Virginia Dignum, and Andreas Abecker

4 Summary and Outlook The goal of this paper was twofold, i) to clarify the relationship between typical characteristics of Knowledge Management environments and core features of software agents as a basic technology to support KM, and ii) to provide a framework for the analysis and description of agent–based KM systems. In Section 1 we emphasized four main characteristics of Knowledge Management which in our opinion fundamentally account for the suitability of agent–based systems for supporting KM: – The distributed nature of knowledge may — from a technical point of view — raise special challenges, but for an organization and its individuals it is the only way to cope with the complexity of knowledge and should therefore be seen as an imperative and not as a nuisance. Agents are a natural form to represent that knowledge is created and used by various actors with diverse objectives. Socially– enabled agents can also help to tackle derived questions like accountability, trust, etc. – The inherent goal dichotomy between business processes and KM processes leads to the fact that knowledge workers typically do not adopt KM goals with a high priority. Proactive agents may be able to stand in for (or at least remind the knowledge worker) when KM tasks fall behind. – Knowledge work as well as KM in general is “wicked problem solving” without a fixed a-priori description of goals and solution paths. Reactive and proactive behavior of agents help to reach the necessary degree of flexibility. Social skills of agents can facilitate the management of the complexity of interactions that are typical for wicked problem solving. – The continuously changing environments are not entirely an intrinsic KM characteristic, but nevertheless any IT support for KM has to deal with this given factor. Agent approaches allow for extensibility and openness in situations where it is impossible to know at design time exactly which components and uses the system will have. In Section 2 we developed a framework for the description of agent–based KM systems with the main dimensions – system development level (analysis, design, implementation), – macro–level structure (single agent, heterogeneous, or homogeneous MAS), and – KM application area (knowledge distribution, generation, use, etc.). The synopsis of exemplary agent–based KM systems in Section 3 with respect to these dimensions showed how the design space is covered by today’s research approaches, prototypes and systems. Though most applications are not entirely agent–based from organizational analysis to system implementation, the potential of agent technology in all phases was demonstrated. On the other hand it is a fact that the vast majority of KM applications nowadays is not explicitly agent–based. Thus, there is still much work to be done in order to fathom the capability of agent technology for KM information systems.

Towards Agent-Mediated Knowledge Management

25

As the development of a comprehensive “Agents–in–KM Roadmap” is well beyond the scope of this paper, we just briefly sketch a couple of directions that may be interesting for future research: 1. Socio–technical: How can the teamwork of human knowledge workers and artificial agents (that might act “on behalf of” people) be balanced? Questions from human– computer interaction arise here, but also questions of trust, responsibility, etc. 2. Agent technology and KM functionality: What agent models and architectures are needed for what kind of KM application? Should concepts of trust, responsibility, rights, obligations be integrated in the models? How can the flexibility of reactivity and proactivity be better exploited for KM tasks? Which new functionalities can agent–based systems offer to KM? 3. Methodological and engineering aspects: Which functionalities can be provided as a kind of “KM middleware” or as modules for building KM applications? How should agent–orientation of design and implementation be reflected in an “agent– based KM methodology” in order to facilitate transitions between different phases in the development cycle? 4. Evaluation of agent–based KM: How well does the integration of (non agent– based) legacy systems into agent environments work in real–world applications (case studies)? How easily can new agent–based components really be integrated into an existing system? Which evaluation paradigms can be used to make different KM applications more comparable (agent–based vs. agent–based, but also agent– based vs. “traditional”)? At the moment it is hard to argue (and indeed not aimed at in this paper) that agent– based systems can do things that could not also be done using conventional technology, especially when only the implementation level is considered. However, we believe that agent technology helps building KM systems faster and more flexibly. We think that the results presented in this paper and in the other contributions in this book have the potential to strengthen the hope that an agent–oriented view (regardless of the implementation technology) leads to a more human–centered, more agile, and more scalable KM support.

References [1] A. Abecker, A. Bernardi, K. Hinkelmann, O. K¨uhn, and M. Sintek. Toward a technology for organizational memories. IEEE Intelligent Systems, 13(3):40–48, 1998. [2] A. Abecker, A. Bernardi, K. Hinkelmann, O. K¨uhn, and M. Sintek. Context–aware, proactive delivery of task–specific knowledge: The KnowMore project. International Journal on Information System Frontiers, 2(3/4):253–276, 2000. [3] A. Abecker, A. Bernardi, and L. van Elst. Agent technology for distributed organizational memories. In Proceedings of the 5th International Conference On Enterprise Information Systems, Vol. 2, pages 3–10, 2003. [4] A. Abecker and L. van Elst. Ontologies for knowledge management. In St. Staab and R. Studer, editors, Handbook on Ontologies in Information Systems, pages 453–474, Berlin, 2003. Springer Verlag. To appear.

26

Ludger van Elst, Virginia Dignum, and Andreas Abecker

[5] A. Abecker, K. Hinkelmann, H. Maus, and H. J. M¨uller, editors. Gesch¨aftsprozessorientiertes Wissensmanagement. Springer, 2002. [6] A. Angehrn, T. Nabeth, L. Razmerita, and C. Roda. K-inca: Using artificial agents for helping people to learn new behaviours. In Proceedings of the IEEE International Conference on Advanced Learning Technologies (ICALT 2001), pages 225–226, Madison USA, August 2001. IEEE Computer Society. [7] I. Becerra-Fernandez. The role of artificial intelligence technologies in the implementation of people-finder knowledge management systems. Knowledge Based Systems, 13(5):315– 320, 2000. [8] D. Beneventano, S. Bergamaschi, G. Gelati, F. Guerra, and M. Vincini. Miks: An agent framework supporting information access and integration. In M. Klusch, S. Bergamaschi, P. Edwards, and P. Petta, editors, Intelligent Information Agents - The AgentLink Perspective, volume 2586 of Lecture Notes in Computer Science, pages 22–49. Springer, 2003. [9] F. Bergenti, A. Poggi, and G. Rimassa. Agent architectures and interaction protocols for corporate management systems. In Proceedings of the ECAI Workshop on Knowledge Management and Organisational Memories, pages 39–47, Berlin, Germany, 2000. [10] M. Bonifacio, P. Bouquet, G. Mameli, and M. Nori. Kex: a peer–to–peer solution for distributed knowledge management. In Proceedings of the Fourth International Conference on Practical Aspects of Knowledge Management (PAKM02), volume 2569 of Lecture Notes in AI, pages 490–500. Springer, 2002. [11] M. Bonifacio, P. Bouquet, and P. Traverso. Enabling distributed knowledge management. managerial and technological implications. Novatica and Informatik/Informatique, 3(1):23–29, 2002. [12] U.M. Borghoff and R. Pareschi. Information Technology for Knowledge Management. Springer, Berlin, Heidelberg, New York, 1998. Some articles of this volume also appeared in a special issue of the Journal of Universal Computer Science, vol. 3, no. 8, Springer, 1997. [13] J. M. Bradshaw, S. Dutfield, P. Benoit, and J. D. Woolley. Kaos: Toward an industrialstrength generic agent architecture. In J. M. Bradshaw, editor, Software Agents, pages 375–418. AAAI Press/The MIT Press, 1997. [14] J. M. Bradshaw, S. Dutfield, B. Carpenter, R. Jeffers, and T. Robinson. KAoS: A generic agent architecture for aerospace applications. In T. Finin and J. Mayfield, editors, Proceedings of the CIKM ’95 Workshop on Intelligent Information Agents, Baltimore, Maryland, 1995. [15] S. Buckingham Shum. Representing hard-to-formalise, contextualised, multidisciplinary, organisational knowledge. In Proceedings of the AAAI Spring Symposium on Artificial Intelligence in Knowledge Management, Stanford University, Palo Alto, CA, 1997. AAAI Press. [16] J. Budzik and K. Hammond. Watson: Anticipating and contextualizing information needs. In Proceedings of the Sixty-second Annual Meeting of the American Society for Information Science, 1999. [17] J. Budzik, K. Hammond, and L. Birnbaum. Information access in context. Knowledge based systems, 14(1-2):37–53, 2001. [18] J. Budzik, M. Ruberry, H. Stein, and K. J. Hammond. Proactively searching the web: Can we agree on what’s relevant? In Demo and Poster Proceedings of ACM HyperText, 2002. [19] J. M. Bradshaw R. Carpenter, R. Cranfill, R. Jeffers, L. Poblete, T. Robinson, A. Sun, Y. Gawdiak, I. Bichindaritz, and K. Sullivan. Roles for agent technology in knowledge management: Examples from applications in aerospace and medicine. In B. R. Gaines and R. Uthurusamy, editors, Papers from the AAAI Spring Symposium on Artificial Intelligence in Knowledge Management, pages 9–16. AAAI Press, Menlo Park, 1997.

Towards Agent-Mediated Knowledge Management

27

[20] J. Chen, S. Wolfe, and S. D. Wragg. A distributed multi-agent system for collaborative information management and sharing. In Proceedings of the 9th ACM International Conference on Information and Knowledge Management (CIKM), pages 382–388. ACM Press, New York, NY, USA, 2000. [21] E.J. Conklin and W. Weil. Wicked problems: Naming the pain in organizations. White Paper of Group Decision Support Systems, Inc. [22] T.H. Davenport, S.L. Jarvenpaa, and M.C. Beers. Improving knowledge work processes. Sloan Management Review, 37(4):53–65, 1997. [23] R. Dieng and J. Vanwelkenhuysen, editors. Tenth Knowledge Acquisition for KnowledgeBased Systems Workshop KAW‘96, Special Track on Corporate memory and Enterprise Modeling. November 1996. [24] R. Dieng-Kuntz and N. Matta. Knowledge Management and Organizational Memories. Kluwer Academic Publishers, 2002. [25] V. Dignum. A Model for Organizational Interaction: Based on Agents, Founded in Logic. PhD thesis, Utrecht University, to appear in 2004. [26] V. Dignum and F. Dignum. The knowledge market: Agent–mediated knowledge sharing. In V. Marik, J. M¨uller, and M. Pechoucek, editors, Multi–Agent Systems and Applications III, volume 2691 of LNAI. Springer-Verlag. [27] V. Dignum and F. Dignum. Modelling agent societies: co-ordination frameworks and institutions. In P. Brazdil and A. Jorge, editors, Progress in Artificial Intelligence: 10th Portuguese Conference in Artificial Intelligence, Lecture Notes on Artificial Intelligence 2258, pages 191–204. Springer-Verlag, 2001. [28] V. Dignum and P. van Eeden. Seducing, Engaging, and Supporting Communities at Achmea. In Proceedings 4th European Conference on Knowledge Management, Oxford, UK, September 2003. [29] V. Dignum, H. Weigand, and L. Xu. Agent societies: Towards framework-based design. In M. Wooldridge, G. Weiss, and P. Ciancarini, editors, Agent-Oriented Software Engineering II, Lecture Notes on Artificial Intelligence 2222, pages 33–49. Springer-Verlag, 2001. [30] L. van Elst and A. Abecker. Ontologies for information management: balancing formality, stability, and sharing scope. Expert Systems with Applications, 23(4):357–366. [31] L. van Elst and A. Abecker. Domain ontology agents for distributed organizational memories. In [24], 2002. [32] L. van Elst, F.-R. Aschoff, A. Bernardi, H. Maus, and S. Schwarz. Weakly–structured workflows for knowledge–intensive tasks: An experimental evaluation. In IEEE WETICE Workshop on Knowledge Management for Distributed Agile Processes: Models, Techniques, and Infrastructure (KMDAP2003). IEEE Computer Press, to appear 2003. [33] M. Eppler. Glossary definition: Knowledge management. NetAcademy: www.knowledgemedia.org/modules/glossary/view\_def.php?id= mediamanagemen%t-20, 2002. [34] P. Feltovich, J. M. Bradshaw, R. Jeffers, and A. Uszok. Order and KAoS: Using policy to represent agent cultures. In Proceedings of the AAMAS 03 Workshop on Humans and Multi-Agent Systems, Melbourne, Australia, 2003. [35] S. Franklin and A. Graesser. Is it an Agent, or just a Program?: A Taxonomy for Autonomous Agents. In Intelligent Agents III. Agent Theories, Architectures and Languages (ATAL’96), volume 1193 of LNCS, Berlin, Germany, 1996. Springer-Verlag. [36] F. Gandon. Distributed Artificial Intelligence and Knowledge Management: ontologies and multi-agent systems for a corporate semantic web. PhD thesis, INRIA and University of Nice - SophiaAntipolis, 2002. [37] F. Gandon. A multi–agent architecture for distributed corporate memories. In Third International Symposium From Agent Theory to Agent Implementation at the 16th European

28

[38]

[39] [40]

[41]

[42] [43] [44]

[45]

[46]

[47] [48]

[49]

[50]

[51]

[52] [53]

Ludger van Elst, Virginia Dignum, and Andreas Abecker Meeting on Cybernetics and Systems Research (EMCSR 2002), pages 623–628, Vienna, Austria, April 2002. F. Gandon and R. Dieng–Kuntz. Distributed artificial intelligence for distributed corporate knowledge management. In M. Klusch, S. Ossowski, and O. Shehory, editors, Cooperative Information Agents VI : 6th International Workshop, CIA 2002, volume 2446 of LNCS, Berlin, Germany, 2002. Springer-Verlag. F. Gandon, R. Dieng, O. Corby, and A. Giboin. A multi–agent system to support exploiting an xml–based corporate memory. In [66], 2000. F. Giunchiglia, J. Mylopoulos, and A. Perini. The tropos software development methodology: processes, models and diagrams. In M. Gini, T. Ishida, C. Castelfranchi, and W.L. Johnson, editors, Proceedings of the First International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS’02), pages 35–36. ACM Press, July 2002. N. Good, J. B. Schafer, J. A. Konstan, A. Borchers, B. M. Sarwar, J. L. Herlocker, and J. Riedl. Combining collaborative filtering with personal agents for better recommendations. In Proceedings of the 1999 Conference of the American Association of Artifical Intelligence (AAAI-99), pages 439–446, 1999. G. van Heijst, R. van der Spek, and E. Kruizinga. Organizing corporate memories. In [23], 1996. G. van Heijst, R. van der Spek, and E. Kruizinga. The lessons learned cycle. In [12], 1998. L. Kerschberg. Knowledge rovers: Cooperative intelligent agent support for enterprise information architectures. In P. Kandzia and M. Klusch, editors, Proceedings ot the First International Workshop on Cooperative Information Agents, volume 1202 of LNAI, pages 79–100, Berlin, February 26–28 1997. Springer. L. Kerschberg. Knowledge management in heterogeneous data warehouse environments. In Y. Kambayashi, W. Winiwarter, and M. Arikawa, editors, Proceedings of the 3rd International Conference on Data Warehousing and Knowledge Discovery, volume 2114 of LNAI, pages 1–10. Springer, 2001. M. Koch. Knowledge management and knowledge agents in Campiello. In B. Lees, H. M¨uller, and C. Branki, editors, Proceedings of the Workshop on Intelligent Agents in CSCW, pages 44–52, Dortmund, Germany, 1998. O. K¨uhn and A. Abecker. Corporate memories for knowledge management in industrial practice: Prospects and challenges. In [12], 1998. H. Lieberman. Letizia: An agent that assists web browsing. In Chris S. Mellish, editor, Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence (IJCAI-95), pages 924–929, Montreal, Quebec, Canada, August 1995. Morgan Kaufmann publishers Inc.: San Mateo, CA, USA. H. Lieberman. Personal assistants for the web: An MIT perspective. In M. Klusch, editor, Intelligent Information Agents: Agent–Based Information Discovery and Management on the Internet, pages 279–292. Springer-Verlag, 1999. H. Lieberman, N. Van Dyke, and A. Vivacqua. Let’s browse: A collaborative web browsing agent. In Mark Maybury, editor, Proceedings of the 1999 International Conference on Intelligent User Interfaces (IUI-99), pages 65–70, N.Y., January 5–8 1999. ACM Press. M. Luck, P. McBurney, and C. Preist. Agent technology: Enabling next generation computing — a roadmap for agent–based computing, version 1.0, 2003. available at http://www.agentlink.org/roadmap/. A. Maedche, S. Staab, N. Stojanovic, R. Studer, and Y. Sure. SEAL – A framework for developing SEmantic Web PortALs. Lecture Notes in Computer Science, 2097:1–22, 2001. P. Maes. Agents that reduce work and information overload. Communications of the ACM, 37(7):30–40, July 1994.

Towards Agent-Mediated Knowledge Management

29

[54] G. Mentzas, D. Apostolou, R. Young, and A. Abecker. Knowledge Asset Management: Beyond the Process-centred and Product-centred Approaches. Advanced Information and Knowledge Processing. Springer-Verlag, 2002. [55] L. Moreau, N. Gibbens, D. De Roure, S. El-Beltagy, W. Hall, G. Hughes, D. Joyce, S. Kim, D. Michaelides, D. Millard, S. Reich, R. Tansley, and M. Weal. SoFAR with DIM Agents: An agent framework for distributed information management. In Jeffrey Bradshaw and Geoff Arnold, editors, Proceedings of the 5th International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM 2000), pages 369– 388, Manchester, UK, April 2000. The Practical Application Company Ltd. [56] T. Nabeth, A. Angehrn, and C. Roda. Towards personalised, socially aware and active knowledge management systems. In Proceedings of E-2002 e-Business and e-Work Annual Conference; Prague, Czech Republic, 2002. [57] W. Nejdl, B. Wolf, C. Qu, S. Decker, M. Sintek, A. Naeve, M. Nilsson, M. Palmer, and T. Risch. Edutella: A P2P networking infrastructure based on RDF. In Proceedings of WWW-2002. ACM Press, 2002. [58] I. Nonaka and H. Takeuchi. The Knowledge–Creating Company. Oxford University Press, 1995. [59] G. Papavassiliou, G. Mentzas, and A. Abecker. Integrating knowledge modelling in business process management. In Proceedings of the Xth European Conference on Information Systems (ECIS-2002), Gdansk, Poland, 2002. [60] H. Van Dyke Parunak. What can agents do in industry, and why? An overview of industrially-oriented R&D at CEC. Lecture Notes in Computer Science, 1435:1–18, 1998. [61] H. Van Dyke Parunak. Industrial and practical applications of DAI. In Gerhard Weiss, editor, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence, chapter 9, pages 377–421. The MIT Press, Cambridge, MA, USA, 1999. [62] E. Plaza and J.-L. Arcos. Context–aware personal information agents. In M. Klusch and F. Zambonelli, editors, Cooperative Information Agents V, Proceedings of the 5th International Workshop CIA 2001, number 2182 in Lecture Notes in Artificial Intelligence, pages 44–55. Springer-Verlag: Heidelberg, Germany, September 2001. [63] A. D. Preece, K. Hui, W. A. Gray, P. Marti, T. J. M. Bench-Capon, D. M. Jones, and Z. Cui. The KRAFT architecture for knowledge fusion and transformation. Knowledge Based Systems, 13(2-3):113–120, 2000. [64] G. Probst, S. Raub, and K. Romhardt. Wissen managen. Wie Unternehmen ihre wertvollste Ressource optimal nutzen. Th. Gabler, Wiesbaden, 1998. [65] U. Reimer. PAKM-98: Practical Aspects of Knowledge Management. Proc. of the Second Int. Conference. October 1998. [66] U. Reimer, editor. Proceedings of the Third International Conference on Practical Aspects of Knowledge Management, PAKM2000, Basel, Switzerland, 2000. http://SunSITE.Informatik.RWTH-Aachen.DE/Publications/CEUR-WS/Vol-34/. [67] U. Reimer, B. Novotny, and M. Staudt. Micro–modeling of business processes for just–in– time knowledge delivery. In R. Roy, editor, Industrial Knowledge Management: A Micro– Level Approach. Springer–Verlag, 2001. [68] U. Remus. Prozessorientiertes Wissensmanagement: Konzepte und Modellierung. PhD thesis, University Regensburg, 2002. [69] B. J. Rhodes. Just-In-Time Information Retrieval. PhD thesis, MIT Media Lab, May 2000. [70] B. J. Rhodes and P. Maes. Just-in-time information retrieval agents. IBM Systems Journal, 39(3/4):685–704, 2000. [71] C. Roda, A. Angehrn, and T. Nabeth. Matching competencies to enhance organisational knowledge sharing: An intelligent agent approach. In Proceedings of the 7th International Netties Conference, pages 931–937, Fribourg, Switzerland, September 2001.

30

Ludger van Elst, Virginia Dignum, and Andreas Abecker

[72] C. Roda, A. Angehrn, T. Nabeth, and L. Razmerita. Using conversational agents to support the adoption of knowledge sharing practices. Interacting with Computers, 15:57–89, 2003. [73] H.-P. Schnurr and S. Staab. A proactive inferencing agent for desk support. In S. Staab and D. O’Leary, editors, Proceedings of the AAAI Symposium on Bringing Knowledge to Business Processes, pages 44–52. AAAI Press, Menlo Park, 2000. [74] G. Schreiber, H. Akkermans, A. Anjewierden, R. de Hoog, N. Shadbolt, W. van de Velde, and B. Wielinga. Knowledge Engineering and Management: The CommonKADS Methodology. MIT Press, 1999. [75] S. Staab and A. Maedche. Knowledge portals: Ontologies at work. The AI Magazine, 22(2):63–75, 2000. [76] S. Staab, H. Schnurr, R. Studer, and Y. Sure. Knowledge processes and ontologies. IEEE Intelligent Systems, 16(1):26–34, 2001. [77] S. Staab and H.-P. Schnurr. Smart task support through proactive access to organizational memory. Knowledge–based Systems, 13(5):251–260, 2000. [78] G. Stefanidis, D. Karagiannis, and R. Woitsch. The PROMOTE Approach: Modelling Knowledge Management Processes to Describe Knowledge Management Systems. In Proc. of the third European Conference on Organizational Knowledge, Learning, and Capabilities (OKLC 02), Athens, Greece, April 2002. [79] J. V´azquez-Salceda and F. Dignum. Modelling Electronic Organizations. In Multi-Agent Systems and Applications III: 3rd. International/Central and Eastern European Conference on Multi-Agent Systems (CEEMAS’03), Lecture Notes on Artificial Intelligence 2691, pages 584–593. Springer-Verlag, 2003. [80] G. Wiederhold and M.R. Genesereth. The conceptual basis for mediation services. IEEE Expert, 12(5):38–47, 1997. [81] K. Wiig. Perspectives on introducing enterprise knowledge management. In [65], 1998. [82] M. Wooldridge. Intelligent agents. In Gerhard Weiss, editor, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence, chapter 1, pages 27–78. The MIT Press, Cambridge, MA, USA, 1999. [83] M. Wooldridge and N. R. Jennings. Intelligent Agents: Theory and Practice. The Knowledge Engineering Review, 10(2):115–152, 1995. [84] M. Wooldridge and N.R. Jennings. Pitfalls of agent-oriented development. In K.P. Sycara and M. Wooldridge, editors, Proceedings of the 2nd International Conference on Autonomous Agents (Agents’98), pages 385–391, New York, May 9–13, 1998. ACM Press. [85] M. Wooldridge, N.R. Jennings, and D. Kinny. The Gaia methodology for agent-oriented analysis and design. Autonomous Agents and Multi-Agent Systems, 3(3):285–312, 2000. [86] B. Yu, M. Venkatraman, and M.P. Singh. An adaptive social network for information access: Theoretical and experimental results. Journal of the Applied Artificial Intelligence, 17(1):21–38, 2003.

Peer-Mediated Distributed Knowledge Management Matteo Bonifacio1,2 , Paolo Bouquet1,2 , Gianluca Mameli2 , and Michele Nori2 1

Department of Information and Communication Technologies University of Trento – Trento, Italy [email protected],[email protected] 2 ITC-Irst – Povo, Trento, Italy {mameli,nori}@itc.it

Abstract. Distributed Knowledge Management is an approach to knowledge management based on the principle that the multiplicity (and heterogeneity) of perspectives within complex organizations should not be viewed as an obstacle to knowledge exploitation, but rather as an opportunity that can foster innovation and creativity. Despite a wide agreement on this principle, most current KM systems are based on the idea that all perspectival aspects of knowledge (including the process of its creation) should be eliminated in favor of an objective and general representation in a sort of corporate knowledge base. In this paper we criticize this approach, and propose a peer-to-peer architecture (called KEx), which implements a distributed approach to Knowledge Managament in a quite straightforward way: (i) each peer (called a K-peer) provides all the services needed to create and organize “local” knowledge from an individual’s or a group’s perspective, and (ii) social structures and protocols of meaning negotiation are introduced to achieve semantic coordination among autonomous peers (e.g., when searching documents from other K-peers). A ﬁrst version of the system, called KEx, is implemented as a knowledge exchange level on top of JXTA.

1

Introduction

Distributed Knowledge Management (DKM), as described in [10], is an approach to knowledge management (KM) based on the principle that the multiplicity (and heterogeneity) of perspectives within complex organizations should not be viewed as an obstacle to knowledge exploitation, but rather as an opportunity that can foster innovation and creativity. The fact that diﬀerent individuals and communities may have very diﬀerent perspectives, and that these perspectives aﬀect their representation of the world (and therefore of their work) is widely discussed – and generally accepted – in theoretical research on the nature of knowledge. Knowledge representation in artiﬁcial intelligence and cognitive science have produced many theoretical and experimental evidences of the fact that what people know is not a mere collection of facts, as any “fact” always presupposes some (typically implicit) interpretation schema, which provide an essential element of sense-making (see, for L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 31–47, 2003. c Springer-Verlag Berlin Heidelberg 2003

32

Matteo Bonifacio et al.

example, the notions of context [25,20,2], mental space [19], partitioned representation [15]); studies on the social construction of knowledge stress the social nature of interpretation schemas, viewed as the outcome of a special kind of “agreement” within a community of knowing (see, for example, the notions of scientiﬁc paradigm [23], frame [22]), thought world [17], perspective [5]). Despite this large convergence, it can be observed that the high level architecture of most current KM systems in fact does not reﬂect this vision of knowledge (see [9,10,7] for a detailed discussion of this claim). The fact is that most KM systems embody the assumption that, to share and exploit knowledge, it is necessary to implement a process of “knowledge extraction and reﬁnement”, whose aim is to eliminate all subjective and contextual aspects of knowledge, and create an objective and general representation that can then be reused by other people in a variety of situations. Very often, this process is ﬁnalized to build a central knowledge base, where knowledge can be accessed via a knowledge portal. In our opinion, this centralized approach – and its underlying objectivist epistemology – is one of the reasons why so often KM systems are deserted by users. In this paper we describe a peer-to-peer (P2P) architecture, called KEx, which is coherent with the vision of DKM. Indeed, P2P systems seem particularly suitable to implement a DKM system. In KEx, each community is represented by a knowledge peer (K–peer), and a DKM system is implemented in a quite straightforward way: (i) each K–peer provides all the services needed by a knowledge node to create and organize its own local knowledge, and (ii) social structures and protocols of meaning negotiation are introduced to achieve semantic coordination (e.g., when searching documents from other peers). A ﬁrst version of KEx has been implemented on top of JXTA, a P2P open source project started in 2001 and supported by Sun (see http://www.jxta.org/). The paper goes as follows: ﬁrst, we brieﬂy discuss the centralized vs. distributed paradigm in KM; second, we describe the main features of KEx, a peerto-peer system for knowledge discovery and exchange, and argue why it provides a suitable system for distributed KM; then we describe the implementation of KEx; ﬁnally, we draw some conclusions and future work.

2

Social and Technological Architectures for KM

The starting point of our analysis is the wide agreement – in the organizational and sociological literature – on the fact that the creation, codiﬁcation, and sharing of knowledge within complex organizations is a process that can be described along two qualitatively diﬀerent dimensions: – on the one hand, knowledge is developed within communities, namely groups of people that share a common perspective (e.g., because they have a common goal, a common education, a common culture). This process, called perspective taking in [5], corresponds to the incremental development of knowledge within a community, an idea closely related to the notion of normal science within a paradigm proposed by the philosopher T. Kuhn with respect to the development of scientiﬁc theories [23];

Peer-Mediated Distributed Knowledge Management

33

– on the other hand, knowledge is developed as a consequence of the interaction between diﬀerent communities. This process, called perspective taking in [5], corresponds to a discontinuity in a community’s development (in science, Kuhn would call it a scientiﬁc revolution). It is not as common (and is deﬁnitely harder) of the ﬁrst one, as it requires the ability of “mapping” the point of view of another community into another community’s perspective, an operation that presupposes the cognitive ability of “transcending” a local perspective and making explicit the assumptions that, within a community, all take for granted. An important assumption underlying our work, which we share with the structurationist approach [27] in organization sciences, is that technology and organization are tightly interrelated dimensions, which need to be reciprocally coherent. The more an organizational process involves high level human activities, the stronger the interdependence between technological and organizational dimensions is. In particular, since each approach to cognition makes speciﬁc assumption on the role of communication, a technology that strongly structures social communication implies a particular model on how cognition occurs [6]. From this point of view, we suggest that the main problem of most current KM systems is that they do not support the two social (“pre–technological”) knowledge processes described above, but rather tend to impose a process of a very diﬀerent nature, namely a process whereby people: – generate knowledge through peripheral socialization in communities of practices [33,14]: through work practice, employees generate implicit knowledge in terms of working solutions that can be fruitfully made explicit and thus reusable; – contribute with their knowledge through a codiﬁcation process: knowledge is categorized and validated by experts according to a corporate language; – retrieve knowledge using a uniﬁed access to the organizational memory: through the use of manuals, procedures, routines, or the access to formal training, people have access to corporate knowledge. Technological architectures are then designed in accordance with this view of organizational cognition. The result is a semantically centralized KM architectures which aims at: – creating and enabling communication within formal and informal groups and communities (e.g., through “virtual communities” and groupware applications, which allow individuals to interact and produce their “raw” peripheral knowledge); – collecting “raw” peripheral knowledge through participation. Workers can contribute to create and feed knowledge using automatic document management tools, clustering, text mining, and information retrieval applications to explicit and collect knowledge; – categorizing and storing knowledge in databases and repositories according to a common and shared system of meaning, this way distilling knowledge that is useful for the entire organization;

34

Matteo Bonifacio et al.

– designing a corporate system of meanings in the form of a common language, an ontology, a knowledge map, a categorization, or a classiﬁcation system that is necessary to codify knowledge according to a shared interpretative schema; – creating an Enterprise Knowledge Portal (EKP) that provides a unique, standard access point to corporate knowledge for the members of diﬀerent organizational units. Typically, people access the KB through various forms of personalization tools (e.g., individual or group proﬁles, views, chats, and so on). Through the analysis of a paradigmatic case study (a worldwide consulting ﬁrm), [9] shows that centralized KM systems are often deserted by end-users; indeed, as Bowker and Star argue in [13], any approach which disregards the plurality of interpretative schemas is perceived either as irrelevant (there is no deep understanding of the adopted and centralized schema), or as oppressive (there is no agreement on the unique schema, which is therefore rejected). Recently, diﬀerent groups of researcher are starting to realize that we need technological architectures that are more coherent with the social model of organizational cognition. In [9,8], a distributed approach is proposed, in which organizational cognition is viewed as a distributed process that balances the autonomous knowledge management of individual and groups, and the coordination needed to exchange knowledge across diﬀerent autonomous entities; from this perspective, technology is viewed as a way enabling distributed control, differentiation, customization, and redundancy. In such a vision, technology should mainly support the autonomous creation and organization of knowledge locally produced by individuals and groups and, on the other hand, support coordination processes among autonomous entities, in order to exchange and share knowledge. In particular this means: – giving each community the possibility to represent and organize knowledge according to its goals and interpretative perspective. The building blocks of the system are the so-called knowledge nodes [7] (KNs), namely the organizational units - either formal (e.g. divisions, market sectors) or informal (e.g. interest groups, communities of practices, communities of knowing) - which exhibit some degree of semantic autonomy3 ; – providing tools to support the exchange of knowledge across diﬀerent KNs without assuming shared meanings, but rather enabling the dynamic translation of diﬀerent meanings. The KNs are thus materialized by local technologies that represent a semantically autonomous expression of local knowledge owned by an individual or a group; – setting mechanisms and protocols to enable the emergent and bottom-up formation of informal communities and communication practices (such as ﬁnding or addressing people to trusted individuals/communities). Here a DKM 3

Semantic autonomy means the ability to develop autonomous interpretative schemas (perspectives on the world) to interpret, organize, and store useful information.

Peer-Mediated Distributed Knowledge Management

35

system supports the formation of groups and knowledge discovery/propagation through social cooperation. In the following section we show how this architecture has been implemented in a P2P system, which provides a peer-mediated support to a distributed approach to designing KM systems. In the conclusions, we will make some remarks on the relation between P2P and agent–mediated knowledge management, and why we went for the ﬁrst.

3

KEx: A P2P Architecture for DKM

KEx is a P2P system which allows each KN (be it an individual or a community) to build its own knowledge space within a network of autonomous K–peers, to make knowledge in this space available to other K–peers, and to search relevant knowledge in the knowledge space of other K–peers. We stress the fact that a K–peer may contain not only a structured collection of documents, but also relational knowledge, such as references to experts in some domain, links to other K–peers, to external resources, and so on. In the following sections, we describe the high-level architecture of KEx, and explain the role that each element plays in a DKM perspective. 3.1

K–peers

K–peers are the building blocks of KEx. From an organizational perspective, each K–peer represents a Knowledge Node [7], namely the reiﬁcation of an organizational unit – either formal (e.g. divisions, market sectors) or informal (e.g. interest groups, communities of practices, communities of knowing) – which exhibits some degree of semantic autonomy, namely the ability to develop autonomous interpretative schemas (perspectives on the world) to interpret, organize, and store useful information. In KEx, each K–peer can play two main roles: provider and seeker. A K– peer acts as a provider when it “publishes” in the system a body of knowledge, together with an explicit semantic view on it (called a context, in the sense deﬁned in [11]); a K–peer acts as a seeker when it searches for information that matches some part of its context. Each K–peer has the structure shown in Figure 1. Below we illustrate the main modules. Document Repository The Document Repository is the place where the data of a knowledge node are stored. In general, it can be viewed as a private space in which document and other data are organized according to some local semantic schema (e.g., a directory structure, or a database schema) and managed by some local application (e.g., a DBMS, a HTTP server, a ﬁle system, a document management system).

36

Matteo Bonifacio et al.

Fig. 1. KEx main components

Context Repository A context is an explicit semantic schema over a body of local knowledge. More that one context can be used to classify local knowledge. The contexts in use in a K–peer are stored in a context repository. We observe that local knowledge deﬁnitely includes documents from the document repository, but it may also include links to other resources and mappings to contexts stored in the context repository of other KNs. To make contexts usable in KEx, we use a web-oriented syntax for them, called CTXML [11]. CTXML provides an XML–Schema speciﬁcation of a context; currently, contexts are concept hierarchies, whose nodes are labelled with words and phrases from natural language, arcs are Is-A, Part-Of or generic relations between nodes. From an organizational point of view, a context is the manifestation of a KN’s semantic autonomy. Even though a context can be a newly deﬁned schema, in typical situations a context is a “translation” in CTXML of the local application schemas. For example, a context can be the representation of a user’s ﬁle system (where directory names are used as concept names and the sub–directory structure is used as the structure of the concept hierarchy); or a context can be the representation in CTXML of the taxonomy of a document management system (where the taxonomy is used as a structure, and relations are Is-A relations). From the standpoint of DKM, contexts are relevant in two distinct senses: – on the one hand, they have an important role within each KN, as they provide a dynamic and incremental explicitation of its semantic perspective. Once

Peer-Mediated Distributed Knowledge Management

37

contexts are reiﬁed, they become cognitive artifacts that contribute to the process of perspective making [5], namely the consolidation of a shared view in a KN, continuously subject to revision and internal negotiation among its members; – on the other hand, contexts oﬀer a simple and direct way for a KN to make public its perspective(s) on the information that it can provide. Therefore, as we will see, contexts are an essential tool for semantic coordination/negotiation among diﬀerent KNs. Context Management Module. The context management module allows a users to create, manipulate, and use contexts. This module has two main components: – Context editor: the context editor provides users with a simple interface to create and edit contexts, and to associate documents and other information with respect to a context. This happens by allowing users to create links from a resource (identiﬁed by a URI) to a node in a context. Examples of resources are: documents in local directories, the address of a database access services, addresses of other K–peers that provide information that a KN wants to explicitly classify in its own context. – Context Normalization and Enrichment: this module provides two important services for achieving semantic coordination among K–peers. The ﬁrst, called normalization, uses NL techniques (e.g., deleting stop words, tokenizing, tagging part-of-speech, etc.) on user deﬁned contexts; the second, called enrichment, provides an interface with an external linguistic resource (in the current version of the system we use WordNet) or with an ontology to add semantic information to concept labels (for example, that in a given context “apple” means a fruit and not a computer brand). Both steps are described in detail in [24]. This notion of enrichment is not equivalent to introduce a shared (universal) semantics in KEx. Indeed, the intuition is that the meaning of a concept label in a context has two components: – the ﬁrst is the linguistic component, which means that the words or phrases used as concept labels have a standard meaning (or, better, a set of meanings) in a “dictionary”. This helps, for example, to distinguish between “apple” as a fruit and “apple” as a tree; – the second is a sort of pragmatic component, which is given by its position in a context (e.g., in a concept hierarchy in CTXML). This helps in understanding what the user means on a particular occasion with a word (e.g., “apple” in a path like “computer/software/apple” is diﬀerent from “apple” in a path like “computer/hardware/printers/apple”, even though “apple” has the same dictionary meaning’). The ﬁrst component is public, namely is shared among those who speak a given language. The second is highly contextual, and cannot be computed a priori, namely independently from the context in which it appears. In this sense,

38

Matteo Bonifacio et al.

contexts are not to be thought of as an alternative to the use of ontologies in a KM system, but as a necessary integration4 . Indeed, a typical context does not provide semantic information on the terms used in a concept hierarchy nor on their relations, but only on how these terms are combined to create more complex concepts in a classiﬁcation schema. Consider again the path “computer/hardware/printers/apple”. Even if we can decide that “apple” is used in the sense of a computer brand, this path does not provide a deﬁnition of what a computer brand is, whereas this deﬁnition could be found in a general ontology. However, it is clear that understanding what is the concept associated to such a path requires to use a lot of ontological knowledge about computers, hardware, printers, and their relation in the domain of computers (e.g., that printers are a kind of hardware, that there are diﬀerent printer makers, that Apple is one of them, and so on). Indeed, in the context matching algorithm we developed for coordinating K–peers (see below), both ontological and contextual information are used to ﬁnd relations over concepts in diﬀerent contexts. It is also important to notice that diﬀerent linguistic resources or ontologies can be used to enrich a context. So far, we’ve been using WordNet, but there’s no reason why other resources can’t be used to replace WordNet. From the standpoint of KM, this is an interesting feature, as we can imagine that diﬀerent communities may decide to adopt a diﬀerent linguistic resource or ontologies to enrich their contexts, namely those which better suit their needs. This fact has a signiﬁcant impact on the mechanisms for sharing knowledge across K–peers, as we will discuss later in this paper.

4

Roles of K–peers in KEx

Each K–peer can act as a provider, as a seeker, or can play simultaneously both roles. The main components of the two roles are described in Figure 1; the interaction between seekers and providers is described in Figure 2. The following sections explain in details the components needed for implementing the two roles, and their interaction. 4.1

Seeker

The seeker module of KEx allows users to search and retrieve useful knowledge from other K–peers. The main module of the seeker component is the query maker. A query is built as follows: – the user opens a context from the context repository; – the user browses the context until a concept (i.e., a node in the context) is found which describes the topic of interest for that query; – the user clicks on that concept, this way telling the system that she is interested in ﬁnding documents about that concept; 4

See [12] for a preliminary investigation of how this integration can be done.

Peer-Mediated Distributed Knowledge Management

39

– the system extracts a focus, namely the portion of the selected context which contains relevant information about the selected concept (in the current version of the query maker, the focus is the entire path from the concept to the root of the concept hierarchy, (see [24] for a formal deﬁnition of focus); – if needed, the user can add one or more keywords to the query. When the user submits the query, the seeker activates a session that is associated to that query and is in charge of resolving it (step 1 in Figure 2). The query is propagated to the P2P system (see below for details). The active session can receive asynchronously several incoming replies from those providers that have been selected/suggested by other peers, and collects results that are composed by the aggregation of all those that have been received; each result is made up of a list of document descriptors (name of the document, short description, and so on) and the indication of the part of the providers context that has been used in order to interpret the meaning of the query and provide a resolution. Finally the seeker allows the user to access the K–Peer downloading service; if the user ﬁnds in the result set one or more interesting documents, she can contact the providing K-Peer to download it. 4.2

Provider

The provider contains the functionalities needed to accept and resolve a query, and to identify the results that must be returned to the seeker. When a K–peer receives a query (keywords and focus), it instantiates a provider, conﬁgured to use a set of contexts and some documents (a particular directory), and to resolve the query. The main modules needed for the provider role are the following: – Query Solver: the query solver takes a contextual query as an input and returns a list of results to be sent back to the seeker. A contextual query is resolved: • by the Semantic Query Solver, if a focus is associated to the query itself; • by the Lexical Query Solver, if a list of keywords is associated to the query. If the query contains both a focus and a list of keywords, then both solvers are invoked and the result set is the intersection of the two result sets. – Query propagation: this module is in charge of propagating a query to other K–peers (see K–services below for the two modes of propagation allowed in KEx). The Semantic Query Solver invokes a context matching algorithm, namely an algorithm that ﬁnds relations between concepts belonging to diﬀerent contexts. The details of the algorithm are described in [30]. Here we only say that the algorithm is based on two main ideas: – that the information provided in a context presupposes a lot of implicit knowledge, which can be extracted automatically from an external resource

40

Matteo Bonifacio et al.

(in our case, WordNet). In the current version of the algorithm, the result of this extraction, combined with the explicit information of the context, is represented as a logical formula associated to each node in the concept hierarchy of a context; – that the problem of discovering what the relationship is between concepts of diﬀerent contexts can be codiﬁed as a problem of propositional satisﬁability of a set of formulae, namely the formulae associated to the concepts to be matched plus a set of “axioms” obtained by extracting relevant facts from the external resource. Given two concepts in two diﬀerent contexts, the current version of the algorithm we implemented returns one of the following relations as its output: (i) the two concepts are equivalent, (ii) the ﬁrst is more general than the second, (iii) the ﬁrst is less general than the second, (iv) the two concepts are disjoint. In KEx, if one of the ﬁrst three relations holds, then the URIs of the resources associated to the concept on the provider side are returned to the seeker, together with the focus of the concept in the provider’s context. This is important, as users on the seeker side may have the opportunity to learn how users on the provider side classify a document in their context. It is important to observe that the semantic match is performed on the provider side. This means that the result reﬂects the provider’s interpretation of the seeker’s query. This explains why we can match contexts normalized and enriched with diﬀerent linguistic resources or ontologies. The intuition is that, in this case, the provider will normalize and enrich the query’s focus using its own resource, this way assigning to it a meaning from the perspective of its users. Of course, the seeker’s users can disagree with the resulting match (if any). However, this is not dissimilar from what happens among human agents, as it may happen that an hearer’s answer is completely incompatible with the speaker’s intended meaning for the question. The interaction between seeker and provider is depicted in Figure 2. The seeker sends a query to a provider (step 1). When a provider receives a query, it starts a query resolution session and selects relevant documents (step 2). The provider sends back to the seeker the result set (step 3). A provider can propagate a query to other providers that, from its perspective, are “experts” about the query’s topic (step 4). Each provider to which the query is propagated activates a query resolution session, and sends the results to the seeker. 4.3

K–services

KEx provides a collection of knowledge related services that have an important role in supporting knowledge exchange in a network of autonomous K–peers. The more important among them are described in the following sections. K–federations. KEx provides a federation management service. A K–federation is a group of K-Peers that agree to behave like a unique entity when other K– peers perform a search. In other words, each K–federation is a “social” aggregation of K–peers that display some synergy in terms of content (e.g., as they

Peer-Mediated Distributed Knowledge Management

41

Fig. 2. The KEx system: interaction between Seeker and Provider roles

provide topic-related content or decided to use the same linguistic resource to create a common “vocabulary”, thus providing more homogeneous and speciﬁc answers), quality (certify content) or access policies (certify members). In this sense, the addition of federations to KEx is not trivial, as they embody a ﬁrst (simple) form of semantic-driven aggregation. To become a member of a K–federation, a K-Peer must provide a K–federation Service (quite similar to that required by the Provider role) that implements the required federation protocol (reply to queries sent to the K–federation) and observes the federation membership policy. Each K–peer can be member of more than one K–federation. Currently, we do not make any assumption on how K–federations are formed (see, for example, [29,16] for two diﬀerent methods of automatic and dynamic formation of communities). However, we anticipate two principal methods of constitution: Top–down: a group of K–peers decide to create a federation for organizational, commercial, or strategic reasons. Depending on the type of agreement, a new K–federation can be created and membership policies can be deﬁned; Bottom-up: there are tools that observe the interactions among K–peers and detects the emergence of synergies among K-peers. In a corporation, this may suggest the existence of a new (informal) community, and lead to the creation of a K–federation to support it. From a technological point of view, a K–federation becomes a new possible target for a seeker’s queries. And indeed, as we can see from Figure 1, the

42

Matteo Bonifacio et al.

federation module uses the provider module, the main addition being that queries are forwarded to all members of the federation. Currently, the result of sending a query to a K–federation is the very similar to the result of sending the query directly to each member of the federation. Two are the main diﬀerences: on the one hand, each K–peer can select which of its contexts and documents can be used to answer a query that comes from each federation to which it belongs; on the other hand, K-peers that reply to a query notify seekers that they replied as members of that K–federation. In the future, we plan to implement smarter policies for handling queries within K–federations, like a query pre-processing or a selection of the members to which the query should be forwarded. Discovery. Discovery is a mechanism that allows users to discover resources in the P2P network. Users need to discover K–peers or K–federations available in the network as potential targets of a query. Each K–peer may advertise the existence of a resource by publishing an XML document (advertisement). In KEx, two type resources are currently advertised: – K–peers that have a provider service to solve queries. The main elements of the advertisement are: a description of the peers contexts, and an address to contact the K–peer in order to send it a query or retrieve documents; – K–federations, namely groups of peers that have a federation service to solve queries. The main elements of the advertisement are: the federation domain description, contact information, membership policy. To discover resources in a P2P network, K–peers can send a discovery request to an already known K–peer, or send a multi-cast request on the network, and receive responses (list of advertisements) that describe the available services and resources. It is possible to specify search criteria (such as a keyword or textual expression) that are then matched against the contents provided by the advertisement related to each K–peer or K–federation description. Query Propagation. When a provider receives a query, it can decide to forward it to another Provider that is considered “expert” about the query’s topic. To decide to which peers the query is to be forwarded, a peer has two possibilities: – physical “neighborhood”: the query will be sent to peers known through the discovery functionality. This way, providers that are not directly reachable by the Seeker, or have just joined the system, can advertise their presence and contribute to the resolution of queries; – semantic “neighborhood”: if the provider computes some matching between a query and a concept in its own context, the system will look for addresses of other K–peers that are linked to that concept. If some are found, then the query is propagated to them, based on the assumption that K–peers classiﬁed under a concept may possess relevant information about it.

Peer-Mediated Distributed Knowledge Management

43

Obviously, there are several parameters and mechanism controlling the scope of the search and prevent a message “ﬂooding”: setting a time to live (TTL), limiting the number of hops, storing in the query the name of peers that already received the query, and so on. Learning. When the matching algorithm ﬁnds a semantic correspondence between concepts of diﬀerent contexts, the Provider can store this information for future reuse. This information is represented as a semantic “mapping” between concepts (see [11]), and can be used in three ways: – when the K–peer receives a query from a seeker, it can reuse the corresponding stored mapping to avoid running the matching algorithm; – a provider can use the existing mapping to forward the query to other peers that present a semantic relation with the topic of the query (see semantic propagation above); – the seeker can search into the mapping relations in order to suggest the user a set of providers with which it had past interactions and are classiﬁed as qualiﬁed with respect to the meaning of the concept selected in a query. Using this mechanisms, the K–Peer network will deﬁne and increase the number and quality of the semantic relations among its members, so that it becomes a dynamic web of knowledge links.

5

Development Framework

KEx is built on top of JXTA, a set of open, generalized peer-to-peer protocols that allow devices to communicate and collaborate through a connecting network. This P2P framework provides also a set of protocols and functionality as a decentralized discovery system, an asynchronous point-to-point messaging system, and a group membership protocol. A peer is a software component that runs some or all the JXTA protocols; every peer has to agree upon a common set of rules to publish, share and access “resources” (like services, data or applications), and communicate among each others. Thus, a JXTA peer is used to support higher level processes (based, for example, on organizational considerations) that are built on top of the basic peer-to-peer network infrastructure; they may include the enhancement of basic JXTA protocols (e.g. discovery) as well as user-written applications. JXTA tackles these requirements with a number of mechanisms and protocols: for instance the publishing and discovery mechanisms, together with a message-based communication infrastructure (called “pipe”) and peer monitoring services, supports decentralization and dynamism. Security is supported by a membership service (which authenticates any peer applying to a peer group) and an access protocol (for authorization control). The ﬂexibility of this framework allows to design distributed systems that cover all the requirements of a DKM application, using the JXTA P2P capabilities, completed and enhanced through the implementation of user-deﬁned

44

Matteo Bonifacio et al.

services. As shows in the previous sections, in the Kex system we combine the P2P paradigm (characterizing a KN network as a network of distributed peers) and JXTA as an implementation infrastructure in a coherent vision with the DKM paradigm. These features of a peer-to-peer system seem to match the spirit and the main non-functional aspects of a KN in a DKM application, and suggest a P2P systems as a natural architectural solution (see [3] for a discussion of this idea). In particular: – autonomy is guaranteed by the fact that each KN can be seen as a peer which owns local knowledge, stored and organized through local technologies and applications; – coordination is guaranteed by enabling peers to collaborate with each other, using a set of dynamic and heterogeneous services that peers provides to each other, in order to support both communication features (as the discovery functionality) and semantic services (e.g. exchange information without imposing a common interpretation schema, but through a meaning negotiation service that automatically maps concepts among diﬀerent systems of meanings).

6

Conclusions and Research Issues

In this paper, we argued that technological architectures, when dealing with processes in which human communication is strongly involved, must be consistent with the social architecture of the process itself. In particular, in the domain of KM, technology must implement a principle of distribution that is intrinsic to the nature of organizational cognition. This distributed approach is becoming more generally accepted, and other groups are working in the direction of building distributed organizational memories (see e.g. [18,32]). Here we also suggest that P2P infrastructures are especially suitable for distributed KM applications, as they naturally implement the principles of autonomy and distribution. it is interesting to observe that also other research areas are moving toward P2P architectures. In particular, we can mention the work on P2P approaches to the semantic web [1,31], to databases [21], to web services [28]. We believe this is a general trend, and that in the near future P2P infrastructure will become more and more interesting for all areas where we can’t assume a centralized control. A number of research issues need to be addressed to map aspects of distributed cognition into technological requirements. Here we propose two of them: – social discovery and propagation: in order to ﬁnd knowledge, people need to discover who is reachable and available to answer a request. On the one hand, broadcasting messages generates communication overﬂow, on the other hand talking just to physically available neighbors reduces the potential of a distributed network. A third option could be for a seeker to ask his neighbors who they trust on a topic and, among them, who is

Peer-Mediated Distributed Knowledge Management

45

currently available. Here the question is about social mechanisms through which people ﬁnd – based on trust and recommendation – other people to involve in a conversation. A similar approach could be used in order to support the propagation of information requests; – building communities: if we consider communities as networks of people that, to some extent, tend to share a common perspective [5], mechanisms are needed to support the bottom-up emergence of semantic similarities across interacting KNs. Through this process, people can discover and form virtual communities, and within organizations, managers might monitor the evolving trajectories of informal cognitive networks. Then, such networks, can be viewed as potential neighborhoods to support social discovery and propagation. To this end, not only techniques based on the explicitation of semantic schemas can be used (e.g., contexts in KEx), but also techniques based on the observation of what users do; a possible extension in this direction in under development using the notion of implicit culture described in [4]. A ﬁnal remark concerns the relation between agent–based and P2P platforms for distributed KM. We believe that P2P infrastructures are a very straightforward way of mapping social architectures onto technological architectures, this way guaranteeing the coherence between the two structures. From a conceptual point of view, peers are much simpler than agents, and do not allow to exploit the potential of reasoning tools, coordination and collaboration, planning, that agents can provide. However, we need to be very careful in introducing software which can have a signiﬁcant impact on the way people work and manage their own and corporate knowledge. In other words, we need conceptual tools for designing agent-based applications which are coherent with the social architecture. An interesting attempt to provide such a methodology can be found in [26], where intentional modeling techniques are used to analyze organizations as sets of actors that cooperate (or compete) to achieve their goals. This analysis is applied in particular to applications of distributed KM.

7

Acknowledgments

This work is part of EDAMOK (Enabling Distributed and Autonomous Management of Knowledge), a joint project of the Institute for Scientiﬁc and Technological Research (IRST, Trento) and of the University of Trento funded by Provincia Autonoma di Trento.

References 1. M. Arumugam, A. Sheth, and I. Budak Arpinar. The peer-to-peer semantic web: A distributed environment for sharing semantic knowledge on the web. In WWW2002 Workshop on Real World RDF and Semantic Web Applications. Honolulu, Hawaii (USA), 2002.

46

Matteo Bonifacio et al.

2. M. Benerecetti, P. Bouquet, and C. Ghidini. Contextual Reasoning Distilled. Journal of Theoretical and Experimental Artificial Intelligence, 12(3):279–305, July 2000. 3. D. Bertolini, P. Busetta, A. Molani, M. Nori, and A. Perini. Peer-to-peer, agents, and knowledge management: a design framework. Technical Report tr0207, ITCirst, Jun 2002. 4. E. Blanzieri, P. Giorgini, F. Giunchiglia, and C. Zanoni. Personal agents for implicit culture support. In L. van Elst, V. Dignum, and A. Abecker, editors, Proceedings of the AAAI-03 Spring Symposium 2003 on Agent-Mediated Knowledge Management (AMKM-03), Stanford University (CA, USA), 2003. AAAI Press. AAAI Technical Report SS-03-01. 5. J.R. Boland and R.V.Tenkasi. Perspective making and perspective taking in communities of knowing. Organization Science, 6(4):350–372, 1995. 6. J.R. Boland, R.V.Tenkasi, and D. Te’eni. Designing information technology to support distributed cognition. Organizational Science, 5(3):456–475, August 1994. 7. M. Bonifacio, P. Bouquet, and R. Cuel. Knowledge Nodes: the Building Blocks of a Distributed Approach to Knowledge Management. Journal of Universal Computer Science, 8(6):652–661, 2002. Springer Pub. & Co. 8. M. Bonifacio, P. Bouquet, G. Mameli, and M. Nori. Kex: a peer-to-peer solution for distributed knowledge management. In D. Karagiannis and U. Reimer, editors, Fourth International Conference on Practical Aspects of Knowledge Management (PAKM-2002), Vienna (Austria), 2002. 9. M. Bonifacio, P. Bouquet, and A. Manzardo. A distributed intelligence paradigm for knowledge management. In AAAI Spring Symposium Series 2000 on Bringing Knowledge to Business Processes, Stanford University (CA). AAAI, 2000. 10. M. Bonifacio, P. Bouquet, and P. Traverso. Enabling distributed knowledge management. managerial and technological implications. Novatica and Informatik/Informatique, III(1), 2002. 11. P. Bouquet, A. Don` a, L. Seraﬁni, and S. Zanobini. Contextualized local ontologies speciﬁcation via ctxml. In P. Bouquet, editor, Working Notes of the AAAI-02 workshop on Meaning Negotiation. Edmonton (Canada). July 28, 2002. AAAI, AAAI Press, 2002. 12. P. Bouquet, F. Giunchiglia, F. van Harmelen, L. Seraﬁni, and H. Stuckenschmidt. C-owl – contextualizing ontologies. Technical report, DIT – University of Trento, 2003. Submitted to the Second International Semantic Web Conference (ISWC03). 13. G. C. Bowker and S. L. Star. Sorting things out: classification and its consequences. MIT Press., 1999. 14. S.J. Brown and P. Duguid. Organizational learning and communities-of-practice : Toward a uniﬁed view of working, learning and innovation. Organization Science, 2(1), 1991. 15. J. Dinsmore. Partitioned Representations. Kluwer Academic Publishers, 1991. 16. J.M. Dodero, S. Arroyo, and V.R. Benjamins. Dynamic generation of agent communities from distributed production and content driven delivery of knowledge. In L. van Elst, V. Dignum, and A. Abecker, editors, Proceedings of the AAAI-03 Spring Symposium 2003 on Agent-Mediated Knowledge Management (AMKM-03), Stanford University (CA, USA), 2003. AAAI Press. AAAI Technical Report SS03-01. 17. D. Dougherty. Interpretative barriers to successful product innovation in large ﬁrms. Organization Science, 3(2), 1992.

Peer-Mediated Distributed Knowledge Management

47

18. M. Ehrig, C. Schmitz, S. Staab, J. Tane, and C. Tempich. Towards evaluation of peer-to-peer-based distributed knowledge management systems. In L. van Elst, V. Dignum, and A. Abecker, editors, Proceedings of the AAAI-03 Spring Symposium 2003 on Agent-Mediated Knowledge Management (AMKM-03), Stanford University (CA, USA), 2003. AAAI Press. AAAI Technical Report SS-03-01. 19. G. Fauconnier. Mental Spaces: aspects of meaning construction in natural language. MIT Press, 1985. 20. F. Giunchiglia. Contextual reasoning. Epistemologia, special issue on I Linguaggi e le Macchine, XVI:345–364, 1993. Short version in Proceedings IJCAI’93 Workshop on Using Knowledge in its Context, Chambery, France, 1993, pp. 39–49. Also IRSTTechnical Report 9211-20, IRST, Trento, Italy. 21. F. Giunchiglia and I. Zaihrayeu. Making peer databases interact – a vision for an architecture supporting data coordination. In 6th International Workshop on Cooperative Information Agents (CIA-2002), Universidad Rey Juan Carlos, Madrid, Spain, September 18 - 20, 2002, 2002. Invited talk. 22. I. Goﬀaman. Frame Analysis. Harper & Row, New York, 1974. 23. T. Kuhn. The structure of Scientific Revolutions. University of Chicago Press, 1979. 24. P. Bouquet B. Magnini, L. Seraﬁni, and S. Zanobini. A SAT–based algorithm for context matching. In P. Blackburne, C. Ghidini, R.M. Turner, and F. Giunchiglia, editors, Proceedings of the 4th International and Interdisciplinary Conference on Modeling and Using Context (CONTEXT-03). Stanford University (CA), June 2325, 2003, volume 2680 of Lecture Notes in Artificial Intelligence. Springer Verlag, 2003. 25. J. McCarthy. Notes on Formalizing Context. In Proc. of the 13th International Joint Conference on Artificial Intelligence, pages 555–560, Chambery, France, 1993. 26. A. Molani, A. Perini, E. Yu, and P. Bresciani. Analyzing the requirements for knowledge management using intentional analysis. In L. van Elst, V. Dignum, and A. Abecker, editors, Agent–Mediated Knowledge Management (AMKM-03). AAAI, 2003. AAAI 2003 Spring Symposium Series, Stanford University (CA). 27. W. J. Orlikowski and D. C. Gash. Technological Frames: Making Sense of Information Technology in Organization. ACM Transactions on Information Systems, 12(2):174–207, April 1994. 28. M.P. Papazoglou, J. Jang, and B.J.Kraemer. Leveraging web-services and peer-topeer networks. 2002. 29. S. Schulz, R. Kalckloesch, T. Schwotzer, and K. Herrmann. Towards trust-based knowledge management for mobile communities. In L. van Elst, V. Dignum, and A. Abecker, editors, Proceedings of the AAAI-03 Spring Symposium 2003 on AgentMediated Knowledge Management (AMKM-03), Stanford University (CA, USA), 2003. AAAI Press. AAAI Technical Report SS-03-01. 30. L. Seraﬁni, P. Bouquet, B. Magnini, and S. Zanobini. An algorithm for matching contextualized schemas via SAT. Technical report, ITC-IRST, January 2003. 31. SWAP. Semantic web and peer-to-peer. http://swap.semanticweb.org/, 2002. Project funded by the EC under the IST Programme. 32. L. van Elst and A. Abecker. Domain ontology agents in distributed organizational memories. In Proceedings of the Internationa Joint Conference on Artificial Intelligence (IJCAI’01). Morgan Kaufmann, August 2001. 33. E. Wenger. Communities of Practice. Learning, Meaning, and Identity. Cambridge University Press, 1998.

The Impact of Conversational Navigational Guides on the Learning, Use, and Perceptions of Users of a Web Site Art Graesser1, G. Tanner Jackson1, Matthew Ventura1, James Mueller2, Xiangen Hu1, Natalie Person2 1

University of Memphis, Department of Psychology, 202 Psychology Building, University of Memphis, Memphis, TN 38152-3230, {a-graesser, gtjacksn, mventura, xhu} @memphis.edu 2 Rhodes College, Department of Psychology, Memphis, TN, 38112 {muejn, person} @rhodes.edu

Abstract. Knowledge management systems will presumably benefit from intelligent interfaces, including those with animated conversational agents. One of the functions of an animated conversational agent is to serve as a navigational guide that nudges the user how to use the interface in a productive way. This is a different function from delivering the content of the material. We conducted a study on college students who used a web facility in one of four navigational guide conditions: Full Guide (speech and face), Voice Guide, Print Guide, and No Guide. The web site was the Human Use Regulatory Affairs Advisor (HURAA), a web-based facility that provides help and training on research ethics, based on documents and regulations in United States Federal agencies. The college students used HURAA to complete a number of learning modules and document retrieval tasks. There was no significant facilitation of any of the guides on several measures of learning and performance, compared with the No Guide condition. This result suggests that the potential benefits of conversational guides are not ubiquitous, but they may save time and increase learning under specific conditions that are yet to be isolated.

1 Introduction Knowledge management systems are expected to be facilitated by intelligent interfaces that guide users who vary in cognitive abilities, domain, knowledge, and computer literacy. Some users will not have the patience to learn systems that are not used very often. These users will need fast and easy guidance. Some prefer to talk with agents in a conversational style rather than reading dense printed material on a computer screen and typing information via keyboard. Therefore, there has been serious interest in intelligent interfaces that have speech recognition and animated conversational agents. These agents incorporate synthesized speech, facial expressions, and gestures in a coordinated fashion that attempts to simulate a conversation partner. An ideal interface would allow the user to have a conversation with the computer, just as one would have a conversation with a person. Animated conversational agents have been explored in the context of learning environments and help systems during the last decade [2], [3], [4], [11], [12], [14], L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 48-56, 2003.  Springer-Verlag Berlin Heidelberg 2003

The Impact of Conversational Navigational Guides

49

[19]. There is some evidence that AutoTutor, a tutoring system with an animated conversational agent, improves learning when college students learn about computer literacy or conceptual physics by holding a conversation with the computer tutor [11], [21]. However, it is still unsettled what aspects of a conversational agent might be effective, and under what conditions [2], [19], [23]. Is it the voice, the facial expressions, the responsiveness to the user, the gestures, the content of the messages, or some combination of these features? Whittaker (2003) has concluded that the voice is particularly effective in promoting learning and in engaging the user’s attention, but the other components of the agent may be effective under specific conditions that are not yet completely understood. One potential function of an animated conversational agent is to serve as a navigational guide to offer suggestions on how the user might use the interface in a productive way. This is an entirely different function from delivering the content of the material that would otherwise be read. The purpose of the present study was to investigate different types of conversational navigational guides that are available to adults when they use a new web site. Do these guides saving time for the user when the agents offer suggestions on what to do next? Does the user acquire more information because of the time that is allegedly saved? What are the perceptions of users toward conversational navigational guides? Do the like them, or are the suggestions irritating? It is widely acknowledged that the Microsoft’s Paperclip irritated many users because of its intrusiveness and the difficulty of getting rid of it. Perhaps a better designed, more conversationally appropriate, agent would be more appreciated by the user. We conducted a study on 155 college students who used a web facility in one of four navigational guide conditions: Full Guide (speech and face), Voice Guide, Print Guide, and No Guide. The web site was the Human Use Regulatory Affairs Advisor (HURAA), a web-based facility that provides help and training on the ethical use of human subjects in research, based on documents and regulations in United States Federal agencies [9]. The college students used HURAA to complete a number of training modules and document retrieval tasks.

2 Different Types of Navigational Guides The Full Guide was a talking head with synthesized speech, facial expressions, and pointing gestures. The Agent told the user what to do next when the user first encountered a web page. For example, when the user entered the “Explore Issues” module, the Agent said, “Select the issue that you would like to explore.” The talking head also moved to direct the user’s attention to some point on the display. For example, the talking head looked down when he said “You may select one of the options below me.” The talking head told the user what each primary and secondary module was supposed to do, after the user rested the mouse pointer over a module link for more than 2 seconds. The Agent was designed to project an authoritative persona and to help the user navigate through the interface more quickly. Many novice users are lost and don’t know what to do next when they encounter a page. The Agent was designed to reduce this wasted time.

50

Art Graesser et al.

In order to directly test the influence of the Agent as a navigational guide, participants were randomly assigned to one of the following four conditions: Full Guide. There is the full talking head. Voice Guide. There is a voice that speaks, but no head. Print Guide. The guidance messages are printed at the location where the talking head normally is. No Guide. There are no messages of navigational guidance, either spoken or in print. If a navigational guide is important, then the completion of the various tasks should be poorer in the No Guide condition than the other three conditions: d < min{a, b, c}. When considering the three conditions with the guidance, there is a question of what medium is effective. If speech reigns supreme, then c < min{a,b}. This would be predicted by available research that has compared the impact of spoken versus printed text on comprehension and memory [2], [19], [23]. If print is superior, then the prediction would be that c > max{a,b}. If the presence of the face provides a persona effect that improves interactivity [14], then the prediction is a > b. However, if the face is a distraction from the material in the main display, then the prediction is a < b.

3 Human Use Regulatory Affairs Advisor (HURAA) HURAA is a web-based facility that provides help, training, and information retrieval on the ethical use of human subjects in research. The content of HURAA is derived from Federal agency documents and regulations, particularly the National Institutes of Health [20], the Department of Defense [6], [7], and particular branches of the US military. The targeted users of HURAA focus on fundamental ethical issues, but not the detailed procedures and paper work associated with gaining approval from Institutional Review Boards. The design of HURAA was guided by a number of broader objectives. The layout and design of the web facility incorporate available guidelines in human factors, human-computer interaction, and cognitive science [5], [17]. The architecture of the HURAA components needed to be conformant with the ADL standards for reusable instructional objects, as specified in the Sharable Content Objects Reference Model [22]. The primary objective of having these standards is to allow course content to be shared among different lesson planners, computer platforms, and institutions. HURAA was designed to optimize both learning and information transmission. Adult users are likely to have very little time, so it is important to optimize the speed and quality of learning in web-based distance learning environments. This requires careful consideration of the pacing of the information delivery, the selection of content, and design of the tasks to be performed. The web site was supposed to be engaging to the use, so there was persuasive multimedia intended to hook the user to continue on the website. Finally, HURAA incorporated some of the sophisticated pedagogical techniques that have been implemented in advanced learning environments with intelligent tutoring systems and animated conversational agents.

The Impact of Conversational Navigational Guides

51

HURAA has a number of standard features of conventional web facilities and computer-based training, such as hypertext, multimedia, help modules, glossaries, archives, links to other sites, and page-turning didactic instruction. HURAA also has more intelligent features that allegedly promote deeper mastery of the material, such as lessons with case-based and explanation-based reasoning, document retrieval though natural language queries, animated conversational agents, and contextsensitive Frequently Asked Questions (called Point & Query, [10]). Additional details about HURAA can be found in Graesser, Hu et al., 2002. This paper directly focuses on some of the tasks users would complete with HURAA and what impact the 4 different guides had on the completion of these tasks and the users’ perceptions of the learning environment.

4 Materials and Procedure The experiment included three benchmark tasks that participants completed while interacting with HURAA. This was followed by a series of tests and surveys that were completed after they interacted with HURAA. We refer to these two phases as the HURAA acquisition phase and the post-HURAA test phase, respectively. The next section describes the modules and HURAA facilities that are directly relevant to the performance evaluation. The participants were 155 undergraduate students at the University of Memphis and Rhodes College who participated for course credit or for money ($20). 4.1 HURAA Acquisition Phase Introduction. The Introduction Module is a multimedia movie that plays immediately after a new user has logged in. It is available for replay for users who want to see a repeat. The Introduction is intended to impress the user with the importance of protecting human subjects in research. It introduces the user to the basic concepts of the Common Rule [6], [20], of the Belmont Report’s coverage of beneficence, justice, and respect for persons, and of the Seven Critical Issues that must be scrutinized when evaluating any case [8]: Social and scientific value, accepted scientific principles, fair subject selection, informed consent, minimizing risks and maximizing benefits, independent review, and respect for subjects. The Introduction was prepared by an accomplished expert in radio and web-based entertainment industries, after rounds of feedback from a panel of DoD personnel. Lessons. This module has four lessons that teach the user about the Seven Critical Issues identified by Emmanuel et al. (2000) and how to apply them to particular cases that involve ethical abuses. This is a form of case-based reasoning [1], [15]. The first lesson presented the user with descriptions of the Seven Critical Issues, a summary of the Tuskegee Syphilis Study, and an explanation of how each of the Seven Critical Issues was violated in the Tuskegee study. The second lesson presented the user with a description of a study on post traumatic stress disorders. The user was then presented with the Seven Critical Issues and must decide, on a six-point scale, the

52

Art Graesser et al.

extent to which there potentially is a problem with each issue in that case. The six point scale is: 1 = Definitely not a problem, 2 = most likely not a problem, 3 = undecided, guess it’s not a problem, 4 = undecided, guess it’s a problem, 5 = most likely a problem, and 6 = definitely a problem. The user then received feedback comparing his/her responses with those of a panel of experts from the DoD, along with a brief explanation. Discrepancies between the learner’s decisions and the judgments of the experts were highlighted. Lesson 3 followed the same procedure as Lesson 2, except there was another case on a routine flight test with an experimental helmet. Lesson 4 presented two additional cases, following the same procedure. One was on helmet-mounted devices and the other on chemotherapy. Signal detection analyses were performed on the learner’s decisions as a measure of performance. There are four categories of decisions when signal detection analyses are applied. Hit (H). Both the learner and expert agree that an issue is potentially problematic for the particular case. Correct rejection (CR). Both the learner and expert are in agreement that an issue is not potentially problematic for a case. Miss (M). The expert believes there is a potential problem, but the learner does not. False alarm (FA). The learner believes there is a problem, but the expert believes there is no problem. The experts were 7 experts on research ethics in the military. A d’ score was also computed that assesses how well the learner can discriminate whether a case does versus do not have a problem with respect to an issue. A d’ score of 0 means the learner is not at all discriminating whereas the score increases to the extent that the user is progressively more discriminating (with an upper bound of about 4.0). Query Documents. This module allows the user to ask a natural language question (or description) and then generates an answer by retrieving high matching excerpts from various documents in the HURAA web site. For each document that the user selects, the highest matching paragraph from the document space is selected by the computational linguistics software and is displayed in a window. Beneath this window, the headings for the next four results appear. If the top choice is not the one that the user needs, s/he can click on the headings to read those excerpts. The search engine that was available to identify the optimal matches was latent semantic analysis [16], [13]. In the search task, the participants were instructed to search the document space in order to find answers to 4 test questions. The participants recorded the answers in a test booklet. If the answer to a question was lengthy, they were instructed to write down the fetched document and section number where the answer was found. Performance was measured by retrieval time and the likelihood of retrieving the correct paragraph out of the large document space. If the natural language query facilities are useful, then there should be facilitation in the speed and likelihood of accessing the correct documents.

The Impact of Conversational Navigational Guides

53

4.2 Tests in the Test Phase The test consisted of three parts: (1) Memory (a test on the important ideas from the Introduction and Lessons), (2) Issue comprehension (a test on the participant’s ability to identify potentially problematic issues in cases), and (3) Perception ratings (ratings on how the participants viewed the learning experiences). Memory for Important Ideas. This phase tests memory for the central, core ideas from the Introduction and Lesson material. These core concepts are those that all users should take away from the learning experience. Memory was assessed in three subtests: Free recall, cued recall, and the cloze task. The free recall test presented a series of concepts that the participants were asked to define or describe off of the top of their head. After finishing the free recall task, the cued recall test was administered on the next page. The cued recall test had more retrieval cues than the free recall test. The cloze procedure has the most retrieval cues. It took verbatim segments of the introductory text and left out key words, which the participant filled in. There were progressively more retrieval cues for content to be retrieved as one goes from free recall to the cloze task. Issue Comprehension. This test assessed how discriminating the participants were in identifying potentially problematic issues on two cases. The cases were selected systematically so that 6 of the issues were problematic in one and only one of the two cases; one of the issues was problematic in both cases so it was not scored. This test is functionally a transfer test from the case-based, explanation-based reasoning task in the HURAA acquisition phase. The participants simply read each case and rated the seven issues on the 6-point scale (as to whether issue I was problematic for case C. Perception Ratings. The participants gave ratings on their perceptions of the learning environments. The four rating scales that were included in all three experiments are presented below. The values on each rating scale were: 1 = disagree, 2 = somewhat disagree, 3 = slightly disagree, 4 = slightly agree, 5 = somewhat agree, and 6 = agree. Examples are as follows: You learned a lot about human subjects protections.” and “It was easy to use and learn from these instructional materials.”

5 Results and Discussion Table 1, on the last page, presents means and standard deviations of the dependent measures in the four experimental conditions. The most striking finding from the experiment is the lack of significant differences among conditions. In fact, there were no significant differences among the conditions for any of the 13 dependent measures in Table 1. This null result is incompatible with all of the above predictions. It should be emphasized that the sample size was quite large so the likelihood of a type II error was not high. The practical implication of the result is that the animated conversational agent did not facilitate learning, usage, and perceptions of the interface. In essence, the agent and the conversational guidance had no bang for the buck. Perhaps the web facility was designed extremely well, so well that a navigational guide was superfluous. The

54

Art Graesser et al.

navigational agent might prove to be more effect when the information on the screen is more complex, congested, and potentially confusing. Knowledge management systems often have complex information displays so the value of these agents may increase as a function of the complexity, ambiguity, and perplexity of the system. Perhaps there are special conditions when a navigational guide of some form will be helpful, whether it be print, voice, or a talking head. However, these precise conditions have yet to be discovered and precisely specified in the literature. It is appropriate to acknowledge that the results of the present study on agents as navigational guides does not generalize to other learning environments. Animated conversational agents have proven to be effective when they deliver information and learning material in monologues and tutorial dialogues [2], [19], [21], particularly when the test taps deep levels of comprehension. However, only a handful of empirical studies has systematically investigated the impact of these conversational agents on learning, so more research is definitely needed. One intriguing finding is that the amount of information that a person learns and remembers from a learning system is not significantly correlated with how much the learner likes the system [18]. Simply put, learning is unrelated to liking. It this result is accurate, then it is not sufficient to simply ask users and individuals in focus groups what they like or do not like about agents and navigational guides. There also needs to be a serious, deep, research arm that goes beyond intuitions of users, designers, and managers.

6 Acknowledgements This research was directly supported by contracts from the Office of Naval Research (N61339-01-C1006) and the Institute for Defense Analyses (AK-2-1801), and was partially supported from grants by the National Science Foundation (REC 0106965) and the DoD Multidisciplinary University Research Initiative (MURI) administered by ONR under grant N00014-00-1-0600. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ONR, IDA or NSF.

References 1. 2. 3. 4. 5. 6.

Ashley, K.D. (1990). Modeling legal argument: Reasoning with cases and hypotheticals. Cambridge, MA: MIT Press. Atkinson, R.K. (2002). Optimizing learning from examples using animated pedagogical agents. Journal of Educational Psychology, 94, 416-427. Baylor, A.L. (2001). Investigating multiple pedagogical perspectives through MIMIC (Multiple Intelligent Mentors Instructing Collaboratively). Workshop on agents at the AI-ED International Conference. San Antonio, TX. Cassell, J. (2001). Embodied conversational agents: Representation and intelligence in user interfaces. AI Magazine, 22, 67-83. Collins, A., Neville, P., & Bielaczyc, K. (2000). The role of different media in designing learning environments. International Journal of Artificial Intelligence in Education, 11, 144-162. 32 CFR 219 (1991). Protection of Human Subjects, Department of Defense.

The Impact of Conversational Navigational Guides 7. 8. 9.

10.

11. 12. 13. 14. 15. 16. 17. 18.

19. 20. 21.

22. 23.

55

DoD Directive 3216.2 (1993). Protection of Human Subjects in DoD supported research, Department of Defense. Emmanuel, E. J., Wendler, D., & Grady, C. (2000). What makes clinical research ethical? Journal of the American Medical Association, 283, 2701-2711. Graesser, A.C., Hu, X., Person, N.K., Stewart, C., Toth, J., Jackson, G.T., Susarla, S., Ventura, M. (2002). Learning about the ethical treatment of human subjects in experiments on a web facility with a conversational agent and ITS components. In S. A. Cerri (Ed.), Proceedings of Intelligent Tutoring Systems 2002 (pp. 972-981). Berlin, Germany: Springer. Graesser, A. C., Langston, M. C., & Baggett, W. B. (1993). Exploring information about concepts by asking questions. In G. V. Nakamura, R. M. Taraban, & D. Medin (Eds.), The psychology of learning and motivation: Vol. 29. Categorization by humans and machines (pp. 411-436). Orlando, FL: Academic Press. Graesser, A.C., Person, N., Harter, D., & TRG (2001). Teaching tactics and dialog in AutoTutor. International Journal of Artificial Intelligence in Education, 12, 257-279. Graesser, A.C., VanLehn, K., Rose, C., Jordan, P., & Harter, D. (2001). Intelligent tutoring systems with conversational dialogue. AI Magazine, 22, 39-51. Graesser, A.C., Wiemer-Hastings, P., Wiemer-Hastings, K., Harter, D., Person, N., & TRG (2000). Using latent semantic analysis to evaluate the contributions of students in AutoTutor. Interactive Learning Environments, 8, 129-148. Johnson, W. L., & Rickel, J. W., & Lester, J.C. (2000). Animated pedagogical agents: Face-to-face interaction in interactive learning environments. International Journal of Artificial Intelligence in Education, 11, 47-78. Kolodner, J. (1984). Retrieval and organizational strategies in conceptual memory: A computer model. Hillsdale, NJ: Erlbaum.. Landauer, T.K., Foltz, P.W., Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 25, 259-284. Mayer, R.E. (1997). Multimedia learning: Are we asking the right questions? Educational Psychologist, 32, 1-19. Moreno, K.N., Klettke, B., Nibbaragandla, K., Graesser, A.C., & the Tutoring Research Group (2002). Perceived characteristics and pedagogical efficacy of animated conversational agents. In S. A. Cerri, G. Gouarderes, & F. Paraguacu (Eds.), Intelligent Tutoring Systems 2002 (pp. 963-971). Berlin, Germany: Springer. Moreno, R., Mayer, R.E., Spires, H.A., & Lester, J.C. (2001). The case for social agency in computer-based teaching: Do students learn more deeply when the interact with animated pedagogical agents? Cognition & Instruction, 19, 177-213. NIH 45 CFR 46 (1991). National Institutes of Health Code of Federal Regulations, Protection of Human Subjects. Person, N.K., Graesser, A.C., Bautista, L., Mathews, E., & TRG (2001). Evaluating student learning gains in two versions of AutoTutor. In J.D. Moore, C.L. Redfield, and W.L. Johnson (Eds.) Artificial Intelligence in Education: AI-ED in the Wired and Wireless Future (pp. 286-293). Amsterdam: OIS Press. Sharable Content Object Reference Model (SCORM), Versions 1.1 and 1.2. www.adlnet.org. Whittaker, S. (2003). Theories and methods in mediated communication. In A.C. Graesser, M.A. Gernsbacher, and S.R. Goldman (Eds.). Handbook of discourse processes. Mahwah, NJ: Erlbaum.

56

Art Graesser et al. Table 1. Means (and Standard Deviations) for Dependent Measures.

--------------------------------------------------------------------------------------------------------DEPENDENT Full Voice Print No MEASURES Guide Guide Guide Guide --------------------------------------------------------------------------------------------------------Number of participants 40 39 38 38 Memory for Core Concepts Free recall proportion .45 (.21)

.43 (.20)

.42 (.21)

.44 (.20)

Cued Recall proportion

.51 (.24)

.50 (.23)

.45 (.26)

.53(.23)

Cloze recall proportion

.44 (.15)

.42 (.18)

.39 (.17)

.47 (.19)

Introduction study time (minutes)

6.6 (3.3)

9.3 (19.0)

7.4 (4.7)

10.7 (19.8)

Problematic Issue Identification Hit proportion .58 (.10)

.58 (.10)

.56 (.12)

.60(.13)

False alarm proportion

.40 (.32)

.39 (.30)

.38 (.34)

.38(.31)

d’ score (discrimination) .30 (.46)

.31 (.36)

.14 (.75)

.27 (.74)

Task completion time (minutes)

23.7 (8.1)

20.8(5.2)

22.9(4.1)

.50 (.49)

.47(.23)

.52(.25)

27.1(11.4)

24.4(9.6)

23.5(9.2)

24.8(7.9)

Perception ratings Amount learned

4.75(1.06)

4.62(1.16)

4.61(1.08)

4.53(1.48)

Interest

3.85(1.63)

4.08(1.46)

4.11(1.41)

3.79(1.82)

Enjoyment

3.50(1.47)

3.46(1.48)

3.58(1.41)

2.89(1.47)

Ease of learning

4.13(1.40)

3.95(1.45)

3.71(1.71)

3.61(1.57)

23.7 (6.2)

Search for Information Correct document retrieval .55(.25) (Proportion) Search time (minutes)

---------------------------------------------------------------------------------------------------------

Agent-Oriented Knowledge Management in Learning Environments: A Peer-to-Peer Helpdesk Case Study Renata S.S. Guizzardi1, Lora Aroyo1,2, Gerd Wagner3 1

Computer Science Department – University of Twente (UT) P.O. Box 217 – 7500 AE Enschede – The Netherlands {souza, aroyo}@cs.utwente.nl 2 Department of Computer Science – Eindhoven University of Technology (TU/e) P.O. Box 513 – 5600 MB Eindhoven – The Netherlands [email protected] 3 Department of Information & Technology – Eindhoven University of Technology (TU/e) P.O. Box 513 – 5600 MB Eindhoven – The Netherlands [email protected]

Abstract. In this paper we present an analysis and modeling case study for agent mediated knowledge management in educational environments: Help&Learn, an agent-based peer-to-peer helpdesk system to support extraclass interactions among students and teachers. Help&Learn expands the student’s possibility of solving problems, getting involved in a cooperative learning experience that transcends the limits of classrooms. To model Help&Learn, we have used Agent-Object-Relationship Modeling Language (AORML), an UML extension for agent-oriented modeling. The aim of this research is two-fold. On the one hand, we aim at exploring Help&Learn’s potential to support collaborative learning, discussing its knowledge management strategy. On the other hand, we aim at showing the expressive power and the modeling strengths of AORML.

1 Introduction As we enter the new millennium, we realize a shift in the business model from the old static model, based on hierarchic organizations, towards a dynamic model, built on top of continuously changing and knowledge-based organizations. This new business model requires that the twenty-first century professionals have a set of characteristics, such as creativity, flexibility and ability to cooperate and work in teams. The hierarchical educational model is not appropriate to educate these professionals [8]. Methods based on collaboration, viewing students as consumers but also as providers of knowledge [15] can lead to better results because they aim at motivating active participation of the individual in the learning process, which often results in the development of creativity and critical thinking [8]. Knowledge Management (KM) deals with the creation, integration and use of knowledge [6]. These processes are directly related and can be very beneficial to collaborative learning. In fact, we can say KM systems support some kind of unintentional learning, since users can learn, while sharing knowledge. Although the L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 57-72, 2003.  Springer-Verlag Berlin Heidelberg 2003

58

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

benefits of KM for education have been acknowledged many times [13], its full exploration in educational settings is still to be seen. In this context, it is important to deliver knowledge in a personalized way, with respect to user’s preferences regarding content and presentation. Another issue here is dealing with logically and physically dispersed actors and knowledge sources. The system must provide flexible ways to access these sources, helping the user to find the required knowledge in the right time [6]. Agent-mediated KM comes as a solution in such dynamic environments [5]. Agents can exhibit this type of flexible behavior, providing knowledge both “reactively”, on user request, and “pro-actively”, anticipating the user’s knowledge needs. They can also serve as personal assistants, maintaining the user’s profile and preferences. This paper presents Help&Learn (H&L), a Web-based peer-to-peer helpdesk system to support extra-class discussions among students and teachers. In this environment, knowledge emerges as a result of the ongoing collaboration process. As the peer-to-peer architecture suggests, the relationship between teachers and students is non-hierarchical. Instead, all users are peers who collaborate in knowledge creation, integration and dissemination. In H&L, personal assistants are used to maintain a user profile and knowledge base, and other software agents are responsible for finding the best peer to answer to a question, and for managing knowledge resources. From a software engineering perspective, the analysis and design of the distributed processes involved in knowledge management become increasingly sophisticated and require an agent-oriented approach, such as the Agent-Object-Relationship Modeling Language (AORML) [16], an extension of UML to model agent-oriented information systems. The strengths of AORML with respect to KM systems are: 1) it considers the organizations and actors of a domain as agents in the modeling process. In this way, it allows to model business processes on the basis of the interactions between (human and artificial) agents working on behalf of their organizations. Related work is mentioned in [5]. Although norms and contracts are not directly supported by AORML, it provides deontic modeling constructs such as commitments and claims with respect to external agents, and obligations and rights with respect to internal agents. 2) the fact that ‘mentalistic’ concepts of agents, such as beliefs and commitments, are explicitly considered in the system model, supports the software engineer to reason about and to model the behavior of agents, both internally and in interaction with other agents of the system; 3) it captures the behavior of agents with the help of rules. Besides these strengths, since AORML is an extension of UML, preserving its principles and concepts, it is an accessible language, and it is likely to face less resistance for industrial acceptance and use. The aim of this research is two-fold. On the one hand, we aim at exploring H&L’s potential to support collaborative learning, discussing its KM strategy. On the other hand, we show that AORML is an appropriate language to model such a system, as well as other KM environments. In section 2, KM is described in connection with the educational context. Section 3 presents a description of H&L, introducing the main problems and activities in focus. Section 4 introduces AORML and its modeling constructs, which are then applied in section 5, presenting part of the Help&Learn system’s modeling. Section 6 acknowledges some work related to Help&Learn and to AORML. Finally, some directions for future work and conclusions are presented in sections 7 and 8, respectively.

Agent-Oriented Knowledge Management in Learning Environments

59

2 Knowledge Management in Education Collaborative Learning mediated by network-based environments have been the focus of many recent research initiatives and experiments [8,12], especially within the CSCL and the E-learning communities. The need for a different kind of learning approach has been also noted within the KM literature, like in [6]: “The traditional education paradigm is inappropriate for studying the types of open-ended and multidisciplinary problems that are most pressing to our society. These problems, which typically involve a combination of social and technological issues, require a new paradigm of education and learning skills, including selfdirected learning, active collaboration, and consideration of multiple perspectives.” Maçada and Tijiboy (1998) [8] consider three essential elements for collaborative learning to succeed in network-based environments: a) cooperative posture, which involves: non-hierarchical relationship between the participants, collaboration, constant negotiation, open-mindedness, etc.; b) collaborative technological infrastructure; and c) a non-hierarchical method, i.e. it is very important that all the participants get involved in the constant organization and re-organization of the environment dynamics (meaning the establishment of goals, norms, roles, priorities of tasks, etc.). Especially focused in b), this work is based on the assumption that KM can be generally beneficial for learning [13]. KM can, for instance, motivate learners to be more active and to collaborate. While feeding a KM system, the users need to create artifacts, externalizing their knowledge, in order to make it available for other users (user-based approach for knowledge creation, similar to the one adopted in [6]). This process of externalization is an important step for learning. Supporting this idea, Constructionist learning theories emphasize the importance for the learner to produce something concrete, which he can share with his peers [3]. In other words, externalizing knowledge by means of a sharable artifact will help the learner to perform synthesis and learn, and at the same time it may motivate him for peer collaboration. The knowledge resources exchanged in a learning environment cannot be much differentiated from those exchanged for other purposes. In this context: i) there is a share of physical resources, such as: books, articles, and other educational artifacts; ii) with the growing use of information technology and the Internet in these settings, there are plenty of electronic documents, references, and web links; and, finally, iii) there is also tacit knowledge [6], i.e. knowledge that is contained in people’s minds and that is usually informally exchanged among them by different means, for instance, in person, through messages, or via Internet communication tools integrated in virtual learning environments [12]. All these forms of knowledge need to be properly integrated and managed in order to bring about positive changes in the teaching/learning process. Exemplifying the common difficulties of this context, we mention the fact that all these resources are distributed among people and that it is not easy to find out who has the right piece of information, knowledge or advice. The nature of these problems suggest that KM systems (KMSs) can be highly recommendable for learning settings. In addition to that, software agents’ specific characteristics turn them into promising candidates in providing a KMS solution [5]. These agents can be used both as a

60

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

metaphor to model the domain in which the system will be deployed, and as software components to develop the actual KMS. Targeting the above highlighted problems, we propose H&L to support the organization and sharing of distributed knowledge. The specifics about the KM strategies applied in H&L are presented in the next session.

3 Help&Learn: A Peer-to-Peer Architecture for Knowledge Management in Learning Settings Peer-to-peer introduces a set of concepts that takes a human centered view of knowledge as residing not just in people’s minds but also in the interaction between people and between people and documents [15]. H&L expands the student’s possibility of solving their doubts, getting involved in a cooperative learning experience that transcends the limits of classrooms. By collaborating with other peers, the students learn with the doubts of others, besides developing cognitive abilities, such as to state clearly their doubts and thoughts; to interpret questions; to mediate discussions; and to solve problems. In this open context, other interested parties may join the learning community, such as business employees and online organizations. They bring different perspectives to the discussions, making the cooperation richer. Figure 1 shows the peer-to-peer architecture of the proposed scenario.

Fig. 1. A teacher, a student, and employees of a company, interacting to ask and answer questions in the proposed peer-to-peer environment

We use the metaphor of a helpdesk, where somebody asks for help (the helpee) and somebody provides the needed help (the helper). Each peer in the network is seen as a source of knowledge. The agents of the system are responsible for managing the exchanges between these sources. This includes: a) handling a peer request for help and delivering help in a personalized way; b) finding the best peer to answer to a help request; and c) searching through previously asked questions/answers [12]. In H&L, knowledge is created and integrated in use-time, including users participation in these processes, and not in design time with the help of a knowledge engineer. This model has many advantages, such as: avoiding that knowledge artifacts become obsolete, for being dependant on the knowledge engineering; and motivating

Agent-Oriented Knowledge Management in Learning Environments

61

the users of the system to engage in collaboration and learning, while creating and sharing the artifact [6]. The knowledge in H&L is exchanged by the system peers in the form of HelpItems. These HelpItems can be what-is or how-to-do explanations, bibliographic or Web references, electronic documents, or even hardcopies, depending on the peers setting (e.g. inside a school or a company, hardcopies can be exchanged in addition to electronic copies). The users are not required to perform knowledge formalization. The exchanged questions and answers are expressed and stored in natural language. Besides mediating this exchange of help, the system agents are responsible for searching through previously asked questions and answers to provide the users with suitable help. The quality of knowledge artifacts is an important issue in KMSs [6]. In H&L, this is measured by the peers themselves. The help provided is annotated by the helpee, and this information is shared among the agents of the system, to be considered in future helper indication. As in a typical peer-to-peer application [15], a key issue here is finding the best peer to satisfy a certain help request. A helper is selected if she can fulfill a help request, by providing the helpee with appropriate HelpItems. Besides expertise, the time and availability of the peer are also considered for the best helper indication. As an example, a teacher may know the answer to a student’s question but she may have less time than an advanced student to spend on it. A common problem in KM settings is motivating the users of the system to use it in its full potential [6,7]. The peer motivation to participate in discussions and answer to helpees’ questions can be given by a sense of belonging to a learning community, or by the desire of having a good social status [7]. However, this motivation can also be caused by external factors, like teacher’s reinforcements or an external grading system.

4 Agent-Object-Relationship Modeling The Agent-Object-Relationship (AOR) modeling approach [16] is based on an ontological distinction between active and passive entities, that is, between agents and objects. This helps to capture the semantics of complex processes, such as the one that involves teachers and students, owners and employees of a company, and other actors involved in a KM environment. The agent metaphor subsumes both artificial and natural agents. This way, the users of the information system are included and also considered as agents in AOR modeling. Intuitively, some connections can already be identified between the knowledge artifacts in a KMS and objects, and between the KMS users and human agents. The KMS itself can also be composed of multiple software agents, which perform different tasks, accomplishing various goals, in order to mediate the processes of knowledge creation, integration and sharing. These agents can be identified and modeled with the aid of AORML. AOR distinguishes between agents and objects according to these two main points: 1) while the state of an object in OO programming has no generic structure, the state of an agent has a ‘mentalistic’ structure: it consists of mental components such as

62

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

beliefs and commitments. 2) while messages in object-oriented programming are coded in an application-specific ad-hoc manner, a message in Agent-Oriented Programming is coded as a ‘speech act’ according to a standard agent communication language that is application-independent. In AORML, an entity is either an agent, an event, an action, a claim, a commitment, or an ordinary object. Agents and objects form, respectively, the active and passive entities, while actions and events are the dynamic entities of the system model. Commitments and claims establish a special type of relationship between agents. These concepts are fundamental components of social interaction processes and can explicitly help to achieve coherent behavior when these processes are semi or fully automated. Only agents can communicate, perceive, act, make commitments and satisfy claims. Ordinary objects are passive entities with no such capabilities. Besides human and artificial agents, AOR also models institutional agents. Institutional agents are usually composed of a number of human, artificial, or other institutional agents that act on its behalf. Organizations, such as companies, government institutions and universities are modeled as institutional agents, allowing to model the rights and duties of their internal agents. There are two basic types of AOR models: external and internal models. An external AOR model adopts the perspective of an external observer who is looking at the (prototypical) agents and their interactions in the problem domain under consideration. In an internal AOR model, we adopt the internal (first-person) view of a particular agent to be modeled. This paper is focused on the exemplification of external AOR models, which provide the means for an analysis of the application domain. Typically, these models have a focus, that is an agent, or a group of agents, for which we would like to develop a state and behavior model. Figure 2 shows the elements of an external AOR model, in which the language notation can be seen. Object types belong to one or several agents (or agent types). They define containers for beliefs. If an object type belongs exclusively to one agent or agent type, the corresponding rectangle is drawn inside this agent (type) rectangle. If an object type represents beliefs that are shared among two or more agents (or agent types), the object type rectangle is connected with the respective agent (type) rectangles by means of an UML aggregation connector. As it can be seen in Figure 2, there is a distinction between a communicative action event (or a message) and a non-communicative action event. Also, AOR distinguishes between action events and non-action events. The figure also shows that a commitment/claim is usually followed by the action event that fulfills that commitment (or satisfies that claim). An external model may comprise one or more of the following diagrams: •

•

Agent Diagrams (ADs), depicting the agent types of the domain, certain relevant object types, and the relationship among them. An AD is similar to a UML class diagram, but it also contains the domain’s artificial, human and institutional agents. Interaction Frame Diagrams (IFDs), depicting the action event types and commitment/claim types that determine the possible interactions between two agent types (or instances).

Agent-Oriented Knowledge Management in Learning Environments

63

Fig. 2. The core elements of AOR external models

•

Interaction Sequence Diagrams (ISDs), depicting prototypical instances of interaction processes. • Interaction Pattern Diagrams (IPDs), focusing on general interaction patterns expressed by means of a set of reaction rules defining an interaction process type. Reaction rules are the chosen component by AOR to show the agent’s reactive behavior and it can be represented both graphically and textually. These diagrams will be exemplified in the following section. For further reference, we refer to [16] and to the AOR website: http://aor.rezearch.info/.

5 Help&Learn Modeling AORML can be used throughout the whole development cycle of a system. In this paper, we will focus on the analysis phase, in which we applied AOR external models. Figure 3 depicts the agent diagram, which includes all human, artificial and institutional agents (distinguished by UML stereotypes) involved in the helpdesk, and their relationships. Note that this diagram is very similar to the UML class diagram, showing the system’s classes and relationships between them. For clarity purposes, the attributes of agents and objects are omitted in this diagram. However, they can be expressed following the traditional UML syntax. As the above diagram shows, H&L brings together students, teachers and general business professionals as peers of a learning community. Below, we give a brief description of each artificial agent of the system. Help&Learn Infrastructure Server (IS). This agent addresses the management of the H&L system itself. It provides the other artificial agents of the system, as well as periodic updates. Peer Assistant (PA). In order to start participating on discussions in the system, a Person downloads the Peer Assistant (PA) from the H&L IS. This way, this Person becomes one of the system Peers, being able to act both as a helpee and as a helper for other Peers.

64

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

Fig. 3. Helpdesk System Agent Diagram

Directory Server (DS) and Broker (B). Every time the PA goes online, it registers with the Directory Server (DS), becoming available to answer help requests. When doing this, the PA will provide the DS with a minimal Peer profile, indicating what topics can be answered by him. On the other hand, the Broker creates his own Peer profile by contacting the PAs and also by applying data mining techniques on the DS profiles, in order to make rankings and classifications. The Broker ranks the Peers based on expertise, availability and reliability and it classifies them based on interests. This way, when queried by the PAs, it can provide information on the most appropriate Peers to answer a certain help request. The DS also maintains a repository of previously provided explanations, along with their respective request (typically, a question). This way, the PA consults this agent every time a question is forwarded to it by a helpee, to check whether or not this question has been already answered. If so, the answer is immediately recovered to the helpee; otherwise, the PA consults the Broker for a best helper indication. In this repository, Information Retrieval Techniques are used in order to group similar questions and aid the retrieval of the relevant ones, as well as to support the creation of an automatic FAQ, according to the proposals of a previous work [12]. SIG Assistant. Special Interest Groups (SIGs) are also allowed to participate in the system (this is indicated by the inclusion of the institutional agent SIG in the agent diagram of Fig. 3). These SIGs usually pre-exist the system, but can also be created

Agent-Oriented Knowledge Management in Learning Environments

65

by suggestion of the Broker. It is not necessary that all the members of a SIG are Peers, only one member is enough (note this, again in the agent diagram, which generalizes a SIG Member as a Person, instead of relating it with the Peer class). The Broker has a representation of the SIGs and can also suggest that a PA contacts one of the SIG Assistants in order to ask the SIG for help. The SIG Assistant broadcasts the message to all members of the SIG. Then, the answers are sent back to the PA. Today, there are many SIGs advertised in the Web, specialized in several different areas. By introducing them to the helpdesk system, we hope to broaden their interaction scope, at the same time that we give the opportunity for other Peers to have their help request answered by an expert on the topic. Resource Manager (RM). The Resource Manager brings to the system existing knowledge bases, which can be databases, document repositories etc. This way, HelpItems that are not owned by any of the system Peers can also be considered and consulted by the PAs. These knowledge bases can be consulted through keyword or query search. A System Peer does not directly contact a RM. Instead, this is done through the PA. In the case of a query search, the Peer uses an interface, based on established query languages such as SQL, XML-Query, or RDF-Query, which will then be translated by the contacted RM to the query language of the specific knowledge base. These agents are typically downloaded by the owners of existing knowledge bases, who will create the translation specification. 5.1 Interactions in Help&Learn The next step after defining the agents in the system is to model their interactions using Interaction Sequence Diagrams (ISDs) for concrete examples. In H&L, a Peer can request explanation, or for a document (reference, electronic copy or hardcopy). For reasons of lack of space, only the first one is exemplified in this paper. A prototypical interaction sequence triggered when a Peer issues a request for an explanation is shown on Figures 4 and 5. Such sequence should generally be maintained in the same ISD, integrating the whole process. This is especially useful for automatic code generation. Here, we chose to divide the sequence in two phases in order to facilitate our exploration of the modeling language specifics. Moreover, this way the general understanding of the interaction sequence may be eased. Figure 4 shows the Peer request and the best helper indication by the Broker. Here, Anna, a system Peer, issues a request for help to her PA, asking “what is p2p?”. The PA attempts first to find out if this question has already been asked, by querying the DS maintaining the Explanation Case (see Fig.3). Since this question is asked for the first time, the PA cannot provide a direct answer and asks the Broker to find the best helper to answer this question. The Broker returns a ranked list of possible Peers for the PA to select. In our example, this list contains only one indication: Mark. Having the Broker’s indication, the PA will then try to get the HelpItem that fulfills its user’s request. This is depicted in Fig. 5. It starts with Anna’s PA contacting Mark’s PA with the request for help. Mark’s PA replies with an acknowledge message, confirming it received the request. In this moment, a commitment is established from Mark’s PA towards Anna’s PA, fixing that the first will try to get help (from its peers) to answer to the latter’s request.

66

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

Fig. 4. Interaction Sequence Diagram, showing a help request being issued by Anna, a H&L Peer, and the best helper indication by the Broker

Fig. 5. Interaction Sequence Diagram, showing how a PA deals with the request for an explanation, on behalf of its user

This commitment is also represented in the ISD of Fig. 5. It is created by the acknowledge message (dashed arrow along with a “C”, for “Create”) and it has two arguments, a provideHelp and a noHelpAvailable message. These two messages compose an Or-Split (diamond containing an “x”), which represent the possible outcomes if the commitment is fulfilled. If any other possibility occurred, it would mean the commitment had been broken. Proceeding in the ISD, we will see this is not the case in this example. Mark’s PA forwards the request to Mark, who provides the following answer: “p2p is a distributed technology…”. This message is then forwarded to Anna’s PA (note an arrow from this message to the commitment, indicating its fulfillment). At last, the help (i.e. the explanation) gets to its destination: Anna.

Agent-Oriented Knowledge Management in Learning Environments

67

Fig. 6. Interaction Sequence Diagram, showing the interaction process when a request for a keyword search is issued by a peer

The use of commitments supports situations in which the communication between two agents is asynchronous, as in this case. Mark’s PA confirms it is going to provide the help. However, Anna’s PA knows this can take some time, depending on Mark’s availability and willingness to respond. Commitments are also good constructs to treat agent’s autonomy. If it were useful, we could represent, for example, a commitment between Mark and its Peer, establishing that Mark commits to answer to the help request. At first sight, this does not seem very natural, since Mark is a human and, as such, has full autonomy over the system. In other cases, though, dealing with lifethreatening situations and, of course, with artificial agents, this can be rather a good approach. Furthermore, commitments can be used as triggers for exception handling. For example, what should Anna’s PA do in case Mark’s PA does not meet its commitment? In H&L, this agent tries to find another Peer to answer to the help request. Besides requesting an explanation, a Peer can also ask for documents, providing its PA with a list of keywords. Figure 6 depicts the interactions between Help&Learn’s agents, when a request of this type is issued. In the ISD of Fig 6, Anna requests its PA for documents about “peer-to-peer”. The PA asks the Broker who are the best helpers to answer to this request. The Broker returns a list of ranked Peers to answer to the request. In this case, this list contain two Peers: Mark and Joanna. Next, Anna’s PA contacts the PA of both Peers, forwarding the request for keyword search to them. The PAs search through the documents owned by their Peers, returning the available documents to Anna’s PA. Finally, Anna’s PA forwards the documents to Anna.

68

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

Fig. 7. Interaction Pattern Diagram, showing the PA’s internal behavior when receiving a HelpItem on behalf of its user

Note that the sequence shown in Fig. 4, 5 and 6 depicts just one of many possible interactions. The software engineer should make a number of ISDs in order to capture various interaction perspectives. This way he can afterwards generalize the interactions in Interaction Frame Diagrams (IFDs), which depicts the action event types and the commitment/claim types that determine the possible interactions between two agent types (or instances) [16]. Further, the interactions can be detailed in Interaction Patterns Diagrams (IPDs). These diagrams depict general interaction patterns expressed by means of a set of reaction rules, defining an interaction process type. Reaction rules are the chosen component by AOR to show the agent’s reactive behavior, and they can be represented both graphically and textually. Figure 7 depicts an example of this type of diagram. The IPD of Fig. 7 depicts only the two agents involved in this specific process: the PA and the DS. When the DS receives a checkIfExistingExplanation message, it immediately reacts, checking if the sent Question can be found in the Explanation Case (i.e. if the question is similar enough to one or more previously asked ones, according to DS’s internal algorithms). In the affirmative case, the DS sends back the respective answer to the PA. Otherwise, it simply “says no”. This is modeled with the rule R1, which is textually represented (See Table 1). After the external model has been completed, the modeling can proceed to the design stage, in which, for each type of agent system to be designed, the external model is internalized according to the perspective of the respective agent, and subsequently further refined. For instance, an action event, if created by the agent to be designed, is turned into an action, while it is turned into an event if it is perceived by it. Using such an internal perspective and the corresponding indexical terms (such

Agent-Oriented Knowledge Management in Learning Environments

69

as actions and outgoing messages versus events and incoming messages), leads to a natural terminology for designing and implementing agents. H&L internal models will be the subject of future publications. Table 1. Textual Representation of the R1 Reaction Rule ON IF THEN ELSE

Event

RECEIVE checkIfExistingExplanation (?Question) FROM ?PeerAssistant Condition ExplanationCase (?Question,?Answer) Action SEND replyIfExistingExplanation (?Answer) TO ?PeerAssistant Action SEND replyIfExistingExplanation (?Answer=“no”) TO ?PeerAssistant

6 Related Work Regarding Help&Learn, it is important to mention other initiatives on developing peer-to-peer architectures to support knowledge sharing. One of these initiatives is the EDUTELLA project [10], which aims at providing a peer-to-peer networking infrastructure to support the exchange of educational material. In order to accomplish this, peers can make their documents available in the network, specifying metadata information as a set of RDF statements. Bonifacio et al. [2] have developed KEx, a peer-to-peer system to mediate distributed knowledge management. KEx allows each individual or community of users to build their own knowledge space within a network of autonomous peers. Each peer can make documents locally available, along with their context, i.e. a semantic representation of the documents’ content. When searching documents from other peers, a set of protocols of meaning negotiation are used to achieve semantic coordination between the different representations (contexts) of each peer. Both EDUTELLA and KEx are specifically concerned with the exchange of documents and do not address peer collaboration through the exchange of messages, which is one of the targets of H&L. On the other hand, the work proposed by Vassileva [14] proposes a peer-to-peer system to support the exchange of messages between students. A student needing help can request it through his agent, which finds other students who are currently online and have expertise in the area related to the question. As in H&L, there is a centralized matchmaker service, which maintains models of the users competences and matches them to the help-requests. This work is particularly concerned with user motivation to collaborate. Thus, the system rewards users who contribute to the community, by providing them with a better quality of service. Concerning agent-oriented modeling, we should mention AUML [11] and Message/UML [4] since both propose UML extensions to model agent-based systems. AUML has especially extended UML sequence diagrams to model interaction protocols involving agent roles. Message/UML proposes 5 views: Organization,

70

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner

Goal/Task, Agent/Role, Interaction and Domain views, each of them modeling a specific aspect of the multi-agent system. In comparison with AORML, these two approaches do not target domain modeling, being both design-oriented. Besides, both of them lack the ‘mentalistic’ concepts (commitments, claims and beliefs) presented by AORML. It is also important to acknowledge the efforts of Molani et al. [9] in the direction of providing a system analysis and design methodology specific for the Knowledge Management domain. They claim that, in order to develop effective KM solutions, it is necessary to analyze the intentional dimension of the organizational setting, i.e. the interests, intents, and strategic relationships among the actors of the organization. Like AORML, they take an agent-oriented approach to model the domain. The major difference when compared to AORML is the adopted i* framework. Instead of the AORML constructs of agents, objects, relationships, messages, commitments, etc., this framework models the organization as a set of actors, goals, ‘soft goals’, dependencies, tasks and resources.

7 A Few Directions for Future Work In the process of defining H&L’s architecture, guided by the use of AORML, we have elicited many questions, whose answers will be important for the future development of the system. For instance, we intend to address issues related to: a) the structuring of the questions and respective answers, present in the Explanation Case (EC); b) the organization of the personal knowledge assets owned by each peer; c) and the management of HelpItems by the Resource Managers (RMs). Targeting a), we aim at investigating, for instance, how the techniques applied in a previous work [12] can be enhanced (and, perhaps, new techniques applied) in order to provide suitable structuring and retrieval of the EC’s questions and answers (refer to IPD of Fig. 6). This investigation, along with some extra studies, can indicate possibilities for addressing b) and c) as well. Inspired by current research on the Semantic Web [1], we intend to incorporate Ontologies into the H&L architecture. A preliminary study suggests that these ontologies can be aimed at making knowledge explicit, supporting interaction among the system peers. Another possibility is applying contexts to organize the peer’s HelpItems, as suggested in [2], where context is defined as an explicit semantic schema over a body of local knowledge. Another important research focus related to Help&Learn is Personalization, i.e. issues regarding how a PA should balance reactiveness, acting on user request, and pro-activeness, delivering content to the user. In this context, two important questions appear to be important: a) how much should be delegated to the PA during the search for help? The precision of the peer help request should be balanced with the PA’s responsibility to search for the appropriate HelpItem; and b) how should the PAs balance what they know about the peers and what they disclose to the DS and the Broker? As mentioned in H&L’s modeling section, each of these agents (PA, DS and Broker) have different representations of the peer, as they have different goals in the system. Thus, the disclosure of information should be appropriate both for the achievement of the agent’s goals and for guaranteeing the right level of user’s privacy.

Agent-Oriented Knowledge Management in Learning Environments

71

8 Conclusions In this paper, we have described our work in progress with Help&Learn, a peer-topeer agent-oriented architecture, aimed at providing its users with a rich environment for both collaborative and individual use of knowledge. In order to do so, results on collaborative learning [8,12] and KM related research [5,6,7] have been used in the conceptualization and modeling of H&L. We take an agent-oriented perspective on system architecture, where agents play a crucial role in supporting the effectiveness, flexibility and personalization of the whole process. Following, we apply an agentoriented modeling approach (AORML), which proved to be an effective modeling language for our purposes. On the one hand, AOR models have led us thoroughly to this specification of H&L, aiding us on a system’s requirements specification, analysis and initial design cycles. On the other hand, this experimentation has also provided us with feedback on how AORML can be extended, adding new constructs to facilitate agent-oriented modeling. Finally, this work has led us to the elicitation of relevant research focuses and questions (presented in section 7), which form our research agenda for the future. Acknowledgement. This paper has been written in the context of Agent Academy: A Data Mining Framework for Training Intelligent Agents, a project funded by the European Union 5th Framework IST Programme. We would like to thank the anonymous referees for their valuable suggestions, which have helped to improve this paper.

References 1. 2. 3. 4. 5. 6. 7. 8.

Benjamin, R., Contreras, J., Corcho, O., Gomez-Pérez. A.: Six Challenges for the Semantic Web. In KR2002 Semantic Web Workshop. (2002) Bonifacio, M., Bouquet, P., Mameli, G., Nori, M. Peer-Mediated Distributed Knowledge Management. In Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management, Stanford University, California, USA (2003) Bruckman, A.: MOOSE Crossing: Construction, Community, and Learning in a Networked Virtual World for Kids. PhD Thesis, MIT Media Lab, at: http://asb.www.media.mit.edu/people/asb/thesis/ (1997) Caire, G. et al.: Agent Oriented Analysis using MESSAGE/UML In: Proceedings of the Workshop on Agent Oriented Software Engineering (2001) Dignum, V. : An Overview of Agents in KM, at: http://www.cs.uu.nl/~virginia/ #Publications (2002) Fischer, G., & Ostwald, J.: KM - Problems, Promises, Realities, and Challenges. In IEEE Intelligent Systems, Vol. 16, No. 1, Jan/Feb’01. (2001) Ferran-Urdanet, C.: Teams or Communities - Organizational Structures for KM. In Proc. of ACM SIGCPR’99 (1999) Maçada, D. L., Tijiboy, A. V.: Aprendizagem Cooperativa em Ambientes Telemáticos (In Portuguese) (Cooperative Learning in Telematic Environments). In Proc. of IV RIBIE Conference, Brasilia, Brazil (1998)

72 9.

10. 11. 12.

13. 14. 15. 16.

Renata S.S. Guizzardi, Lora Aroyo, and Gerd Wagner Molani, A., Perini, A., Yu, E., Bresciani, P. Analyzing the Requirements for Knowledge Management using Intentional Analysis. In Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management, Stanford University, California, USA (2003) Nejdl, W., Wolf, B., Qu, C., Decker, S., Sintek, M., Naeve, A., Nilsson, M., Palmer, M., Risch, T. EDUTELLA: P2P Networking Infrastructure Based on RDF. In Proceedings of WWW2002, May 7-11, Honolulu, Hawaii, USA. (2002) Odell, J., Parunak, H., Bauer, B.: Extending UML for Agents. In Proceedings of the Agent-Oriented Information Systems Workshop at the 17th National conference on Artificial Intelligence (2000) Souza, R. S., Menezes, C. S.: Aplicando Técnicas de RI no apoio às Interações Mútuas de uma Comunidade Virtual de Aprendizagem – Um Ambiente Orientado a Agentes (in Portuguese) (Applying Information Retrieval Techniques to support mutual interactions in a Learning Virtual Community – An Agent Oriented Environment). MSc. Thesis. Federal University of Espirito Santo, Brazil. (2001) Ulbrich, A., Ausserhofer, A.: Aspects of Integrating E-learning Systems with KM. In Proc. of IASTED 2002, Cancun, Mexico. (2002) Vassileva J. Supporting Peer-to-Peer User Communities. In R. Meersman, Z. Tari (Eds.): CoopIS/DOA/ODBASE 2002, LNCS 2519, pp. 230-247. Springer-Verlag. (2002) Viant Innovation Center: The Human Side of P2P: where technology and conversation come together, at: http://www.viant.com/downloads/innovation_p2p.pdf Wagner, G.: The Agent-Object-Relationship Meta-Model: Towards a Unified View of State and Behavior. Information Systems, 28:5 (2003)

Towards Evaluation of Peer-to-Peer-Based Distributed Knowledge Management Systems Marc Ehrig, Christoph Schmitz, Steffen Staab, Julien Tane, and Christoph Tempich Institute AIFB, University of Karlsruhe, 76128 Karlsruhe, Germany http://www.aifb.uni-karlsruhe.de/WBS/ {ehrig,schmitz,staab,tane,tempich}@aifb.uni-karlsruhe.de

Abstract. Distributed knowledge management systems (DKMS) have been suggested to meet the requirements of today’s knowledge management. Peer-to-peer systems offer technical foundations for such distributed systems. To estimate the value of P2P-based knowledge management evaluation criteria that measure the performance of such DKMS are required. We suggest a concise framework for evaluation of such systems within different usage scenarios. Our approach is based on standard measures from the information retrieval and the databases community. These measures serve as input to a general evaluation function which is used to measure the efficiency of P2P-based KM systems. We describe test scenarios as well as the simulation software and data sets that can be used for that purpose.

1 Introduction Many enterprizes have spent large amounts of money to implement centralized knowledge management systems to keep in business in today’s knowledge-based economy, often with little success. Among others [1] suggest a distributed approach to knowledge management which better fits organizations and their employees. Participants can maintain individual views of the world, while easily sharing knowledge in ways such that administration efforts are low. The distributed environment is implemented by a peer-to-peer network (which is basically equivalent to a system of distributed agents) without any centralized servers. P2P systems have been used for collaborative working or file sharing, but knowledge sharing applications herein mostly relied on keyword search and very basic structures. Modern (centralized) knowledge management systems are based on ontologies which have shown to be the right answer for problems in knowledge modelling and representation [2]. An ontology [3] is a shared specification of a conceptualization. Through their structure ontologies allow answering a wider range of queries than standard representations do. Semantic Web technologies can augment this [4]. Current research projects1 attempt to exploit the best of the two worlds. Specifically, we want to do semantic information retrieval in a peer-to-peer environment - resulting in a Distributed Knowledge Management System (DKMS). 1

SWAP (swap.semanticweb.org) and Edutella (edutella.jxta.org)

L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 73–88, 2003. c Springer-Verlag Berlin Heidelberg 2003

74

Marc Ehrig et al.

In this work, we suggest a framework for evaluation of such distributed knowledgebased systems. Only through a thorough evaluation we can gain the insights to further develop and enhance ideas and systems. Evaluation is either possible through userbased evaluation or system evaluation. User-based evaluation measures the users satisfaction with a system, system evaluation compares different systems with respect to a given measure. While system evaluation permits a more objective confrontation of different approaches, the correlation of the results with the final user satisfaction is not always clear. However, user-based evaluation is expensive, time-consuming and it is difficult to eliminate the noise which is due to user experience, user interface and other human specific factors. Tools developed within the cited projects focus on the technical aspects of knowledge management. Thus, we use the system evaluation approach. The need for a standard evaluation mechanism is also recognized in other papers in this book e.g. [5]. Techniques from traditional Information Retrieval [6] and networking research [7] will have to be combined with ontology specific measures to gain meaningful results. This paper is structured as follows: in the first section we will introduce a set of use cases to illustrate the different dimensions of the problem at hand. A definition of evaluation measures will be given in the second section. In the following section we want to give a notion of tools which can be used. A part on generation of test data for these simulations follows in the succeeding section. Further we give a practical view describing the test parameters. Related and future work conclude this paper.

2 Scenarios The field of possible applications for peer-to-peer computing is huge. Currently running systems include file sharing (e.g. Gnutella2 [8]), collaboration (e.g. Groove3), computing (e.g. Seti@home4 ), knowledge management [9], to name but a few. For this reason we provide some scenarios for DKMS we examine. Various conclusions for our ontology based KMS will be drawn from these scenarios. 2.1 Application Scenarios This part will describe some real life situations in order to find characteristics which influence the distribution of information within the examined scenarios (Figure 1). The purpose of the scenarios is not to give a detailed impression of the entire IT-structure within the scenario. But rather to emphasize the challenging points for an ontology based KM-system realized by a peer-to-peer network. Corporation With their organization in many different units, entire corporations impose the most complex situation, with respect to number of domains, conceptualizations and 2 3 4

http://www.gnutella.com http://www.groove.net http://setiathome.ssl.berkeley.edu/

Towards Evaluation

75

Fig. 1. Scenarios Overview

documents, we want to consider here. Typically these units are distributed according to organizational tasks, like accounting or human resources, or more product related such as development and marketing. The product related units, for example, work on one product (topic) with diverse perspectives or on varying products with similar views, viz. use the same vocabulary. We assume each peer5 has its own ontology, but with the addition that employees working in similar business units use ontologies which have some concepts in common while ontologies in unrelated units describing e.g. the same product are not easily comparable, viz. use a different hierarchy and vocabulary. Our evaluation has to show which techniques make best use of existing ontologies in order answer queries according to the user needs. These demands will be examined precisely in the future. Therefore queries must reach quickly the peers which can answer them, without flooding the network. The answers should be relevant with respect to the query. Further the demand for computer resources like storage and processor time has to be monitored. Working group A special case within a big company is the single department. In this case the domain is predefined and terms with the same meaning occur more often in each ontology. However, the demand in terms of retrieval accuracy increases. A major research question here is, how to capitalize on ontologies from other peers. viz. Selforganization is often cited as one of the advantages of peer-to-peer systems. If every peer partly conceptualizes information the combination will result in a more detailed description for everybody, because each peer can add concepts from other peers to its own structure. 5

A peer can be the computer system of one user or a general database.

76

Marc Ehrig et al.

Very structured department A department of the kind using a very structured process. In this case it is possible to define and implement a single ontology which any employee has to follow. 2.2 Summary To summarize the single cases from an ontological point of view we distinguish two dimensions. The number of domains which are conceptualized and conceptualizations used for one domain. From the combination four possibilities evolve. This observation is in line with the suggestions in [10]. nm ontologies Each peer uses its own ontology. These ontologies conceptualize different domains. n1 ontologies Each peer uses its own ontology, but all peers conceptualize the same domain. 1m ontologies There is one general ontology, but it conceptualizes many domains. The peers use only parts of the entire ontology. But they can be merged from a top level perspective. In this case two different possibilities evolve: disjoint The peers commit to a particular part of the ontology. Hence two peers use either the same or a different ontology. overlapping Each peer has parts of the ontology without respect to the ontologies others are using. 11 ontology Each peer uses the same ontology in one domain. From a technical point of view we consider networks with a small numbers of peers to huge corporate networks. This means, that different routing strategies have to be analyzed. The evaluation criteria are straightforward. In all cases the relevance of the answer should be high and response time low using little resources of the peers. Further aspects are the network behavior if single peers fail or return wrong answers. Ontologies provide means to define contexts. The effects on these criteria through incorporation of meaning will be evaluated. The case studies have demonstrated the kind of peer-to-peer system we focus on. To evaluate our techniques we use well established measures from the Information Retrieval and Peer-to-Peer community, but we also have to introduce new ones which take the use of ontologies into account. These measures are described in the following section.

3 Evaluation Functions and Their Parameters This section presents a theoretical model of evaluation. In a general overview we define the evaluation function followed by its premises. Additionally we present ideas of which input and output parameters can be of interest in a DKMS.

Towards Evaluation

77

3.1 Evaluation as a Function One can imagine our DKMS as a black box doing information retrieval in a Semantic Web environment. This black box is supposed to have a certain behavior from which evaluation figures result giving us insight into the DKMS. To test and measure this behavior we can adjust different input parameters and collect the output figures. This can be modelled as a function. The function (F ) describes the setting and the basic algorithms used, that is, the interior of our system. Different parameters are used as input (in ) e.g. the number of peers. Specific output figures (om ) result from it, e. g. relevance or performance measures. Input and output in this context are not queries and answers of the peer network, they rather are parameters of the DKMS and its methodologies. F → (o1 , o2 , . . . , om ) (i1 , i2 , . . . , in ) − Having discussed the correlation between input and output one can adjust the parameters until an optimal solution is found. This approach is designed along an implementation line with the function representing the hard-coded program and the parameters being variables of it. 3.2 Function Modelling The function depends on the algorithms and other properties which will be described further. Topology. The topology is crucial for the network load imposed by each query. Do we want to evaluate random graphs, the star topology, or the HyperCuP environment [11]? Further the content of each peer (and its semantic context) could be used for building a network structure. Document distribution. The distribution of the documents in a real peer-to-peer system is hardly random. The influence of different document distributions on the output figures will be evaluated. Query language. The query language defines the expressiveness of queries. It can be interesting to compare performance results of the peer-to-peer system between query languages which only allow conjunctions or disjunctions and query languages which allow complex recursive queries. Selection function. Having a peer structure and a formulated query the next step is to find good ways of matching them. How to select and route to the best peers is a core component [12]. For the reader it might be confusing why the mentioned points belong to function rather than to input parameters. In a way the function premises are also input parameters. The difference lies in fact that they are explicitly modelled in the algorithm and can not be changed easily. Input parameters on the other hand are more flexible and can be adjusted by changing the value of a variable of the algorithm, the algorithm itself will stay the same. The next paragraph will show this. 3.3 Input Parameters A list of possible input parameters that can be entered into the system will follow: Number of peers. The size of the peer-to-peer network affects the results of the system. The scalability of the system is represented by this number.

78

Marc Ehrig et al.

Number of documents or statements. Another type of scalability is checked with this parameter. Whereas peers are physical locations, this parameter describes content objects. They represent the smallest entities in the system. Structure. Most of the decisions around topology directly influence the function. But depending on the chosen topology different parameters can be used for further adjustment. When using indexes an important figure is the index size. How much content will eventually be stored in the network and how detailed is the knowledge about other peers. Slightly different is the level of connectivity or the size of the routing table. These are figures representing the characteristics of the network. 3.4 Output Figures The output figures of evaluation functions ensure comparability to other systems. As the area of semantic peer-to-peer systems is rather new, there are no established standard evaluation functions which makes it difficult to fulfill the first mentioned requirement. The following list will provide well-known evaluation functions from related research fields. Relevance. Relevance is the subjective notion of a user deciding whether the information is of importance with respect to a query. Approximations can be done using, e. g., keywords. For comparison purposes one could imagine to have a rating between 0 and 1 for each answer. Recall R. Recall is a standard measure in information retrieval. It describes the proportion of all relevant documents included in the retrieved set. R = |relevant∩retrieved| |relevant| Precision P. Precision is also a standard measure in information retrieval. It describes the proportion of a retrieved set that is relevant. P = |relevant∩retrieved| |retrieved| F-measure F. Several combinations of the two first mentioned measures have been developed. The most common one is the F-measure [13] describing the normalized symmetric difference between retrieved and relevant documents. 2 R F = (ββ 2+1)P P +R with β = P/R Information loss. A measure to evaluate the loss of information which occurs when a query must be generalized on the answering peer. This might happen if the queried ontology does not contain a specific concept, but one which is more general and included in the ontology of the requesting peer [14]. Reliability. This can be split into two sub-parameters. Fault tolerance describes which degree of failures and problems are still tolerated until the system finally breaks down. Failures in a DKMS can be a peer leaving the network or unacknowledged messages. The failure rate specifies the percentage of actual breakdowns of the whole system. Real time. This measures the time from sending off the query to getting a result. As this figure is critical for end users, we take it into consideration as well. It was used in [15]. Network load. This technical figure can be measured with different sub-parameters. This is especially important for internal technical measurements [16]. Messages per

Towards Evaluation

79

query traces to what extent the network is being flooded by one query. The number of average hops can indicate how goal-oriented a query is routed and how fast a answer may be returned. Time to satisfaction. It is a combination between relevance and real time, with relevance having to exceed a certain value[17]. Again this is a very subjective figure. 3.5 Output Combination We have set up a theoretical model for evaluation. The benefit of semantic peer-topeer lies not on its single areas but its strength actually is the combination of them. Just like the input parameters come from the different areas of peer-to-peer, Semantic Web, and information retrieval, it is also necessary to unite the output figures to achieve meaningful results. A possibility would be to arrange linear combinations. Normalized vectors represent another. The combination of different output figures will finally allow us to decide upon the quality of the new system. The output figures will be provided using a simulation package.

4 P2P Network Simulation P2P systems are not set up and maintained by a central authority; thus, creating and observing a non-trivial network and measuring the evaluation functions as described in the previous section is a hard task. Simulation can help to gain insight into the behavior of the system. Many research contributions such as Freenet [18] and Anthill [19] have used simulations in order to demonstrate the performance of their systems. Simulation is a core component for evaluation. 4.1 Discrete Event Simulation (DES) Discrete Event Simulation observes the behaviour of a model over time [20]. The model has a state described by variables of the model that completely define the future of the system. The state of the model is usually encapsulated into a set of entities (cp. objects in OOP). Discrete Events changing the state of the system occur at discrete points in time (as opposed to continuous state changes). Events may trigger new events. Statistical Variables define the performance measure the user is interested in. This could be something like “average load on the server” or “maximum queue length”. Event oriented DES describes the dynamic behavior of the system solely by a sequence of events; the actions triggering the events are not considered. Process oriented DES combines the entities containing the state of the system and the actions that cause events (cf. OOP). Typical Components of Simulation Software Packages DES software typically includes abstractions for entities, connections between entities, and events transmitted on those connections (see fig. 2), which corresponds well to the P2P scenario. Process oriented packages also include an abstraction for processes running on entities. Some simulators provide a glue language which can be used to compose models easily.

80

Marc Ehrig et al.

Entity e1

e2

Events

Processes

e3 Connection

Fig. 2. Abstractions in DES packages 4.2 Simulation Packages Several simulators, all of which contain the abstractions mentioned above, were examined in more detail: SSF in its incarnations DaSSF6 (C++-based) and Raceway SSF7 (Java); OMNet++8 ; JavaSim9 ; Simjava10 ; JADE11 . Simulators Feature Matrix We only present a short overview of the details, as the features of the above mentioned packages are similar. Name Raceway SSF DaSSF Omnet++ JavaSim SimJava JADE

Language Java C++ C++ Java Java Java

Distributable Glue language DML MPI DML PVM/MPI NED Tcl/Python CORBA

Table 1. Overview of different simulations systems

6 7 8 9 10 11

http://www.cs.dartmouth.edu/ jasonliu/projects/ssf/ https://gradus.renesys.com/exe/Raceway http://www.hit.bme.hu/phd/vargaa/omnetpp.htm http://javasim.cs.uiuc.edu/ http://www.dcs.ed.ac.uk/home/hase/simjava/ http://sharon.cselt.it/projects/jade/; JADE is a special case here because it is an agent platform rather than a simulation package, but nevertheless it contains the same abstractions as the others.

Towards Evaluation

81

4.3 Selection Criteria for Simulation Infrastructures Performance considerations Qualitative experiments concerning model sizes and performance were conducted on a commodity PC in order to find out how large models could become. JavaSim could handle about 6000 entities, while all other Java-based systems are restricted to less than 1024 active entities. The C++ systems can handle hundreds of thousands of entities and process tens of thousands of events per second. Ease of implementation Undoubtedly, a model based on a Java-based API and simulator is much easier to program and debug than C++. The exception handling and debugging capabilities of the Java language facilitate a rapid generation of models. Glue languages, graphical environments While glue languages and graphical editors are useful in order to get started, they may not be able to cope with complicated and/or large models. In that case, a clean programming interface on the C++/Java level is crucial. 4.4 Conclusion A Java-based simulator would be much easier to get started with. On the other hand, for large models the C++ systems are able to handle numbers of entities that are two orders of magnitude larger than those of the Java systems. We are currently implementing a simulation environment with JAVA using the SSF framework.

5 Data Generation - Evaluation Datasets No evaluation can be done without using a dataset which we can query on the semantic level. The choice of this dataset will be influenced by different criteria. First, we need to consider what type of semantic data we want to query. Then we explore the problem of how the data should be distributed on the network. 5.1 Data Understanding One can see the peer-to-peer network as a network of repositories called peers. Each provides a set of resources, which we will call documents. Every single document is then described by some sort of schema shared across the network. In usual peer-to-peer systems, the metadata is provided with a very simple schema of fields containing plain text (e.g: title, author or format of the document considered). The query must then have the following form: return all instances (approximately) having the following values v1 , ..., vk in the fields f1 , . . . , fk . A semantic query in the peer-to-peer network is a query using the semantic metadata available on the given objects. The structure of the metadata is enriched in order to address certain issues. We list them here with a few words of explanation as well as the kind of query that might require it. 1. identity problem: simple values do not determine identity e.g return all the documents written by the CEO aka Mr Johnson

82

Marc Ehrig et al.

2. relational problem: values in fields are not links to other objects e.g return all the papers of scientist who had an accepted paper at the SIGMOD conference 3. the classification problem: a given classification needs to be shared to be useful on the network 12 ? e.g return all documents written by university professors 4. the inference problem: inclusion of classes also need to be shared e.g return all documents on semantics theory in Computer science ( not equal to the intersection of “semantics theory” and “Computer science”) Moreover, these different types of queries can be combined, for example: e.g return the different names of all oil companies quoted at the Stock Exchange. An ontology is a metadata schema whose semantic description addresses these issues.

Ontology Hierarchy Relations Classification Fields Fig. 3. Different types of metadata schema

5.2 Generating the Data - Existing Data A possible approach for the evaluation of a semantic peer-to-peer network is to define a generation mechanism, which generates data with a semantic-like structure. However, for evaluation purposes the difficulty of deciding whether the data generated corresponds to typical real data might turn out to be a drawback. A second approach is to use existing datasets and distribute them over the network. However, for certain types of datasets of very specialised domains, the drawback here will be that one might have some difficulty to interpret the results (for example MEDLINE dataset). Table 2 summarizes the different datasets considered. Corpus name nb of Docs DBLP 310000 Reuters 21578 21,578 Reuters 2002 806,791 DMOZ 190,000,000 Medline 1141893

Text? type no relational yes classification yes small hierarchy yes hierarchical no ontology

Table 2. Datasets of different types 12

the difference to keyword-based is that the values belong to predifined values, shared in the network

Towards Evaluation

83

Each row corresponds to a given corpus we considered, whereas the columns are criterias of interests for our purpose. DBLP13 is a computer science article bibliography database. Medline14 offers the same purpose for medicine. Both of the Reuters datasets15 are newswire collections from Reuters. DMOZ16 is a collection of internet links organized in a hierarchy. The following criteria have been considered for our evaluation: the number of documents, whether the texts of the documents are available, and then the kind of metadata schema used. 5.3 Distributing the Dataset Once the dataset has been chosen, it must be distributed on the peers of the network. For this, different possibilities might be chosen depending on the structure of the network and of the datasets. Of course, it is always possible to distribute the content among the peers randomly. However, this is probably not going to be the case in a peer-to-peer network. For instance, on a given peer it is more likely to find the similar content than on any other peer. Thus, other data distribution schemes have to be chosen according to the test scenarios we want to cover.

6 Test Scenarios In this section we suggest several test scenarios to evaluate different retrieval strategies. As discussed before the function modelling as well the input parameters influence the performance of the peer-to-peer system. It is essential that we examine each parameter separately to obtain meaningful results. Therefore we now outline the dimensions of our special interest. 6.1 Ontology The first dimensions is the number of ontologies as discussed in the scenarios. Since already available ontologies are rare we will use different approaches to generate different kinds of ontologies. To generate different ontologies out of one general ontology it is possible to take the existing one and to exchange concept names with synonyms and deleting other concepts completely. However, we keep in mind that the generating strategy will certainly influence our results. The availability of ready made ontologies will certainly increase as projects like [21] proceed. 6.2 Matching Algorithms In the test cases with more than one general ontology mappings must be applied to identify the concepts with the same meaning on different peers. Some strategies are 13 14 15

16

http://www.informatik.uni-trier.de/ ley/db/ http://www.nlm.nih.gov/bsd/sample records avail.html http://kdd.ics.uci.edu/databases/reuters21578/reuters21578.html http://about.reuters.com/researchandstandards/corpus/index.asp http://www.dmoz.org

84

Marc Ehrig et al.

already available like [22, 23, 24] which are based on lexical, textual and structural matchings. Other are focusing on statistical information [25]. An interesting approach was presented at the symposium. [26] uses knowledge maps and keyword assignment to documents to identify keyword assignment differences within communities. 6.3 Network Topology and Routing The network topology directly influences the available routing algorithms. In the firework routing model [27], for example, a query is forwarded until a peer knows something about the query. The query is then distributed to all peers in the neighborhood. Within a semantic context the network topology must support that peers with similar domains know each other to use this model efficiently. 6.4 Query Language Query languages will be tested which support simple keyword based search but also complex recursive queries. Besides, it is also important to consider the construction of a query. Possibilities are to take just the actual chosen keyword and concepts or to expand them with various techniques. 6.5 Number and Distribution of Documents or Statements Our scenarios do not impose restrictions to the number of documents within the peer-topeer network. However, the distribution of the documents will influence our results. In distributed database research documents are generally distributed uniformly [28]. When semantics come into play this does not seem appropriate. A short analysis of the Yahoo! categories suggests, that not only single words in documents are following a Zipf distribution [29], but also the allocation of documents in categories. Besides statistical distribution functions we also consider to distribute clusters. Different methods can be used to cluster our documents [30]. 6.6 Number of Peers The number of peers surely influences the behavior of any peer-to-peer system. Therefore we will use ranges from small (about 10) to big (about 106 ) scenarios in our test cases. As with documents, the distribution of peers in a network at large follows powerlaws [31]. 6.7 Structure The available resources on each peer influences the information a peer can hold about the others. In this early phase of our projects we will not pay too much attention to this parameter but rather put as much (within reasonable limits) information on each peer as we need. The same holds for the computational effort on the peers. Therefore we distinguish three cases

Towards Evaluation

85

1. All peers offer only a limited resources to the network. 2. Some peers offer a lot some peers offer limited resources. 3. All peers offer a lot of resources to the network. Figure 4 gives an overview of the different scenarios.

Fig. 4. Test scenarios Overview

7 Related Work We found that there are different communities coping with the task of retrieving information from knowledge sources. They use either system evaluation or user-based evaluation. Classical information retrieval from text documents is mostly affected by technical changes to the system. Therefore they predominantly use system evaluation to compare different methodologies [6]. Closest to the peer-to-peer approach in information retrieval are the results from e. g. [28, 32] to search distributed databases. The focus has been to choose from a known set of databases where the structures are known. The selection was made on keyword based criteria. As a testing environment the TREC dataset17 was chosen and the different documents where distributed uniformly according to their creation date. Our approach adds new dimensions to these results since the total number of peers is not known; neither are the information structures on the peers. Further we introduce new methods to distribute data on different peers. Research in Ontology based search in distributed environments has been conducted with systems like OBSERVER [14]. The focus was rather to find strategies for better 17

http://trec.nist.gov

86

Marc Ehrig et al.

information retrieval in one particular case in our scenario than on comparing different strategies for many scenarios as is proposed here. The first user-based evaluation of an ontology based KM system was realized by [33]. It delivers encouraging results about the use of ontologies to retrieve knowledge. In contrast to our scenarios the tests were accomplished on a centralized system using one ontology. Efficient file allocation with hashing algorithms in peer-to-peer networks has been the focus in research such as [34]. However, this approach is feasible only for rather simple knowledge representation as necessary for music-file search, where keyword matching on a file name may be sufficient. [7] introduced a function to calculate search costs in peer-to-peer networks and algorithms to optimize the function with respect to varying parameters. The peer selection is based on rather simple meta information such as response time. We want to advance this approach including more content based meta information. [35] have summarized evaluation attempts regarding the economic-financial and the socio-organizational viewpoint of KM. The research in this area is complementary to our approach. Furthermore there is ample research on evaluation methods to classify response times of databases, e.g. TPC18 and other technical aspects of information retrieval. This is also complementary to our suggestions. Our impression is that there is a lot of research dealing with certain aspects of peerto-peer systems and knowledge management but no general framework to compare the different systems.

8 Conclusion We have examined the problem of evaluating a distributed knowledge management system. While evaluation of a centralized KM system is a challenging task in itself, the distributed case adds more parameters to the evaluation function. The well-known notions of precision and recall are not sufficient to evaluate the performance of a DKMS. A performance measure for DKMS must include semantic retrieval quality as well as measures from the P2P field like the number of hops needed to answer a query. Simulation packages for testing have been investigated. For traditional database and information retrieval systems the generation of test data has been examined, and standardized data collections are supplied. In our case of P2P knowledge management, neither standardized data generation methods nor test data sets are available. We have made suggestions on how that problem may be tackled; it will have to be verified that the test data we generated are valid in the sense that they resemble real-world data from use cases like ours according to certain similarity measures. Different application scenarios have shown a variety of possible uses for a DKMS which have different impacts on the performance of the system, and thus on the evaluation process. 18

http://www.tpc.org/

Towards Evaluation

87

Our suggested framework for evaluation can be used as a basis for future research and development of distributed knowledge management systems. Acknowledgements. Research reported in this paper has been partially financed by EU in the IST project SWAP (IST-2001-34103) as well as by the German Federal Ministry of Education and Research (BMBF) in the PADLR project.

References [1] Bonifacio, M., Bouquet, P., Traverso, P.: Enabling distributed knowledge management: Managerial and technological implications. Novatica and Informatik/Informatique III (2002) [2] OLeary, D.: Using ai in knowledge management: Knowledge bases and ontologies. IEEE Intelligent Systems 13 (1998) 34–39 [3] Gruber, T.R.: Towards Principles for the Design of Ontologies Used for Knowledge Sharing. In Guarino, N., Poli, R., eds.: Formal Ontology in Conceptual Analysis and Knowledge Representation, Deventer, The Netherlands, Kluwer Academic Publishers (1993) [4] Maedche, A.: Emergent semantics for ontologies. IEEE Intelligent Systems (2002) [5] Moreale, E., Watt, S.: An agent-based approach to mailing list knowledge management. [36] To appear 2003. [6] Voorhees, E.M.: The philosophy of information retrieval evaluation. In Peters, C., Braschler, M., Gonzalo, J., Kluck, M., eds.: Evaluation of Cross-Language Information Retrieval Systems, Second Workshop of the Cross-Language Evaluation Forum, CLEF 2001. Volume 2406 of Lecture Notes in Computer Science., Darmstadt, Germany, Springer (2002) 9–26 [7] Yang, B., Garcia-Molina, H.: Efficient search in peer-to-peer networks. In: Proceedings of the International Conference on Distributed Computing Systems (ICDCS). (2002) [8] Kan, G.: Gnutella. In: Peer-to-Peer: Harnessing the Power of Disruptive Technologies. O’Reilly (1999) 94–122 [9] Bonifacio, M., Bouquet, P., Mameli, G., Nori, M.: Peer-mediated distributed knowldege management. [36] To appear 2003. [10] Wache, H., Voegele, T., Visser, T., Stuckenschmidt, H., Schuster, H., Neumann, G., Huebner, S.: Ontology-based integration of information - a survey of existing approaches. In Stuckenschmidt, H., ed.: IJCAI-01 Workshop: Ontologies and Information Sharing. (2001) 108–117 [11] Schlosser, M., Sintek, M., Decker, S., Nejdl, W.: HyperCuP - hypercubes, ontologies and efficient search on P2P networks. In: Proceedings to the Eleventh International World Wide Web Conference, Honolulu, Hawaii, USA (2002) [12] Carzaniga, A., Wolf, A.L.: Content-based networking: A new communication infrastructure. In: Proceedings of the NSF Workshop on an Infrastructure for Mobile and Wireless Systems, Springer-Verlag (2002) [13] Van Rijsbergen, C.J.: Information Retrieval, 2nd edition. Dept. of Computer Science, University of Glasgow (1979) [14] Mena, E., Kashyap, V., Illarramendi, A., Sheth, A.P.: Imprecise answers in distributed environments: Estimation of information loss for multi-ontology based query processing. International Journal of Cooperative Information Systems 9 (2000) 403–425 [15] Nodine, M., Bohrer, W., Ngu, A.H.H.: Semantic brokering over dynamic heterogeneous data sources in infosleuth. Technical report, MCC (1998)

88

Marc Ehrig et al.

[16] Yang, B., Garcia-Molina, H.: Designing a super-peer network. Technical report, Stanford (2002) [17] Yang, B., Garcia-Molina, H.: Improvig search in peer-to-peer networks. Technical report, Stanford University (2002) [18] Clarke, I., Sandberg, O., Wiley, B., Hong, T.W.: Freenet: A distributed anonymous information storage and retrieval system. In: Proceedings of the Workshop on Design Issues in Anonymity and Unobservability, Berkeley, California (2000) [19] Montresor, A., Meling, H., Babaoglu, O.: Towards adaptive, resilient and self-organizing peer-to-peer systems. In: Proceedings of the International Workshop on Peer-to-Peer Computing, Pisa, Italy (2002) [20] Ball, P.: Introduction to discrete event simulation. In: Proceedings of the 2nd DYCOMANS workshop on ”Management and Control : Tools in Action”. (1996) [21] Pease, A., Li, J.: Agent-mediated knowledge engineering collaboration. [36] To appear 2003. [22] Magnini, B., Serafini, L., Speranza, M.: Linguistic based matching of local ontologies. In: Workshop on Meaning Negotiation (MeaN-02), Edmonton, Alberta, Canada (2002) [23] Modica, G., Gal, A., Jamil, H.M.: The use of machine-generated ontologies in dynamic information seeking. In: Proceedings of the Ninth International Conference on Cooperative Information Systems (CoopIS 2001), Trento, Italy (2001) [24] Melnik, S., Garcia-Molina, H., Rahm, E.: Similarity flooding: A versatile graph matching algorithm. In: Proc. 18th ICDE Conf. (2001) [25] Steels, L.: The origins of ontologies and communication conventions in multi-agent systems. Autonomous Agents and Multi-Agent Systems 1 (1998) 169–194 [26] Novak, J., Wurst, M., Fleischmann, M., Strauss, W.: Discovering, visualizing and sharing knowledge throught personalized learning knowledge maps. [36] To appear 2003. [27] Ng, C.H., Sia, K.C.: Peer clustering and firework query model. Technical report, The Chinese University of Hong Kong (2002) [28] Callan, J.P., Connell, M.E.: Query-based sampling of text databases. Information Systems 19 (2001) 97–130 [29] Zipf, G.K.: Human Behavior and the Principle of Least Effort: An Introduction to Human Ecology. Addison-Wesley, MA (1949) Reading. [30] Sebastiani, F.: Machine learning in automated text categorization. ACM Computing Surveys 34 (2002) 1–47 [31] Faloutsos, M., Faloutsos, P., Faloutsos, C.: On power-law relationships of the internet topology. In: SIGCOMM. (1999) 251–262 [32] Gravano, L., Garc´ıa-Molina, H.: Generalizing GlOSS to vector-space databases and broker hierarchies. In: International Conference on Very Large Databases, VLDB. (1995) 78–89 [33] Iosif, V., Sure, Y.: Exploring potential benefits of the semantic web for virtual organizations. In: PAKM to appear. (2002) [34] Stoica, I., Morris, R., Karger, D., Kaashoek, F., Balakrishnan, H.: Chord: A scalable PeerTo-Peer lookup service for internet applications. In: Proceedings of the 2001 ACM SIGCOMM Conference. (2001) 149–160 [35] Dieng, R., Corby, O., Giboin, A., Ribiere, M.: Methods and tools for corporate knowledge management. Int. Journal of Human-Computer Studies 51 (1999) 567–598 [36] van Elst, L., Dignum, V., Abecker, A., eds.: Proceedings of the AAAI Spring Symposium “Agent-Mediated Knowledge Management (AMKM-2003)”. Lecture Notes in Computer Science (LNCS), Stanford, CA, USA, Stanford University, Springer (2003) To appear 2003.

TAKEUP Trust-Based Agent-Mediated Knowledge Exchange for Ubiquitous Peer Networks Stefan Schulz, Klaus Herrmann, Robert Kalckl¨ osch, and Thomas Schwotzer Berlin University of Technology, Institute for Telecommunication Systems, Sekr. EN6, Einsteinufer 17, 10587 Berlin, Germany {schulz, kh, rkalckloesch, thsc}@ivs.tu-berlin.de

Abstract. Agent-mediated Knowledge Management is a promising approach to handle and maintain knowledge, especially in a distributed and mobile environment. One example for such an environment is a mobile community: mobile individuals that group together because they share the same interests. The members of a mobile community maintain a common knowledge base by exchanging and managing information via mobile devices (e.g. PDAs). Trust is essential in such an environment for the assessment of communication partners and their knowledge. This paper presents our conceptual framework for trust-based agentmediated knowledge exchange with respect to the highly distributed environment of mobile communities. We extend an existing knowledge management system by using mobile agents that serve as autonomous delegates of mobile users. These mobile delegates spread out onto other devices and trade knowledge while the user is not present. When the user and one of his1 delegates get in contact again, a knowledge reconciliation takes place. We propose a trust-based approach to the process of autonomous knowledge acquisition. Trust is organized in a distributed way among the participants of a mobile community and serves as a basis for the decisions of mobile delegates and mobile users.

1

Introduction

After several years of research and application, knowledge management is still one of the most important challenges in IT infrastructures. Especially, as knowledge became a proﬁt criterion for businesses, it seems to be a critical factor to organize it and make it intelligibly and ubiquitously accessible. This gets even harder as the society becomes increasingly decentralized and mobile. A promising approach for solving distributed and mobile management tasks is the paradigm of software agents. Several current projects apply this paradigm to the area of knowledge management. Software agents are used for user support and personalization [8], organizing and sharing information over the Web 1

For the remainder of this paper, the male form is used for readability reasons, implying that male or female may perform the role.

L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 89–106, 2003. c Springer-Verlag Berlin Heidelberg 2003

90

Stefan Schulz et al.

[24], building and maintaining Organizational Memories [12], and supporting the fusion of knowledge [20]. Moreover, software agents are often applied as an intermediate layer to provide ubiquitous access and a uniﬁed information presentation for distributed, heterogeneous information sources. Our approach of trust-based agent-mediated knowledge exchange for ubiquitous peer networks (TAKEUP) can be seen as an extension to existing agentbased systems for mobile and stationary knowledge management systems: Firstly, to ease the process of ﬁnding and exchanging information and, secondly, to raise the acceptance of knowledge management by providing beneﬁts for both the organization and the individual. We aim at integrating the process of knowledge management into the users’ natural behavior. Managing knowledge exchange between mobile users that connect to each other in an ad hoc fashion via short-range radio technologies like Bluetooth is a tough challenge. However, individuals and organizations can beneﬁt from mobile knowledge management systems since modern citizens adopt mobile communication devices and get increasingly mobile in their daily life. Additionally, communication devices become more and more invisible and integrated in the environment, which results in ubiquitous peer-to-peer networks for information exchange. We propose the extension of an existing knowledge management framework (Shark) [27] for mobile ad hoc networks [13] with mobile agent technology to decouple the task of distributing and collecting knowledge from the physical presence of the user. Mobile agents autonomously position themselves in speciﬁc locations and act as virtual representatives of their users. They exchange knowledge with other users and agents while their own user resides in a diﬀerent physical location. To let mobile agents decide autonomously which pieces of information to exchange with whom, we introduce the notion of statistical trust management. Trust information is accumulated and propagated among the agents to enable them to gauge communication partners and the information received from them. We discuss the concept of a trust network that is motivated by recent discoveries in the area of complex social networks. We argue that the structure inherent to these networks of human relationships have a strong inﬂuence on mobile knowledge management and trust management because a person’s mobility is correlated with his social relationships. We exploit the tendency of social networks to self-organize into highly clustered structures with very eﬃcient communication paths between individuals. This phenomenon was ﬁrst discovered in the late sixties and has gained some popularity under the term small-world phenomenon. In the following section we shortly describe our previous work in the area of software agents and knowledge management. In section 3 we introduce the underlying concepts adopted from the domain of social networks and our conception of trust levels. In section 4 we present the technical aspects of TAKEUP, especially regarding the organization of knowledge and the integration of trust. This is followed by an example scenario in section 5 which illustrates the application of TAKEUP. Finally, conclusions and an outlook on future work are given in sections 6 and 7.

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

2

91

Previous Work

The basic technologies and mechanisms used to realize TAKEUP have been developed in various research activities. In this section we give an overview on each of the relevant technologies including a mobile agent-based middleware platform and a knowledge management framework. 2.1

MESHMdl – A Mobile Agent Platform for Mobile Devices

To cope with the fundamental challenges of knowledge management in mobile adhoc networks (MANETs) [13], we employ MESHMdl [16] as a mobile agent platform for mobile devices. MESHMdl allows developers to implement distributed applications for ad-hoc networks as groups of cooperating mobile agents. These agents are able to move from one mobile device to another at their own will if both devices are within each other’s transmission range. To deal with the dynamics of a MANET, MESHMdl decouples mobile agents by introducing tuple spaces as an asynchronous communication medium. Agents may write tuples (data objects) into the space and read or take (read and delete) tuples from the space. Read and take are available as blocking and non-blocking primitives. Moreover, an agent may request to be notiﬁed when a speciﬁc tuple is put into the space. Tuples are addressed associatively. They are read from the space by specifying the contents of the desired tuple partially in a so-called template. The space matches the existing tuples in the space with this template and returns matching tuples. On each participating device a MESHMdl runtime environment (called engine) is installed. Each engine runs its own space for the agents that are locally present. Agents on remote devices in transmission range have restricted access to an engine’s space. This simple communication paradigm allows agents to communicate and coordinate their actions in a very ﬂexible way without tying them too close together. This decoupling is very important in a MANET. Agents need to be ﬂexible enough to use their mobility in a timely fashion when a migration to another mobile device seems advantageous. MESHMdl agents are mobility-aware. Each engine presents a view of the other engines (devices) that are currently within transmission range. Thus, an agent may react directly to changes in the neighborhood. It may decide to migrate to a neighboring device or access its space to deposit tuples for the agents on that device. In TAKEUP, the ability to migrate, the ﬂexibility, the decoupling, and the asynchronous space communication are exploited by specialized mobile agents. These agents explore a mobile community and position themselves on key devices from where they can coordinate a user’s knowledge management process. 2.2

Shared Knowledge

The basis for TAKEUP is the Shark system [27]. We developed this technology for sharing knowledge in mobile peer-to-peer networks. In Shark, we apply

92

Stefan Schulz et al.

TopicMaps [30,26] as a knowledge representation language. In this approach, knowledge is assumed to consist of documents, which contain arbitrary data in arbitrary formats, and a layer describing the semantics of these documents. These semantics are deﬁned by a network of topics. By assigning documents to one or more topics, the documents’ semantics (context) is given.

Fig. 1. Two KEP scenarios with Bluetooth. Shark is a mobile extension to TopicMaps and the Semantic Web [29]. It provides a peer-to-peer knowledge exchange protocol (KEP) that is similar to KQML [11], but specialized in knowledge exchange and restricted in its functionality with respect to the limited resources of mobile devices. KEP comprises two phases: negotiation and exchange. During negotiation, the peers inform each other about the topics on which each of them is willing to receive or disseminate knowledge. Knowledge is exchanged, if one or more topics match. Figure 1 illustrates two KEP scenarios with Bluetooth. The Local Station is a stationary PC with an XML TopicMap engine. Mobile Stations run J2ME with a small Shark Knowledge Base. Whenever two Shark stations are close enough to establish a Bluetooth connection (about 10 meters) knowledge exchange can take place. In this example, one Mobile Station exchanges knowledge with both a Local Station and another Mobile Station.

3

Informal Networks and Trust

Several surveys and articles have examined the existence of informal networks in organizations (e.g. [25], [28]). The beneﬁt of such networks for the performance

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

93

of the whole organization is immense and has been analyzed extensively (e.g. [19], [31]). The key of informal networks is a natural process in which people become a part of communities. The general goal of such a community is to share knowledge on a common interest. Thereby, each of the members actually gains knowledge and the common knowledge in the community increases through the exchange of information. For each member, the community represents a collaboration platform, where both the organization and the individual beneﬁt. 3.1

Organized Anarchy

Self-organization plays a major role in our approach since communities are anarchic networks of individuals who share information. Anarchic, in this context, means that associations between people do not necessarily obey the structures imposed by superordinate organizations or institutions. Although, usually companies are hierarchically structured, the information ﬂow does not always follow these structures. The same holds for TAKEUP, where each individual deﬁnes his own interest proﬁle and exchange policies. What seems contradictory at ﬁrst glance does not thwart the original goal of managing knowledge. From our knowledge management point of view, individuals and communities (groups of individuals) behave in much the same way. Similar to a single individual, a community has its own interests and its own rules for exchanging knowledge. For example, a company may deﬁne a set of policies for knowledge management stating which content is relevant, conﬁdential, or inappropriate for the whole company. When an employee represents his company, he must obey these rules like any other employee. Thus, all members of the company community act in a consistent way and the community as such may be viewed as an individual in terms of knowledge exchange. However, the role as employees may diﬀer from the role (or proﬁle) they have in their private life. Therefore, we support role-based knowledge dissemination and allow users to take on diﬀerent roles. Moreover, the anarchic aspect of social relationships that do not follow the paths imposed, for example, by a company structure, turns out to be wellstructured in itself. This structure is embodied by a social network among users as we will see in section 4.2. In TAKEUP, knowledge management is performed by means of agents. Such agents act on behalf of a network member. Thus, a user takes along his personal agent and may also carry agents of other members. Communication in TAKEUP always is peer-to-peer, where peers are agents. If two individuals physically meet, with their mobile devices, actually several network members (i.e., their agents) meet to exchange knowledge. These knowledge agents exploit the structure of the social network among the users to self-organize a mobile knowledge network in an eﬃcient way (see section 4.2).

94

3.2

Stefan Schulz et al.

Levels of Trust

Related to the aspect of mobility and self-organization are the questions of ﬁnding actually needed or requested knowledge within the community and exchanging knowledge with the community members. The aspect of trust in individuals and acquired information plays a major role for communities and, thus, for exchanging knowledge. Trust has been studied as a means for evaluating the relationship between individuals (cf. [14]). In most cases an individual as such is regarded as being trustworthy. For knowledge exchange, we adopt the viewpoint of Marsh [23], who identiﬁed three aspects (dimensions) of trust: basic, general, and situational trust. In TAKEUP, the value of initial trust reﬂects the basic trust, i.e., the general willingness of an individual to share knowledge with others. With trustability we represent the general trust an individual has in a speciﬁc community member. Finally, with the notion of competence we map the situational trust to an individual’s trust in a communication partner regarding a speciﬁc topic (or domain), i.e., reﬂecting a member’s reputation with respect to that topic. Hence, a level of trust reﬂects the combination of initial trust, trustability, and competence. As trust is not a binary decision, we use a continuous interval between [-1, +1] for the trust values (cf. [23]). Naturally, a high level of trust in a person also renders information given by that person trustworthy, whilst a low level of trust usually does not. On the other hand, it is rather diﬃcult to judge information given by a person who is either not trustworthy or not competent. However, such information is not necessarily incorrect. If trustability is negative, a person can be regarded as being hostile. Therefore, the most diﬃcult case occurs when there is a lack of trustworthiness or competence. In this situation, trusted sources are needed to judge given information.

4

Trust-Based Knowledge Management

An environment for trust-based knowledge management must be able to cope with mobility and ﬂuctuation. For these requirements, an agent-based system oﬀers a technical platform, where mobile agents represent community members. They encapsulate a mechanism to manage topics and related information and implement trust-based exchange protocols. Equally important is the fact that the agent paradigm provides a natural way for representing the diﬀerent roles a person takes depending on his current situational context. 4.1

Principle of Roles

With very few exceptions, each individual in our society is part of at least one community. Communities are, e.g., companies, sport teams, fan clubs, and cliques. The context in which a communication takes place has a strong impact on communication behavior of the participants. People usually behave diﬀerently, e.g., at business talks or when playing tennis, because they act in diﬀerent roles.

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

95

The diﬀerence in communication behavior is twofold. Firstly, people ﬁlter information based on its relevance. The measurement of relevance is more or less subjective and strongly depends on the context. For instance, a weather forecast will be less interesting during working hours but more interesting at leisure time. Secondly, communities evolve information policies that deﬁne what kind of information is allowed for dissemination, to whom, and under which circumstances. These policies are more formal and restrictive in functional diﬀerentiated groups like companies, and more informal in private social groups. Nevertheless, people follow these (formal or informal) policies when acting as member of a community. Figure 2 illustrates the relationship of relevance, information policy, role and context. The person in the middle communicates with his environment in two diﬀerent roles due to diﬀerent situational contexts: business talk and leisure activity. Both, impression of relevance and communication policy depend on the person’s current role. This means, he will receive and disseminate diﬀerent information depending on the context.

Fig. 2. Relationship of relevance, information, policy, role, and context.

Applying agent technology, we encapsulate the roles of communities in separate knowledge operating agents. Such agents hold the role-speciﬁc topic proﬁle, communication policies, and trust values. The bases for these knowledge operating agents are the core components and protocols of the Shark system. 4.2

Organizing Knowledge in a Community

To extend Shark’s scope and to enable autonomous knowledge acquirement and management, we have developed two classes of MESHMdl agents: KnowGents (knowledge agents) reside on the users’ mobile devices and control their knowledge management processes. They acquire knowledge from other

96

Stefan Schulz et al.

users and maintain his local knowledge base. To enable a user’s knowledge management system to expand into the MANET of his community, the second class of mobile agents is needed. DeleGents (delegate agents) are proxies of a KnowGent (and thus, of the user) that may be sent into the community MANET. A KnowGent can autonomously take the decision to create and dispatch a new DeleGent to another device. While KnowGents represent the classical knowledge management application component, DeleGents explore a mobile community and position themselves on key devices from where they can coordinate the users’ knowledge management process. The basic assumption behind the idea of DeleGents is that the users within a community do not move and meet other people by chance. They are situated inside a social network. They have friends and colleagues that deﬁne their social environment and thus their communities. Therefore, the people a user meets are an indication for the structure of his social network. In addition, Shark provides a sophisticated mechanism for deﬁning and maintaining the topics that are of interest to a user. Finally, as the following sections will explain, information exchange among mobile users is based on trust. Thus, in order to autonomously acquire and distribute knowledge on behalf of its user, a DeleGent may use patterns of encounters with other users, those users’ topic proﬁles, and the accumulated trust information. Using Patterns of Encounter. In recent years the research on complex self-organized networks resulted in some remarkable insight in the structures underlying all sorts of diﬀerent networks. Among others, social networks, networks of humans and their relationships, turned out to be so-called small worlds. These small worlds are characterized by a low average path length between nodes (e.g. humans), a high tendency to build clusters or cliques, and the formation of hub nodes [2]. This structure allows for a very eﬃcient communication in such networks. Experiments proved that humans are quite eﬀective at using their acquaintances (i.e. only local information) to transport a message towards an individual of which only sparse information is given. Thus, incomplete local knowledge can be used to eﬃciently organize the ﬂow of knowledge. This is exactly the idea of using patterns of encounter. KnowGents maintain a local model of the world they perceive through their MESHMdl engine. Upon encountering another KnowGent (mobile user) a KnowGent records certain data about the other KnowGent’s user. Essentially, it registers which topics the other user is interested in. It then updates its local encounter profile (EP) in which it stores information, e.g., on which topics it encountered how often and for how long. Hub users will get in touch with many diﬀerent topics while users with a more restricted range of interests are more likely to meet other people who are interested in the same topics. Thus, the inspection of the recorded proﬁles lets a KnowGent or a DeleGent assess which other types of users are currently around. Moreover, when two users meet, their KnowGents also exchange their

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

97

encounter proﬁles. The receiver of an encounter proﬁle incorporates it into his own proﬁle. As a result, each encounter proﬁle reﬂects not only the interests of the user’s ﬁrst-order (direct) acquaintances, but also, to a certain degree, the interests of the higher-order (indirect) acquaintances. This enables our mobile agents to disseminate and collect knowledge in an altruistic fashion. Users can be employed to transport knowledge between communities without personally being interested in its content. In our experiments on information diﬀusion in mobile communities we were able to proof that a simple strategy with egoistic and altruistic elements based on encounter proﬁles is superior to the pure egoistic approach in which users only store information that is of interest to themselves. In these experiments we have shown that an information ﬂow can be established that spans large parts of a social network and organizes the information in the network in a way that optimizes its accessibility. DeleGents on the Move. Apart from the direct interaction among KnowGents based on KEP, a KnowGent is able to create and dispatch a DeleGent when it comes across another device with an interesting EP. This DeleGent may then be equipped with subsets of the topic proﬁle, the EP, and also the knowledge base before dispatching it to the other device. Policies may be speciﬁed beforehand to let the system decide what exactly the DeleGent carries, based, for example, on the target device’s capabilities and security restrictions. Once the DeleGent is on the target device, it begins to live a life of its own. It may contact other DeleGents or KnowGents and exchange knowledge (ﬁgure 3).

Fig. 3. A mobile device encounters a stationary knowledge server. The user’s personal KnowGent (PK) creates a DeleGent (PD) and dispatches it to the Local Station. DeleGents employ encounter proﬁles to decide about an optimal position in the community network. They may autonomously migrate towards devices

98

Stefan Schulz et al.

Fig. 4. Another mobile user encounters the local station. The user’s personal KnowGent discovers the foreign DeleGent and engages in a knowledge exchange.

that share many interests with their owner. These positions are strategically advantageous because they represent focal points for interesting topics. With high probability the DeleGents in these position are able to acquire knowledge that their user would not have been able to collect otherwise. Besides the EP and the topic proﬁle, the trust information is vital to make decisions with respect to knowledge acquisition and distribution. Several DeleGents can represent a single user at diﬀerent locations while he is not locally present. Based on the EPs, the user’s KnowGent will choose target devices, which the user meets frequently for dispatching DeleGents. Consequently, the KnowGent and its DeleGents will get in contact again and be able to reconcile their proﬁles and knowledge bases (ﬁgure 4). The real novelty of this approach lies in the fact that dispatching a DeleGent and thus creating a virtual presence, enables individuals to exchange knowledge that would otherwise not even have met. Figures 3 and 4 depict a scenario that is explained in more detail in section 5. In this scenario two users engage in a knowledge exchange indirectly without the need to meet each other. Thus, by spreading out DeleGents, a user may increase the scope of his knowledge management activities considerably. By exploiting the properties of the underlying social network via the EP, this can be done automatically. Over time, a personal knowledge network develops when the user’s DeleGents ﬁnd the best places to get in contact with the user and other community members. This network is also able to adapt by reacting to changes in the EPs and thus in the social network. 4.3

Exchanging Knowledge on Trust

For exchanging knowledge we extend the protocol deﬁned in the Shark framework with the notion of a level of trust as described in section 3.2. Therefore, the KnowGent owner can specify thresholds or strategies for the trust values

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

99

describing the minimal competence he requires his communication partners to have. This information is then used to decide about the subset of the topic proﬁle that becomes visible to the partner. Gathered information then will be assessed for its true value by the level of trust. Modeling and exchanging trust as a form of knowledge makes it possible to use the same mechanisms for exchange, storage, and organization of trust that are used for knowledge. Through these mechanisms, trust will be propagated within a community like any other form of knowledge. Additionally, the trustbased exchange protocol allows for an automated adjustment of the level of trust for communication partners. Initial Trust. Initially, there is a default level of trust for users one has never met before. If user A has no information on the trustworthiness of user B, A uses a default value depending on his preferences. This default value represents his general willingness to put trust in other individuals and in their competence (cf. [23]). Furthermore, user A may have community-related trust information on B. If B works for a well-known company, A may assign the company’s speciﬁc initial trust value to B. This is common practice in human interaction. Previous experience with an individual provides a basis for an initial level of situational trust for topics that have not been discussed before. This might be a trust value for a similar topic. For example, user A may have established a trust value for user B on the topic “tennis” and uses this value to infer an initial trust value for information on a speciﬁc female player (ﬁgure 5).

Fig. 5. Competence proﬁle for a speciﬁc communication partner and derivation for a new topic. The competence value for the new node Davenport will be determined by a function on its superordinate node tennis.

100

Stefan Schulz et al.

Adjusting Trust. To adjust the level of trust, we identiﬁed three aspects continuously inﬂuencing trustability and competence: knowledge evaluation, information consistency, and trust exchange. Knowledge evaluation depends on the result of direct exchange. According to this result, the trust value for the competence of its provider will be adapted. This applies to information being accurate and of good quality as well as to information being incorrect and incomplete. A (manual) decision on the quality of each piece of information is needed to achieve this. Information consistency complements the knowledge evaluation. In a consistency check, new information is (manually) compared with existing information. If a piece of information is found to be inconsistent with a number of already existing information, the provider’s trustability value is degraded and, vice versa, if consistency is conﬁrmed, his trustability is increased. This review process implies that adjusting the trust value of a knowledge provider does not necessarily require direct interaction. Finally, it is possible to exchange the trust values. Similar to the knowledge exchange, policies decide upon which trust values to exchange and with whom. Hence, gathered trust values are used to adjust the receiver’s trust in knowledge providers. However, the trust information about person P received from a communication partner CP is also subject to a trust-based evaluation. Like any other piece of information received from other users, trust information is gauged based on the receiver’s trust value with respect to the sender. Hence, an adjusted trust is a function of the current trust value for P, the exchanged trust value for P, and the trust in CP (see ﬁgure 6). The resulting trust values not only represent the experiences gained from direct knowledge exchange, but also, to a certain degree, the experiences of other individuals. Obviously, an appropriate user interface that integrates manual evaluation seamlessly into the process of consuming knowledge is necessary. However, the fact that trust is propagated and that many users beneﬁt from one evaluation helps in decreasing the personal overhead associated with manual evaluation. It is not necessary that every user evaluates every piece of information he gets. Moreover, information evaluation is always a natural part of consuming information. Therefore, the overhead should be restricted to pressing the appropriate button in the user interface if an information seems to be of low quality or inconsistent.

Trust Propagation. The mechanism of adjusting trust as described before has a side eﬀect regarding knowledge and trust in communities. By exchanging trust values, the level of trust will adjust and propagate throughout the social networks and especially within the communities an individual belongs to. Trust propagation self-organizes a network of trust and provides an overall statistical trustworthiness based on continuous and repeated interactions with other community members. As a consequence, an individual will be able to deal with each member of the social network on an appropriate basis of propagated lev-

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

101

Fig. 6. Adjustment of trust values (simpliﬁed): exchanging competence information with communication partner CP on a speciﬁc person (P) and topic (T).

els of trust since the trust network provides him with trust information about community members he never met.

5

Tavern-Based Knowledge Exchange

Taverns present a favorite hot spot for personal communication. In such a place, people with various social backgrounds meet to enjoy the company of others, sharing similar interests, arguing about politics, sports, etc. It is one of the most natural but, at the same time, challenging scenarios for a mobile knowledge management system. As stated before, in section 4, people usually do not meet by chance. They regularly return to places visited before, to exchange information matching their topic proﬁle. A tavern represents a naturally occurring information hub for different communities as people tend to visit the same taverns periodically. Assume the tavern called “Advantage” is well-known as a meeting point for tennis fans. Here, knowledge is shared and discussed on the latest games, single players, and of course the latest gossips on tennis stars. This scenario of physical mobility and peer-to-peer knowledge exchange already is covered by Shark, which electronically supports human communication. Using Shark, images, news reports etc. might be exchanged in addition to the vocal knowledge exchange in a conversation. The mobile devices of guests may also exchange knowledge on common topics that are not actively discussed. The agent-based approach extends Shark by introducing the notion of a virtual presence (DeleGent) in addition to the physical one. To support the

102

Stefan Schulz et al.

communities meeting in the tavern, its publican sets up a tavern server. This server acts as a Local Station, maintaining a knowledge base as well as a tavern KnowGent. Besides exchanging knowledge with guests, this server also allows for guests to leave a DeleGent in the tavern (see ﬁgures 1 and 3). As described above, a KnowGent will decide to create a DeleGent, if the tavern server provides an interesting EP. The KnowGent creates a subset of its topic proﬁle describing the information that is to be gathered. The DeleGent will be equipped with this proﬁle subset (ﬁgure 7), an excerpt of the KnowGent’s knowledge base, and information on trust and exchange policies. Finally, the DeleGent migrates to the tavern server (ﬁgure 3) to act as a virtual presence and to interact with DeleGents or KnowGents of other guests (ﬁgure 4).

Fig. 7. The personal KnowGent (PK) builds a DeleGent (PD) with a subset of its topic proﬁle, only containing information on tennis. By the time, the KnowGent’s owner returns to the tavern, the KnowGent recognizes its DeleGent and synchronizes the knowledge on tennis, communication partners, and trust values (ﬁgure 8). If necessary, the DeleGent is updated with knowledge, policies, and new topics. Not only Local Stations can carry DeleGents. If capabilities allow it, DeleGents can also migrate to Mobile Stations. Therefore, members of a community can carry DeleGents of each other and act as a delegate for other members. Synchronization happens each time, the KnowGent meets one of its DeleGents or two DeleGents meet. Because of this and the small world paradigm discussed before, new relevant information quickly reaches interested users.

6

Conclusions

TAKEUP provides the means to further automate and evaluate the exchange of knowledge between individuals. It provides a framework for exchanging knowl-

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

103

Fig. 8. The personal KnowGent (PK) synchronizes with the DeleGent (PD) and updates it with the latest knowledge, i.e. topic proﬁle subset, trust information, etc.

edge in a highly dynamic and mobile environment. To achieve this, we make use of the advantages introduced by mobile agent technology. Mobile agents are a well-accepted means for the development of highly distributed, dynamic, and decoupled systems like the one presented in this paper. We use mobile agents to extend the Shark system, a knowledge management framework for mobile ad hoc networks. We demonstrated how DeleGents can spread out into the network of a mobile community and use trust information to serve as a virtual presence of a mobile user. By extending Shark with this agent-based approach, we enable mobile users to be virtually present in several locations at the same time. The sphere of knowledge acquisition can extend beyond the limits set by mobility and physical presence. Applying software agents as technology for mediating knowledge exchange in our concept provides the abstractions needed to model, develop, and implement a system that is modular and scalable to the needs of a social network. The agents autonomously act on behalf of their users, providing support for decisions upon knowledge, strategies, and policies. They encapsulate the users’ personalized interest proﬁle and allow for implementing separate strategies and policies for each community member. On the other hand, the resulting AMKM framework provides a basis for the extension of existing knowledge management approaches by means of mobility and distributed management and introduces mechanisms to support quality measurement and a well-organized distribution of knowledge. These abilities strengthen the autonomy of software agents and therefore the agents’ usefulness for managing knowledge on behalf of their users.

104

7

Stefan Schulz et al.

Outlook

Regarding the formalization of trust, several approaches have been taken (e.g. [7]), also regarding distributed systems (e.g. [4]) and security issues in agentbased systems [18]. The use of trust is closely related to reputation management, where approaches are mainly targeting security aspects [33], especially in decentralized systems [1]. We are currently working on an appropriate adaptation for trust and reputation, focusing on the exchange of knowledge. In this paper, we left out discussions regarding an individual’s personal communication strategy (cf. [6]) and exchange policies. Users may implicitly follow diﬀerent general strategies with respect to their communication habits. In our view, the concept of trust may be extended to include information on these strategies in order to reach an even more precise validation of knowledge. The notion of context is a very important aspect for TAKEUP. For example, [5] describes an approach similar to Shark in which context descriptions are used for categorizing documents. Besides extending the current topic-based contexts in Shark to include other context-related information, the overall concept has to consider aspects of context management. In [21] the authors provide means for the representation of an agent’s mental state. These concepts could be applied to KnowGents and DeleGents. And for identifying a context, [3] provides some initial ideas on how to combine multiple sensual inputs to constitute a perspective as a context representation container. As described in section 4.3, there still is a need for manual decisions on the quality and consistency of gathered information. It is a matter to evaluation whether current approaches on natural language processing and document decision (e.g. [9]) at least support and at best automate the decision process. A ﬁrst step to support document decision and a step to abstract from operating on a generalized ontology is the application of classiﬁcation tools, which largely automate the process of associating knowledge to topics. For example, [22] provides a framework to generate such classiﬁers.

References 1. Aberer, K. and Despotovic, Z.: Managing trust in a Peer-2-Peer information system. In Proceedings of Tenth International Conference on Information and Knowledge Management, 310-317. ACM Press, New York (2001). 2. Albert, R., and Barab´ asi, A.-L.: Statistical mechanics of complex networks. Reviews of Modern Physics, 7J:47-97 (2002). 3. Bailin, S. C., Truszkowski, W.: Perspectives: An Analysis of Multiple Viewpoints in Agent-Based Systems. In this Volume. 4. Beth, T., Borcherding, M., and Klein, B.: Valuation of trust in open networks. In ESORICS 94. Brighton, UK (1994). 5. Bonifacio, M., Bouquet, P., Mameli, G., Nori, M.: Peer-mediated Distributed Knowledge Management. In this Volume.

TAKEUP: Trust-Based Agent-Mediated Knowledge Exchange

105

6. Broersen, J., Dastani, M., Huang, Z., and van der Torre, L.: Trust and Commitment in Dynamic Logic. In Proceedings of The First Eurasian Conference on Advances in Information and Communication Technology (EurAsia ICT 2002). LNCS, Springer (2002). 7. Burrows, M., Abadi, M., and Needham, R. M.: A Logic of Authentication. ACM Transactions on Computer Systems, 8:1:18-36 (1990). 8. CoMMA: http://www.si.fr.atosorigin.com/sophia/comma 9. Cooper, J. W. and Prager, J. M.: Anti-Serendipity: Finding Useless Documents and Similar Documents. In Proceedings of HICSS-33. Maui, HI (2000). 10. Dunn, J.: Trust and Political Agency. In Gambetta, Diego (ed.) Trust: Making and Breaking Cooperative Relations, electronic edition, 5:73-93. Department of Sociology. University of Oxford (2000). 11. Finin, T., Weber, J., Wiederhold, G., et. al.: Specification of the KQML AgentCommunication Language (Draft, 1993). http://www.cs.umbc.edu/kqml/. 12. FRODO: http://www.dfki.uni-kl.de/frodo. 13. Giordano, S.: Mobile ad hoc networks. In Handbook of Wireless Networks and Mobile Computing, ed. I. Stojmenovic. John Wiley & Sons (2002). 14. Hardin, R.: Trust and Trustworthiness. Russell Sage Foundation (2002). 15. Harmelen, F. v., Patel-Schneider, P. F., Horrocks, I., (eds.): Reference description of the DAML+OIL ontology markup language (2001). http://www.daml.org. 16. Herrmann, K.: MESHMdl - A Middleware for Self-Organization in Ad hoc Networks. In Proceedings of the 1st International Workshop on Mobile Distributed Computing (MDC’03) May 19, 2003. Providence, Rhode Island USA (2003). 17. Herrmann, K., Geihs, K.: Integrating Mobile Agents and Neural Networks for Proactive Management. In Proceedings of IFIP International Working Conference on Distributed Applications and Interoperable Systems (DAIS 01). Chapman-Hall, Krakow/Poland (2001). 18. Kagal, L., Finin, T., Peng, Y.: A Delegation Based Model for Distributed Trust. In Proceedings of the IJCAI-01 Workshop on Autonomy, Delegation, and Control: Interacting with Autonomous Agents, 73-80. Seattle (2001). 19. Krackhardt, D., Hanson, J.: Informal Networks: The Company Behind the Chart. Harvard Business Review (1993). 20. KRAFT: http://www.csc.liv.ac.uk/∼kraft/ 21. Lou¸ca ˜, J.: Modeling context-aware distributed knowledge. In this Volume. 22. de Magalh˜ aes, J. A. P., de Lucena, C. J. P.: Using an Agent-Based Framwork and Separation of Concerns for the Generation of Document Classification Tools. In this Volume. 23. Marsh, S. P.: Formalising Trust as a Computational Concept. University of Sterling (1994). 24. Mathe, N. and Chen, J. R.: DIAMS: Distributed Intelligent Agents for Information Management and Sharing. In Proceedings of the workshop ”Adaptive Systems and User Modeling on the World Wide Web”. Sixth International Conference on User Modeling, Chia Laguna, Sardinia (1997). 25. Ouchi. W. G.: Markets, bureaucracies and clans. Administrative Science Quarterly, 25:129-141 (1980). 26. Pepper, S., Moore, G.: XML Topic Maps (XTM) 1.0. http://www.topicmaps.org/xtm/1.0/ 27. Schwotzer, T., Geihs, K.: Shark - a System for Management, Synchronization and Exchange of Knowledge in Mobile User Groups. In Proceedings of the 2nd International Conference on Knowledge Management (I-KNOW ’02), 149-156. Graz, Austria (2002).

106

Stefan Schulz et al.

28. Scott, W. R.: Organizations: Rational, Natural, and Open Systems. Prentice-Hall, Inc., New Jersey (1992). 29. SemanticWeb. http://www.w3.org/2001/sw/ 30. TopicMaps. ISO/IEC 13250. 31. Wassermann, S. (Ed.): Advances in social network analysis. Sage Publications, Thousand Oaks, California (1994). 32. White, T. and Pagurek, B.: Emergent behaviour and mobile agents. In Proceedings of the Workshop on Mobile Agents in Coordination and Cooperation at Autonomous Agents. Seattle (1999). 33. Yu, B. and Singh, M.P: A social mechanism of reputation management in electronic communities. In Cooperative Information Agents, 154-165. Boston, USA (2000).

Knowledge Management Framework for Collaborative Learning Support Mizue Kayama1, and Toshio Okamoto2 1 Senshu University, School of Network and Information, Higashimita 2-1-1, Tama, Kawasaki, Kanagawa, 214-8580, Japan [email protected] 2 Graduate School of Information Systems, University of Electro-Communications, Chofugaoka 1-5-1, Chofu, Tokyo, 182-8585, Japan [email protected]

Abstract. The purpose of this study is to support the learning activity in the Internet learning space. In this paper, we examine the knowledge management and the knowledge representation of the learning information for the collaborative learning support. RAPSODY-EX (REX) is a distributed learning support environment organized as a learning infrastructure. REX can effectively carry out the collaborative learning support in asynchronous/synchronous learning mode. Distributed learning is a learning style where individual learning and collaborative learning are carried out on the multimedia communication network. In the distributed learning environment, arrangement and integration of the learning information are attempted to support the decision making of learners and mediators. Various information in the educational context is referred and reused as knowledge which oneself and others can practically utilize. We aim at the construction of an increasingly growing digital portfolio database. In addition, the architecture of the learning environment that includes such a database is researched.

1 Introduction The development of the recent information communication technology is remarkable. As an effect of this, the education environment is being modified to a new environment which differs qualitatively from the previous one (Kuhn 1962). The new education environment contains not only computer but also communication infrastructures such as the information communication network represented by the Internet (Cumming et. al 1998) (Elliot 1993). We call this learning environment the Internet learning space. Information is transmitted for the learner in this learning space from the external space. The information quantity that is available to the learner is enormous. However, there is a limit to the information quantity, which the learner can process. The imbalance of this information processing quantity is a peculiar phenomenon in postmodern ages. Secondary phenomena are also triggered by this problem. These phenomena become factors which inhibit the sound transmission of knowledge and the progress of learning (McNeil et al. 1998) (Chen et al. 1997). In asynchronous L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 107-117, 2003.  Springer-Verlag Berlin Heidelberg 2003

108

Mizue Kayama and Toshio Okamoto

learning, the transformer of knowledge and the transformee of knowledge communicate with a time lag. In such a situation, more positive support is required to realize an effective and efficient learning activity. We need to build a learning infrastructure with learning spaces with various functions.

2 The Purpose of This Study We investigate the mechanism of transmission and management of knowledge for the development of the knowledge community in the learning space, within the educational context. In this paper, we examine the knowledge management and the knowledge representation of the learning information for the collaborative learning support. The purpose of this study is to support the learning activity in the Internet learning space. REX is a distributed learning support environment organized as a learning infrastructure (Okamoto et. al 2000). REX can effectively carry out collaborative learning support in asynchronous/ synchronous learning mode. Distributed learning means using a wide range of information technologies to provide learning opportunities beyond the bounds of the traditional classroom. Some examples of distributed learning technologies include the World Wide Web, email, video conferencing, groupware, simulations and instructional software. A distributed learning environment facilitates a learner-centered educational paradigm and promotes active learning. Distributed learning is a learning style where individual learning and collaborative learning are carried out on the multimedia communication network. In this environment, arrangement and integration of the learning information are attempted to support the decision making of learners and mediators. Various information in the educational context is referred and reused as knowledge which oneself and others can practically utilize. We aim at the construction of an growing digital portfolio based on the agent technology. In addition, the architecture of the learning environment including such a database is researched.

3

Distributed Education/Learning System

Distributed education/learning support systems are classified into 2 types. One type is systems using the advanced information network infrastructure to realize smooth communication for distributed education/learning. The other type represents systems, which positively support various activities of distributed education/learning. As examples of the former, we can indicate the Open University, which provides a correspondence course in the United Kingdom (Kaye 1994) and the other online universities (UoP) (JIU). In addition, ANDES which is a satellite communication distance education system using digital movies (Shimizu 1999) and the synchronous distributed education system which apply Web browser sharing technology(Kobayashi et al. 1998) should also be mentioned. As examples of the latter type, the "TeleTOP" at the University of Twente(Collis 1999), and a distributed education system which does not utilize the WWW technol-

Knowledge Management Framework for Collaborative Learning Support

109

ogy at the University of Trento(Colazzo, and Molinari 1996) are pointed out. The "Electronic Learning" project at the University of Southern California (Johnson and Shaw 1997) provides an intelligent tool which supports designing of the learning course and authoring of the teaching material for distributed education. A distributed education system with synchronous and asynchronous learning mode at IBMJapan (Aoki and Nakajima 1999), applies the Web operation recording technology . For both types of systems, some agent-based distributed learning environments are also reported (IPAJ 1999) (JIFETS 2000) (JILR 1999). Some of them are implemented with some standardizing technologies (CORBA including FIPA (FIPA), OMG(OMG) and Agent Society (Agent Society)). Agent technologies are useful to realize multi-users learning environment. We have explored an intelligent media oriented distributed e-Learning environment. In this system, we developed a learning management system and some learning control mechanisms based on agent technologies. Some collaborative learning media controlled by agents are also developed.

4 Collaborative Learning with REX A learner group that guaranties the smooth transmission of knowledge can form a community (the knowledge community) by sharing and reusing common knowledge. The image of collaborative/ group learning with REX is shown in figure 1.

Learner B

（６）Read article posted by learner B

(5)I think it’s enough to certify congruence between △ABC and △CDA

Learner C （７）△ABC　is not congruent with △CDA

Learner A (4)Draw an additional line CD onto the collaborative workplace

(3)question to new logged-on user

Collaborative Workplace

Learning Resources

（８）Any questions? (as you have no prompting)

agent agent

Communication tool （２）Certify ∠ABC＝∠DEC

Learner D

（１）Assignment: circular angle and triangle similarity

（９）Discussion monitoring

（10）All learners completed Assignment. Next please.

Coach

Fig. 1. Collaborative learning with REX

Learning activities that occur within this group are as follows: − achievement of learning objectives as a group, − achievement of the learning objectives of each learner, − achievement of the learning objectives of the learner group which consists of multiple learners.

110

Mizue Kayama and Toshio Okamoto

REX supports the transmission of knowledge in the learner group and the promotion of the learning activity. It is indispensable that REX has the following functions: − function which controls learning information for the individual learner and the group, − function which manages learning information of the learner for mediation. The learner and group information are produced from the learning space. This information will be stored in the collaborative memory (CM). This information is defined as learning information. We also define the method of information management of such information and the structure of the CM.

5 The Management of Learning Information 5.1 Information Management Mechanism Scheme of REX The simple mechanism of the management of learning information developed in this study is shown in figure 2. The processing mechanism consists of two components. The first one is a module that offers the learning environment. The second one is the CM that controls various information and data produced from the learning environment. In the learning environment, 2 types of functions are offered. One is the monitoring function for the learning progress. The other is the tool/application for the collaborative learning. The former function controls the learning history/record of individual learners and the progress of the collaborative group learning. The latter tool/application becomes a space/workplace for collaborative synchronous /asynchronous learning. The learning information, which emerged from such a learning environment, is handed to the CM. The CM offers 2 types of functions. One is the knowledge processing function, and the other is the knowledge storage function. In the former, input learning information is shaped to the defined form. In the latter, for the formatted information, some attributes related to content are added. The complex information processing takes place in the CM. 5.2 The Knowledge Management in REX In this study, the processing described in the previous section is considered as a process of the knowledge management in the learning context. Knowledge management is defined like follows (Davenport 1997). The knowledge management is "the systematic process of finding, selecting, organizing, distilling and presenting information in a way that improves an employee's comprehension in a specific area of interest."

Knowledge Management Framework for Collaborative Learning Support Client system Tools/applications for the collaborative learning Ex: Web-based CAI/ITS, Simulator, ILT, CHAT, Draw

Learning activity management tools Management tools for learner/group

Learning information management unit Storage module

111

Learning Environment Producing Layer Collaborative memory

Reference module

arrangement

reference

visualization

housing/ classification

Learning Information Management Layer

inquity

delivery

(knowledge processor in the collaborative memory)

Learner / group profile DataBase Portfolio Information Personal/ Group member Information

Security Information Preference Information

Achievement Information Performance Information

Storage Layer (knowledge storager in the collaborative memory)

Server system RAPSODY-EX : learning environment in INTERNET

Fig. 2. The mechanism schema of the REX

Nonaka arranged the process of knowledge management as a SECI model (Nonaka 1995). The SECI model is expressed as a conversion cycle between tacit knowledge and expressive knowledge. Tacit knowledge has a non-linguistic representation form. Expressive knowledge is a result of putting tacit knowledge into linguistic form. Tacit knowledge is shared with others by converting it into expressive knowledge. In the SECI model, socialization (S) / externalization (E) / combination (C) / internalization (I) of knowledge is expressed. The knowledge management in educational context is defined as follows: "the systematic process of finding, selecting, organizing, distilling and presenting information in a way that improves a learner's comprehension and/or ability to fulfill his/her current learning objectives." REX aims to support participants’ activities in the C (combination of knowledge) phase. Moreover, it affects not only the process of knowledge conversion from the C phase to the I (internalization of knowledge) phase, but also from the E (externalization of knowledge) phase to the C phase. The information of learning entity contains the expressed knowledge by learners. This overt knowledge can be represented by natural language as verbal information. So, we can regard this knowledge as one that would be elicited from the learner's tacit knowledge. In this situation, what we have to consider is as follows:

112

Mizue Kayama and Toshio Okamoto

Who are the subjects of our knowledge management work? Learners and the persons who support the learners are our subjects. Learners’ task is to acquire the ability/skill for the problem solving. On the other hand, supporters’ tasks are to support for acquisition of ability/skill of the learner, and to support of the problem solving by the learner. Supporter means a facilitator/tutor/coach/ organizer etc. What are the knowledge resources in the learning group? For learners, the knowledge for the effective and efficient problem solving is their knowledge resource. On the other hand, for the supporters, the knowledge on problem setting and activity assessment is their knowledge source. What is the gain for the learning group? The gains for learners are to acquire the ability in which to effectively and efficiently solve the problem, and to acquire the meta-cognition ability. For supporters, the acquisition of the ability of supporting the ability acquisition of the learner is their gain. How are the knowledge resources controlled to guarantee the maximum gain for the learning group? By the information processing to relate common knowledge of the CM and learning context, we try to manage the knowledge in the collaborative learning. To create the collaborative portfolio between individual and group learning, extension of acquired knowledge of learners, knowledge extraction from learning history under the problem solving and making outline of problem solving process.

6 The Collaborative Memory 6.1 Functions of CM In the CM, information generation / arrangement / housing / reference / visualization are the management processes of expressive knowledge in the learning space. REX is a learning environment, which possesses a knowledge management mechanism. In this environment, 1. the review of the learning process, 2. the summarization of the problem solving process, 3. the reference of other learners' problem solving method are realized in the learning space. Learning information is expressed by an unified format. Then, that information is accumulated in the CM. This information becomes the reference object of the learner. The generation and the management of the information on the learning performance and the portfolio of the learner and group are main objects of the knowledge management. In this study, learning information is obtained from the application tools for the collaborative learning. It is necessary to control the learning record, the reference log of the others' learning information and

Knowledge Management Framework for Collaborative Learning Support

113

the log of problem solving and learning progress. To realize this control not only techniques based on symbolic knowledge processing approach, but also techniques based on sub-symbolic knowledge processing approach are used. The CM consists of two layers. One is the information storage layer and the other one is the management layer of the stored information. At the information storage layer, 4 kinds of information are mainly processes. 1. 2. 3. 4.

Learning information, Information on the learner, Information on the setting of the learning environment, Information on the learning result.

At the information management layer, the reference/arrangement/integration of learning information are processed. The individual learner profile information is composed of information following the IEEE Profile information guidelines (IEEE 2000). The group information is expressed by the expansion of the individual learner profile information. The conversion from the learning log data to learning information is necessary to develop this profile database. The information, which should apply in learning information, is as follows: − − − − −

information and/or data on its learning context and/or learning situation, information about the sender and the sendee of the information, significance and/or outline in the educational context, information on the relation structure of the learning information, reference pointer to individual learner and group who proposed or produced the information, − relation with other material. By adding these information, the learning information is arranged into an unique form. If a learner requires some information related to his/her current learning, REX shows the (estimated) desired information to the learner. 6.2 Communication Scheme of the RAPgents Figure 3 shows a communication scheme in the REX. Three types of agents are existed in the REX. These agents perform each mission with user (user-RAPgent), collaborative learning tool (application-RAPgent) and the CM in REX server (CMRAPgent). Communication protocol between RAPgents is defined based on the FIPA ACL communicative act. The missions of each RAPgent are to transform information adaptively to create group portfolio, to maintain learning contexts in the group member and to let refer information in CM. To realize the knowledge management in REX, application-RAPgents develop some learning contexts by using learning information in the CM. Then they refer the suitable learning information for the collaborative tool/application.

114

Mizue Kayama and Toshio Okamoto

User RAPgent

Operation for tool_B

Application RAPgent

Broadcast for group members

CM RAPgent

Operation + learning context

Application RAPgent Operation for tool_A REX Server User RAPgent

Fig. 3. Communication scheme in the REX

7 Examples of the Knowledge Management in REX Figure 4 shows the window images of the collaborative applications on REX. Two types of applications are loaded. One is a chat tool for the text communication among the group member. Another application is a collaborative simulator for the nutrition treatment. Each application has each application-RAPgent. By the functions of these RAPgents, learning history data at this session is stored in the CM and formulates a set of the group portfolio. The examples of knowledge management at this session are shown in the figure 5. A log data of this dialog is visualized by three kinds of methods. These results are produced by three application RAPgents. The first method is visualization of the dialog structure. The dialog layers are reasoned based on the dialog proceeding model (Inaba 1997) and the utterance intention information that were given to the dialog log. The result is shown as tree structure. The second method is visualization of transition of the contents of a dialog. An appearance of the important term that is in a dialog is searched for using the term dictionary about the current discussion/learning domain (Chiku 2001). This result and the timing connection of each utterance are considered to detect a transition of the contents of a dialog. The result is shown as graph structure. The third method is visualization of transition of problem solution process. One utterance can be unified as meaningless unit for the problem solving process from the first and the second processing result and an educational mentor's expertise/educational intentions. The result re-constituted as problem solution process is shown with the structure that imitated the dendrogram.

Knowledge Management Framework for Collaborative Learning Support

115

④

Collaborative simulatorfor the training of the nutrition the miedical care treatment

CHAT Tool

③

① ④ ③ ② ①

②

Fig. 4. Two types of collaborative learning applications : a chat tool and a collaborative simulator for the nutrition treatment

8 Conclusion The purpose of this study is to support the learning activity in the Internet learning space. We examine the knowledge management and the knowledge representation of the learning information for the collaborative learning support. REX is a distributed learning support environment organized as a learning infrastructure. In this paper, the management of learning information in REX is described. REX is an integrated distributed learning environment and supporting tools/ applications for the collaborative learning. Also, in this paper, the knowledge management mechanism in the educational context is showed. The detailed knowledge management technique is exploring by using the semantic web approach, and it will be integrated with the current learning support environment.

References 1. Baldonado, M., Chang, C.-C.K., Gravano, L., Paepcke, A.: The Stanford Digital Library Metadata Architecture. Int. J. Digit. Libr. 1 (1997) 108–121.

116

Mizue Kayama and Toshio Okamoto

Log data of dialog

Dialog structures 自グループの過去の問題解決過程自グループの現在の問題解決過程

Transition of the 自グループの現在の contents of問題解決過程 a dialog

他グループの問題解決過程

他グループの問題解決過程

Transition of problem Solving process

Fig. 5. Examples of Knowledge Management in REX

2. Bruce, K.B., Cardelli, L., Pierce, B.C.: Comparing Object Encodings. In: Abadi, M., Ito, T. (eds.): Theoretical Aspects of Computer Software. Lecture Notes in Computer Science, Vol. 1281. Springer-Verlag, Berlin Heidelberg New York (1997) 415–438. 3. van Leeuwen, J. (ed.): Computer Science Today. Recent Trends and Developments. Lecture Notes in Computer Science, Vol. 1000. Springer-Verlag, Berlin Heidelberg New York (1995). 5. Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs. 3rd edn. Springer-Verlag, Berlin Heidelberg New York (1996). 6. ADLNet 2000. Shareble Courseware Object Reference Model:SCORM, Ver.1.0, http://www.adlnet/org/. 7. Agent Society, http://www.agent.org/ . 8. Inaba A. and Okamoto T.. Negotiation Process Model for intelligent discussion coordinating system on CSCL environment, Proceedings of the AIED 97 (1997) 175-182.

Knowledge Management Framework for Collaborative Learning Support

117

9. Kaye A. R. : Computer Supported Collaborative Learning in a Multi-Media Distance Education Environment, in Claire O'Malley (Ed.) Computer Supported Collaborative Learning, Springer-Verlag (1994) 125-143. 10. Collis B. : Design, Development and Implementation of a WWW-Based Course-Support System, Proceedings of ICCE99 (1999) 11-18. 11.FIPA, http://drogo.cselt.stet.it/fipa/ . 12.Cumming G., Okamoto T. and Gomes L. Advanced Research in Computers in Education, IOS press (1998). 13. Elliott J. : What have we Learned from Action Research in School-based Evaluation, Educational Action Research ,Vol.1, No.1 (1993) 175-186. 14. IEEE 2000. Draft Standard for Learning Technology -Public and Private Information (PAPI) for Learner, IEEE P1484.2/D6, http://ltsc.ieee.org/, (2000). 15. IMS 1998. Learning Resourcde Metadata : Information Model, Best Practice and Implementation Guide, IMS Ver1.0, http://www.imsproject.org/. 16. Nonoka I. : The Knowledge-Creating Company, Oxford University Press (1995). 17.IPSJ Journal : Special Issue on Multimedia ・Distributed and Cooperative Computing, Vol.39 No.02 (1999). 18.Journal of International Forum of Educational Technology & Society. Special issue on Theme: On-line Collaborative Learning Environment, Vol.3, No.3 (2000). 19.Journal of Interactive Learning Research : Special Issue on Intelligent Agents for Educational Computer-Aided Systems ,Vol.10, Nos.3-4 (1999). 20.Colazzo L., and Molinari A.: Using Hypertext Projection to Increase Teaching Effectiveness, International Journal of Educational Multimedia and Hypermedia, AACE (1996). 21.Chiku M. and Okamoto T.: A dialog visualization tool : Gijiroku”, Proceedings of the 62th Annual Conference of the Information Processing Sciety of Japan (2001) 241-244. 22. Kobayashi M.: Collaborative Customer Services Using Synchronous Web Browser Sharing, Proceedings of CSCW 98 (1998) 99-108. 23.OMG, http://www.omg.org/ . 24. The United Kingdum Government : Connecting the Learning Society, The United Kingdum Government's Consultation paper (1997). 25.McNeil S.: Technology and Teacher Education Annual, AACE (1998). 26. Chan T.et al. : Global Education ON the Net, Springer-Verlag (1997). 27. Davenport T. : Working Knowledge, Harvard Business School Press (1997). 28. Kuhn T. : The structure of scientific revolutions, University of Chicago Press (1962). 29. Okamoto T., Cristea A.I. and Kayama M. : Towards Intelligent Media-Oriented Distance Learning and Education Environments, Proceedings of ICCE2000 (2000). 30. University of Phoenix Home Page: http://www. uophx.edu/ . 31. Jones International University Home Page : http://www.jonesinternational.edu/ . 32. Johnson W. L. and Shaw E.: Using Agents to Overcome Deficiencies in Web-Based Courseware, Proceedings of the workshop "Intelligent Educational Systems on the World Wide Web" at the 8th World Conference of the AIED Society, IV-2 (1997). 33. Aoki Y. and Nakajima A.: User-Side Web Page Customization, Proceedings of 8th International Conference on HCI, Vol.1 (1999) 580-584. 34. Shimizu Y. : Toward a New Distance Education System, Proceedings of ICCE99 (1999) 69-75.

An Agent-Based Approach to Mailing List Knowledge Management Emanuela Moreale1 and Stuart Watt2 1

Knowledge Media Institute (KMi), The Open University, Walton Hall, Milton Keynes, England, MK7 6AA [email protected] 2 School of Computing, The Robert Gordon University, St. Andrew Street, Aberdeen, Scotland, AB25 1HG [email protected]

Abstract. The widespread use of computers and of the internet have brought about human information overload, particularly in the areas of internet searches and email management. This has made Knowledge Management a necessity, particularly in a business context. Agent technology – with its metaphor of agents as assistants – has shown promise in the area of information overload and is therefore a good candidate for Knowledge Management solutions. This paper illustrates a mailing list Knowledge Management tool that is based on the concept of a mailing list assistant. We envisage this system as the first step towards a comprehensive agent-based Knowledge Management solution.

1 Introduction As recently as twenty years ago, it was thought that computers would bring in the age of leisure, with considerable shorter working weeks and overall better quality of life [1]. On the other hand, nowadays computers are more often associated with frustration and tasks reportedly “taking longer than they should”. What is certain is that the increasingly widespread use of computers has meant an exponential increase in the amount of stored documents, making the task of locating and retrieving useful information rather complex and time-consuming. On the other hand, today’s economy is knowledge-based, so the main asset of companies, and one on which their competitive advantage rests, is their stock of knowledge [2][3][4]. Document management is thus a must in today’s organisations. Most work in this area has focused on web pages. These efforts range from information retrieval (IR) to information extraction (IE) and wrapper generation [5]. One of the most important types of document is email. According to a survey commissioned by O2 (then BT Cellnet), UK employees spend up to eight hours per week on email [6]. Most of us feel that there is just too much email to deal with and that better support for this essential working tool is needed. Yet, email is complex: it often contains ‘noise’ (e.g. parts of earlier emails, signatures) and it displays several different formatting conventions (such as paragraphs and signature layouts). Within mailing lists, the need for information management for email is even more felt: although the ‘noise-to-information’ ratio varies across lists, the large number of L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 118-129, 2003.  Springer-Verlag Berlin Heidelberg 2003

An Agent-Based Approach to Mailing List Knowledge Management

119

postings often results in subscribers being unable to keep up and/or follow the flow of argument. Postings prior to one’s subscription can often be found through archives: yet, these are rarely used, perhaps because of the low perceived success rate of this operation, the time needed and unhelpful archiving conventions (e.g. by subject line). Thus, often queries are asked again and again on a list. It is also likely that humans prefer ‘dialectic’ discovery involving interaction with active entities to mechanical and repetitive sifting through “passive material”. What is certain is that humans prefer to ask someone a question to doing the searching themselves [7]. This paper explores the application of the agent assistant metaphor to mailing list Knowledge Management. Our Sentinel system works with several lists, giving users archiving and retrieval assistance through an intuitive and dialectic interface: users can email their query directly to the agent and receive a prompt reply day or night. Alternatively, users can post their query publicly to a forum (monitored by the agent) or run a web-like search over the monitored lists. Through the application of IE, IR and a novel information integration (II) technique to the mailing lists, Sentinel automatically links email into a tangled network of stories, and arranges them in a meaningful way (digests, queries asked to date), also providing details of contributors and their postings. Because it allows users to notice relationships between/among pieces of information and people, Sentinel is a useful tool for “community support” (see section 1.3) that can be successfully employed as part of an organisation’s knowledge management strategy. This paper will first introduce the concept of agents as assistants and briefly point to research on email management and touch on community genre, then describe our approach to the problem of mailing list management: the Sentinel system. An evaluation of the system and future work conclude the paper. 1.1 Agents as User Assistants Agent systems have been proposed as solutions to the problem of information overload, particularly regarding email [8] and internet searches [9]. Most of the current implementations aiming to ease the burden of dealing with email are text classifiers [10][11] or keyword extractors [12], often working as email clients plug-ins [12][13]. Unlike our system, these solutions target general email in the users’ inbox and not specifically mailing lists. The basic idea behind this paradigm is that software agents are to perform tasks similar to those that a human assistant would carry out. Gruen [14] conducted field studies and an analysis of the types of assistance provided by human assistants. These were found to include: pre-processing, filtering/prioritising, adding relevant information, performing a number of steps in response to a single request and peripheral awareness/pointing out information. Most of these functions, as applicable to mailing lists, have been included in our Sentinel system. Murch [1] dedicates a chapter of his book to email agents. He suggests that they may successfully be employed to perform the following eight activities: a) Controlling any unwanted email or “spam” b) Alerting users by voice if a certain message arrives c) Automatic mail forwarding d) Consolidating mail from numerous sources

120

Emanuela Moreale and Stuart Watt

e)

Searching the internet for new sources of news, stocks and deals and then delivering them by email f) Distinguishing between private/personal and corporate/business email g) Automatically answering email and responding according to conditions h) Carrying out regular administrative tasks such as archiving and indexing for future searching While b) is no more useful for email in general than for mailing lists and c) and g) are easily achieved through user-end programming, Sentinel can be said to address these issues, while in many respects being more than a traditional assistant agent. 1.2 Email Analysis and Management While considerable effort has taken place in the area of document management (from company document warehouses and intranets to efficient IR on the web), document analysis and management techniques obviously depend on document characteristics. Web pages and emails might be considered to be fairly similar types of documents, but mailing lists are characterised by smaller volumes of documents (than the Web), more complex item structure and presence of noise. Although email is sometimes said to be particularly suitable for knowledge management because “it has a fair amount of metadata attached to it” [15] (e.g. headers and threading information), the latter can give information that is misleading (for instance when people hit ‘reply’ to send an email on a new subject). Emails are largely unstructured documents: while headers are structured, the message body – the text written by the sender – is unstructured1. This suggests that headers and message body should be treated differently by the text mining operation. Because of these characteristics of email, it is best not to apply IR to the whole document, but instead minimise noise first. This means totally removing irrelevant emails (e.g. ‘out-of-office replies’) and then carrying out straightforward IE on the header fields and a more sophisticated IE on the body text (e.g. remove salutations and signatures). The ‘clean’ email text can then be stored in a database, optionally undergoing some kind of information integration. In any case, once in a database, the text is easily searchable and IR techniques can then be successfully used. It is argued that a combination of II, IE, and IR represents the best text mining solution for email. Email is a dynamic type of document: since changes in employees’ interests are reflected in their emails, email gives an up-to-date snapshot of a community’s activities and, for instance, current distribution of expertise within a company. Email is also “where coworkers trade stories, ask questions, propose new methods, debate techniques” [15] and where knowledge is created through interaction [16]. Email thus constitutes an ideal target for knowledge discovery once the problems relating to its unstructured nature are overcome. Section 2 illustrates such a project, in which we applied the techniques discussed above. 1

E.g. an email often contains parts of previous emails; paragraphs may be separated by blank lines or not, signatures can have different layouts or be absent.

An Agent-Based Approach to Mailing List Knowledge Management

121

1.3 Community Support An often talked-about concept in Knowledge Management is that of Community of Practice (CoP): a group of individuals who work, learn or socialize together sharing insights and developing a shared knowledge as a consequence of participation. These communities evolve, develop and merge around shared interests and expertise [17]. A Knowledge Management approach for online CoPs is that described in [18]: IDIAG is an attempt to distil messages by a large number of people into a more succinct, durable knowledge. It forms an interactive or dynamic “book” where the corpus is constructed iteratively and collaboratively by people with different opinions. IDIAG has three classes of users: people that enter comments, moderators controlling the discussion and distillers distilling the archived materials, usually to produce a more concise site. While in I-DIAG the distilling is carried out by humans, in our system the distilling process is carried out automatically by a “community-aware” indexing program. Mailing lists are online communities of practice whose group’s tacit knowledge is reflected in the postings, the artefacts created and shared by the community. It has been pointed out that these artefacts are characterised by their own genre [19]. The authors recommend that tools to support a community in sharing and retrieving information be sensitive to community genre in order to properly reflect that community’s (group) tacit knowledge. Our proposed solution, Sentinel, uses heuristics based on the community’s specific genre to identify important information. The next section will illustrate this system in more detail.

2 The Sentinel System We were asked to develop an analysis and management system for email. Our data consisted in files containing several mailing lists arranged as Microsoft Outlook public folders, giving a total of several thousands of complex email messages. Our original task was to develop a tool that interacts with mailing lists, extracting information and answering user queries by email. A later requirement was that of arranging information in an easily searchable and semantically meaningful way. Sentinel builds on previous work [20]: the key recommendations it implements are listed in Table 1. 2.1 System Characteristics The Sentinel system has the following characteristics: • It extracts and stores important information from emails through text mining; • It links discussions occurring in different lists; the process of feeding knowledge from one forum to another amounts to knowledge discovery for the latter forum; • It is an information agent system with access to one or more mailing lists. It is able to collate and manipulate information obtained from them in order to answer queries about it such as “Does anybody know about X?” [23]. In fact, users can email such queries to this “virtual participant” in the lists and the agent will send a useful reply back to them.

122

Emanuela Moreale and Stuart Watt Table 1. Recommendations Implemented in the Sentinel System

Issue 1: Public vs. Private Interaction Recommendation: both should be allowed. If all interaction is forced to be public, the total amount of interaction will be reduced. [20] Issue 2: Anthropomorphism vs Mechanomorphism Recommendation: these systems are more acceptable to users when mechamorphised, i.e. presented as an “Active Archive” rather than as an anthropomorphic character (“Uncle Derek”) [20]. This was overturned (due to issue 4). Issue 3: Closeness vs. Openness / Visibility Recommendation: it is best to open up the system to users as a series of threads, thus contextualising content to the current discussion. [20] Issue 4: Fit into the Company Culture: Groupware sold off-the-shelf is doomed Groupware needs to be customised [21] and must fit into the company culture[22]. Our target company requested an anthropomorphic character (Figure 1), thus recommendations to avoid anthropomorphism (Issue 2) were overturned.

Fig. 1. Screenshot illustrating some of the functionality of the Sentinel System, in particular the result of the query “Does anybody have experience of reducing sand erosion in gas wells?” – some text has been omitted. Uncle Derek is a specific incarnation of Sentinel. Notice the use of anthropomorphism, as requested by the target company

• • •

Alternatively, human users can access a browsable version of the agent’s “digested knowledge”: the structure of information contained in the list discussions is displayed as a simple semantic network; By showing all contributors to the monitored lists as well as their contributions, it allows people to easily identify each contributor’s area of expertise; It contains an automated “FAQ-Maker” which extracts discussion digests and identifies previously-asked questions with replies. This low-cost alternative to manually-crafted ‘Frequently-Asked Questions’ (FAQs) is a partial

An Agent-Based Approach to Mailing List Knowledge Management

123

application of case-based reasoning (CBR) [24]: problems are identified by initial questions (usually starting a new thread), while replies (often containing ‘Best Practices’) are the source of case solutions and outcomes. Sentinel can store multiple solution options for each case; • It uses concepts, a basic “ontology” for the application domain. Ontologies help with knowledge organisation [25] and domain concepts provide useful “entry points” for browsing the information network, particularly by people new to the domain and/or list; • It is presented as an anthropomorphised agent system: although previous work in this area had suggested that CSCW systems are best presented as mechanomorphised systems, this recommendation was overturned to meet the target company’s requirements. • It is a customised system tailored to the target company and fitting into its culture. Customisation is an essential requirement for groupware system [21]. The next section will illustrate the steps we followed in developing the system. 2.2 Document Analysis and Text Classification Pre-filtering The first step consisted in examining the structure of the public folders and several hundred messages. It was evident that some emails should be ignored (e.g. automated ‘out-of-office’ replies). Text classification was successfully employed here using pairs or triplets of adjacent terms in a manner similar to [26]. 2.3 Text Mining: Information Integration and Extraction IE extraction rules were then applied to the ‘non-irrelevant’ emails to obtain a set of purged email messages. Examples of cleaning rules include: splitting email into simple email chunks and trimming text of main email chunk from both top (to remove salutation) and bottom (to remove signature). The clean text, together with important information about the email (such as threading information), was then stored in a database. Thanks to the hand-crafted rules, the process is remarkably accurate and system recall very high. An integration step followed: clean email chunks were threaded into ‘stories’ or coherent sequences of chunks. This story-weaving provides some measure of contextualisation and mimics humans’ way of organising information [27]. We then identified queries asked to date and mapped queries to stories as a means to provide a simple procedure for automated generation of CBR cases. Information integration is central to Sentinel. Stories are linked by author, by subject, by questions, and by domain concepts. Given a story, a user can use the web interface to find stories by the same author, addressing a similar question, or touching on similar concepts. Stories are woven together using a semantic network index linking stories to objects through different kinds of relationship (‘written by’, ‘about’, and ‘asks’ are some of the relationships implemented by Sentinel). This low-cost automated hyperlinking technique, simplified from that of [28], turns Sentinel’s storybase into a tangled web of related stories, which can be browsed through the web.

124

Emanuela Moreale and Stuart Watt

Sentinel’s linking approach depends, through Cleary and Bareiss’s work, on Schank’s “Conversational/ Associative Category” model [29]. This is a psychological model of conversational dialogue, and in particular, of topic change in conversations. We are applying this model to explore the extent to which an agent can support the information needs of a community of users in a dialogic style, and as such, Sentinel is a psychological agent as well as a computational one. Although using this is a promising line of work, the current system remains limited in its application of this theory, because the IR techniques it employs are still based on a “query-response” model, and don’t yet support the change in topics recommended by this theory. This is a major component of our current work on refining the Sentinel concept.

Fig. 2. The agent’s prompt email reply – some text has been omitted

2.4 Information Retrieval and Knowledge Discovery The last step was to devise an appropriate method to do text mining out of the database as well as to design a suitable query interface. Unlike typical CBR systems, Sentinel does not try to provide definitive answers but, rather, selects the most relevant

An Agent-Based Approach to Mailing List Knowledge Management

125

Fig. 3. “The Tangled Web of Information”: simple semantic network providing a useful way of visualizing the information contained in the monitored mailing lists. The edges represent entry points into the system

In d e x in g

A ll M e ssa ge s

N e g a tiv e F ilte rin g (R e m o v e Ir re le v a n t M essages)

In fo rm a tio n E x tra c tio n ( IE )

U s e fu l M e ssa ge s

C le a n M essages

In fo rm a tio n In te g ra tio n ( II )

Stories, digests, cases

In fo r m a tio n R e trie v a l ( IR )

R e s u lt

R e trie v a l

Fig. 4. Text and information mining flow in Sentinel (operations/techniques; items)

cases and likely solutions, leaving the final decision to the user. Any story retrieved by Sentinel forms a starting point for further browsing through the web of stories. Sentinel uses an important additional technique to enhance information retrieval. Communities converge on communicative genres, and we observed several distinct classes of message in the communal lists. Sentinel is principally intended to respond to one of these – the open question inviting response from the community. Sentinel indexes and stores other messages, including answers, follow-up questions, and forwarding recommendations. Patterns which can be used to recognise these genres [19] are programmed into Sentinel, and are used as an additional source of evidence when choosing appropriate actions. This hints at the possibility of structuring the system more explicitly and directly as a team of experts with different skills.

126

Emanuela Moreale and Stuart Watt

Technically, Sentinel is implemented in Perl and uses Apache and mod_perl to provide a web interface, and MySQL to store data. Sentinel offers an interactive way to search the archives: users can email a question to the agent privately or post their query to the whole list. The agent monitors the list and, after an appropriate interval of time (to allow human participants to contribute) since a query is posted, can post a reply. In doing so, it will prefer less recent material and postings from other lists. Alternatively, users can browse the system through a web interface. As well as providing a standard search box, this lists all contributors to the mailing lists, with links to their postings, and allows users to discover knowledge about other members’ skills that would otherwise be difficult to gather from reading the postings “manually”.

3 Evaluation The system described in this paper has been in use for approximately a year in the target company. Informal feedback suggests that the software has proved useful and has been successfully received. While a formal qualitative evaluation of the system is planned, it is significant that an operationalised version of the system was requested and recently introduced into the company. It is worth noting that evaluation of large-scale groupware systems like Sentinel is problematic (it is one of Grudin’s challenges for developers of systems such as this [21]), and most evaluations inevitably resort to surveys. We feel that a more ‘utilization-focused’ approach [30] is appropriate in this context, directing evaluation at improving use and uptake of the system, and even assisting key stakeholders in conducting their own evaluations. Our criteria for success, therefore, can be summed up as “do people use it?”, and although some do, there is definite room for improvement. In keeping with this approach, we have conducted informal evaluations of the relevance of postings from Sentinel with domain experts. The results have been mixed: while some matches have highly relevance, others much less so. One important result from our evaluation is that assessing relevance for posting messages is a novel challenge for IR: IR techniques traditionally rank retrieved matches, but are less well developed in the area of providing a confidence measure associated with these relevance measures. A separate evaluation of these IR aspects of Sentinel is planned for the near future, although the methodological issues it raises are different to most IR evaluation work (e.g., that in the TREC series of conferences). The planned formal evaluation will explore the wider presentation and use of the Sentinel system. This is because some aspects of the implemented system (e.g. management of information about people) evolved as the system was introduced to the users and were not part of the original concept.

4 Future Work Sentinel has been applied with success to internal mailing lists that are part of a corporate intranet. Consideration is being given to expanding its knowledge capture net:

An Agent-Based Approach to Mailing List Knowledge Management

127

being built in a modular fashion, Sentinel can be easily extended to integrate with other document management initiatives to perform more extensive knowledge discovery. Work is also taking place on a separate Sentinel implementation in which both the filtering and relevance matching stages are carried out by an agent team [31], using a peer-to-peer architecture similar to that of [32] and [33], as this should allow easier implementation of certain features such as customisation and personalisation. More work on the IR side is being conducted to support this, developing techniques based on the probabilistic model [34] to provide the kinds of confidence measures that will allow multiple agents to reason about judgements in an effective manner. Further agents supporting querying through different channels (e.g. WAP, Instant Messaging) are also planned. Other likely upgrades include extending the concept network to form an ontology and make more use of CBR and learning (in particular, it would be useful to provide for integration of user feedback).

5 Conclusion This paper has illustrated how a combination of shallow text processing techniques (II, IE, and IR) can be successfully used to build an agent software system that alleviates the problem of information overload and management in mailing lists and helps with a company’s overall knowledge management strategy. The case is made for Sentinel, a software tool that uses II, IE and IR techniques, agent technology and CBR to manage information in email within several lists. Sentinel is not just an email-mining agent tool offering a convenient way to archive emails and search over them, but is also a community support tool and allows knowledge discovery within mailing lists and organisations.

6 Acknowledgements We are grateful to BP for data and support, and to Trevor Collins for valuable input.

References 1. Murch, R.: Intelligent Software Agents. Prentice-Hall (1999) 2. Allee, V.: The Knowledge Evolution: Expanding Organizational Intelligence, ButterworthHeinemann (1997) 3. Barchan, M.: How Celemi Ensure Strategic Gains by Measuring Intangible Assets, Knowledge Management Review, September–October 1998 (1998) 4. Uit Beijerse, R.P.: Questions in Knowledge Management: defining and conceptualising a phenomenon, Journal of Knowledge Management, 3(2):94–109 (2000) 5. Eikvil, L.: Information Extraction from World Wide Web – A Survey, July 1999 (1999) 6. Sturgeon, W.: Eight hours per week lost to email, Dec 2001, www.silicon.com (2001) 7. Ackerman, M.: Augmenting the Organizational Memory: A Field Study of Answer Garden, Conference on Computer-Supported Cooperative Work, 243–252 (1994)

128

Emanuela Moreale and Stuart Watt

8. Maes, P.: Agents that Reduce Work and Information Overload. Communications of the ACM, 37(7), July 1994 (1994) 9. Caglayan, A.K. and Harrison, C.G.: Agent Sourcebook: A complete Guide to Desktop, Internet, and Intranet Agents (1997) 10. Segal, R.B. and Kephart, J.O.: MailCat: An Intelligent Assistant for Organizing E-Mail. Proceedings of the 3rd International Conference on Autonomous Agents, Autonomous Agents ’99, Seattle, WA, USA (1999) 11. Takkinen, J. and Shahmehri, N.: CAFÉ: A Conceptual Model for Managing Information in Electronic Mail. Copyright 1998 IEEE. Published in the Proceedings of the Hawaii International Conference on System Sciences, HICSS98, January 6-9, 1998, Kona, Hawaii (1998) 12. Abu-Hakima, S., McFarland, C. and Meech J.F.: Proceedings of the 5th International Conference on Autonomous Agents, AGENTS’01, Montreal, Quebec, Canada (2001) 13. Mock, K.: An Experimental Framework for Email Categorization and Management. Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR’01, New Orleans, Louisiana, USA (2001) 14. Gruen, D., Sidner, C. and Boettner C.: A Collaborative Assistant for Email, Proceedings of Human Factors in Computing Systems, Extended Abstract (1999) 15. Weinberger, D.: Tacit Knowledge, KMWorld, 22nd Nov 1999 (1999) 16. Nonaka, I. and Konno, N.: The Concept of ‘Ba’: Building a Foundation for Knowledge Creation, California Management Review, 40(3) (1998) 17. Lave, J. and Wenger, E.: Situated Learning: Legitimate Peripheral Participation. New York: Cambridge University Press (1993) 18. Ackerman, M., DeMaagd, K., Cotterill S. and Swenson, A.: I-DIAG: From Community Discussion to Knowledge Distillation, Technical Report SS-03-01, AAAI Press (2003) 19. Collins, T.D., Mulholland, P. and Watt, S.N.K.: Using Genre to Support Active Participation in Learning Communities. In the Proceedings of Euro-CSCL 2001, Maastricht, NL (2001) 20. Masterton, S. and Watt, S.N.K.: Oracles, Bards, and Village Gossips, or Social Roles and Meta Knowledge Management, Information Systems Frontiers 2(3/4):299–315 (2000) 21. Grudin, J.: Groupware and social dynamics: Eight Challenges for Developers. Communications of the ACM, 37(1):92–105 (1994) 22. Beyer, H. and Holzblatt, K.: Contextual Design, Morgan Kaufmann (1998) 23. Wooldridge, M.: An Introduction to Multi-Agent Systems. John Wiley & Sons (2002) 24. Watson, I. and Marir, F.: Case-Based Reasoning: A Review, Knowledge Engineering Review, 9(4):355–381 (1994) 25. Mayfield, J.: Ontologies and Text Retrieval. The Knowledge Engineering Review, 17(1):71-75, Cambridge University Press (2002) 26. Kushmerick, N., Johnston, E. and McGuinness, S.: Information Extraction by Text Classification, IJCAI-2001 Workshop on Adaptive Text Extraction and Mining (2001) 27. Schank, R.C.: Tell Me a Story: Narrative and Intelligence, Rethinking Theory. Third edition. Evanston, Illinois: Northwestern University Press (2000) 28. Cleary, C. and Bareiss, R.: Practical Methods for Automatically Generating Typed Links, Seventh ACM Conference on Hypertext (Hypertext '96) (1996) 29. Schank, R.C.: Rules and Topics in Conversation. Cognitive Science, 1:421–441 (1977) 30. Patton, M.Q.: Utilization-Focused Evaluation: The New Century Text. Sage Publications (1997) 31. Tambe, M.: Agent Architectures for Flexible, Practical Teamwork. National Conference on Artificial Intelligence (AAAI-97) (1997) 32. Cranefield, S.J.S., Moreale, E., McKinlay, B. and Purvis, M.K.: Automating the Interoperation of Information Processing Tools, Proceedings of the 32nd Hawaii International Conference on System Sciences (HICSS-32), Maui, Hawaii, IEEE (CD-ROM) 10 pages (1999)

An Agent-Based Approach to Mailing List Knowledge Management

129

33. Guizzardi, R., Aroyo, L. and Wagner, G.: Agent-Oriented Knowledge Management in Learning Environments: A Peer-to-Peer Helpdesk Case Study, Technical Report SS-03-01, AAAI Press (2003) 34. Sparck Jones, K., Walker, S., and Robertson, S.: A Probabilistic Model of Information Retrieval: Development and Status. Information Processing and Management, 36(3):809-840 (1998)

Information Fields in Organization Modeling Using an EDA Multi-agent Architecture Joaquim Filipe Escola Superior de Tecnologia de Setúbal Instituto Politécnico de Setúbal Rua Vale de Chaves, Estefanilha, 2910-761 Setúbal, Portugal [email protected]

Abstract. The EDA model (Epistemic-Deontic-Axiological) is an agent model based on the social psychology theoretical classification of norms and corresponding attitudes: Ontological, Epistemic, Deontic and Axiological. EDA agents are situated in normative information fields, and are described in terms of the basic attitudes aforementioned. Information fields are used as the basis for coordinating an organization, which is seen here as a collective agent composed of other individual and/or collective agents and/or roles, that encompasses multiple embedded information fields. The paper discusses coordination and the representation of social structures based on using the EDA agent model combined with the notion of information field.

1 Introduction The norm-based framework that we propose here assumes that organizations can be modeled as abstract specifications, irrespective of the actual agents that will populate them, which in many cases can be both human and artificial. We believe that some of the many roles human agents perform in an organization – namely the less creative, less prone to exceptions and more repetitive ones – can be partially delegated to artificial agents, although we consider that keeping the ultimate responsibility in the human agent is unavoidable. The essential units in our model are the organizational roles and their relationships. The EDA model, described in this paper, was mainly conceived to facilitate the creation of social environments in terms of normative intelligent multi-agent systems (Filipe, 2000). The focus of our approach differs from other multi-agent systems approaches (Cohen and Levesque, 1990; Rao and Georgeff, 1991; Jennings, 1994), which mainly focus on the design of the internal (mental) structures of single agents instead of the normative (social) shared structures that underlie multi-agent cooperation. Here, we seek to describe how to build an organizational model based on the multi-agent system metaphor using the EDA agent model for providing a full lifecycle method that guides the designer in the model development all the way from the L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 130-142, 2003.  Springer-Verlag Berlin Heidelberg 2003

Information Fields in Organization Modeling

131

conceptual level to implementational level. In that process we adopt an organizational semiotics perspective (Stamper, 1973).

2 Semiotics and Business Process Modeling Semiotics is the science of signs (Peirce, 1931-58). Signs are social constructs that require meaning assignment. The semiotic approach to computing in organizations (Stamper, 1973, 2000), adopts a constructivist perspective and emphasizes the importance of the integration of computers in social reality. It is very important to make computer-based systems fit into a business organization and integrate information technology with the social aspects that enable the successful fulfillment of business goals. Sometimes highly sophisticated technology is applied without a clear understanding of the information circuits and information systems already in place. Norms are social constructs that represent business rules, social goals, constraints and other structural aspects of the organization and are essential for defining an agent’s roles, including the specification of its functions and obligations. The adopted approach views a business process as a process-oriented network of autonomous normative agents. Agents can represent individuals or collectives, including external stakeholders such as customers, regulators or suppliers, and internal entities such as staff, departments, or systems.

3 The EDA Model Using the social psychology taxonomy of norms, and based on the assumption that organizational agents’ behavior is determined by the evaluation of deontic norms given the agent epistemic state, with axiological norms for solving eventual conflicts of interest, we propose an intentional agent model, composed of three main components: the epistemic, the deontic and the axiological. • Beliefs are incorporated in the Epistemic component, • Obligations, rights and behaviors are incorporated in the Deontic component, and Axiological Component ∑ (values)

perception

action Deontic Component

Ψ

∆ (behavior)

Epistemic Component Κ (knowledge)

5: The EDAcomponent agent model relationships. Fig. 1: Theigure EDA model

132

•

Joaquim Filipe

Values (using a partial order relation of importance) are incorporated in the Axiological component.

Figure 1 depicts the EDA model and its component relationships. Ψ is a pragmatic function that filters perceptions, according to the agent ontology, using perceptual and axiological norms, and updates one or more model components. ∑ is an axiological function that is used mainly in two circumstances: to help decide which signs to perceive, and to help decide which goals to put in the agenda and execute. Κ is a knowledge-based component, where the agent stores its beliefs both explicitly and implicitly, in the form of potential deductions based on logical reasoning. ∆ is a set of plans, either explicit or implicit, the agent is interested in and may choose to execute. The detailed description of each component, including its internal structure, is provided in (Filipe, 2000).

4 Issues Involved in Building Normative Multi-agent Systems The EDA model is a normative intentional model developed for facilitating the analysis and design of coordinated behavior in organizations. More specifically, we propose the use of the EDA model in collaborative multi-agent environments. Since, according to our definition of an agent, agents must choose in a rational, autonomous and pro-active way their next actions (Wooldridge and Jennings, 1995), all agents are assumed to be rational decision systems, although the utility they tend to maximize may not necessarily be of an economic nature. The collaborative behavior requirement places several practical constraints on the multi-agent structure: • Agents need to have an information discovery mechanism through which they discover the existence, the location, and the roles of other agents, especially their capabilities, controlled resources and power relationships. This can be done using a special agent with whom other agents register. Such an agent acts simultaneously as a namespace server and a yellow-pages agent and will be referred hereafter as a facilitator. • Agents need a standard communication environment, including a standard language that establishes a communication channel through which agents are able to transmit and understand (syntactically) their messages. • Agents require an internal inference machine, that permits them to reason and make their choices, based on their EDA model state in each moment. • Agents need to have a common conceptual framework, with a shared representation and understanding (semantics) of the common domain concepts. We assume agents use a shared ontology socially constructed (offline) using

Information Fields in Organization Modeling

133

well-known methods of semantic analysis from organizational semiotics (Liu, 2000). We are particularly interested in organization modeling; therefore in the remaining of this paper we will consider the application of the EDA model to organizations. We also postulate that organizations are structured in terms of roles and agents who perform those roles (Biddle, 1979).

5 Organizational Multi-agent Architecture Roles are structured descriptions of agent behaviors. A role includes the specification of what an agent is able to do, is authorized to do and is obliged to do. In this paper we sometimes refer to roles as abstract agents, i.e. as agent shells, that to become active require instantiation by an active entity (either human or artificial) that can actually play the role. A role is thus necessarily defined prior to the assignment of an agent to fulfill it. Figure 2 depicts the assignment of a role to an agent as an EDA model composition process. The composition process that operates in each model component is essentially the merge of the two sets of norms: the set that existed in agent with the set that is provided in the role. Potential conflicts between the existing and the new EDA model components are avoided by keeping each knowledge statement indexed to the corresponding role. This indexing means that an agent may behave differently in different roles thus, inter-agent relationships such as conversations must always specify, either explicitly or normatively (by default), the roles in which each agent is participating. Role 1 E

D

Agent E

A1

E+E1

α

D

Agent

α D+D1

A+A1

A

Fig. 2: Role Instantiation / EDA Model Composition

5.1 A Role-Based Organizational Model The role-based organizational model we propose originates the kind of structure depicted in figure 3.

134

Joaquim Filipe

Collective Agents (Organizations)

Roles (Serv., Policies)

Individual Agents (human/artificial)

O1 [EDA]

R1 [EDA]

A1 [EDA]

O2 [EDA]

R2 [EDA]

A2 [EDA]

O3 [EDA]

R3 [EDA]

A3 [EDA]

…

…

On [EDA]

Rk [EDA]

… Aj [EDA]

gure 3: Organization-Role-Agent Architecture (3 classes)

Fig. 3: Organization-Role-Agent Architecture (3 classes)

This figure suggests that organizations are composed by roles and that roles are played by agents. There are three types in this diagram: one is the organization, another is the role and a third one is the agent. However, the relationships between them are not trivial. For example, in the figure we have depicted an organization instantiating a role; we have also tried to suggest that an agent can instantiate several roles, yet the following question may arise: “Can a role be played by more than one agent, simultaneously?” In role theory this is an open issue. However, due to the impact of the answer in design and implementation issues, we had to analyze the problem and make a decision. Before indicating our decision in this matter, we need to clarify what we mean by conceptual role hierarchy. 5.2 Conceptual Role Hierarchies The most salient feature here is that roles in conceptual role hierarchies are defined in terms of specialization relationships of the type class-subclass or class-instance, (simply denoted as “is-a” relationships). This kind of role hierarchy is independent of the power relationships that also relate different roles in an organization. Conceptual relationships are useful for modeling different abstraction levels and decomposing the different roles in such a way that they can automatically inherit the properties from more general roles instead of repeating the same properties in many roles. Inheritance also provides an efficient way of ensuring consistency whenever a general property is changed and all subsumed roles must be changed accordingly. Figure 4 shows a simple example. We could easily extend this example to add more roles under “Lecturer”: one for each different course lectured at the Teaching Institution being modeled. If the CS 1.0.1 course had more than one lecturer (e.g. one for theory and another one for laboratories) then we would add two sub-roles to the “Lecturer of CS 1.0.1” role, denoted perhaps by “Lecturer of CS 1.0.1 theory” and “Lecturer of CS 1.0.1 labs”.

Information Fields in Organization Modeling Staff member

135

is-a Lecturer is-a Lecturer of CS 1.0.1

Fig. 4: Conceptual role hierarchy

Finally, we add a distinction that is relevant to the current discussion concerning the role-based organizational structure: we designate as “Role instances” all the roles that are “leaves” of the conceptual role hierarchy and we call “Role classes” the remaining ones. Furthermore, we postulate that our organizational model can only associate agents with role instances (not role classes). For example, although it makes sense to speak and reason about the concept of lecturer and its relationships with other organizational roles, it is not allowed to assign a particular agent to play the role of lecturer (e.g. the role of a lecturer of a particular subject, to a particular class might be adequate, admitting it would be a leaf in the role hierarchy).

6 Organizations, Roles, and Agents For the sake of simplicity, and without loss of generality or expressiveness, we postulate that a role can only be played by one agent at a time, although it can be an individual agent or a collective agent. Collective agents are co-ordinated entities, that have their own goals and their own knowledge, i.e. they are organizations. The instantiation of an organization is, however, different from the simple instantiation of a role, because in the case of an organization its instantiation requires filling in all its roles or at least a sufficient number of roles for enabling its functioning. For example, the program committee of a conference is a collective agent, whose decisions concerning paper acceptance are taken using specific co-ordination methods, which plays a role in the conference organization. Another example: The scientific council of a School plays a role within the School organizational structure; it has a number of obligations and rights, and performs a number of legally instituted functions, which require the co-ordination of its members, typically using voting mechanisms. Each member of this collective agent performs a specialised role and the collective agent can be described as a composition of roles, which are instantiated by single agents. The examples illustrate the coherency between the collective agent concept, in our role-based multi-agent architecture model, and the concept of organization, as suggested above. A relevant aspect of this architecture is that an organization belongs simultaneously to two classes: it belongs to the class of roles because it must be instantiated by (multiple) agents in order to become active and it belongs to the class of agents because it can instantiate a role in another multi-agent system; to enable this we assign an EDA model to every collective agent. In this way it becomes possible for an organization to maintain its knowledge even after a complete change of the agents that instantiate it, which would not be the case if all the knowledge would be kept at the

136

Joaquim Filipe

individual agents’ EDA models. This important aspect is depicted in figure 5, indicating that organizations, roles and agents have different EDA models. Assigning an EDA model to a role is required because the agent that instantiates it may change during the enactment of a business process and yet the process ought to continue without any interruption, as if the agent was the same. For example, if a customer places a complaint concerning a certain defective product to a company representative, and it calls back the next day to change a detail in its complaint, it is not relevant that it talks with the same or other representative – ideally the customer should be able to continue the conversation without knowing whether it is talking with the same representative or not. 6.1 Multi-tiered Organizational Layer The highest normative layer is the organizational layer. However, since we postulate that any collective agent is an organization, the organizational layer may actually be composed by many organizational layers. Each layer corresponds to an information field since there is a one-to-one relationship between organizations and information fields. Therefore, to be precise it is necessary to establish a priority order within this multi-tiered layer. By definition, an organization Org1 subsumes another one Org2 (Org1! Org2) if is situated at a higher layer. For example, the department of Informatics at the School of Technology of Setubal is a collective agent (organization) that is subsumed by the School of Technology of Setubal. Therefore, it is possible that the department may have some specific beliefs, goals or values in addition to those prescribed at the School level. However, the department ought to be consistent with the norms defined by the School, otherwise a norm violation occurs. We postulate – coherently with the principle of the minimisation of conceptual distance, proposed by Touretzky (1984) and also according to what is usual in human organizations – that in case of conflict between two organizational layers, the agents at lower levels ought to assign a higher priority to the hierarchically immediate organization.

EDA Model Components Epistemic

Deontic

Axiologic

Org-beliefs and plans

Org-goals

Org-values

Role-beliefs and plans

Role-goals

Role-values

Priv-beliefs and plans

Priv-goals

Priv-values

Fig. 5: Normative knowledge levels

Information Fields in Organization Modeling

137

6.2 Role Resolution and Role Relationships Since we adopt the view that organizational co-ordination depends essentially on role interaction as much as the particular agent that instantiates it, organizational agent communication and co-ordination requires role resolution. In our multi-agent system architecture we assign this task to the facilitator. The facilitator is also a domain name server where all the agents must register whenever they enter the network, indicating the role(s) they are playing. The facilitator has access to the organizational ontology and all role descriptions. Whenever an agent requires an interaction with a certain role player it is necessary to identify the agent that is playing the role and ensure that the message is channeled to it. In some circumstances an agent who needs a particular service may request the facilitator to find out those agents that can provide the required service. This is a twostep process, involving firstly the identification of the list of roles that can provide the service and, secondly, the selection of one or more agents that play one of these roles, which the client is authorized to access.

7 A Collaborative Communicative Environment In figure 6 we show the typical collaborative and communicative environment that we use for organization modeling and implementation. Below, we show how the EDA paradigm can be effective in the modeling of organizational multi-agent systems, bringing together several notions previously described. 7.1 The Pragmatics of the EDA Model Conversations are meaningful sequences of speech acts that pragmatically modify the EDA models being used by both agents (sender and receiver). This is consistent with communication theories such as the Speech-Act Theory (Searle, 1969) or the Theory of Communicative Action (Habermas, 1984). However, it is not obvious whether the modification produced by a speech act is made at the agent level, at the role level or at the organization level. Consider the following example: an agent A1 is playing role R1 in organization O1, whereas agent A2 is playing role R2 in organization O2; if A1 needs to buy 100 screws from A2, then where is this fact represented? • It may be represented in the epistemic component of A1, if no other agent will participate in the acquisition process; • It may be represented in the epistemic component of R1 if another agent, A3, from the same organization O1, can (e.g. in another shift) play the same role as A1, and continue the transaction with A2. • It may be represented in the epistemic component of organization O1, if handling this kind of requests can be performed by more than one role and there is a dynamic binding of requests to roles.

138

Joaquim Filipe Facilitator (yellow pages)

Person (Human Agent)

agent addresses and capabilities

H1

Role Description (services,policies)

Role Agent A1

Role Agent A2

A

Conversation C1 1 Conversation

.. . Conversation

Role Agent A4

C2A1

CnA1

Comm. chan-

Conversation

C1A2

Conversation

C2A2

nel

Role Agent A3 Conversation

C1A3

Fig. 6: Collaborative Multi-Agent System Architecture

The procedure we propose for speech act pragmatic processing in organizational agents is the following: The speech act is always channeled first through the agent, so the decision procedure starts at that level: the agent must then decide, based on private and inherit deontic rules, whether the speech act should be processed at its private level and let it modify its private EDA model and/or it is of interest at a higher role level and, in that case, sending it up the nested EDA model hierarchy. Each role level would perform a similar decision process. The speech act upward movement can be blocked at any level, to avoid cluttering the higher levels, closer to central control. In communicative action theory, Habermas (1984) postulates the existence of three worlds: • The subjective world (how the speaker perceives the world) that is constituted by the feelings, beliefs, desires, experiences and intentions of the agent,

Information Fields in Organization Modeling

139

•

The common social (inter-subjective) world that is constituted by norms, commitments, agent relationships, and institutions to which the agents belong themselves, and which defines how agents stand towards each other, and • The objective world of objects and states of affairs (external world) that describes “how things are”. The pragmatics of speech acts may impact any or several of these worlds, therefore we need to address this problem in terms of the EDA model application and the information space concept. Since we are interested in using the organizational models for partially automating certain organizational tasks, we are interested mainly on aspects that can be formalized, because formalization is a precondition for automation. In organizational settings the subjective world seems to be too complex to be formalized and the objective worlds seem to be relatively easy to model using conventional methods, thus we will address here the social world (inter-subjective world). 7.2 The Representation of Social Objects In the inter-subjective world a shared ontology or inter-subjective reality that defines the social context (information field) where agents are situated. This kind of social shared knowledge is not reducible to individual mental objects (Conte and Castelfranchi, 1995). For example, in the case of a commitment violation, sanction enforcement is explicitly or tacitly supported by the social group to which the agents belong, otherwise the stronger agent would have no reason to accept the sanction. This demonstrates the inadequacy of the reductionist view. Once again, we look at human organizational models for designing multi-agent systems: contracts in human societies are often written and publicly registered in order to ensure the existence of socially accepted, and trusted, witnesses that would enable the control of possible violations at a social level. Non-registered contracts and commitments are often dealt with at a bilateral level only and each concerned agent has its internal contract copy. This observation suggests two possible representational models: • A distributed model: Every agent keeps track of social objects in which that agent is involved and may also be a witness of some social objects involving other agents. • A centralized model: There is an Information Field Server (IFS) that has a social objects database, including shared beliefs, norms, agent roles, social commitments, and institutions. The distributed model is more robust to failure, given the implicit redundancy. For example, a contract where a number of parties are involved is kept in all concerned agents’ knowledge bases, therefore if an agent collapses the others can still provide copies of the contract. It is also more efficient assuming that all agents are honest and sincere; for example, commitment creation and termination involved in business transactions would not need to be officially recorded – a simple representation of a social commitment at the concerned agents EDA model would suffice. However, since these assumptions are often unrealistic, the distributed model cannot completely replace the role of certified agents, trusted by society to keep record of

140

Joaquim Filipe

shared beliefs and social commitments. We assume here that these social notions are part of the ontology that is shared by all members of an information field; that’s why we call these trusted repositories of the shared ontology “Information Field Servers”. These servers have the following characteristics: • Different information fields must have different IFS because the shared ontology may differ among specific information fields. • Each information field may have several non-redundant IFS, each representing a small part of the shared ontology. • The robustness problems of IFS are minimized by reliable backup (redundant) agents. Communication bandwidth is another relevant factor to consider: if all social objects were placed in central IFS agents these agents might become system bottlenecks. In figure 7 the architecture of the inter-subjective level is depicted with respect to the localisation of social objects in addition to showing an example of how social objects are used at the subjective level. This figure emphasizes commitments. Commitments are actually very important coordination instruments. They consist of conditional obligations to do some action (e.g. contracts), thus they are represented in the deontic part of an EDA model. In the social EDA model we propose they are first class objects that can be represented either in the agents’ EDA models (which we designate as the agents’ space) or in the IFS’ EDA model (which we designate as the Information Field Server’s space). In the example above, agents A1 and A2 have only an internal representation (in each EDA model) of a shared commitment C1, whereas Agents A2 and A3 do not have an internal representation of commitment C2 because this commitment is represented in IFS1. All agents A2 and A3 need is a reference (i.e. a pointer) to that shared commitment. The latter solution is preferred for commitments that the agents intend to make public, and in that way make the commitment stronger. Inter-subjective Level Objects Possible Localization Regular agents’ space

Information Field Servers’ space IFS1

Commitment C1

Subjective Level

Agent A1

Commitment C2

Agent A2

Fig. 7: Social objects representation and usage

Agent A3

Information Fields in Organization Modeling

141

8 Conclusions and Future Work Organizational co-ordination is an activity that requires viewing an organization as made up of normative, role-playing, agents with social obligations and personal interests, which they eventually convert into agent-level intentions and persistent goals that could be pursued using a network of relationships with other agents. This paper addressed the coordination issue using an intentional normative agent model (EDA) that is supported by a norm classification theory from social psychology, and the concept of information field as the normative support of coordinated groups of agents. We have described how the EDA model can be used together with information fields to create role-based organizational models and to describe social structures which are commonly seen in practice for the coordination of human activities, namely based on private and public commitments. The advantage of the EDA model is that it can be used uniformly to model individual agents, collective agents and even abstract agents such as roles and information fields. Coordination usually requires communication. The meaningful unit of communication is the conversation (a sequence of speech acts). However, to be effective, conversations require not only a common communication language but also a common ontology and the mutual understanding of several normative notions, i.e. the context of an information field. An important aspect of communicative coordination is the pragmatic intake of the conversation. A conversation can change one or several components of an agent model (Epistemic, Deontic or Axiological) and it can also produce effects at several levels: agent level, role level or organizational level. Although a small case study has been implemented, based on a toy problem, using the Java-based communication environment Jini combined with the knowledge representation tool Jess, the next step in the research work that was described here is the implementation of a full scale case study, in order to assess in practice the effectiveness of the proposed theoretical model. The links to other work in the area of social agent systems will be developed in the near future, especially in relation to the work of Carles Sierra and colleagues, such as (Vasconcelos et al., 2001; Esteva et al., 2001), and also in relation to the work of Dastani et al. (2003) and Dignum (2002a; 2002b), especially related to organizational roles, norms and deontic logic.

References 1. 2. 3. 4.

Biddle, B., 1979. Role Theory: Expectations, Identities and Behaviors. Academic Press, New York. Cohen, P. and H. Levesque, 1990. Intention is Choice with Commitment. Artificial Intelligence, 42:213-261. Conte, R., and C. Castelfranchi, 1995. Cognitive and Social Action, UCL Press, London. Dastani, M., V. Dignum, F. Dignum, 2003. Role Assignment in Open Agent Societies. In: Proceedings of AAMAS’03, Second International Joint Conference on Autonomous Agents and Multi-agent Systems, Melbourne, Australia.

142 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.

15. 16. 17. 18. 19. 20.

Joaquim Filipe Dignum, F., 2002. Software agents and e-business, Hype and reality. In J. Filipe, B. Sharp and P. Miranda (eds.) Enterprise Information Systems III, Kluwer, Dordrecht, pages 2235. Dignum, F., 2002. Abstract Norms and Electronic Institutions. In: Proceedings of International Workshop on Regulated Agent-Based Social Systems: Theories and Applications (RASTA'02), at AAMAS, Bologna, Italy. Esteva, M., J. Padget and C. Sierra, 2001. Formalizing a Language for Institutions and Norms. In Proceedings of the 8th ATAL, Seattle, USA. Filipe, J., 2000. Normative Organisational Modelling using Intelligent Multi-Agent Systems. Ph.D. thesis, University of Staffordshire, UK. Habermas, J. 1984. The Theory of Communicative Action: Reason and Rationalization of Society. Polity Press. Cambridge. Holt, A., 2000. Tutorial Series on Organised Activity, Escola Superior de Tecnologia, Setúbal, Portugal. Jennings, N., 1994. Cooperation in Industrial Multi-Agent Systems, World Scientific Publishing, Singapore. Liu, K., 2000. Semiotics Information Systems Engineering. Cambridge University Press. Cambridge, United Kingdom. Peirce, C., 1931-1958. Collected papers of Ch. S. Peirce (8 vols.), C. Hartshorne and P. Weiss (Eds). Cambridge, Harvard University Press. Rao, A. and M. Georgeff, 1991. Modeling Rational Agents within a BDI architecture. Proceedings of the 2nd International Conference on Principles of Knowledge Representation and Reasoning (Nebel, Rich, and Swartout, Eds.) pp.439-449. Morgan Kaufman, San Mateo, USA. Searle, J. 1969. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press. Stamper, R. 1973. Information in Business and Administrative Systems. John Wiley & Sons. Stamper, R., 2000. New Directions for Systems Analysis and Design. In J. Filipe (Ed.), Enterprise Information Systems, Kluwer Academic Publishers, Dordrecht. Touretzky, D., 1984. The Mathematics of Inheritance Systems, Ph.D. thesis, Carnegie Mellon University, USA. Vasconcelos, W., J.Sabater, C. Sierra and J. Querol, 2001. Skeleton-based agent development for electronic institutions. In proceedings of UKMAS’01. Wooldridge, M. and N. Jennings, 1995. Agent Theories, Architectures and Languages: A Survey. In Wooldridge and Jennings (Eds.), Intelligent Agents, LNAI 890, SpringerVerlag.

A Quantum Perturbation Model (QPM) of Knowledge Fusion and Organizational Mergers William F. Lawless1 and James M. Grayson2 1

Paine College, 1235 15th Street, Augusta, GA 30901-3182 [email protected] 2 Augusta State University, 2500 Walton Way, Augusta, GA 30904 [email protected]

Abstract. Future multiple agent system (MAS) missions are at risk whenever agent interactions occur faster than humans can intervene. Yet, no first principles exist that can be applied to improve mission success, solve illdefined problems (idp’s), or increase autonomy. To address this issue, agent mediated knowledge management (AMKM) must determine the fundamentals of information, I, and knowledge, K, generation plus their fundamental relations with agent organizations, decision-making, trust, cooperation, and competition. With the discovery of a group process rate equation, an ab initio approach to a new rational perspective of organization formation based on first principles was linked with K fusion and organizational mergers. Coupled Kolmogorov equations for I and K required distinguishing stable procedural-algorithmic K from unstable interaction-belief-expectations K (Kχ). Results suggest organizations use endogenous feedback from perturbations to control, “tune”, or defend against competition; conversely, exogenous I modifies competitive attacks against an organization. The quicker to respond survives.

1 Introduction This research is designed to construct not only a computational theory of K generation from argument, relating it to other lines of research such as Markovian models, but also to illustrate how important is K and argument for a theory of autonomy, especially when interactions within an MAS occur between agents at a distance too far to permit timely human intervention (e.g., beyond Mars), or between two MAS’s opposed on a battlefield or in industry (possibly as an extension of Robocup). Without the tool of argument to generate new K, agents are restricted to “rational” theory space, constraining them by logic to the solution of a problem based on the set of available propositions or existing knowledge (e.g., [69], pp. 24-26). From a traditional rational perspective, groups serve to reduce the workload on an existing problem using teamwork and adaptive learning, but as a consequence, promoting the idea that command or consensus decision-making (CDM) is superior to democratic decisionmaking. Under this traditional, rational perspective, autonomy becomes a channel to implement existing solutions, making autonomy ineffective (e.g., Tambe’s experience with managing the schedules of his colleagues at USC with artificial agents; [60]). However, from a different direction, we have found that argument is more applicable L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 143-161, 2003.  Springer-Verlag Berlin Heidelberg 2003

A Quantum Perturbation Model (QPM)

145

In our first extension, organizations grow when new recruits or acquisitions reduce combined energy, E, to form a lower joint ground state; bonding increases between interactants as vocal frequencies converge (resonance; e.g., given the frequency difference between two individuals at their average E ground state, resonance occurs as their joint E state converges and reactance as it diverges); interaction success depends directly on the size of its cross-section, a new concept (see Equation 4); and the likelihood of an interaction varies inversely with the E it requires (four respective examples: a successful versus a failed merger; a highly respected business manager; multiple supply chains versus single ones; and AMKM software purchased contingent on its functional capability). From the perspective of human research, the “groupness” problem arises by recognizing that once the members of a group have been surveyed with questionnaires or polls (e.g., attitudes), aggregating the individual data does not reconstitute the group (adapted from [76]). For example, a study in 1991 (pp. 102-3 [67]) with questionnaires led to the prediction that the negative images associated with a nuclear waste repository at Yucca Mountain would harm the Las Vegas economy by reducing tourism; however, ten years later the authors admitted that tourism continued to make Las Vegas the fastest growing community in the U.S. To avoid “groupness” effects in bargaining situations, Nash [55] assigned zero social value to groups with internal dissent, as a consequence focusing game theory on rational groups where values might be summated (for a review, see [42]). But even for stable, dissent-free groups, Lewin [45] famously recognized that a group is different from the sum of its parts. From the perspective of an autonomous artificial agent, the “groupness” problem can be illustrated with an example of Predator: Currently, a staff of 20 humans is now required to operate a single Predator drone, yet the crash rate is 100 Predators to one piloted USAF aircraft (e.g., [58]). To reverse this relationship between staff and artificial agents, or to link agents like Predator into an MAS or organization that broadly shares I, is a goal that will require the rational control and optimization of group processes, beginning with “groupness”, a revolution in computing foundations that permits autonomous teams of computational agents to solve problems better than the current generation of remotely controlled unmanned systems [19]. To achieve this goal requires understanding the biological function of “groupness” as well as cooperation, conflict and turmoil among autonomous agents. In contrast to traditional models, QPM suggests that combining competition and cooperation with argumentation produces a robust model of decision-making that increases in computational power with N [37]. This is not a new perspective in AI. While cooperation reflects K, competition generates I by breaking symmetry (e.g., Robocup). What is new is that an analytical approach for decision-making and group formation can be extended to K fusion and organizational mergers, and used in equations adapted from atmospheric science and biology that suggest the possibility of being able to computationally “tune” the control of an organization. In SQM, “groupness” is entanglement, occurring whenever forces endogenous to an interaction are stronger than context. Given the absence of entanglement, human preferences in laboratory game experiments are robust and stable, but once entangled, preferences change dramatically even as subjects subsequently deny that their choices had changed or that social influence had affected their choices. Kelley [33] attributed this phenomenon as the cause of why game theory could not be validated.

144

William F. Lawless and James M. Grayson

to solving idp’s [43],[47], like those confronted in the courtroom, political arena, market economy, or the practice of science—all occurring in social settings. From this social or group perspective, argument between autonomous agents is critical to generate and process I into new K, underscoring that argument between autonomous human agents is the only path to the solution of difficult problems (idp’s), but also as a consequence, promoting democracy as the seed for self-organization and selfregulation [12],[63], the key ingredients of autonomy. More broadly for agent-mediated knowledge management (AMKM) systems, our thesis extends Dignum and Weigand [20]. Not only is K an organization’s most important asset, but also how the organization reacts to this K is important to its competitors. Just as argument determines the stronger belief among many beliefs, competition between organizations can determine the K that survives (see Figure 5). One reason given to avoid argument or conflict is that it leads to poor outcomes for all (p. 7 [7]), specifically a lack of trust, indicating both that trust is a function of the cooperation among agents and that competition reduces social welfare. But this assumption has never been validated. According to Wendt [73], cooperation is inadequate to generate I about others to produce trust. Supporting Wendt but not Axelrod, our research compared a Citizen Advisory Board at the DOE Savannah River Site (SAB), in Aiken, SC, using majority rules to foster “a competition of ideas” to cleanup nuclear wastes with the Hanford Citizens Advisory Board (HAB) in Richland, WA, using consensus rules to foster cooperation; we found significantly more conflict in the former but also more trust (similar results were found between nations; [39],[40]). Further, the downside of cooperation (e.g., corruption; a reduction in computational power; terrorism; in [42]) limits the number of computational agents that can cooperate to solve ill-defined problems, idp’s [30],[47]. Another reason given by traditionalists to avoid argument is the belief that the best societies should avoid turmoil and can do so with adaptive learning instead of competition, exemplified, according to Skinner [66], by Communist China; unawares, Skinner drew his conclusion shortly after millions of Chinese had died from famine and cannibalism during China’s “Great leap forward” (1958-62; in [10]). These beliefs seem predicated on devaluing turmoil and democracy, but we have found that turmoil and democracy are correlated with better living standards, science, and health [40], in part because deciphering intent during interaction is more easily inferred under competition than cooperation. As Wendt (p. 360 [73]) explains, paradoxically, intent inferences occur when differences become evident. Dating from the 1920’s [1], the major unsolved problem in social interaction, from either social psychology experiments or the mathematics of game theory, remains the inability to distinguish an aggregate of autonomous individuals from a disaggregated group of those individuals [44]. This fundamental “groupness” problem arises primarily from theories constructed for individual agents who process I based on rational choices; e.g., in game theory, the choice to cooperate or compete pits opponents over a set of independent choices with maximum individual utility as the game solution [50]. In contrast, our quantum perturbation model (QPM) has made progress on understanding “groupness” with the primary interaction factors of action and observational uncertainty that constrain a group to act as a decision center; as such, maximum utility becomes a group’s ability to generate and process I [41]. We have since made extensions to organizational [42] and argument theory [43].

146

William F. Lawless and James M. Grayson

In the second extension of our theory, we associated quantum-like square E wells with argument during decision-making [42]. After finding that interaction crosssections are related to vocal frequencies, we speculated this also applies to brain waves: if gamma waves (≈ 40 Hz) mediate the binding of sensory features into objects and concepts [24], transitions between opposing views in an argument act as K concept reversals that reflect the time required for humans to sufficiently grasp and apply difficult K concepts to solve idp’s, linking solution “detection” to signal detection (for a review of the quantum model as a satisfactory alternative to signal detection theory, see [48],[49]; see also [47]); e.g., forming an ad hoc team to discuss and solve a perplexing customer problem (p. 69 [28]), reflecting that open organizational approaches encourage innovation [20]. We postulate that the flexibility to permit a series of concept reversals during argumentation is necessary for decisionmakers originally neutral to a set of arguments to determine whether one particular argument can be defended “against all contestations” (p. 76 [53]). In contrast to using force or learning with CDM to moderate social turmoil, in democracies, as long as a majority of decision-makers remain in the neutral middle [43], the competition for their support by opposition groups generates I but also moderates conflict sufficient to process I. Without this vital neutral center [63], violence and social breakdown follow as has happened recently in Columbia, South America [35]. Thus, too little conflict, no self-organization, new K or innovation; too much conflict, and problem solution processes break down. Taken together, these findings illustrate the theory behind Western systems of justice, markets and science is that the same data can lead to orthogonal or incommensurable interpretations that can be exploited to power I processing in observers neutral to argument, evolving a social system [43], producing a striking similarity between quantum computation and democracy in politics, markets or science [41]. According to Lloyd [46], with temperature, T, governing the number of operations that can be performed each period, the more E available to a quantum computer, the more computations it can perform. E can be ‘borrowed’, ‘invested’ in a computation, and returned to storage with a net ‘profit’ as processed I. For example, notwithstanding his contribution with game theory to the study of cooperation, Von Neumann [72] wanted mathematicians to act more like physicists: they signal the limits of rational thought with conflict (rising social T), they never avoid conflict (social computation), and their resolution of conflict creates the largest advances in rational thinking (social profit). As a speculation, to meld Nash’s [55] criteria of avoiding conflict as a prerequisite for negotiation with the quantum I approach, the two orthogonal states of cooperation and competition should act simultaneously to moderate turmoil to process I. If opinions are diametrical (180 deg out of phase; e.g., “Concept A is right” versus “Concept A is wrong”, promoting division and conflict), versus orthogonal (“Concept A is right” versus “Concept B is right”, promoting more opportunities to search for McBurney and Parson’s “winning” argument), diametrical concepts represent conflict and less opportunity for negotiation, while orthogonal concepts represent the I processing that engages an entire system, promoting negotiation and compromise.

A Quantum Perturbation Model (QPM)

147

1.1 Introduction. Summary When it comes to solving idp’s, we have found that the weakest decisions are made by individual rational logic, teams, or consensus seekers, the underlying rationale to CDM (e.g., authoritarian decisions, military failures such as the USS Vincennes shoot-down of an Iranian airbus, and bureaucracies; in [43]). There is nothing wrong with arriving at consensus. The danger is in making decisions by seeking consensus [32], thereby increasing the power of the weakest members in a group while reducing the ability of the group to process I. Even the previously consensus-minded European Council has rejected consensus decision-making as inefficient (p. 29 [74]). As an alternative, assuming that social reality is bistable (i.e., action I and observation I are conjugate), to eliminate redundancies and gaps in large complex systems and reduce their overall cost, a system of agents with multiple factions of complementary beliefs and actions expressed as argument, producing something akin to a virtual K characterized by increasing belief strength associated with decreasing observational accuracy about agent actions [39], strangely, will increase I processing power as the number of decision-makers increase [37], precisely the opposite of what happens as traditional computational power increases: As systems get larger and more complex, there is evidence that utility and productivity are increasingly falling off the curve that tracks pure processor size and speed leading to the conclusion that no amount of pure computational power will afford us the kind of intelligent computations that we need to address new problems, thus investing in more of the same will not get us where we need to go [19]. Traditional models cannot easily account for I processing among agents, trust, the value of emotion between agents, or what it is about groups that make them superior or inferior decision makers, but these are naturally derived with conjugate or quantum models [37]. Nor can traditional models explain the power of a plan that succeeds, like the U.S. Constitution—an imperfect plan written by imperfect men that has allowed imperfect leaders and imperfect power centers to compete and excel—with its system counterbalanced to maximize autonomy among multiple deciders yet constrain autonomy with checks and balances (from Montesquieu), yielding a system to reduce corruption, increase trust, and balance cooperation and competition with tension [12]. In contrast (e.g., [41]), QPM helps us to see that in contrast to the “efficiency” of CDM, and with it Plato’s model of the “ideal leader”, the social tumult associated with argument to persuade the neutral middle during decision-making in democracies and decision centers continuously tunes a decision with feedback, converting argument into a source of bifurcations that increase choices, whereas a single leader or bureaucracy slows the rate of evolution by dampening argument in favor of consensus and homogeneity [43]. To computationally simplify cognitive models, Simon believed that rationality was bounded, but bounded rationality mischaracterizes conjugate I. The chief characteristic of an optimum decision-making system is one that can exploit the conjugate I that exists in every social interaction, yet at the same time accepts that conjugate I precludes participants from accurately articulating their own decision processes [42], e.g., legal decisions sometimes become highly valued as precedents even as the best rational justifications for them have virtually no social value [59]; similarly in physics, Planck spent years in a failing attempt to justify his accidental discovery of discrete energy packets that ended the traditional view of causality so

148

William F. Lawless and James M. Grayson

important to his prior mechanical view of reality, R; and, in decision theory, Tversky and his colleagues [65] found no relationship between a decision and its justification.

2 Mathematical Approach 2.1 “Groupness”, Entanglement, and Social Decision-Making The primary weakness in traditional rational theory, exemplified by game theory, is that uncertainty for action and observation in the interaction is treated independently (pp. 147-8 [72]). To overcome this weakness requires bistable interdependence ([40]; derived from [13]) between uncertainty in action I, ∆a (where I = -∑ p(x) log2 p(x), and I flow = a = ∆I/∆t), and I uncertainty, ∆I, to give ∆a∆I > c (1) Since the constant c is unknown, boundary conditions are necessary to solve Equation (1). As ∆a -> 0, ∆I -> ∞, implying that observational I by observers is unbounded as skills are maximized. Alternatively, to process I and choose the optimal action for an idp requires orthogonal differences between entangled groups, like Republicans and Democrats after the last presidential election, to create tension sufficient for emotional responding in order to convert I into K yet insufficient to precipitate violence [37]. Previously we had found that, compared to consensus processes, emotions generated during decision-making if managed are associated with better decisions, more trust, and less conflict at local and national levels [43]. (Inversely, as ∆I -> 0 to form ideology or strong beliefs, ∆a -> ∞; this inverse case disproves Simon’s predicted correspondence between expert K and expert behavior [39],[40]). Conversely, maximum deception occurs when deceivers cooperate with opponents to reduce emotional responses to the K of their politics. Similarly, to maintain power with deception, dictators block I flow. Once severed, entangled interactions render accounts into individual histories that cannot recreate the interaction, explaining the lack of validity with self-reports in social science [22], making the intent of agents uncertain (e.g., p. 360 [73]). One approach to detecting intent is to study a group’s reactance to I change [14], analogous to inertia, thus revising Equation (1) to time, ∆t, and energy uncertainty, ∆E [39] to give: ∆a∆I = ∆ (∆I/∆t) • ∆t/∆t • ∆I = j• ∆ (∆I/∆t)2 • ∆t = ∆t∆E > c (2) Equation (2) predicts that as time uncertainty goes to zero, E becomes unbounded (e.g., big courtroom cases or science); inversely, when ∆E goes to zero, time becomes unbounded (e.g., at resonance voice boxes operate a lifetime). Quantizing the interaction results in E wells localized around beliefs as potential E set points, accounting for j. As increasing E levels approach these set points, emotions increase, forcing a return to stability (e.g., in set point theory, an “insult” provokes an agent’s response as its set points are engaged, or a group when its “laws” are broken; for a review, see [37]; our use of set point theory follows Chomsky’s idea that meaning exists only with function; from [57]). Axtell [6] questions the validity of QPM. But his belief that application follows validation is a test that game theory has never passed (for strengths and weaknesses of game theory, see [42]). Several reasons besides Bohr’s exist to support QPM. The

A Quantum Perturbation Model (QPM)

149

brain acts as a quantum I processor, converting photons hitting the retina into usable I [26]. Unlike the continuous model of traditional signal detection theory (i.e., ROC curves), the Bèkèsy-Stevens quanta model is based on detecting discrete stimulus differences from background, producing a linear relationship between threshold and saturation [49]. The dichotomous choices in decision making have been quantized (e.g., [23]). And evidence for the validity of Equation (2) comes from Penrose [29]: if the brain is one unit and c Planck’s constant, h, then ∆E∆t = ∆(hω/2π)∆t > h/2π reduces to ∆ω∆t > 1. Choosing ω for human brain gamma waves (≈40 Hz) associated with object awareness gives a reasonable minimum ∆t of 25 ms [18]. 2.2 Information Density Functional Theory (IDFT) IDFT models the function of I density and discrete E effects in an organization from adding or removing members. It also measures the need to reorganize; e.g., organizations gain capability in part by reorganizing to execute better (Jack Smith of GM, in [28] p. 74). A group forms by entangling I from an aggregation of individuals to solve an idp, like designing a complex weapon [3]. The chief characteristic of an idp is that K concepts do not correspond to objects or actions in R. Once solved, however, an idp becomes a well-defined problem, wdp, characterized by a correspondence between K, skills and R [47],[62] (cooperation implies low I density, maximum K, and the E ground state). In the solution of wdp’s, individuals function in roles as part of a stable network oriented by a shared emotional potential E field. The potential E surface (EPES) represents the hierarchy and geo-cultural differences across a group, organization, or society [62]. A recruit moves across the E surface of an organization, Rorg, where ETOT is the ground state and PES the minimum total E along the z coordinate of the organizational configuration (its hierarchy) until reaching a minima (stability), EPES (x,y) = minz,R-org ETOT (x,y,z,Rorg). (3) By construing the group as a series of interdependent interactions among individuals represented approximately as vocal harmonic I resonators, different processes, P, of an organization, like its growth rate from diffusing or adsorbing recruits, are given by: ΓP = nAnB a σAB exp (-∆A/kBT). (4) where nA and nB are the numbers of recruits and leaders interacting; a = ∆I/∆t; σAB is the cross-section (the probability that an interaction succeeds is an area determined by the vocal frequency reflecting the field forcing function FE(t) (see Equation 11) of say indoctrinators, ωE, and recruits, ω0, increasing rapidly as vocal “matches” increase (with frequency differences determined at the average E ground state, resonance occurs as the joint E state converges to a joint E baseline, with reactance proportional to its divergence); i.e., f(ωE4/(ωE2-ω02)2)); exp(•) is the probability of sufficient free E, ∆A, for the interaction to occur, with kB as Boltzman’s constant and T as agitation, turmoil, or emotional temperature (T = ∂E/∂I); here exp (•) becomes the probability for interaction, decreasing with the increasing free E required, ∆A, and increasing for rising T (instead of Boltzman’s constant, a normalization factor could be used or combined with T to form an estimate of variance, such as q2). Equation (4) indicates that the more ∆A required for an interaction, the less likely it occurs; that friendship is optimal for those who listen to synchronize, similar to resonance between harmonic

150

William F. Lawless and James M. Grayson

oscillators; and that terrorists manipulate their cross-sections by cooperating with field agents to preclude warning observers about their hidden intent. Once a bond forms between two members, A and B, the ground E state of the group becomes less than the aggregate of its members, the difference being the binding E, W, calculated from the configuration of barriers and nearest and nextnearest neighbors (see Figure 1). E

2A + B

²A W

A 2B

x,y (PES surface)

Fig. 1. The binding E to form a group or break it up. Shown here, two followers (2A) bind together with each other and to one leader (B) to form a group (A2B).

The E required to reverse the process and break apart the group becomes ∆A + W. Assuming that two recruits (A) bind to one another and to one leader (B), the Hamiltonian (Lyapounov function) consists of a site contribution, H0, and an interaction term, Hint, with: H0 = EbA ∑knk + EbB ∑kmk + VA-B ∑knkmk (5) where k as a role site, nk, is either 0 or 1 if k is empty or filled, mk is the same for leader sites, V is an interaction parameter, and Hint = 1/2V1nA ∑k,anknk+a + 1/2V2nB ∑k,bnknk+b +1/2V1nB ∑k,amkmk+a + 1/2V2nB ∑k,bmkmk+b + 1/3 VtrioB ∑k,a,a’mkmk+amk+a’+ … (6) Here k + a and k + b denote nearest and next nearest sites. 2.3 Mergers and Knowledge As an aggregate of individuals with independent sets of beliefs interact, agitation (turmoil) first increases as they aggregate then reduces over time to a new joint E ground state if the interactions serve and stabilize the group to form interdependent emotional fields that orient shared K and skills to correspond with objects in reality, the amount of effort (training) determining E well-depth. Effort expended to convert I into K, and to recruit, indoctrinate, and train new members departs from resonance. K fusion also requires a rise in agitation [45] which we now define as emotion T (where T = ∂E/∂I). While physical T has no analog in social systems, assuming that T is the average activation (emotion) experienced in an I field (FE(t) from Equation 11), T affects the strength, f(E), of an interaction structure, or f(E) = 1/(1 + exp(-2E/kBT)); as T drops, interaction structures become more reactant to I change. Baseline T or minimum E occurs at resonance, the point of greatest cooperation between agents (see [37]). Conversely, as emotion T increases, I eventually becomes random, causing organizational bonds to dissipate, illustrating creative destruction [64].

A Quantum Perturbation Model (QPM)

151

The USS Vincennes, while engaged in battle with Iranian gunboats in 1988, inadvertently shot down an Iranian Airbus (www.crashdatabase.com), illustrating that the danger of CDM is the belief that there always exists a single rational decision superior to a democratic solution [11]. But democratic decision-making can minimize operational mistakes when orthogonal arguments are used to generate I (alternatives) by strengthening opposing arguments until neutral decision-makers reach an optimum decision or K [43]. To fuse data into K requires a T sufficient for neutrals to process an argument, but not so high as to impair the interaction. Possibly the acquisition by neutrals of the concept being forwarded by one of the discussants followed by the reverse concept from the second discussant helps to capture the essence of a problem, and that a certain number of concept reversals are required to decide on the solution path, indirectly measuring the degree of the problem and the E required to achieve K fusion. We speculate that using these “reversals” will reduce tragic blunders. Andrade and Stafford [4] noted that mergers from an excess of profit can reduce social welfare (e.g., gaining price control across a market) and promote organizational fragmentation; e.g., criticisms from Senator R. Shelby, Republican: “We've made some adjustments, but the cultures have not changed between all the intelligence agencies making up the community. I don't believe they're sharing information. There's no fusion, [no] central place yet to do it” (www.cbsnews.com/sections/ftn). Similar to optimum decision-making or K fusion, reducing fragmentation in an organization requires a rise in emotional agitation or T to break barriers between groups (e.g., the inability to integrate is the putative cause of the failure in 2003 of the AOL-Time Warner merger). In contrast, mergers for survival occur in a consolidating market (e.g., the loss of purchasing power across a market sector); they are highly threatening, spontaneously causing a rise in T, but also reducing the likelihood of cultural clashes; e.g., in past merger plans between AOL Time Warner’s CNN and Walt Disney’s ABC News (www.latimes.com), internal conflicts over star anchors and content control derailed merger talks meant to help both compete against Fox News; however, as the plunge in advertising revenue exacerbated, merger talks resumed. Mergers to survive are strikingly similar to slime molds and ants when their environments are threatened (pp. 33, 236 in [56]), suggesting that Markovian master equations already developed in biology can be applied to organization mergers and K fusion. The processes above (Equations 3-6) can be used to model mergers between heterogeneous groups. They model the stresses that accrue from a mismatch between an organization and its merger target. As a heterogeneous merger target nucleates on the surface of the merging organization, the tension on the target to be absorbed by the lead organization relaxes the larger the target island grows, separating the lead organization and target island’s leaders, the release of E freeing and driving island members to build a hierarchy of leaders less like those in the lead organization, consequently placing a premium on a joint merger strategy and joint implementation team to fully and quickly integrate the leadership and members of both organizations (e.g., the Hewlett-Packard and Compaq merger of 2002-3 is one year ahead of schedule and has already saved over $3 billion; www.hp.com/hpinfo/newsroom).

152

William F. Lawless and James M. Grayson

2.4 Integrating Cooperation, Competition with “Groupness” and Entanglement I produced between members within a group (MAS) or between MAS’s is as follows: I(x1,x2) = I(x1) + Ix1(x2). (7) Equation (7) is joint entropy [17]. It has a range of {min(I(x1) or I(x2)) to max(I(x1) + I(x2))}. Thus, minimum entropy occurs during cooperation or when agents or groups are dependent or slaved to a leader; maximum entropy occurs when all agents or groups are autonomous. Looking at I transmitted between two or more agents gives: IT(x1:x2) = I(x1) + I(x2) - I(x1,x2) = I(x2) - Ix1(x2). (8) I transmitted ranges from {0 to the maximum of min(I(x1), I(x2))}. This range means that I transmitted is minimum when agents or groups are autonomous and maximum when fully interdependent or slaved together. When two or more agents agree (consensus, cooperation, minimum conflict), I is minimum but K is maximum. However, when agents disagree (dissensus, competition, maximum conflict), I is maximum but K is minimum. From Equations (7) and (8), these results for argument and autonomy follow. 1. Maximum autonomy leads to maximum I but with zero I transmitted. 2. Autonomy promotes argument, which in turn constrains autonomy; competition maximizes both the generation of I and the I processed into K. 3. Oppositional autonomy between agents generates conflict, polarizes neutral audiences, and reduces I processed into K. The result is maximum reactance or resistance between agents or groups--the inverse of resonance. 4. Autonomous agents as argument leaders are unable to persuade neutral audiences; to persuade and to build maximum cooperation or rapport with audiences, autonomy must be reduced. 5. Thus, transform autonomy into an orthogonal relationship between argument leaders by providing alternative views to increase I variance yet permit argument leaders to build rapport with audiences, minimizing conflict between the three groups while maximizing competition without conflict between argument leaders. The very problem Nash [55] sought to avoid can be exploited to generate the maximum force between two argument leaders, driving concept reversals to reduce I into K (near zero social forces, or ∑F = 0 = F1 –F2), and, by constraining autonomy, promoting opportunities for compromise between the leaders or within neutral audiences. 6. The latter condition occurs when the three groups become entangled. Nonentangled groups are unable to solve a problem, remaining “deaf” to the other’s argument [70]. Entanglement occurs when interaction forces between agents or groups are greater than the external forces during problem solution. In this fundamental state, K at the local level of an interaction becomes more relevant than global K, producing the self-organization unique to a democracy but not CDM. The entanglement of neutrals with the driving force of an argument leader can be modeled with a Markovian process. The special case of conflict entanglement can be modeled with a limit cycle for each side, the winner’s cycle narrowing over time. 7. At the beginning or argument, an entangled MAS exists in a state of superposition simultaneously equal to both positions of the argument leaders. The subsequent decision or outcome renders the group into a single state that represents its K or profit from its expenditure of effort.

A Quantum Perturbation Model (QPM)

153

2.5 Randomness, Social Structures, and a Master Equation QPM is congruent with dissonance learning theory (e.g. argument). In contrast to traditional social learning theory (e.g., adaptation or reinforcement), dissonance learning theory offers a link between machine learning and awareness. Assuming that first interactions before structure exists are approximately Markovian, then social (e.g., organizations) and psychological structures (e.g., stable beliefs) reduce randomness by increasing predictability. An organization becomes a set of constraints to assemble, retain, or marginalize agents, reducing management to controlling I to make rational decisions (e.g., [25]; we assume that rational business decisions are characterized by less variance in the available choices than the mean). From the global perspective of Nicolis and Prigogine ([56] p. 255), by construing society as a dissipative system with chaotic attractors, predictability increases along the direction of flow in contracting phase space (e.g., axes of I and I flow), such as consolidation mergers, while variety and choice increase during economic expansions. This leads us to postulate that outside of social structure, randomness is more likely (e.g., [51]). Further, Kimura [34] leads us to expect more rapid structural change among less successful organizations (Shumpeter’s creative destruction). But even in stable structures, with dissonance learning theory, randomness can be mindfully injected within structures (we assume that under dissonance learning theory, variance in the available choices is greater than their mean of zero; see [55]). Examples are plans that reflect instability (e.g., Fiat in 2003); disagreements and argument (e.g., [43]); and a wide range of many others (war, art, entertainment, innovation, technology), including mergers and K fusion (e.g., [42]). Randomness can be mindfully marginalized from structure with CDM (e.g., dictatorship, consensus, bureaucracy), but by proportionately slowing its evolutionary rate. Oppositely, increasing uncertainty among autonomous agents or MAS’s increases their emotional T, producing more I and anxiety, thereby motivating efforts to constrain and reduce uncertainty with structures designed to process I to increase K, but that once established, slow change [37]. The value of adding a Markovian approach can be easily seen by contrasting a sketch (Figure 2) of Equation 1 (using an arbitrary value for c) against the Congressional Budget Office Forecasts for GNP (representing ∆I) and actual GNP (representing ∆a). Figure 2 above Suggests that Equations 1 and 2 applied to the human interaction must be modified to account for feedback. When humans see the effect of real world outcomes on their predictions, feedback “tunes” the results to look like a limit cycle (e.g., p. 90 [52]). This contrast suggests that agents struggle in their attempt to couple the factors in both Equations 1 and 2, with the result being a nonlinear limit cycle. Then argument can be modeled as a double square-well (Figure 3) to capture the stability of two alternative-choice opponents attempting to convince neutral decisionmakers who act as a barrier that must be overcome by the “winning” argument that withstands “all contestations” [53].

154

William F. Lawless and James M. Grayson

Fig. 2. Comparison of CBO forecasts of two-year average growth rates for nominal GNP output for the USA, years 1976 to 1992 ([15]; in 1992, CBO switched to GDP). The estimated limit cycle is for GNP data; it contracts as it moves towards the origin (increasing predictability), and expands moving away from the origin (increasing choice). For the curve ∆a∆I ≈ c, the value for the constant c has been arbitrarily chosen. (We have not calculated the dimensions of this phase space or the attractor to see if it is chaotic, but we expect that contrasted with a CDM economy, a market economy exists in a higher dimension space; e.g., p. 281 [56].)

V

V0

Fig. 3. The double square well model represents an energy barrier between two views (at the bottom of the left square well for belief B1 with B2 on the right; see below), often overcome in a democracy by compromise (resonance) between polarized views, or by proponents of the views persuading more neutral participants (in the raised middle with relatively higher V [i.e., potential E] between the two square wells, or B0) of the “correctness” of their belief.

A Quantum Perturbation Model (QPM)

155

Argument produces bifurcations or multiple solutions, implying a regulatory function (p. 98 [56]). Transition times between alternative choices are estimated by τ ≈ exp (N ∆V) (9) where ∆V = V(B0) – V(B1 or 2) is the “potential barrier” to be overcome by the fluctuations, B is the game-theory choice that represents the attractiveness of either alternative belief B1 or B2, and N is the total number in a group required for a decision. For example, data from Lawless and colleagues [39],[40] indicate that one group using consensus to make nuclear waste environmental cleanup decisions (HAB, N = 32 – 4 as their minimum consensus) took about 2 hrs, giving ∆V ≈ .0248 units. To contrast with a second group using majority rule (SAB, with N = 25/2, rounded off to 13), not knowing its ∆V but to model its quicker decisions attributed to resonance (compromise; see Figure 3), which reduces barriers, we kept the same numerical value of ∆V but changed its sign, calculating τ = 0.72 hrs, close to actual results [21]. The Kolmogorov equation begins to look like: dI/dt = I F(I,K,X) dK/d = K G(I,K,X) dX/dt = X H(I,K,X) (10) where X is the number recruited into an organization. Considering only I and the rate of K, initial results indicate that K subsists on I, but as it “consumes” I, K rapidly declines. This disturbing result can be appreciated with a new interpretation of K not solely as the reduction of I, but as composed of algorithmic K and local beliefs, and where both forms of K can become obsolete or useless [20]. Then we can readily account for organizations highly successful in one time period that fail in the next by continuing to act on obsolete or risky beliefs (e.g., the hedge fund, Eifuku, posted a 76% gain in 2002 but collapsed over 7 days in January 2003). This leads us to pose that K = Kprocedural-algorithmic + Kbelief = Kalg + Kχ, where Kalg includes technical and design K [68], and Kχ includes beliefs, expectations and predictions. In that we don’t have enough information to write the master equations for the underlying probability distributions, we assume that fluctuations are caused by a random force field, FE(t), such as social noise, giving d(T,I)/dt = -∂U(T,I)/∂(T,I) + FE(t) (11) and τ ≈ exp (∆V/q2) (12) 2 where q is the variance of FE(t). When ∆V is very large (e.g., CDM), the forcing function will have negligible effect on transitions; however, when the period of the forcing function approaches the transition time, τ, the system will respond dramatically even to a weak forcing function, known as stochastic resonance. What is exciting about FE(t) is the possibility that it could account for the differences between CDM and majority rule decision-making processes. In the case of CDM, FE(t) becomes a strong, global forcing function in that to survive, all individuals and groups must align with the global leader (dictator) or consensus, dampening autonomy and reducing factional diversity. Thus, to fulfill self-interest under CDM, choices must align with the field. Extensive literature supports this conclusion for complete alignment (i.e., obedience; in [54]) and partial alignment (i.e., social influence; in [5]). In contrast, the most notable result from majority rule and self-organization is a plethora of groups and organizations, wide-spread diversity (political, ethnic, or

156

William F. Lawless and James M. Grayson

enterprise), and large-scale migration as agents are recruited or marginalized from organizations; thus, in democracies, FE(t) is a weak function globally but with strong effects locally from selection pressures as agents “match” the local field, producing either stochastic resonance when matches occur, or stochastic reactance otherwise. Not only is more I generated in market versus command economies, but also the rate of I processing into K is far greater, reducing corruption [43]. Preliminary data from the U.S. and former U.S.S.R. support this conclusion (e.g., www.transparency.org).

3 From Argument to Perturbation Theory and QPM Joining a group promotes the survival of individuals by reducing E expenditures in exchange for membership: social loafing [36]; audience effects enhance skills [75]; and greater interaction density protects health [31] and belief systems [61]. In exchange, a group exploits the E and skills it collects [3], forming a structure as a network of interactions between roles [62]. Generally at the lowest E state, interaction exchanges —voice, visuals, products, and money— between agents cycle I back and forth in interactions coordinated by common K [73]. Among the groups that gain more E than it costs to survive [16], some gain sufficient free energy, ∆A, to grow in size, experience and wealth, deepening E wells to process more I, while others merge to offset competitive weaknesses (e.g., in 2002 HP merged with Compaq to offset its weakness in computer servers, Compaq’s strength). Most interactions within a stable organization serve to fulfill a mission, defend a worldview, or acculturate members, but interactions to solve idp’s are different. These interactions temporarily shift members into factions of argument leaders and neutrals, where ∆t is the time for the system to evolve to an orthogonal state to reach a decision [2], signified by the existence of multiple views (for the value of multiple views, see [8]). For optimal decisions, an MAS neutral to pre-established conclusions and open to argument processes I or uncertainty into K (e.g., political, legal, and scientific dissonance precede optimal decisions; in [41]). Identifying the optimum solution of an idp is analogous to signal detection, the time (∆t) to detect from social noise lasting until a solution signal is separated and adopted; e.g., air-to-air combat, environmental cleanup, disaster recovery, or weather prediction. However, given the unreliability of self-reports (measurement collapses the interaction into individual histories that cannot recreate it), a new approach must be initiated to measure physiological E states, such as vocal energy changes, to contrast normal and dissonant states (see Figure 4). Linear stability analyses of bifurcations due to system perturbations from dissonance during social decision-making appears likely. It should become possible to build a model of a system exhibiting chaotic dynamics where Lyapounov exponents characterize the rate of divergence between Kχ (beliefs, choices, or expectations) during argument (pp. 254-5, 262-3 [56]). Figure 5 is an example of this effect. While in general, organizations favor CDM (e.g., businesses, government agencies, universities), in a democracy there is a greater likelihood that a leader will choose a decision influenced by the public, reflecting self-organization (e.g., recently the editors of the Wall Street Journal charged the New York Stock Exchange of being a monopoly, “front running” to buy stocks ahead of large purchases, and failing to

A Quantum Perturbation Model (QPM)

157

adopt new technology; compare those charges with recent ones against Communist China of covering up the extent of its SARS epidemic).

Neurological

Dissonant

²E

I

Vocal

²E

Fig. 4: Picard’s liquid model of emotion suggests that social perturbations caused by dissonant I produce a spectrum of emotional responses. Significant vocal E changes from normal to angry speech have been confirmed for one subject but not yet for groups [37].

New endogenous K = New defense strategies

Organization

I generation

Dissonant Information

New exogenous K = New attack strategies

Fig. 5. After a perturbation, an organization’s goal is to respond with endogenous feedback to dissonant I by creating new K (Knew = Kalg + Kχ) to design new innovations, strategies or technology to defend the organization (see [47]). Conversely, using exogenous feedback, a competitor’s goal is to devise innovations, strategies or technologies to defeat the organization. In general, the quicker respondent determines which organization wins and evolves; e.g., in 2003 in the war with Iraq, coalition decision-making and implementation of those decisions occurred faster than Iraq’s Defense Forces, causing the latter to panic (i.e., in engineering control theory late feedback is destabilizing; in May, p. 5 [52]).

158

William F. Lawless and James M. Grayson

With the help of May [52], in this last paragraph we speculate on the implications of QPM. First, argument serves to entangle neutrals with pro-con forces sufficient to drive neutrals to process and choose a concept that solves an idp by converting I into K, but insufficient to provoke violence, indirectly promoting the value of limit cycles to manage an MAS in contrast to CDM control (e.g., by managing the divergence between Lyapounov exponents). Second, coupled Kolmogorov equations with forcing functions indicate that stochastic resonance has more potential to produce innovation in a self-organized, democratically governed system than one governed by CDM, but once an idp becomes well-defined, CDM systems are more efficient. Third, if the average distance between agents in an MAS imply organizational wave length, λ, then the more cooperation that exists, the less I density and more Kχ (beliefs) that occur; the converse occurs under competition—our mathematics then reproduce the well-established finding that organizations under threat become physically closer; i.e., a higher E from a threat reduces λ. Finally, under stable environmental conditions, the more diverse are the numbers of groups in a community, the more stable is K and unstable is Kχ; conversely, under environmental instability, the less stable is global K and more stable is Kχ. The reason is that under stable conditions, evolution is driven by the unstable dynamics between competing groups, but under unstable conditions, evolution slows as group dynamics stabilize from a reduction in the number of organizations as they attempt to merge in order to survive.

4 Conclusion QPM offers the possibility that an analytical model that simulates the conjugate aspects of decision-making, I generation, and organizational growth may also simulate K fusion and organizational mergers. The major discovery with QPM is the existence of interaction cross-sections (Equation 4), suggesting new areas for future research (e.g., MAS density-dependent chaos versus random social noise). But QPM also accounts for differences between an aggregation and a disaggregated group of the same individuals, succeeding where game theory has failed [50]; it explains why traditional models based on the individual perspective of rationality fail [44], why traditional ABM’s of social processes are difficult to validate [9], including AMKM systems [47], and why traditional perspectives of cooperation are normative [27], rather than scientific. QPM also suggests how game and other rational theories may be revised; e.g., during negotiations as foreseen by Nash [55] and others but only when driven by the turmoil from competition, cooperation harnesses competition to solve idp’s by precluding convergence (unlike machine intelligence) until after orthogonal arguments have been sufficient to promote optimal mergers of organizations, the generation of I, or the fusion of K. In conclusion, MAS’s should be designed to account for the E and K advantage gained by an organization at the commensurate cost to each agent member of its autonomy, freely exchanged for the K each agent acquires in its search for the E to survive. Bringing Dignum and Weigand [20] full circle, while K is the most important asset for an organization, its competitor, and the agents who are members of these organizations, it is found by randomly exploring the universe of alternatives until a state of stochastic resonance occurs between I and a random forcing field.

A Quantum Perturbation Model (QPM)

159

Acknowledgements. The first author thanks J.A. Ballas, Information Technology Division, Naval Research Laboratory, Washington, DC, where most of this research was conducted with funds from the Office of Naval Research through an American Society of Engineering Education grant.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27.

Allport, F.H. (1962), A structuronomic conception of behav., J Abn. Soc Psych, 64, 3-30. Ahoronov, Y., & Bohm, D. (1961). PR 122: 1649-1658. Ambrose, S. H. (2001). "Paleolithic technology and human evolution." Science 291: 174853. Andrade, G., & Stafford, E. (1999). Investigating the economic role of mergers (Working Paper 00-006). Cambridge, MA, Harvard. Asch, S. (1951). Effects of group pressure upon the modification and distortion of judgement. Groups, leadership, and men. H. Guetzkow(Ed.). Pittsburgh, Carnegie. Axtell, R. (2002). Personal communication. Axelrod, R. (1984). The evolution of cooperation. New York, Basic. Bailin, S.C. & Truszkowski, Perspectives: An analysis of multiple viewpoints in agent based systems. AAAI-Spring Symposium, AMKM, Stanford, pp. 156-8. Bankes, S. C. (2002). "Perspective. Agent-based modeling Proc NAS, 99(3): 7199. Becker, J. (1998). Hungry ghosts: Mao's secret famine. New York, Holt. Benardete, S. (2002), The argument of the action: Essays on Greek poetry and philosophy, Chicago U. Berken, C. (2002). A brilliant solution: Inventing the American Constitution. Harcourt. Bohr, N., Ed. (1955). Science and unity knowledge. Doubleday. Brehm, J.W. (1966). A theory of psych. reactance. New York: Academic Press. CBO (1999). Evaluating CBO's record of economic forecasts (Supplement to: The Economic and Budget outlook: an Update, July 1, 1999); www.cbo.gov, (Congressional Budget Office). Coase, R. (1937), The nature of the firm, Economica, 4: 386. Conant, R. C. (1976). "Laws of information which govern systems." IEEE Transaction on Systems, Man, and Cybernetics 6: 240-255. Crick, F., & Koch, C. (1998). "Consciousness and neuroscience." Cerebral Cortex 8: 97107. Darpa (2002), www.darpa.mil/. Dignum, V., & Weigand, H. (2002). Towards an Organization-Oriented Design Methodology for Agent Societie. Intelligent Agent Software Engineering. V. Plekhanova, Idea Group Publishing. DH (2001), Conversations involved with joint SAB-HAB meetings what alternated between consensus and majority rule; also Lawless et al. 2000a. Eagly, A.H. & Chaiken, S. (1993), The psychology of attitudes, Harcourt. Eisert, J., Wilkens, M., & Lewenstein, M. (1999). "Quantum games and quantum strategies." PRL, 83(15): 3077-3080. Engel, A., Fries, P, Konig, P, Brecht, M, & Singer, W (1999). "Temporal binding, binocular rivalry, and consciousness." Consciousness and Cognition 8(2): 128-151. Farber, D. R. (2002`). Sloan rules: Alfred P. Sloan and the triumph of General Motors. University of Chicago. French, A. P., & Taylor, E.F. (1978). An intro. quantum physics. New York, Norton. Gmytrasiewicz, P.J. (2002), On decision-theoretic approach to game theory, SS-02-02, pp. 1-8, Menlo Park, CA: AAAI.

160

William F. Lawless and James M. Grayson

28. Grayson, J. M., Runger, G.C., & Ibrahim, N.A. (1996). "Object-oriented organizations: Management insights from software principles." Quality Management Journal 3(4): 66-75. 29. Hagan, S., Hameroff, S.R., & Tuszynski, J.A. (2002). "Quantum computation in brain microtubles." Physical Review E 65: 061901-1, 11. 30. Hannebauer, M. (2000), Their problems are my problems, In C. Tessier, H.J. Muller, & L. Chaudron, Conflicting agents: Conflict mgt in multi-agent systems, pp. 63-110, Boston: Kluwer. 31. House, J. S., Landis & Umberson (1988). "Social relationships and healthq." Science 241: 540. 32. Janis, I. L. (1982). Groupthink. Boston, Houghton Mifflin. 33. Kelley, H. H. (1992). "Lewin, situations, and interdependence." Journal of Social Issues 47: 211-233. 34. Kimura, M., Takahata, N. (Ed.), & Crow, J.F. (Designer) (1994). Population genetics, molecular evolution, and the neutral theory: Selected papers, U. Chicago Press. 35. Kirk, R. (2003). More terrible than death. Massacres, drugs, and America's war in Columbia, Public Affairs. 36. Latane, B. (1981), The psychology of social impact, American Psych., 36, 343-356. 37. Lawless, W.F. (2001), The quantum of social action and the function of emotion, Proceedings, pp. 73-78, AAAI Fall Symposium. 38. Lawless, W.F. (2003), Information Density Functional Theory (IDFT): A quantum approach to knowledge fusion and organizational mergers, AAAI Spring Symposium, Stanford, pp. 71-3. 39. Lawless, W.F., Castelao, T., and Ballas, J.A. (2000a), Virtual knowl: Bistable reality and solution ill-defined problems, IEEE Systems Man, & Cybern., 30(1), 119-126. 40. Lawless, W.F., Castelao & Abubucker (2000b), Conflict, in Tessier et al., Conflicting agents, pp. 279-302, Kluwer. 41. Lawless, W.F. & Castelao, T. (2001), Univ Dec. Center, IEEE Techn. Soc., 20, 6-17. 42. Lawless, W. F., & Chandrasekara, R. (2002). Information density functional theory: A quantum approach to intent. Draft Proceedings, AAAI Fall Symp on Intent Inferencing, pp. 26-31, Cape Cod, MA. 43. Lawless, W. F., & Schwartz, M. (2002). "The social quantum model of dissonance." Social Science Computer Review, 20(4), 441-450, Sage. 44. Levine, J.M. and Moreland, R.L. (1998), Small groups, In Gilbert et al., Hdbk Soc Psych, pp. 415-469, McGraw. 45. Lewin, K. (1951). Field theory in social science, Harper. 46. Lloyd, S. (2000). "Ultimate physical limits to computation." Nature 406: 1047-1054. 47. Louçã, J. (2003). Discussion Leader: Directions for Agent-Mediated Knowledge Management. AAAI Spring Symposium 2003, AMKM-03, www.dfki.unikl.de/~elst/AMKM/papers_presentations.html, "directions discussion.ppt". 48. Luce, R.D. (1963), Decision and recognition, Hdbk Math Psych, R.D. Luce et al. (Eds.), New York: Wiley, I, pp. 103-189. 49. Luce, R. D. (1997). "Several unresolved conceptual problems of mathematical psychology." Journal of Mathematical Psychology 41: 79-87. 50. Luce, R.D. and Raiffa, H. (1967), Games and decision, New York: Wiley. 51. Malkiel, B. G. (2000). A random walk down wall street (7th ed.). New York, Norton. 52. May, R. M. (2001). Stability and complexity in model ecosystems, Princeton U. Press. 53. McBurney, P., & Parsons, S. (2001). Agent ludens: Games for agent dialogues. Game Theoretic and Decision Theoretic Agents, AAAI Spring Symposium TR 22-01-03. 54. Milgram, S. (1974). Obedience to authority. New York, Harper and Row. 55. Nash, J.F., Jr. (1950), The bargaining problem, Econometrica, 18, 155-162. 56. Nicolis, G., & Progogine, I. (1989). Exploring complexity. New York, Freeman. 57. Petrie, C. (2003, March 24). Adding AI to web services. Plenary, AMKM, AAAI-Spring03, http://snrc.stanford.edu/~petrie/fx-agents/xserv/AIfWS/index.html.

A Quantum Perturbation Model (QPM)

161

58. Pfister, H. L. (2002). Cooperative control for wide area search with multiple autonomous vehicles using fuzzy logic. Annie 2002, St. Louis, MO, ASME. 59. Polanyi, M. (1974). Personal knowledge. U Chicago Press. 60. Pynadath, D., Scerri, P. & Tambe, M. (2001). MDPs for adjustable autonomy in realworld multi-agent environments. 2001 AAAI Spring Symposium, Technical Report SS10-03, Stanford, CA, AAAI. 61. Rosenblatt, A., Greenberg, J., Solomon, S., Pyszczynski, T., & Lyon, D. (1990). "Evidence for terror management theory." J Personality Soc. Psych. 57: 681-690. 62. Sallach, D. (2002). Modeling emotional dynamics: Currency versus fields. unpublished, U. Chicago. 63. Schlesinger, A. (1949). The vital center. Cambridge, MA, Riverside. 64. Schumpeter, J.A. (1989), A theoretical, historical, and statistical analysis of the capitalist process, Porcupine Press. 65. Shafir, E., Simonson, I. and Tversky, A. (1993). Cognition 49: 11-36. 66. Skinner, B. F. (1978). Reflections on Behaviorism and society. Prentice Hall. 67. Slovic, P., Layman, M., Kraus, N., Flynn, J. Chalmers, J., & Gesell, G. (2001). Perceived risk, stigma, and potential economic impacts of a high-level nuclear waste repository in Nevada. Risk, Media and Stigma. J. Flynn, P. Slovic, & H. Kunreuther. London, Earthscan. 68. Tacla, C.A. & Barthes, J.P. (2003), A multi-agent architecture for knowledge acquisition, AAAI Spring-03, AMKM, Stanford, pp. 156-8. 69. Tessier, C., Muller, H-J, Fiorino, H., and Chaudron, L. (2001). Agents’ conflicts: new issues. Conflicting agents. C. Tessier, L. Chaudron, and H.J. Muller. Boston, Kluwer. 70. van Eeten, M. J. G. (2002). Dialogues of the Deaf (Environ. Deadlocks). Eburon. 71. Von Neumann, J., & Morgenstern, O. (1953). Theory games & economic beh. Princeton. 72. Von Neumann, J. (1961). The mathematician. Collected works, Pergamon. 73. Wendt, A. (1999). Social theory of international politics. Cambridge U. Press. 74. White Paper (WP) (2002). European governance (COM, 2001, 428 final). Brussels. 75. Zajonc, R. B. (1998). Emotion. The Handbook of Social Psychology. D. T. Gilbert, Fiske, S.T., and Lindzey, G. Boston, McGraw-Hill. 76. Zeilinger, A. (1999), Experiment and foundations of quantum physics, Reviews of Modern Physics, 71: S288.

Improving Organizational Memory through Agents for Knowledge Discovery in Database João José Vasco Furtado, Vinícius Ponte Machado UNIFOR - Universidade de Fortaleza, Mestrado em Informática Aplicada - MIA Washington Soares 1521, Fortaleza – CE, Brazil [email protected], [email protected]

Abstract. In this article we describe a multi-agent system integrated into a Knowledge Management Information System, called MC2, for knowledge discovery in database. MC2 brings together a set of tools that contribute to the knowledge management process by allowing for the creation and maintenance of an organizational memory. To achieve this end, the approach taken by MC2 is aimed at establishing favorable conditions for interaction between the personnel within an organization as well as with the system itself. In particular, we describe agents that attempt to achieve automatic knowledge discovery from organizational databases and from the manner in which that knowledge is integrated into the MC2 environment. By using these agents, organizational development is fostered through the dissemination of knowledge, yet, in such a way so as to be transparent to the holders of that knowledge without requiring additional activities above and beyond those already carried out as part of their day-to-day routine.

1 Introduction Knowledge is a fundamental building block in business and is viewed as a decisive variable in the competitiveness of an organization. Aspects linked to the mechanisms for creating, representing, distributing, commercializing and exploring knowledge must be analyzed and understood in order to make a mark within a competitive environment. To the collective administration of these mechanisms we give the name knowledge management. Another concept corollary to the ever-increasing relevance of the role of knowledge in an organization is that of learning organization. Here the aspect of continuous and collective learning of the personnel within an organization is emphasized. Consequently the development the organization passes through as a result of this learning process. To fully achieve the goals of knowledge management, as well as to oversee, induce and even to accomplish the process of organizational learning itself, a computational support system is needed that offers flexibility and reliability to the process. Knowledge Management Information Systems (KMIS) constitute this support and allow for dealing with large amounts of information as well as for providing the primary means in the formation of organizational memory. L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 162-176, 2003.  Springer-Verlag Berlin Heidelberg 2003

Improving Organizational Memory

163

In this article, we describe a KMIS called MC2 that brings together a set of tools that support the process of knowledge management by allowing for the creation and maintenance of an organizational memory. MC2 is divided into three levels: The data level where data concerning the organization is stored; the IS level, called participation space, which makes computerized procedures available that seek to contribute to the formation of better relationships between the people and sections of an organization; the knowledge level that is composed of software agents that seek to aid in the creation, search and exploration of knowledge that permeates the organization on a daily basis. These two levels are strongly linked to and mutually feed off one another. In particular, we describe a multi-agent system (MAS) which composes the MC2 knowledge level. The MAS accomplishes automatic knowledge discovery from the databases of an organization and show how that knowledge is delivered to a MC2 user. With the use of this tool, organizational development is fostered by disseminating knowledge in such a way so as to be transparent to its originators, thus, it requires no additional tasks on their part above and beyond those already accomplished in their day-to-day activities.

2 The State of Art 2.1 Knowledge Management and Learning Organization We utilize the definition of [1] for the learning organization (LO) as being: “Organizations where people expand their capacity to continually create results that they really want, where new and expansive ways of thinking are encouraged, where collective aspiration is free, and where people are constantly learning to learn collectively”. He emphasizes that LOs are organizations that are constantly expanding their capacity to create and recreate their respective futures. Another perspective in the challenge involved in transforming a company into an LO is that which focuses on the need for companies to “unlearn” or “forget the past” [2]. Common challenges are identified in the software tools that are intended to accomplish the creation and maintenance of organizational memory [3], [4]. In [5] a system is presented that supports the definition of models for the description and storage of documents, thus creating an organizational memory (OM). These models are graphically built and are the result of a collaborative effort among users. To such an end, each user has several of these previously defined models available and can use them interactively when classifying a document by fitting it to the model that best describes it, facilitating its storage and any potential future queries to it. The authors point out that for the success of this type of organizational memory there must be resources available for the storage and distribution of knowledge in compatible formats so that users with different profiles can easily contribute and query information relating to their particular work, thereby facilitating the creation of additional models beyond those already defined. The authors also point out that people must note a direct benefit in the administration of knowledge in order to

164

João José Vasco Furtado and Vinícius Ponte Machado

continue contributing and the effort required by their contribution should be minimized so as not to unduly interfere with the flow of their normal workload. Other authors [6] focus on how organizational memory can be administrated. For them OM is an explicit and persistent representation of the knowledge and information in an organization. Any part of the knowledge or information that contributes to the performance of the organization can be stored in it. In the four knowledge treatment processes dealt with by the authors (knowledge acquisition, storage, distribution, and knowledge combination), we encountered, not by coincidence, difficulties and solutions similar to those encountered by Abecker and his colleagues. We point out those regarding the fact that both approaches indicate that participation must be active and permanent. To do this, the user should be motivated to contribute his knowledge and there should be a return for him in doing so, that is, the person should notice a benefit to his work resulting from the use of organizational memory. Besides the effort of the user’s contribution being minimized to the extent possible, it should also be automated so that he needn’t set aside his professional tasks in order to accomplish those relating to organizational memory. The tools should provide a means to that end so that cooperation is effected automatically and interacts to the maximum with their actual work. 2.2 Inductive Learning Algorithms In order to get assistance in the creation of OM, it is necessary to depend on software tools that can generate knowledge from organizational data and information. In the realm of Artificial Intelligence (AI), inductive learning algorithms [7], [8] support the creation of example-based concepts. These examples represent past situations and are usually described by attribute/value pair lists. It falls on the algorithm to seek out correlation between characteristics (attribute/value pairs) within these examples and to induce from the correlation generic concepts representative of the examples, which can then be used to classify other analogous examples not yet observed. Concept acquisition algorithms have C4.5 [9] as their most representative and popular example. This algorithm type, taken from a group of examples, generates a decision tree that represents concepts within the domain being worked. They are considered supervised algorithms because the examples indicate the class to which they belong, leaving the algorithm to discover a generic concept that represents each class indicated by the examples. In unsupervised algorithms, also known as conceptual clustering, the examples do not pertain to any known predefined class. Here, the task is finding clusters of who is similar to whom. These clusters are said to be conceptual because the characteristics for each cluster defines the concept of the cluster represented. COBWEB [10] and FORMVIEW [11] are instances of this class of algorithms. Besides conceptual clusters, FORMVIEW constructs concepts following different perspectives in order to discover relationships between different contexts, called bridges. The function of both algorithms is knowledge acquisition through concept formation. These concepts, when placed in an organizational context, come to reveal knowledge that until then was merely implicit within the organization, thereby contributing not only to feeding OM but supplementing organizational learning as well.

Improving Organizational Memory

165

2.3 The Problem The brief state of the art accomplished above allowed us to determine that the creation of an OM and by consequence the development of organizational learning requires that those involved in the process be motivated to participate. Without their active and continuous participation, the reception as well as the dissemination of knowledge would be compromised. It is necessary to create the means that will allow users to extract knowledge from their daily tasks and yet in doing so does not unduly interfere with their professional responsibilities so that the process of creating an OM does not become an overly tiresome prospect or an utter waste of time. Additionally, we believe that if everyone in the organization is well aware of the utility of organizational knowledge, they will more readily and more willingly contribute. Therefore, the creation of automatic knowledge discovery tools based on user activities along with their dissemination throughout the organization may be a strong motivating factor toward participation as well. Along these same lines, the use of inductive learning algorithms like concept acquisition and conceptual clustering constitutes an interesting alternative toward creating organizational memory. Our research work concentrates on this aspect of the process and how a tool of this type may be integrated into an environment specifically designed to create an OM. In what follows, we describe the interactive environment of creating an OM called MC2 and how knowledge engineering tools, and more specifically, those designed for database knowledge discovery, are integrated with it. The use of knowledge discovery as a tool for knowledge management is exemplified in [12] where agents are used to track internet behaviors in the context of a pharmaceutical industry and in [13] where a personal agent automatically generates DAML markup for set of documents based on unseen labeled training semi-structured documents. More strongly related to this work, [14] define a system that use agents to discover and categorize knowledge based on group of users. The exploration of documents by users is taken as basis for creating a user-specific profiles. These profiles are useful to cluster documents which are visualized by means of a graphic tool.

3 The MC2 System 3.1 A Global View MC2 is a web-based information tool and a managerial strategy that supports the formalization of a learning culture and an organizational memory (see www.mc2.com.br). In general, the objective of MC2 is to assist the administration of knowledge. People are required to share their experiences, ideas, knowledge and activities linked to learning such as: book summaries, project structures and methodologies dealing with their activities. The aim is to promote positive attitudes and to build a participation culture that result in the creation of important and highly valuable knowledge bases.

166

João José Vasco Furtado and Vinícius Ponte Machado

To encourage participation it is necessary to implant in the organization an effective means and adequate structures where new assignments, roles and responsibilities can be put into actual practice. The general architecture of MC2 can be seen in Figure 1. It is modeled with Intelligent Agents that have the characteristic of using the information handled by individuals in the context of their work processes and convert this information into knowledge for use in decision-making. Database – Information System

Knowledge Base Meta-Knowledge about MC2 Knowledge

Level Intelligent Agents

Information Level

Data Level

MC2 Modules Abilities

Position

People

::: :::

::: :::

Profile

Fig. 1. MC2 Architecture

MC2 is divided into three levels: the data level where data concerning the organization and its personnel is maintained, the IS level that we call the participation space, which tries to make computerized procedures available that try to facilitate relationships between people within an organization; and the knowledge level, where tools are stored in expressions for tasks that are accomplished daily by the individuals of the organization while they are working and undertaking their professional activities. The tools that comprise the participation space can be seen on the screencapture shown in Figure 2: • Corporate University through its modules gives support for personal development of the members of the organization. • Expression Modules support various ways of documenting and accessing useful information for organizational development such as experiences obtained individually or in group, acquired abilities, problem resolutions, topics for further discussion, etc. • Knowledge Circles are modules that allow people interested in certain subjects to learn and to share specific knowledge. This module concentrates on all knowledge contributions available within the system (articles, tests, book summaries, etc). • Queries Modules allow the members of the organization to find answers to frequently asked questions such as those concerning personnel, problems and solutions, abilities, norms and basic concepts of the organization. • Communication Modules allow the individuals of the organization to exchange ideas and experiences from their daily lives, besides serving as a resource for managers of the organization to communicate with employees.

Improving Organizational Memory

167

• Support Modules allow for the customization of the MC2 system to fit the reality of different institutions as well as to support the perfect implementation and use of the MC2 strategy.

Fig. 2. MC2 with their expression modules

3.2 The MC2 Managerial Strategy As we have already mentioned, the MC2 managerial strategy is based on the participation concept. The possibility of participation is permanent in relation to any subject matter or perspective of the company and is accomplished through various possible channels. The result then, in this context, is that when referring to people, we mean not only their technical competence but all their talents, experiences, vision, initiatives and capacity for action when joining together with others in order to develop their personal projects, thus assuring that the entire process translates into permanent organizational learning [15]. That is: people are empowered to grow as the organization develops; to change, in the effort of improving it. For the proper operation of the MC2 software there is an implantation strategy in which MC2 is aided. This strategy makes use of ontology where fundamental concepts of the organization are defined such as the functions, abilities and requirements expected of personnel in the exercise of their jobs. These ontologies are known in MC2 as critical knowledge lists (CKLs). There are three types of CKLs: common, which contain themes that everyone in the organization should know. Job

168

João José Vasco Furtado and Vinícius Ponte Machado

CKLs, designed to relate the indispensable knowledge for the exercise of each position or function, and Personal CKLs, a list of necessary knowledge for the essential development of each collaborator. Besides the formal aspect regarding the necessary concepts to be defined, it is worth pointing out some of the essential roles that should be exercised by people in the organization so that MC2 is implanted satisfactorily. These roles are: • Articulator: is the person responsible for the professional development of others and for managing the resources that will make possible the maximization of this development. • Multiplier: is the person who establishes a formal, permanent and parallel commitment to their normal activities, delving more deeply into and disseminating knowledge relative to a particular theme from one of three CKLs. • MC2 Consultant: is the person to whom the organization assigns the tasks of coordinating, implementing and administrating the MC2 system.

4 Modeling the MAS for Knowledge Discovery Having briefly described the MC2 system, we shall now focus on defining the process of the knowledge discovery agent’s operation as well as its architecture. Agents have been used to automate tasks and to carry out processes not requiring user input. The use of the Knowledge Discovery MAS (KDMAS) in MC2 is intended to contribute to knowledge acquisition within an organizational learning environment by carrying out a database knowledge discovery process. In this process the KDAMAS explores databases containing information on the business activities of an organization. The KDMAS sweeps the company’s business databases in order to, through learning algorithms, detect facts and information not explicitly outlined in its normal rule of operations. Such information can become knowledge, and per chance, may be used in making decisions, thereby forming a knowledge base that can feed organizational memory. To proceed we shall describe how this agent is modeled, describing its architecture, and its operation. 4.1 KDMAS Architecture To model the KDMAS, we followed the model proposed by [16]. To the authors, an agent is anything that is able to perceive its environment through a sensor and is capable of interacting with that environment by means of a determined action. In the case of the KDMAS, Interface Agents (IA) act as sensors to monitor user intervention to various MC2 system events. For each of these interventions the IA tries to find data that might serve as parameters for its action. Based on these parameters, the Data Preparation Agent (DPA) and the Meta-data agent agent will prepare a row of data to be mined. Finally, the Knowledge Discovery Agent (KDA) uses inductive learning algorithms like those described previously. The four type of agents of the MC2 MASKD can be seen in Figure 3.

Improving Organizational Memory

169

Fig. 3. The architecture of the MC2 Multi-agent system for Knowledge Discovery

To proceed we will describe each of these parts and how they contribute to the MAS action as a whole. 4.2 Environment The MASKD environment is divided into two parts: the perception environment and the action environment. The perception environment is where the MASKD carries out its monitoring activities that detect events to be broken down into particular actions. This environment is the participation space of the MC2 system. The MASKD action environment is where information is stored regarding the business of the company, its structure, customers, suppliers and personnel (usually its database). This allows the organization to obtain more knowledge concerning their methods, operations and finances. 4.3 Sensors As sensors, we define the MASKD procedures, done by the Interface Agent, for monitoring the events of some of the modules of the MC2 participation space. These events serve as stimulus so that the IA can carry out its actions. The events constantly monitored by the IA are: • Knowledge Circles; • Groups; • Query Database. The Knowledge Circle concentrates on all contributions carried out in MC2 by its members on a particular theme; such as Forum participation, comments on titles in the collection, queries made, and new ideas, allowing members to learn and share their contributions. IA then monitors the creation, alteration and inclusion of new

170

João José Vasco Furtado and Vinícius Ponte Machado

members in the circle, in such a way so as to try to discover a profile of these members and thus, aid in the selection of new circle participants. Another event that is monitored by the IA is the creation of the Query Database. In this base the user formalizes questions and topics on a particular theme and addresses them to particular people or groups of people who might be able to answer their questions. In this case the agent also monitors the inclusion in the database of queries and tries to anticipate answers to the questions that per chance the user inputs into the database. Different from the Knowledge Circle, the Groups module registers a meeting of people with the purpose of carrying out tasks or deliberations. Yet, like the Knowledge Circle, the MASKD tries to discover the profile of these members, aiding in the selection of new participants. 4.4 Action As mentioned previously, MASKD has as an action environment the organizational database. So that MASKD may take action, some parameters, captured from the perception environment, are necessary. These parameters are actually the attributes the knowledge discovery algorithms implemented in the Knowledge Discovery agent that were needed to accomplish its processing. The choice of these attributes is an important task since the results can mean relevant knowledge to the user that may or may not be related. For each event cited before, the IA identifies data that can serve as parameters for the agent’s action. In the case of the Knowledge Circle and Discussion Groups, the theme of the circle and the individual CKLs are examples of such parameters. For the Query Module, the Interface agent tries to use keywords from questions and uses them as parameters for the algorithms. The agent operates in a similar fashion in the Groups module, where the parameters used by the agents are the themes as well as the CKLs of its members. The attributes collected compose the structure of the examples, which will feed the machine learning algorithms. Such a procedure is performed by the Meta-data Agent which relates keywords and parameters captured by IA with database fields in a very similar way to those developed in the context of meta-data definition to data warehouse creation [17]. After this, the Data Preparation agent generates an example training set from business meta-data representing important concepts of the domain. It is important to emphasize that the agent’s action can also be run manually. If a user so desires, attributes may be chosen at will to carry out knowledge discovery. 4.5 Product The result for the user of the MAS system is obtained by the application of the Knowledge Discovery agent that creates the knowledge base. This knowledge base is made up of production rules and concept hierarchies that are the products of the inductive learning algorithms implemented in the agents. The use of production rules consists of representing the knowledge domain with a group of IF-THEN rules, representing a particular concept through these rules. The

Improving Organizational Memory

171

concept hierarchies are intended to classify a group of observations based on certain attributes. Depending on the event that triggers the agents (knowledge circles, groups, queries), there will be differing utility to the user. In the case of Knowledge Circle and Groups, the main purpose is to discover profiles. 4.6 Communication Each agent needs information that will serve of subsidies for its functioning. With the exception of the interface agent, these information are deriving of the result of the execution of another agent. In this way when an agent finishes its action, it supplies necessary information so that another one carries through its action and continues the process of knowledge discovery. However, these information can contain great volume of data and, depending on the amount of solicitation, the exchange of these information between them can cause a degradation of the performance of the process. That is reason we have used a blackboard as way a mean of communication between agents. In our architecture the direct exchange of messages (agent to agent) is carried through only when an agent activates another one. The necessary information to each agent for its execution will be passed through a blackboard.

5 Example To exemplify our proposal, we shall describe the knowledge acquisition process based on the database of CAGECE (State of Ceará Water and Sewage Company), taking as an example the Human Resources Knowledge Circle module. In this case the agent’s function is to try to determine the profile of the circle members, thereby aiding in finding new members with the same profile. To do this the agent will utilize criteria obtained from the knowledge circle itself along with personnel attributes found in MC2 itself. Thus, we can divide into two groups, the attributes that will be used by the agent to carry out concept formation. We have parameters taken from the knowledge circle itself that are mapped onto a meta-data layer with attributes of the organizational database. The others will be attributes pertaining to the users themselves that can aid in categorizing their profile such as: number of registered articles, forum contributions, registered professional activities, query responses, etc. This way, observations will be formulated based not only on the organizational database, but also on MC2 data itself. We may highlight three steps that the agent will carry out to effect its function: Sensing, where the agent notes the instant of its action and captures the concepts involved in taking action; the linkage of concepts with database attributes, where the agent uses meta-data to link the MC2 concepts to attributes within the organizational database; and the action that involves the execution of processing the algorithm. To proceed, we shall show how the agent will act in this example of a Knowledge Circle with a Human Resources theme.

172

João José Vasco Furtado and Vinícius Ponte Machado

5.1 Sensing As already cited, there are certain events that the MAS constantly monitor; one of which is the Knowledge Circle. For each new circle created the agents are activated and enter into action. In our example the MAS action is to try to determine the profile of the members of the knowledge circle from two perspectives. It aims at creating concept hierarchies, based on criteria captured from the Knowledge Circle, for aiding the identification and selection of people with potential for participation in that circle. The capturing of criteria becomes an important action since it can directly influence the concept formation accomplished by the algorithm triggered by the agent. The sensing function is done by the Interface Agent. Besides deciding the instant of its activation, it captures those criteria correctly. In our Knowledge Circle example there are two types of attributes that the Interface Agent will use and that determine the perspectives to be used: attributes about the organizational database and attributes about the MC2 database. The organizational database attributes, which are mapped onto the meta-data layer, contribute to the creation to the professional perspective. Basically, these attributes are related to the following parameters: Theme and the CKLs of current members. CKLs are composed of keywords that define indispensable knowledge for the execution of each position or function. These keywords compose the group of attributes to be used by the algorithm. Using the Human Resources Knowledge Circle, we take into account that this same circle is composed of three users, for example. These users have the themes in their CKLs indicated in Table 1. Table 1. CKLs of the Knowledge Circle Members

User A B C

CKL Finance Finance, Project Management, Human Resources Psychology, Project Management

Beside the attributes found in the Knowledge Circle (theme and the CKLs of the members), attributes about MC2 database will lead to attempt at creating the profile of a circle member from the participation perspective. These attributes determine the degree of a user’s participation in MC2 like the number of written articles, forum contributions, registered professional activity, query responses, new ideas, and son on. 5.2 Linking Concepts to Database Attributes In order for the MAS (the Meta-data Agent) to correctly correlate concepts captured by sensing from the organizational database, we shall use what is called a meta-data layer. This layer’s main function is to serve as a link between the concepts declared in MC2 and the fields, tables, and attributes in the organizational database. Actually, this layer maps all existing keywords in the MC2 system and links them to related attributes in the organizational database.

Improving Organizational Memory

173

Each organization has its own keywords that is important not only for the action of the agents but also for the operation of the MC2 system itself, since these words are distributed throughout the system to facilitate searches and to attribute concepts. Due to its importance in any implantation of the MC2 system, it is necessary to register the concepts (keywords) used by the organization. It is at this point that the mapping of concepts to database attributes takes place. For each inclusion of a keyword, the MC2 consultant can associate its existing attributes to the organizational database. There are other passive relational concepts with organizational database attributes that may also be used as parameters by the agents: CKLs, abilities, job positions etc. Like keywords, concepts that are unique to a particular organization can also be mapped to attributes in the implementation of the MC2 system. Conforming to the Human Resources Knowledge Circle example, we have the concepts described in Table 2 with their related attributes. Table 2. relationship between concepts and attributes mapped in the meta-data layer

Concept Human Resources Psychology Finance Project Management

Attributes Related to the organizational database Burrow, Sex, Department, Job, Seniority Social Level Salary Range Number of Projects Accomplished

5.3 Action After collecting the necessary parameters (attributes) the Data Preparation agent forms a group of observations that FORMVIEW will use to carry out processing. Tables 3 and 4 show some defined observations related to the professional and participation perspectives. Table 3. Subset of observations of the professional perspective for FORMVIEW processing Extra -time

Project executed

P1

0

4

P2

4

5

P3

2

10

P4

0

20

P5

0

23

P6

0

18

Salary range Up to 1500 Up to 1500 More 2000 Up to 1500 More 2000 Up to 1500

Position

Department

Social Level

Adm. time

Number clients

Direct

Marketing

Middle

Up to 3

4

Employee

HR

Middle

4-7

5

Direct

HR

High

Up to 3

10

Direct

Marketing

High

4-7

20

Direct

Finance

Middle

More 10

23

Employee

HR

Middle

More 10

18

174

João José Vasco Furtado and Vinícius Ponte Machado

Table 4. Subset of observations of the participation perspective for FORMVIEW processing Forum

Article

Answer queries

Professional Activities

New ideas

Votes

Group

Number Participation

P1

0

4

6

4

2

0

5

4

P2

4

5

7

3

0

0

11

5

P3

2

11

4

2

0

2

3

10

P4

0

2

3

2

2

3

10

5

P5

0

3

3

2

1

1

8

5

P6

0

8

5

7

1

1

10

5

Based on these observations the Knowledge Discovery Agent agent triggers the algorithm, in this case, FORMVIEW, to carry out the concept formation. As a result of the processing we have the concept hierarchies (Figure 4) that will be made available to current circle members. The user, if he or she so desires, may observe how other people, holders of the same knowledge are grouped following the two perspectives.

Fig. 4.. Example of concept hierarchies generated by KDA in MC2

Improving Organizational Memory

175

As we can observe in Figure 4, on the right side, there is a concept hierarchy representing the participation perspective and on the left side, there is the professional perspective. Some categories are named by the users like the category “HR older than 10” representing the concept of people working in the Human Resource department with more than 10 years in the organization or the category “High Participation” in the perspective participation representing people with high level of participation in MC2. These two categories are linked by means of bridges which represent the set inclusion relationship. Bridges indicate that entities represented by a concept in a perspective imply another in a different perspective. For instance, the bridge shown in Figure 4, indicates that the personnel from human resource department with more than 10 years of experience have high level of participation in the use of MC2 tools. Observing these concept hierarchies the current circle members can indicate new members based on this profile, and even discover if a particular user has a profile to be able to participate in the Human Resources Knowledge Circle. Additionally, the agent will automatically send messages to people with this profile, inviting them to take part in this Knowledge Circle.

6 Conclusion In this article, we describe agents that aid in the creation and maintenance of organizational memory that is characterized by the use of knowledge engineering and automatic knowledge acquisition algorithms. Among the benefits of this work we may cite that which allows for the automatic generation of knowledge that until now has been implicit within an organization. This knowledge is related to the activity goals of an organization and contributes toward memory creation and organizational learning. Another advantage is to create profiles, categories and classifications to better understand the business end of the company, such as suppliers, customers, employees and products. Another contribution of the approach here presented is to aid in bringing the user closer to MC2 system. By using the algorithms FORMVIEW, C4.5 and COBWEB to exploit company databases—that is to say in his work environment—the user is stimulated to contribute with the idea that the knowledge captured and deposited in MC2 is related to their practical work activities.

References 1 2 2. 3

Senge, Peter. The Fifth Disciple: The Art and Practice of the Learning Organization. New York, Currency/Doubleday, (1990). Gamble, R. Using AI for Knowledge Management and Business Process Reengineering. Technical Report Ws-98-13. AAAI Press, Menlo Park, CA (1998). Beckman, T., Liebowitz. Knowledge Organization: What Every Manager Should Know. St. Luice Publications (1998). van Buren, M. A Yardstick for Knowledge Management. Training and Development, (1999), 71-78.

176 4

5 7. 8. 9. 10. 11. 12. 13. 14.

15. 16. 17.

João José Vasco Furtado and Vinícius Ponte Machado Bidjan Tschaitschian, Andreas Abecker, Joachim Hackstein, and Jamel Zakraoui: Internet Enabled Corporate Knowledge Sharing and Utilization. In (eds) David G. Schwartz, Monica Divitini & Terje Brasethvik 2nd International Workshop on Innovative Internet Information Systems (IIIS-99) at ECIS'99, Copenhagen, Denmark, Hershey (USA) & London (UK): Idea Group Publishing (2000). van Heijst, G. Spek, R, Kruizinga, E.: Organizing Corporate Memories, Knowledge Acquisition Workshops and Archives KAW96, Alberta Canada (1996). Langley, P.: Elements of Machine Learning. Morgan Kaufmann (1996). Mitchell, T.: Machine Learning. McGraw Hill (1997). Quinlan, J.R. Learning Inductions of decision trees. Machine Learnig (1986). Fisher, D. Knowledge Acquisition via Incremental Conceptual Clustering. Machine Learning, v.2,n.2 (1987). Furtado, J.J.V: Formation des Concepts dans les langages de Schémas. Thèse de Doctorat, Université d´Aix- Marseille III (1997). Haimowitz, I. Santo, Nuno: Dynamic System for internet Direct-to-Consumer AgentMediated KM, Proceedings of the Agent-Mediated Knowledge Management Symposium. AAAI Spring Symposia series, Palo Alto, Stanford (2003). Krueger, W. Nilsson, J. Oates, T. Finin, T. Automatically Generated DAML Markup for Semistructured Documents, Proceedings of the Agent-Mediated Knowledge Management Symposium. AAAI Spring Symposia series, Palo Alto, Stanford (2003). Novak, J. Wurst, M. Fleischmann, M. Strauss, W.: Discovering, Visualizing and Sharing Knowledge through Personalized Learning Knowledge Maps. In: Proceedings of the Agent-Mediated Knowledge Management Symposium. AAAI Spring Symposia series, Palo Alto, Stanford (2003). Furtado, J.J.V, Colera, C. "Integrating Information Systems and Knowledge Engineering to Improve Learning Organization". Americas Conference on Information Systems, AMCIS 2000, Long Beach (2000). Russell, S., Norvig, P.: Artificial Intelligence a Modern Approach. Prentice Hall series on AI (1995). Inmon, W.: Enterprise meta-data, Data management Review magazine, Nov, www.dmmagazine.com/master.Efm?/NavID=216&EdID=298, (1998).

Experience in Using RDF in Agent-Mediated Knowledge Architectures Kit-ying Hui , Stuart Chalmers, Peter M.D. Gray, and Alun D. Preece Department of Computing Science Department, King’s College University of Aberdeen, Aberdeen AB24 3UE, Scotland, United Kingdom {khui|schalmer|pgray|apreece}@csd.abdn.ac.uk

Abstract. We report on experience with using RDF to provide a rich content language for use with FIPA agent toolkits, and on RDFS as a metadata language. We emphasise their utility for programmers working in agent applications and their value in Agent-Oriented Software Engineering. Agent applications covered include Intelligent Information Agents, and agents forming Virtual Organisations. We believe our experience vindicates more direct use of RDF, including use of RDF triples, in programming knowledge architectures for a variety of applications.

1

Introduction

Resource Description Framework (RDF) and the associated RDF Schema (RDFS1 ) were introduced by W3C as a portable framework for passing structured data and its associated metadata over the web. To date, much of the work on RDF and RDFS and on higher level languages such as DAML has concentrated on its role in supporting ontologies, rather than on how it helps a programmer working with contemporary agent architectures using Java toolkits. In this paper we wish to distil experience from using RDF in two very diﬀerent agent-based systems (AKT and Conoise) where we needed to use an agent platform (such as JADE) to exchange rich content. We believe that the virtues of RDFS as a sparse extensible metalanguage for describing data moved across the web have been lost sight of in the rush to explore more elaborate languages such as DAML and OWL. One should remember that these languages are themselves layered on RDF/RDFS. Also, languages such as OWL are evolving rapidly, while RDFS has remained surprisingly stable. Lastly, there is now very good support for RDF/RDFS in Java-based class libraries which can easily be integrated with FIPA-compliant Java-based agent toolkits, such as JADE. Thus from the viewpoint of Software Architectures there is much to commend the sparse elegance and stability of RDF/RDFS in a fastchanging world. Indeed, we now believe that it is timely for FIPA agent languages 1

Now working in the School of Computing, Robert Gordon University, Aberdeen AB25 1HG, Scotland, United Kingdom. http://www.w3.org/TR/REC-rdf-syntax

L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 177–192, 2003. c Springer-Verlag Berlin Heidelberg 2003

178

Kit-ying Hui et al.

to move from a LISP-based syntax to an alternative XML version, and to adopt RDF as a standard content language. A particular advantage of RDFS is its use from Java as a convenient accessible metalanguage giving the types of structured data, including possible specialisations. When data is coming across the web in many forms it is essential to have this type information easily available in memory. Thus the programmer does not have to invent their own form of object class for this metadata. They also have Java methods available to populate these classes from input in a widely used interchange format. Contrast this with the intricate methods provided for accessing some fairly basic metadata in JDBC. We have used this metadata for describing schemas (ontologies), quantified constraints for planning applications, and structured data for use in modelling Virtual Organisations. We describe these in more detail below. When working with RDF, we have found it useful to convert it into RDF triples held in a Java table. This is equivalent to the usual pointer-based graph structure, but easier to search and more convenient when merging in extra RDF triples. Searching the table is easier than trying to work directly with an XML tree structure, because equivalent XML abbreviations are reduced to a single form, like a canonical form. Besides using Java, we have developed in-memory Prolog term structures to hold these triples, which are more convenient for Prolog pattern matching and reasoning. This is crucial for semantic web applications. Thus we are not tied solely to handling triples in Java. Lastly, we have experience of using this system with several diﬀerent transport protocols. For web sources, we ﬁnd the HTTP protocol very useful, particularly where there are ﬁrewalls. For remote agents running on diﬀerent platforms we ﬁnd Linda [1] very useful in conjunction with remote procedure calls, while for Java platforms we may use Servlet technology. Because of the well established use of XML, support for these protocols is widely available and tested, which is another plus for RDF/RDFS.

2

The RDF Data Model

RDF is not unlike the Entity-Relational data model in its use of Entity identiﬁer as Subject, and Property or Relationship names as Predicate in RDF triples. However, it also includes features of object data models (such as OQL) in its use of object identiﬁers and subclasses. Although it looks simple, it has all the essential features for mapping other data models or layering extra details, as intended in its design. We have found it easy to map data from our existing FDM (an early semantic data model) and also quantiﬁed constraints which are formulae of logic expressed in this model. One very satisfying feature of our constraint interchange format in RDF (CIF) is that the tags used make a clean separation between information about logical formulae with the usual connectives, and information about expressions denoting objects in the data model. Eﬀectively CIF gives another layer with

Experience in Using RDF in Agent-Mediated Knowledge Architectures

179

richer semantic information, but it is able to use all the processing convenience of RDF. Expressions in our CIF language refer to facts about entities, their subtypes, attributes and relationships, which of course can be expressed in RDF. Note also that this model abstracts over relational storage, ﬂat ﬁles and object-oriented storage, following the principle of data independence. Thus it does not tie one to any particular system, such as Oracle or P/FDM. This is a great advantage to the programmer. Thus we stress the value of using RDFS in place of data structures that tie the programmer to a particular storage schema. The mapping to a particular knowledge source or data source can then take place separately through a wrapper. This makes it very much easier to integrate data from different sources, as is often required over the Web. We illustrate these programming principles below in two rather diﬀerent applications. Firstly we consider the fusion of constraint information from diﬀerent sources on an intranet in the KRAFT project. This uses intelligent information agents as mediators and facilitators. It originally used a textual version of Prolog term structures for interchange. We are now reworking it to use RDF and XML, and to use RDFS in place of P/FDM speciﬁc constructs. This is being used in the AKT (Advanced Knowledge Technologies) collaborative project, so that we can fuse data from other partners. Our second example uses agents that make bids and that come together to form Virtual Organisations. This also uses constraints, but more as a way to give additional desires to BDI agents, whilst allowing them autonomy in how they cope with other conﬂicting goals and desires.

3

Agent Architectures for Information Integration and Fusion

The KRAFT2 system [2] employs an agent-based architecture inspired by the Knowledge Sharing Eﬀort which has proved to be an eﬀective approach to developing distributed information systems. The basic philosophy of the architecture design is to deﬁne a KRAFT domain where certain communication protocols and languages must be respected (ﬁgure 1). Within this domain, agents are assumed to cooperate and connections are made dynamically3 . KRAFT recognises constraints as abstract mobile knowledge which can be extracted, transported, transformed and processed by software components. We use the constraint formalism as a domain-independent framework to represent application problems as constraint satisfaction problems (CSPs). The application domain is modelled by a database schema and domain knowledge is captured as constraints. Constraints distributed in resources form a library of shareable and reusable knowledge blocks which can be combined to compose problem speciﬁcations. Having a semantic data model extended with constraints and mapped into an open interchange format supports a range of applications in which infor2 3

Knowledge Reuse And Fusion/Transformation Hence the absence of explicit connections in the grey area in ﬁgure 1.

180

Kit-ying Hui et al. Non-KRAFT components

R UA

UA UA

R

W W

R

W F

User Agent

Resource

M

M

KRAFT facilities

M F

M

W

Wrapper

M

Facilitator

F

Mediator

W W W R

R UA

Fig. 1. This ﬁgure shows a conceptual view of the KRAFT architecture. KRAFT components are round in shape while non-KRAFT ones are marked as squares. The grey area represents the KRAFT domain where a uniform language and communication protocol is respected. mation needs to be moved across a network with rich metalevel information describing how the information can be used. An example application common in business-to-business e-commerce involves the composition of a package product from components selected from multiple vendors’ catalogues. There are various kinds of constraints which must be aggregated and solved over the available component instances: constraints representing customer requirements, constraints representing rules for what constitutes an acceptable package, and constraints representing restrictions on the use of particular components. In this last case, the ability to store constraints together with data in the P/FDM database system4 allows instructions, which we called “small print constraints”, to be attached to the class descriptor for data objects in a product catalogue database. When a data object is retrieved, these attached instructions must also be extracted to ensure that the data is properly used. Thus the attached constraint becomes mobile knowledge which is transported, transformed and processed in a distributed environment. This approach diﬀers from a conventional distributed database system where only database queries and data objects are shipped. 3.1

Modelling a Domain in RDF Schema

In the original KRAFT system, we model a domain by a database schema, using the functional data model. The schema eﬀectively serves as an ontology that captures knowledge of classes, attributes and subclass relationships in the domain. The following is part of a schema showing the pc and os classes and the memory property in an application domain where components are put together to conﬁgure a workable PC. The complete ER model is shown in ﬁgure 2: 4

http://www.csd.abdn.ac.uk/∼pfdm/

Experience in Using RDF in Agent-Mediated Knowledge Architectures declare ... declare declare declare ...

181

os ->> entity pc ->> entity memory(pc) -> integer has_os(pc) -> os

The functional data model is an extended ER model which can be easily mapped into the RDF schema speciﬁcation. A mapping program reads meta-data from the database and generates the corresponding RDF schema, making this knowledge web-accessible. The following RDFS fragment refers to the schema above:

Mapping a P/FDM schema into RDFS has the advantage of making the domain model available to RDFS-ready software. On the other hand, some semantic information is lost. The cardinality of each attribute, for example, is not expressed in the RDF schema. Information on the key of each entity class is also omitted. However, this information can easily be added to other metadata held in the cif:entmet layer [3] Brieﬂy, entmet is a class of metaobjects whose instances correspond one-toone with entity classes, and whose property values give metadata such as class name, superclass metaobject and RDF URI. One instance is the object in ﬁgure 4 whose ID is entmet pc. This gains us extensibility by adding extra properties, pc model(pc) -> string cpu(pc) -> string memory(pc) -> integer has_os(pc) -> os has_disk(pc) ->> hard_disk

os name(os) -> string size(os) -> integer hard_disk model(hard_disk) -> string size(hard_disk) -> integer

Fig. 2. This schema shows three entity classes. The single arrow means that each pc may have only one os installed. A double arrow means that a pc can have multiple hard-disk.

182

Kit-ying Hui et al.

such as key, to the metaobjects in entmet or in propmet (for properties). There is, of course, some redundancy, as where we record subclass information both in entmet and by using rdfs:subClassOf. However, it is kept consistent, and provides a clean layering of extra semantic information to be used by enabled reasoners. 3.2

Capturing Domain Knowledge in RDF

Domain knowledge in KRAFT is captured as integrity constraints expressed against the database schema, using the constraint language CoLan [4]. CoLan is based on ﬁrst-order logic and has proved to be expressive enough to represent complex constraints [5]. CoLan constraints have evolved from the usual database state restrictors into mobile problem speciﬁcations [6]. The following is an example of a design constraint saying that “the size of a hard disk must be big enough to accommodate the chosen operating system in every PC”: constrain each p in pc to have size(has_os(p)) =< size(has_disk(p))

User requirements are also captured as constraints. The following constraint speciﬁes that the conﬁgured PC must use a pentium 4 CPU: constrain each p in pc to have cpu(p)="pentium 4"

In practice, the human-readable CoLan constraints are compiled into an intermediate format, called Constraint Interchange Format (CIF). CIF expressions are syntactically Prolog terms, which are easier to process by software components. To make CIF portable, we encode CIF constraints into RDF by deﬁning an RDF schema for the CIF language that is layered cleanly on top of RDF, serving as a meta-schema [3]. A constraint encoded in RDF makes explicit references to classes deﬁned in the domain model as well as the CIF language definition in RDF schema (ﬁgure 3). The RDF fragment in ﬁgure 4 shows the cif namespace deﬁnition and references to the CIF language RDF schema resource. In this example, the variable with name uevar1 is restricted to be an instance of the entity class pc which is deﬁned by the domain model in "http://www.csd.abdn.ac.uk/ ∼schalmer/schema/pc schema#pc". References to the domain model can be found as value of the cif:entmet rdfname property. This approach makes no change to the existing RDF and RDF schema speciﬁcations. As constraints are now represented as resources, RDF statements can also be made about the constraints themselves. A detail discussion of the CIF encoding in RDF is presented in [3]. 3.3

Using Knowledge Encoded in RDF

Knowledge encoded in RDF is a web-accessible resource which can be utilised by knowledge processing components. While the RDF model is simple and may not

Experience in Using RDF in Agent-Mediated Knowledge Architectures

RDF schema of the "PC configuration" domain

RDF schema of the CIF language

constraint on the "PC config" domain as an RDF resource

RDF schema of domain "X"

constraint on domain "X" as an RDF resource

183

RDF schema

RDF

Fig. 3. Constraints in RDF make references to the CIF language deﬁnition and domain models in RDF schema. uevar1 http://www.csd.abdn.ac.uk/~schalmer/schema/pc_schema#pc ...

Fig. 4. This fragment of a constraint encoded in RDF shows explicit references to the CIF language deﬁnition and the domain model in RDF schema.

be expressive enough to represent complex and sophisticated knowledge, higherlevel layers can be nicely built above RDF to incorporate the richer semantics. This is demonstrated in our RDF encoding of the CIF constraint language. RDF provides a uniform framework for the representation of knowledge and meta knowledge. Although it is simple enough to be processed by diﬀerent soft-

184

Kit-ying Hui et al.

ware platforms, we believe that the RDF-encoded format should be used for transportation only, not for direct reasoning purposes. Instead, inference and reasoning are better done on the RDF triples; or on Prolog term structures which are mapped from the triples for processing convenience.5 RDF triples may be representing knowledge in a small granularity but they can be easily assembled into knowledge chunks of a convenient size for processing purposes. Thus the triple is a format which can be readily used in reasoning or as a stepping stone to other formats. These triples can conveniently be created in memory by Jena (ﬁgure 9). In this section, we show how constraints, meta knowledge and domain models encoded in RDF can be used by knowledge processing components. – Converting Constraints between CIF and RDF An RDF schema provides knowledge on the class hierarchy and class properties with type information. For example, as we know that the CIF constraint class has three subclasses impliesconstr, unquantified constraint and existsconstr, we have to try all these subclasses when we try to parse an RDF statement as a constraint resource. Similarly when we try to parse a property of a resource, we know what class of resource is valid. The RDF schema of the CIF language alone does not give us enough information to map an RDF-encoded CIF constraint into its Prolog representation (and vice versa). There are two pieces of knowledge missing: • What is the Prolog term structure that corresponds to a certain class in the RDF schema (of the CIF language)? • Given the Prolog representation of a resource of a certain class, how can we ﬁnd each property of the resource (as Prolog terms)? This missing knowledge should not be speciﬁed in the RDF schema as it cares only about the semantics of a constraint resource but not the way that it is represented in Prolog. To solve this problem, we represent this knowledge as mapping rules between a Prolog term structure and its corresponding RDF class (in the CIF language deﬁnition RDF schema). Once the missing knowledge is provided, the mapping process between CIF and its RDF encoding is totally driven by the RDF schema of the CIF language. Once again, structuring the task around RDF has made the programming job easier. – Knowledge Reuse by Constraint Fusion Declarative constraints stored as self-contained knowledge objects in a distributed system form a shared library of building blocks. The key to reusing this knowledge is the process of constraint fusion, which dynamically combines the semantic content to compose problem speciﬁcation instances. While a single piece of constraint may not contain enough information to solve a 5

We use the Prolog “Pillow” library (http://clip.dia.fi.upm.es/Software/ pillow/pillow.html) for XML parsing, with an RDF parser sitting on top of Pillow to map XML trees into RDF triples.

Experience in Using RDF in Agent-Mediated Knowledge Architectures

185

CSP, we hope that by combining constraints together, their total value is enhanced, thus making the problem solvable. The domain independent constraint fusion engine in KRAFT looks for any potential semantic information exchange when constraints are logically conjoined together. Simply speaking, it works by looking for variable pairs which are generated in the same way. Under this condition, constraints that apply to one variable may apply to the other variable. For example, variable x and y share each other’s constraints as they are both instances of the pc class: constrain all x in pc to have cpu(x)="pentium 4" constrain all y in pc to have memory(y)>=128

In fact, sharing of constraint also happens between variables of a superclass and a subclass, as constraints are inherited by the subclass from the superclass. That means the constraint fusion engine needs knowledge of the class hierarchy in the problem domain in order to make the correct inference. In this case, the domain model is readily available as an RDF schema which can be accessed as RDF triples. Given a constraint, the fusion engine can easily retrieve the RDF schema of the domain model by following the explicit links. By collecting RDF triples on the rdfs:subClassOf predicate, the fusion engine then gets a complete picture of the class hierarchy in the problem domain. The small granularity of knowledge represented in the triple form allows the knowledge processing component (in this case, the constraint fusion engine) to selectively access the required knowledge in a quick and convenient manner.

4

Conoise and Virtual Organisations

Virtual organisations (VOs) in Conoise6 are composed of a number of autonomous entities (representing diﬀerent individuals, departments and organisations) each of which has a range of problem solving capabilities and resources at their disposal. These entities co-exist and sometimes compete with one another in a ubiquitous virtual market place. Each entity attempts to attract the attention of potential customers and ultimately tries to sell them its services by describing the cost and quality of the service. Sometimes, however, one or more of the entities may realise there are potential beneﬁts to be obtained from pooling resources: either with a competitor (to form a coalition) or with an entity with complementary expertise (to oﬀer a new type of service). When this potential is recognised, the relevant entities go through a process of trying to form a new VO to exploit the perceived niche. 6

Constraint Oriented Negotiation in Open Information Seeking Environments (http://www.conoise.org)

186

Kit-ying Hui et al.

Given the independent nature of the entities involved, this process may succeed or it may fail. If it succeeds, the collection of independent entities have to start acting as a single conceptual unit. In particular, they need to cooperate and coordinate to deliver the services of the newly formed organisation. In dynamic environments, the context may change at any time, such that the VO is no longer viable. Then it will either need to disband or re-arrange itself into a new organisation that better ﬁts the prevailing circumstances. In Conoise we represent the knowledge and communication between the agents using CIF/RDF. We build on the ability to combine the CIF/RDF knowledge [7] so that we can use many disparate information sources to help in the formation, management and dissolution of these VOs. The ability to combine information from these diﬀerent sources means that our decision making process uses as much knowledge as is available at the time to aid in the process of choosing other agents as VO partners. 4.1

Conoise Agent Design

Conoise uses CIF/RDF constraints to represent the services required in such VO’s, and uses a CSP (Constraint Satisfaction Program) solver to provide the reasoning process for their identiﬁcation, formation, management and dissolution. The reason for using a CSP is that, as in KRAFT: – We get data independence across platforms – It is easy to combine conjunctive First Order Logic constraints compared to imperative code Typically the starting point for this process will consist of an agent receiving a call for bids to provide a service. The agent must then decide what course of action to take to provide that service. This can be either: – To provide a bid based on its own resources. – To provide a bid based on the resources available from its membership of an existing Virtual Organisation. – To provide a bid by creating a new Virtual Organisation, thereby initiating a new call for bids to ﬁnd new VO partners. When deciding on which action to take the agent must be aware of the current status of other agents and their abilities as well as what resources it can itself provide. constrain all t in travel_plan such that travel_method(t) = train and travel_class(t) = sleeper_train to have arrival_time(t) < 0830

Fig. 5. A call for services in CIF

Experience in Using RDF in Agent-Mediated Knowledge Architectures

187

constrain all s in service_description such that name(s) = Euston to have travel_info(s) = cancelled

Fig. 6. Environment information held as CIF Figure 5 shows an example call to provide a service (a travel plan representing a complete travel ’package’ to a speciﬁed destination [8]). The RDF is not shown for reasons of space, but is similar to the example in ﬁgure 4. This constraint determines that if the travel plan includes travel by train and it is a sleeper train, then it must arrive at its destination by 8:30am. The added bonus of such representation is that, once in the CIF format, we can combine this call for services with other information (also held as CIF/RDF), such as previous knowledge, current commitments and environment information in the CLP process, thus providing us with a more detailed constraint problem to manage the VO. For example, this train service information can be combined with information that the agent receives about the availability of trains from London Euston (ﬁgure 6). This means that the travel plan will not only take into account the sleeper train and arrival time speciﬁc constraints, but can also combine this with knowledge on a speciﬁc station it has received. 4.2

Conoise Agent Implementation

Fig. 7. The Conoise Agent Architecture

188

Kit-ying Hui et al.

The Conoise agent architecture is shown in ﬁgure 7. The agents are built using the JADE Java-based agent platform7 . Each agent in Jade communicates using FIPA ACL 8 , leaving the content language to be speciﬁed by the agent (here CIF/RDF).

Do options := option_generator(event_queue,B,D,I) selected-options := deliberate(options,B,D,I) update-intentions(selected-options,I) execute(I) get-new-external-events(); drop-succesful-attitudes(B,D,I) until quit

Fig. 8. BDI Algorithm

The agent architecture is based on the Beliefs, Desires, Intentions (BDI) agent model [9]. This uses data structures to represent the beliefs (what the agent knows), the desires (what the agent wants to do) and the intentions (how the agent achieves its desires) and the agent’s reasoning and decision making process comes from the interaction of these data structures according to a deliberation process (represented as a top level pseudo-code algorithm in ﬁgure 8). The desires of the agent are represented as constraints, and therefore exchangeable between agents as CIF/RDF constructs. To parse the RDF we are using the Jena toolkit9 , a Java API for parsing and manipulating RDF data models. It takes the RDF constraint and stores it as a set of Subject-Predicate-Object triples in a Model object. This object can then be queried using the API methods getSubject(), getPredicate() and getObject(). The Java code fragment in ﬁgure 9 shows the model variable being instantiated with the RDF from the URI speciﬁed by the constraintURI variable. The while loop shows how we can then parse this model and extract the necessary subject-predicate-object values. Once a message has been received by the agent it is passed into the message queue (ﬁgure 7), where it is parsed into a new Jena Model (as shown). The agent then creates a desire object and, from this, creates a set of possible intentions describing ways of completing this desire. It then chooses between these competing intentions (using the CLP solver and its beliefs) and executes the chosen one, which will complete the necessary steps to fulﬁll the desire. As we have all the agent’s other commitments (the other desires it is doing at the same time) stored in desire objects as CIF/RDF parsed Jena models, we can combine these with the current desire when deliberating to provide a choice of intention.

7 8 9

http://sharon.cselt.it/projects/jade/ http://www.ﬁpa.org http://www.hpl.hp.com/semweb

Experience in Using RDF in Agent-Mediated Knowledge Architectures

189

Model model = new ModelMem(); JenaReader r= new JenaReader(); r.read(model, constraintURI); StmtIterator iter = model.listStatements(); while (iter.hasNext()) { com.hp.hpl.mesa.rdf.jena.model.Statement stmt = iter.next(); Resource subject = stmt.getSubject(); Property predicate = stmt.getPredicate(); RDFNode object = stmt.getObject(); ... }

Fig. 9. Using Jena to create RDF triples.

5

Discussion & Related Work

In this paper, we have demonstrated how RDF has a far broader applicability than its original role as a metadata format for web resources. Our usage of RDF places it at the core of an agent-mediated distributed knowledge architecture, in which RDF provides: a carrier for the communication of entity-relational information between web services, a format for mobile and self-describing constraints, and a content language for inter-agent communication. In this section, we summarise our experience in using RDF, and draw comparisons with related work. The most common approach for transport of structured information between web services is to deﬁne the data using XML Schemas (or DTDs) and use an XML-RPC framework such as SOAP. The use of RDF as a layer on top of XML means that the communicated information has a data model (semantics), whereas the XML Schema/DTD-deﬁned data has only a structure (syntax). Most of the work that acknowledges the importance of a semantics for communicated data — essentially the founding principle of the Semantic Web movement — proceeds to deﬁne additional layers on top of RDF and RDFS (DAML+OIL, OWL, etc). O-Telos-RDF [10] even proposes an alternative to RDF, serialisable in XML, but with much clearer axiomatic semantics. In contrast, we view RDF and RDFS as suﬃciently useful and extensible in itself; in defence of this position, we have shown that the RDF data model is adequately expressive to transport data originally stored against the P/FDM semantic data model. Other work within the AKT project further supports this position, where RDF is used to represent a large repository of information on the research activities of UK universities, and to draw inferences from this information [11]. Further evidence for the utility of RDF in knowledge management applications is provided by the EU COMMA project [12]. In all of these approaches, the suﬃciency of RDF and RDFS to represent and communicate both ontological (schema) information and individual instances is demonstrated.

190

Kit-ying Hui et al.

Our original inspiration for using RDF and RDFS as a format for mobile and self-describing constraints was the original OIL proposal [13], wherein the RDF data model was extended with logical connectives (and, or, not, etc) which were themselves syntactically deﬁned using RDFS. The more recent Semantic Web work on DAML+OIL has backed-away from this original position, apparently because the logical apparatus of boolean expressions has been “pushed up” to a higher logic layer, above the DAML+OIL ontology layer 10 . While this is not unreasonable, we have opted to deﬁne our constraint format CIF directly on top of RDFS, as we view RDF itself as a major application of constraints. For example, CIF can be used to deﬁne integrity constraints on RDF classes, even though they are not expressible in RDFS itself. Our CIF work is related to the ongoing RuleML initiative11 (positioned at the Semantic Web logic layer) and, while we would see our work coming into alignment with RuleML in the future, we note that currently RuleML is more concerned with traditional if-then rules rather than declarative constraints. OCL is a declarative expression language for annotating UML class diagrams with invariants, especially more complex cardinality constraints. It has been used to create formal models of conﬁguration problems [14], which is more concerned with formal correctness than with AI problem-solving. However their logical formalism is very similar to ours, and a UML class diagram is just an ER diagram with methods, where we use functions. Our high-level architecture is agent-based, and RDF serves as the content language. Again, this communication framework stands in contrast to the mainstream RPC model of web inter-application communication, where SOAP dominates. However, agent-to-agent communication is far more ﬂexible than RPC, being loosely-coupled, asynchronous, and allowing individual network nodes full autonomy [15]. Regrettably, the current state-of-the-art in agent communication languages is not directly compatible with RDF. Both FIPA and KQML rely primarily on a LISP-based syntax rather than XML and, while FIPA informally allows RDF as a content language, this is not part of the standard. Nevertheless, we assert that the combination of an agent architecture and RDF as a content language is powerful and open, and believe that a fusion of web standards with agent standards is a desirable goal of both communities.

6

Conclusion

An important theme of the conference, enunciated by Ludger van Elst in his opening remarks, was the move towards intelligence-enhanced and integrated solutions. This is exactly the motivation behind our paper. We use FDM because it is a powerful semantic data model, designed for data integration across multiple platforms, and proven in use. By expressing range-restricted constraints in portable form, conforming to the data model but independent of the implementation details of the storage model, we make data semantics available in a form 10 11

http://www.w3.org/2000/Talks/1206-xml2k-tbl/slide10-0.html http://www.dfki.uni-kl.de/ruleml/

Experience in Using RDF in Agent-Mediated Knowledge Architectures

191

suited to mechanised reasoning and intelligent processing at the target site. This is the gateway to intelligence-enhanced solutions. In our case it is done using ﬁnite domain constraint solvers, but other kinds of constraint checker or solver are applicable. This theme was also picked up by Charles Petrie in his invited talk in the opening session. He expressed the need to exchange information in a form that conforms to W3C standards, yet allows AI reasoners to work with it and so gain ﬂexibility. Thus he wants interfaces to be ﬂexible enough to allow code to be generated on the ﬂy to ﬁt them, so that they do not need the intervention of human programmers. We have found that a combination of Prolog and Java, as described in this paper, works very well for this purpose. Prolog is excellent for on-the-ﬂy code generation, and also for pattern-matching type descriptions, possibly held in memory as RDF triples. Java is excellent for creating external interfaces conforming to W3C speciﬁcations. A combination of the two, as in Sicstus, works very well. Prolog is also well adapted to constraint solving and mechanised reasoning, when working with a logic formalism such as that advocated by Petrie or Pease [16]. In conclusion, we have argued that, for agent-mediated knowledge systems, there is a clear need not just to represent and communicate information, but also to reason with it. We have shown how we are using RDF to do both: we can readily map entity-relational data into RDF, communicate it within an agent communication language, and reason with it using constraints also deﬁned in RDF. In conclusion, we see the desirable features of RDF for agent-mediated knowledge systems as being: – simplicity of its triple-based data model; – uniformity in the representation of both knowledge (instances) and metaknowledge (schemas); – portability of the XML serialisation; – web-enabled, compatible with W3C standards; – extensibility, exempliﬁed by the layering of RDFS on RDF.

7

Acknowledgements

This work is supported under the Advanced Knowledge Technologies (AKT) Interdisciplinary Research Collaboration (IRC), which is sponsored by the UK Engineering and Physical Sciences Research Council (EPSRC) under grant number GR/N15764/01. The AKT IRC comprises the Universities of Aberdeen, Edinburgh, Sheﬃeld, Southampton and the Open University. The constraint fusion services were developed in the context of the KRAFT project, funded by the EPSRC and British Telecom. The CONOISE project is funded by BTexact Technologies.

192

Kit-ying Hui et al.

References [1] Gelernter, D., Carriero, N.: Coordination languages and their signiﬁcance. Communications of the ACM 35 (1992) [2] Preece, A., Hui, K., Gray, A., Marti, P., Bench-Capon, T., Cui, Z., Jones, D.: KRAFT: An agent architecture for knowledge fusion. International Journal of Cooperative Information Systems 10 (2001) 171–195 [3] Gray, P.M.D., Hui, K., Preece, A.D.: An expressive constraint language for semantic web applications. In Preece, A., O’Leary, D., eds.: E-Business and the Intelligent Web: Papers from the IJCAI-01 Workshop. AAAI Press (2001) 46–53 [4] Bassiliades, N., Gray, P.: CoLan: a Functional Constraint Language and Its Implementation. Data and Knowledge Engineering 14 (1994) 203–249 [5] Fiddian, N.J., Marti, P., Pazzaglia, J.C., Hui, K., Preece, A., Jones, D.M., Cui, Z.: A knowledge processing system for data service network design. BT Technical Journal 17 (1999) 117–130 [6] Gray, P.M.D., Embury, S.M., Hui, K., Kemp, G.J.L.: The evolving role of constraints in the functional data model. Journal of Intelligent Information Systems 12 (1999) 113–137 [7] Gray, P.M.D., Hui, K., Preece, A.D.: Finding and moving constraints in cyberspace. In: Intelligent Agents in Cyberspace, AAAI Press (1999) 121–127 Papers from the 1999 AAAI Spring Symposium Report SS-99-03. [8] Chalmers, S., Gray, P., Preece, A.: Supporting virtual organisations using BDI agents and constraints. [17] 226–240 [9] Rao, A.S., Georgeﬀ, M.P.: BDI agents: From theory to practice. In: Proceedings of the First International Conference on Multi-Agent Systems. (1995) [10] Nejdl, W., Dhraief, H., Wolpers, M.: O-Telos-RDF: A resource description format with enhanced meta-modeling functionalities based on o-Telos. In: Workshop on Knowledge Markup and Semantic Annotation at the First International Conference on Knowledge Capture (K-CAP’2001), Victoria, B.C., Canada (2001) URL: http://www.kbs.uni-hannover.de/Arbeiten/ Publikationen/2001/kcap01-workshop.pdf. [11] Alani, H., O’Hara, K., Shadbolt, N.: Ontocopi: Methods and tools for identifying communities of practice. In Musen, M., Neumann, B., Studer, R., eds.: Intelligent Information Processing. Kluwer Academic Press (2002) 225–236 [12] Gandon, F., Dieng-Kuntz, R.: Distributed artiﬁcial intelligence for distributed corporate knowledge management. [17] 202–217 [13] Broekstra, J., Klein, M.C.A., Decker, S., Fensel, D., van Harmelen, F., Horrocks, I.: Enabling knowledge representation on the web by extending RDF schema. In: World Wide Web. (2001) 467–478 [14] Felfernig, A., Friedrich, G., Jannach, D.: Conceptual modeling for conﬁguration of mass-customizable products. Artiﬁcial Intelligence in Engineering 15 (2001) 165–176 [15] Vinoski, S.: Putting the ”web” into web services: Web services interaction models, part 2. IEEE Internet Computing (2002) July/Aug, 90–92 [16] Pease, A., Li, J.: Agent-mediated knowledge engineering collaboration. In: Agent Mediated Knowledge Management, AAAI Press (2003) 167–172 Papers from the 2003 AAAI Spring Symposium SS-03-01. [17] Klusch, M., Ossowski, S., Shehory, O., eds.: Proceedings of the 6th International Workshop on Cooperative Information Agents (CIA 2002). Number 2446 in Lecture Notes in Artiﬁcial Intelligence, Madrid, Spain, Springer Verlag (2002)

Using an Agent-Based Framework and Separation of Concerns for the Generation of Document Classification Tools João Alfredo Pinto de Magalhães, Carlos José Pereira de Lucena Departamento de Informática, Pontifícia Universidade Católica do Rio de Janeiro Rua Marquês São Vicente 255, Rio de Janeiro, Brazil {magalha, lucena}@inf.puc-rio.br

Abstract. The present paper describes the Avestruz1 framework, a multi-agent based architecture for generating tools for document classification in specific domains. We have designed and implemented an architecture that separates the concerns related to document search and election (called a concern platform) from those related to the classification algorithm. Not only is it possible to use and experiment with existing classification algorithms but also to generate new algorithms, taking into consideration specific characteristics of the domain

1. Introduction The problem we address in this work is document classification - that is, given a set of documents and a set of classes, to decide which documents best match each class. When a tool for classifying documents is generated, many concerns need to be taken into consideration. These concerns can be divided into two main groups: • The first group deals with concerns related to general functionalities, such as document processing (extraction of important parameters used during classification), document location (finding the places where the documents are stored), detection of duplicates (to avoid the useless re-classification of already classified documents) and report generation (in which way the results will be used), just to mention a few; • The second group deals with concerns related to the document classification algorithm that is used. Due to its intrinsic characteristics, the second group of concerns strongly suggests that it should be handled by someone with accurate skills in algorithm engineering, whereas the first group does not require such a specialization. The first group of concerns can be "frozen" in a framework [1] to allow for algorithm research. Such a framework, in which one can plug and use different algorithms in an easy fashion, 1

the word for ostrich in Portuguese. L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 193-200, 2003.  Springer-Verlag Berlin Heidelberg 2003

194

João Alfredo Pinto de Magalhães and Carlos José Pereira de Lucena

plays the role of a laboratory where one can analyze the overall performance of different algorithms, processing different types of documents (for instance, documents about economy, politics, documents written in HTML, PDF or PS). After testing different algorithms, the best one can be chosen in order to create a complete application that uses the functionalities of the first group. Following this strategy, we have developed Avestruz, a multi-agent framework that generates tools to classify documents. Since their work is automatic, the generated tools can be highly autonomous: the only inputs required are the places where the documents are stored. The existing hot spots [1] are the type of documents supported, the reports generated, the degree of memory and pro-activity of the agent, the document selection strategy and the classification algorithm. Clearly, the last hot spot is directly related to the second group of concerns (which we call classification concerns), whereas the others comprise the first group (which we call concern platform). The main goal of the framework architecture is to decouple the concerns related to the classification algorithm from those related to the platform by using the concept of separation of concerns [2]. Once the concerns platform is defined, a specialist from the documents domain can handle the concerns related to the classification algorithm. The great advantage is the ability to test different algorithms: a specialist in algorithms has a powerful tool for making tests.

2. Methodology The decision of adopting a design approach based on multi-agents for the development of the framework meant to facilitate the understanding of the problem at hand. A classifier agent can be seen as an entity that is constantly searching for documents, while selecting the ones that must be processed. It processes them according to a classification algorithm and reports the results. In order to prevent that concerns related to a specific agent conflict with concerns related to the multi-agent system [2][3] we have adopted a two-layered approach. The software agents that are controlled by the second layer are placed in the first layer, which is the management layer. The strategy used to obtain the desired independence between platform and classification concerns was to concentrate the concerns platform on both the management layer and the software agent layer while the classification concerns are located only in the agent layer. This is done in a way that by dealing with the software agent as a user of external services it hides the software agent concept. Once the platform is instantiated (generated), the framework generator is free from the software agents concerns to concentrate its attention on the classification algorithm. It does not need to know that the framework was built using software agents - this capability is completely encapsulated. All the frozen spots [1] are located in the management layer, and they deal with concerns such as agent cooperation and coordination. Thus, in order to benefit from the proposed two-group approach, it is necessary to instantiate the framework following two steps: the first step deals with the concerns platform, where all the corresponding hot spots are completed; the second step is dedicated exclusively to the classification hot-spot.

Using an Agent-Based Framework and Separation of Concerns

195

3. The Avestruz Framework Avestruz was developed using the object-oriented paradigm and Java as the programming language. The architecture built is completely based on the two-layered approach, and we made extensive use of Design Patterns [5] to achieve the desired flexibility. We discuss now in detail each of the layers. 3.1 Agent Layer Each classifier agent has its own thread of execution, and is composed of six subsystems, as shown in figure 1. The Coordination subsystem encapsulates the multi-thread characteristics of the framework, so it appears in the figure as multiple instances, while the other subsystems are shared by the coordination subsystems.

Fig. 1. Agent subsystems

Each subsystem is responsible for a very specific task: • Document Processor: processes the documents and extracts from them all the relevant information that is to be used during the classification process. Among the classes that compose this subsystem, the most important is the Document interface whose methods define a unique interface that is used by other subsystems, specially the Classifier. When instantiating the framework, for each type of document that is to be processed (for instance, HTMLs, PDFs, etc) there must be a parse class implementing Document. Each different document generates a different instance of the corresponding parse class, and this instance must be capable of providing access to all necessary information of the parsed document. As some documents are able to indicate new document paths to be processed., such as HTMLs or PDFs (that may contain hyperlinks to other documents), there is the getPointedDocuments method, that may return the pointed document paths. This is another hot-spot of the framework: the software agents can be able to discover new documents other than the ones initially indicated.

196

João Alfredo Pinto de Magalhães and Carlos José Pereira de Lucena

• Memory: keeps track of the processed documents. It avoids a document to be processed twice in the same execution. This is a frozen spot of the framework: a document can be processed only once in an execution. However, it may be required that a document is processed again in another execution. That is the difference between a volatile and a persistent memory. For this reason, although the volatile memory is frozen, the persistent memory is a hot-spot. • Report Generator: This is another important hot-spot as different tools require different reports to be generated. For example, it is possible to develop a tool that stores all the reports in a database, whereas another one can generate a report and send it by e-mail. • Classifier: responsible for the classification process. This subsystem is another hotspot as it encapsulates the classification algorithm used, and uses the instances of the classes that implement Document generated by the Document Processor to gather all the required information. • Document Paths List: stores the document paths. When the system starts the execution, this list must contain all document paths to be processed. As new document paths are indicated by the processed documents, they are also put in this list. Avestruz provides a ready-made list, that can be extended to add new features, such as implementing filters or priorities for documents to be processed. For instance, when selecting document paths to be classified, store only those that have the string “cnn.com” in its path. This would imply that only the cnn.com site will be classified. The process continues until this list is empty. • Coordination: coordinates the agent tasks: (i) get a document path from the Document Paths List; (ii) ask the Document Processor to open and parse the document; (iii) add to the Document Paths List all the document paths pointed by the processed document; (iv) check the Memory to be sure the document has not already been processed and if not, ask the Classifier to classify the document, and send the results to the Report Generator. Each classifier agent of the system performs exactly the same tasks in the same order – there is no specialization. Also, by analyzing the Coordination subsystem it possible to understand why the software agent abstraction is completely hidden from the instantiation of the framework: none of the five subsystems require previous knowledge about the Coordination subsystem to be instantiated – all the five subsystems are seen by the Coordination as external services with a well-defined interface. 3.2 Management Layer The concerns related to this layer deal with the overall execution of the system, such as (i) creating the software agents with the necessary parameters; (ii) stopping the software agents prematurely, if necessary; (iii) generating statistics about the

Using an Agent-Based Framework and Separation of Concerns

197

execution of the agents; (iv) providing a framework [1] [5] for the communication and coordination of the software agents. Instances of the framework use this layer to communicate with the underlying multi-agent system – there is no direct communication with the agents. So, it can be seen as the only access point for the multi-agent system. In terms of separation of concerns, this means that it is possible to can think of the Management Layer as the Facade [5] for a classification subsystem. The adopted communication framework is based on blackboards [4], and is present mainly in the Document Paths List and the Memory, as both need to share their content among the agents. Even though part of these two subsystems are related to communication and cooperation concerns, this part is completely isolated from the instantiation process. 3.3 Instantiation The two-step approach proposed means that the first subsystems that must be instantiated are Document Processor, Memory, Report Generator and Document Paths List, as all of them deal with the platform concern. The Classifier subsystem is the one that directly deals with the classification concerns, therefore it must be instantiated in a second step, thus allowing for algorithm research. Figure 2 depicts the complete instantiation process, and indicates in which layer each part is located. There is a third layer called Application Layer, representing code specific for an application, such as user interface or other that is not directly related to Avestruz.

4. The Generated Tools The main purpose for making the tools generated from the framework highly autonomous is to minimize the need for interference from the user during the classification process. Also important is the fact that the instances of the framework are not limited to self-contained applications: it is possible to generate independent components to be connected to other knowledge management applications. There are already four substantially different instances derived from the framework, two applications and two components: • Semantic Probe is an instantiated component that classifies documents in the SHOE format [6], a proposal for marking up documents for the Semantic Web [7]. The implemented algorithm combines the document content (only the textual content) with the author’s original classification. It is an ontology-based algorithm, where the taxonomy elements are treated as nodes in an ontology tree. The algorithm output either validates the existing classification or adapt it according to an external classification taxonomy. As a component, it should be integrated to some KM application that makes use of the probe’s output. The current implementation saves all the output in a XML file, what proves that it can be connected to any application that follows the avestruz framework output format classes.

198

João Alfredo Pinto de Magalhães and Carlos José Pereira de Lucena

Fig. 2. Instantiation Process by Layers

Semantic Probe is currently being used to evaluate how the abundance of metainformation in documents affects the classification process. • KM Probe is a variation of the Semantic Probe, except for the fact that the documents processed are in HTML. So, its classification algorithm is not as sophisticated because there is no abundance of meta-information. • Webclipper is an instantiated application that searches for news over the internet. Everyday, it visits all the new web-pages from a set of three hundred selected information sources (portals, online magazines, online newspapers, etc.), what means approximately sixty-thousand visited pages each day. Webclipper’s classification algorithm is logic-based: its input consists of a list of classification classes, each one containing a set of rules that are applied to each news. If one of the rules is validated against a news document, the document is included in a personalized report automatically generated. We generate one report for each classification class, which is either sent by e-mail (to a configurable set of e-mails, where each classification class has its separate set) or saved in local file, that can be, for example, placed in a web server. The report format is also flexible, and we currently use HTML. • Site Seeker is a simple application that works just like an online search engine, searching for combination of words in web-pages, with one difference: there is no database involved, the search is directly on the web-page. Because of that, it is suitable for small sites.

Using an Agent-Based Framework and Separation of Concerns

199

5. Ongoing Work We are currently working on two independent ways: • Improving existing instances and creating new ones in order to validate our twostep development approach. We are improving the Semantic Probe, to make it compatible with other markup proposals, and improving the classification algorithm implemented in Webclipper. • Trying to generalize our two-step approach for other domains. We believe that the more the concerns related to the multi-agent system and to the agents are independent from the final system functionalities, the more it is possible to separate the generation of the system in one step dedicated to the agent technology, and another dedicated to the system itself.

6. Relevance to Agent Mediated Knowledge Management Anyone that belongs to the AMKM development/research team has always to deal with concerns coming from both Agent and Knowledge Management Technologies. This implies that pure Knowledge Management developers/researchers are excluded from the benefits that multi-agent systems bring to software development. By separating the concerns the way we proposed and implemented in this work, it is possible to let pure knowledge management researchers to deal only with their issues, leaving the platform concerns to somebody else. Another relevant contribution is the creation of an agent-based tool that is suitable for researches on how the abundance of meta-dada in marked-up documents for the semantic web should affect the classification process. By creating a framework where the algorithm can be changed in an easy fashion, a knowledge management researcher has a powerful tool to run extensive tests. Yet another contribution is the possibility of creating an agent-based knowledge miner, that can be defined as a special web-crawler that actually understands the documents it is crawling. This special web-crawler can be used, for example, to gather content for an intra-semantic web, playing the role of the web wrappers agent society generated in [8]. An enhancement to such a tool is using an interpreted-based language for parsing documents (like the one proposed in [9]), allowing the user to change some classification rules for extraction of document parameters.

References [1] W. Pree, “Design Patterns for Object-Oriented Software Development”, Addison Wesley, 1995 [2] Tarr, P.; Ossher, H.; Harrison, W.; and Sutton, S.M. 1999. N Degrees of Separation: MultiDimensional Separation of Concerns. In Proceedings of the International Conference on Software Engineering (ICSE'99).

200

João Alfredo Pinto de Magalhães and Carlos José Pereira de Lucena

[3] Zambonelli, F.; Jennings, N.; Omicini, A.; Wooldridge, M. 2001. Agent-Oriented Software Engineering for Internet Applications. In Coordination of Internet Agents. [4] Silva, O.; Garcia, A.; Lucena, C. J. 2001. A Unified Software Architecture for SystemLevel Dependability in Multi-Agent Object-Oriented Systems. 7th ECOOP Workshop on Mobile Objetcts Systems. [5] Gamma, E. 1995. Design Patterns: Elements of Reusable Object-Oriented Software. Addison Wesley. [6] Helfin, J. D.; Hendler, J.; Luke, S. 1999. SHOE: A Knowledge Representation Language for Internet Applications, Technical Report, TR-99-71, Science Computer Department, University of Maryland.Pree, W. eds. 1995. Design Patterns for Object-Oriented Software Development. Addison Wesley. [7] Buranarach. M. 2001. The Foundation for Semantic Interoperability on the World Wide Web. Ph.D. diss. Science Computer Department, University of Pittsburgh. [8] Cao, T. D.; Gandon, F.: Integrating External Sources in a Corporate Semantic Web Managed by a Multi-Agent System. In Proceedings of the Agent-Mediated Knowledge Management. AAAI Spring Symposium (2003), 123-130 [9] Krueger, W.; Nilsson, J.: Oates, T.; Finin, T.: Automatically Generated DAML Markup for Semistructured Documents. In Proceedings of the Agent-Mediated Knowledge Management. AAAI Spring Symposium (2003), 131-135

Modeling Context-Aware Distributed Knowledge Jorge Louçã UNIDE - ISCTE Instituto Superior de Ciências do Trabalho e da Empresa Av. das Forças Armadas, 1649-026 Lisboa, Portugal [email protected]

Abstract. This paper presents a multi-agent model to support decision-making in organizations. The model is characterized by being interactive, distributed, and incremental and by the use of cognitive maps to represent the knowledge of decision-making actors. The main proposition is to consider the context of concepts belonging to cognitive maps in a way that it represents agent’s mental states, allowing some kind of inference. To do so, context is conceptualized in cognitive maps, defining agent’s mental states from concepts being causally related to their context.

1 Introduction This paper proposes a multi-dimensional reasoning process in a multi-agent system. Software agents are used to support decision making in organizations, representing the knowledge of the actors that participate in the decision making process. In the model, cognitive maps are used as instruments to represent dispersed knowledge sources. This kind of cognitive model is used to compose a collective solution to a goal through a distributed and incremental process, based on agent’s interactions. Rational relations between agent’s mental states are mapped during agent’s internal reasoning processes. Finally, the emergence of collective knowledge, where interactions give rise to some kind of organizational culture, is represented in the artificial agent’s cognitive maps. The main proposition of this paper is to consider the context of concepts belonging to cognitive maps in a way that they can represent mental states, allowing some kind of inference. To do that, the philosophical positioning of functionalism is here assumed, aiming to model relationships (some kind of “functional roles”) between agent’s mental states. Philosophy and artificial intelligence both try to understand, in a physical world, all kinds of perception, action, intelligence and consciousness phenomena. In particular, artificial intelligence is a domain where mental experiments have been conducted with a main goal: starting from a given conception of what can be the mind, controlled mental experiments simulate reasoning and they use it in software programs. Their main advantage remains on the possibility of refutation during the experimentation process. This approach is different from the one generally followed in philosophy, where a given reality is studied – in this case there is no artificial model to experiment. Artificial intelligence is inspired by a functionalist conception of the mind, characterized by the recognition, by a physical system, of a given functional organizaL. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 201-212, 2003.  Springer-Verlag Berlin Heidelberg 2003

202

Jorge Louçã

tion [4]. If this functional idea of the mind is fair, then experiments conducted by artificial intelligence can help us to understand the knowledge representation systems used in our minds. On other side, artificial intelligence can improve upon the understanding of folk psychological constructs such as mental states. Actually, the notions of intelligence and mental states are strongly connected. Intelligence can be seen as the capacity of problem solving and decision. Mental states are representations of the world. There is no problem solving and decision capacity without some representations of the world. These representations can be, in folk psychology terms, intentional states such as intentions, beliefs and desires. According to this point of view, artificial intelligence can be seen as a laboratory where new software architectures, starting from a specified idea of what is the mind, are conceived and tested [14]. Therefore, artificial intelligence experiments can be seen as a way of really doing philosophy, because they search the conditions that make possible cognition in general. I believe that this position is according to the computational ideas that support multi-agent systems. In a multi-agent system we consider the hardware as the brain of an artificial agent, the software as its mind, and following this kind of parallel, intentional agent’s mental states can be defined by the roles they play in the system. A software agent is characterized by its autonomy regarding the user, by its proactiveness – it acts to achieve its goals – and by its intentionality [16]. To represent agent’s intentionality, I make use of some mentalistic notions found in folk psychology, such as beliefs and desires, as they are described for human behavior. Folk psychology allows us to make conclusions from mental states using assumptions. An intentional agent has beliefs, desires and, in a generic way, different kinds of mental states. The folk understanding of mental states has been presented as a theory of mind with an interesting operational content, to study the role of our own mental states in our behavior [28]. According to this approach, the understanding of internal mental states, and their internal cognitive mechanisms gives individuals the capacity to predict and explain their own behavior. Folk Psychology permits also the manipulation of knowledge some kind of data structures representing mental states, which mediates between our observation and our predictions or explanations. Actually this point of view acts as a functionalist theory, identifying mental states in terms of their causalfunctional relations. In a recent research [22, 23] I have proposed an inter-disciplinary approach concerning decision-making in human organizations, cognitive mapping and interaction between intelligent artificial agents. Multi-dimensional reasoning processes were modeled as multi-agent systems. I aimed to process automatically some mental faculties of individuals and groups. To do so, causal cognitive maps [2] were used as instruments to support collective reasoning. Those kinds of cognitive models were used to represent agent’s mental states and to compose a collective solution to a goal through a distributed and incremental process, based on agent’s interactions. I propose now to map rational relations among an agent’s mental states and to use this mapping to study the emergence of collective knowledge, where interactions give raise to some kind of organizational culture. This document is organized as follows. The next section presents the functionalist idea of contextual mental states, from which is based this research. After that, the third section presents the domain of cognitive mapping. The fourth section concerns the main proposition of the article, e.g. to consider the context of concepts in a way that they can represent mental states, allowing some kind of inference. The document concludes with the discussion of some perspectives of research in the domain.

Modeling Context-Aware Distributed Knowledge

203

2 The Functionalist Idea of Contextual Mental States The philosophical positioning assumed in here is that of functionalism. The philosophical doctrine of functionalism holds that ordinary people understand each common mental state descriptor to pick out a distinctive “functional role”, i.e., a set of causal – functional relations as stimulus inputs, behavioral outputs, and other mental states [12]. Two main subjects have been studied regarding functionalism: intentionality and consciousness. I will focus on the first one. To explain where intentionally come from, the base idea of functionalists stands that mental content is identified with causal-functional roles. In general terms, functional analysis decomposes mind in parts each time smaller until they are uniquely reactive. These components belong to a net of mechanical capacities, where the intelligence of the system is in the interaction of its components [4]. Several theories were proposed based on this model, as for instance the Functional Role Theories of Content. According to these theories, a network of inference relationships among states defines mental states. An example of this approach is the Global Workspace Theory [3], where mental states are defined in terms of their functions. The main characteristic of the Global Workspace Theory is that a mental state is activated (e.g., it is a conscious representation) when its message is broadcasted to the whole system. Then, according to the nature of the mental state, some specific receptors will process the message. These receptors are working memories acting in parallel, composing a distributed control structure with their interactions [10]. This picture of the mind, a collection of intercommunicating subsystems where reasoning is done through a set of messages posted in a large blackboard for all cognitive subsystems to read, it is far from being intuitive for most people [12]. Another problem of this approach is its holistic dimension – the content of every belief depends on the content of every other belief in the blackboard. To reduce this holistic dimension to some pertinent and applicable dimension – to distinguish mental states from the hole – the Causal Covariance Theory of Content [1] proposes that mental states get their content by being causally related only to what they are about (e.g., to those mental states belonging to its own specific context). This idea is in here adopted in a general way, to define and operate mental states. To understand this propositions let’s previously present cognitive maps, descriptive cognitive structures used by psychologists to represent the decision making process in organizations.

3 Cognitive Mapping From an epistemological point of view, the decision making process can be studied through an individual perspective, concerning methodological individualism, or through an organizational perspective, relating to holism. The analysis of a behavior or situation in its context is the holistic approach used in social sciences [8]. Cognitive mapping facilitates the adoption of this kind of approach, allowing the representation of knowledge in a holistic perspective, taking into account individual knowledge of deciders in the context of collective decision making in organizations. A

204

Jorge Louçã

cognitive map is a graphical representation of the behavior of an individual or a group of individuals, concerning a particular domain. This kind of graphical representation can help the understanding of different points of view and can evidence conflicts between deciders [2]. A cognitive map is composed by concepts (representing things, attitudes, actions or ideas) and links between concepts. Those links can represent different kinds of connections between concepts, such as causality or influence.

t6 - Create propositions

+

e1 - Age of employees : 30 / 35

+

(-,++)

e2 - Employees notion of "group"

t1 - Interact

(- -,++)

(+,++) b2 - Inovate working processes

++

t7 - Research to improve working processes

++ t8 - Invest in R & D

e3 - Employees thrust

t2 - Internal debates

t9 - Professionnal learning

+ ++

b1 - Adapt employees to changes

t3 - Accept suggestions

t4 - Periodical reunions

+

(-,+)

+

+

+

--

t5 - Dialogue

e4 - Resistance to change working habitudes

Fig. 1. Example of a cognitive map composed by three different kinds of concepts – tasks, goals and states of the world [22]

The main interest of cognitive maps it’s their reflexive character, allowing people to became conscious of their implicit knowledge, through the visualization of all direct and indirect links between concepts. We each construct our private versions of reality and deal only with those constructions, which may or may not correspond to some real world [21]. Cognitive maps are used, mainly by psychologists, as data structures to represent knowledge and to make behavioral analysis in what concerns decision-making in organizations. According to Karl Weick and others, organizations can be seen, at some abstraction level, as systems of construction and interpretation of reality [31] [20]. Following this approach, cognitive maps can be employed at an individual level, to represent individual viewpoints, and at an institutional level through the use of collective cognitive maps. Generally, this kind of cognitive models facilitate the analysis of the graphically represented ideas and lines of thought, facilitating communication inside a group supporting discussion and negotiation between the elements of the group having different points of view. Cognitive maps can also be used to detect conflict situations between deciders. This cognitive map implements a concept typology similar to the one proposed in [6], regarding knowledge representation in an organization. According to this typol-

Modeling Context-Aware Distributed Knowledge

205

ogy, tasks are concepts that describe actions or attitudes (for example, to interact, to make reunions, to dialogue, to invest in R&D). Goals must be achieved by doing tasks (for instance, to achieve the innovation of working processes or to accomplish the adaptation of employees to changes). States of the world represent things or qualities of the environment. All those concepts, tasks, goals and states of the world, are connected by influence links. Influences can be very negative (--), negative (-), positive (+) or very positive (++). For instance, in the example above the link between tasks t8 – Invest in R&D and t7 – Research to improve working processes is (++), meaning that t8 has a very positive influence in t7. On another hand, t7 has a very positive influence on the achievement of the goal b2 – Inovate working processes. A particular case is represented by links representing simultaneously different qualities, as the case between the task t9 – Professional learning, connected to e4 – Resistance to change working habitudes by a (-,+) link. This kind of double links represent the existence of different opinions concerning the nature of the link. In this example, the cognitive map evidences two points of view: one standing for a negative influence between t9 and e4 and another saying that there is a positive influence between those two concepts. Several methodologies in psychology are used to compose cognitive maps, including two main alternatives to extract and represent knowledge from individuals: the phenomenological and the normative ways. The first alternative considers the subjective dimension of behavior - it must be the individual by itself to compose its own cognitive map. The normative methodology stands for the use of observers, specialized in extracting concepts and links between concepts from written texts and oral interviews. These observers are normally psychologists [8].

4 Cognitive Maps Standing for Mental States In a previous research I have proposed a multi-agent model based on multidimensional reasoning processes [22, 23]. In this proposition each artificial agent supports the decision of an individual participating in the collective decision-making process in the organization. Causal cognitive maps are used to represent knowledge of those deciders. Artificial agent’s knowledge is used to compose a collective solution to a goal, through a distributed and incremental process based on agent’s interactions. This distributed and incremental process is represented in Fig.2. When an actor requests its agent to propose a solution to a goal, this one uses the set of concepts represented in its cognitive map to compose the solution. Throw a reasoning process, which is an extension of the Negative-Positive-Neutral Logic [34], named NPNe Methodology [22], the agent become aware of the tasks, goals and states of the world that influence, directly or indirectly, the achievement of the goal. Then the agent

206

Jorge Louçã

Individual solution to a goal

Goal allocation : results from other reasoning processes

Solution to an allocated goal

Solution to an allocated goal

Solution to an allocated goal

----------------Cognitive Map Aggregation-------------------------

Collective Solution Matr ix co mpo sitio n

Interaction Matrix

licts onf ce c den Evi

Collective Solution with conflicts

Fig. 2. An interactive, distributed and incremental resolution process [22]

composes, with those concepts, its individual solution to the goal, represented by a partial cognitive map. The second step concerns the allocation of sub-goals (those goals that belong to the previous solution) to other agents in the system. As represented in Figure 2, each agent that receives an allocation message including a goal, starts its own reasoning process to the sub-goal and, in return, responds with a solution to the sub-goal. This distributed reasoning process allows representing several points of view concerning the sub-goals. Next, agents aggregate their partial solutions in a collective solution, throw the NPNe Methodology of Aggregating Cognitive Maps detailed in [23]. This is done, mainly, throw de composition of the interaction matrix, where are represented links between all concepts that belong to the different partial solutions. The matrix represents all links between to given concepts, including conflicting points of view. Then, according to the NPN Logic [34], only the most acute opinions are considered to compose the collective solution. For instance, lets consider the concepts referred in the previous section, t9 – Professional learning and e4 – Resistance to change working habitudes. Supposing that, after goal allocation, the system obtains three different viewpoints: +, - and --. In this case the link between those two concepts would be (+,-), evidencing a conflict in the organization. This way, the collective reasoning

Modeling Context-Aware Distributed Knowledge

207

mechanism will detect and evidence conflicts in a collective solution, graphically represented in the form of a cognitive map, allowing a clear discussion and negotiation, in the organization, about those conflicts. An extension of the research previously described is the proposition to consider cognitive maps composed, on one hand, by concepts and by causal links between those concepts, in a strictu senso way [30], and on the other hand to consider the context from where we can take the assumptions allowing some kind of inference. The idea of context is fundamental to clarify the collective meaning of a concept.

t6 - Create propositions

+

e1 - Age of employees : 30 / 35

+

(-,++)

e2 - Employees notion of "group"

t1 - Interact

(- -,++)

(+,++) b2 - Inovate working processes

++

t7 - Research to improve working processes

++ t8 - Invest in R & D

e3 - Employees thrust

t2 - Internal debates

t9 - Professionnal learning

+ ++

b1 - Adapt employees to changes

t3 - Accept suggestions

t4 - Periodical reunions

+

(-,+)

+

+

+

--

t5 - Dialogue

e4 - Resistance to change working habitudes

Fig. 3. Example of a scheme included in a cognitive map [23]

I stand for a pragmatic constructivist approach [19] that allows us to understand context following a circular cognitive process that departs from several contextual hypotheses to, interacting with the user, arrive to contexts defining the agent mental states1. This way, we can define mental states from cognitive maps by getting its content (e.g. sub-maps) from the concepts being causally related to their context. More precisely, a mental state is represented by a concept and its context. According to Krippendorf, meaning connects the features of an object and features of his context into a coherent unity [17]. So, we can say that text unifies concepts in its mental states. In cognitive mapping terms we have that: (2) concepts can’t be understood without context, (3) the context of a concept is composed by concepts that influence and that are influenced by the concept, and (3) each concept is coupled to its context, which can be called a scheme [5].

1

A pragmatic adoption of this approach is detailed in [Louçã,02a].

208

Jorge Louçã

Given this notion of scheme as a concept and its context, and the notion of context as the linked concepts that influence and that are influenced by the named concept, the mapping of mental states can take the graphical form represented in Figure 4.

Fig. 4. Mapping mental states that participate in the internal resolution process of an agent

During the internal resolution process of an artificial agent, those mental states that have some concept matching a goal allocated to the agent or matching a task to be achieved by the agent are activated and participate in the aggregation of the local solution. This is a relational perspective of the reasoning process, where the relevance of a mental state is a function of its relation with the other mental states of the system. The agent’s local solution, in the form of a cognitive map, is like a higher-order mental state, representing the result of the conjunction of the activated mental states. On another hand, agents interact by communicating schemes concerning a given situation. As a scheme represents the meaning of a given concept, communicating schemes influence agent’s mental states. This influence represents the spreading of collective knowledge, characterized by schemes shared by the elements of the organization. The mechanism that generates collective knowledge is as follows (see figure 5): when an agent communicates a scheme to another one, the receiver verifies if the concepts belonging to the scheme are comparable (e.g., if they are the same or similar) to the concepts belonging to its own cognitive map. If the concepts are comparable, it means that both agents have points of view about the subject of the scheme. These opinions can be alike or distinct (they can represent a conflict). In this last case, the graphical interface of the software system proposes a discussion between the users that correspond to the artificial agents.

Modeling Context-Aware Distributed Knowledge

209

Structure Cognitive

Modèle Social

Modèle du Domaine

Structure Cognitive

Module d'Analyse du Domaine

Structure de Raisonnement et de Communication

Module de Résolution

Scheme

Module d'Analyse Sociale Module de Communication

e1

- Age du personnel : 30 / 35

Transfert de données

Flux de Contrôle

e2

t1

e3

- Cohésion du personnel

+

- Interagir

- Confiance du personnel

t9

- Faire formati professionnelle

+ b1

- Adapter le personnel au changement

++

t3

If concepts are comparable, then the graphical interface supports discussion (and argumentation) between individuals, to converge to a common scheme

+ +

--

Structure de Raisonnement et de Communication

Module de Résolution

+ t5

- Dialoguer

Module d'Analyse Sociale Module de Communication

-Transfert de données

- Accepter suggestions

t4 - Faire des réunions périodiques

Module d'Analyse du Domaine

+

++

t2 - Faire réflexion interne

Modèle Social

Modèle du Domaine

Flux de Contrôle

e4 - Résistances à changer d'habitudes de travail

Différents niveaux de détail dans la visualisation d’une solution collective avec mise en évidence de conflits

If concepts are not comparable, then “CM aggregation”

Fig. 5. Convergence to a common scheme in STRAGENT

The aim of this discussion is to converge to a common scheme. Finally both agents actualize their cognitive maps according to the result of the discussion, leading to the emergence of a collective point of view about the subject. On the other hand, if the concepts belonging to the communicated scheme are not comparable to those existing in the cognitive map of the receiver, this means that this agent don’t have an opinion about the given subject. In this case the receiver improves its knowledge through the aggregation of the scheme to its own cognitive map. In conclusion, collective knowledge is mainly the result of a dynamic process of interaction (argumentation, negotiation) in the interior of the group. The quality of the results and the cohesion of a group depend largely on the quantity and quality of its collective knowledge [5]. This approach has several advantages: (1) the expressability in verbal behavior of a plan of actions, (2) the transmission of the content of some mental state to some other mental state in the system, and (3) a higher-order state which reflects on the target state [24]. These propositions above were tested in STRAGENT, a distributed software system to support decision-making in human organizations (Figure 5). This prototype was applied in an industrial enterprise in the domain of telecommunications and electronics, to support the collective decision making process. Cognitive maps were design from documents and interviews. The main goal of this application was to model collective discussions, to represent actor’s knowledge, to support conflict resolution and to identify and understand the interactions between agent’s mental states during the decision making process.

210

Jorge Louçã

5 Related Work This research can be compared with other propositions about intentional systems and mental states concerning artificial agents, as well as bayesian and semantic networks concerning knowledge representation. Finally a reference is made to previous systems that have already contributed with some notions used in this research, such as the aggregation of cognitive maps or the use of an interaction matrix. The notion of intentional system was originally proposed by Daniel Dennet, to whom those systems are defined by beliefs and intentions [9]. The notion of intentional agent is largely used in multi-agent systems literature. Mental states such as beliefs, desires and intentions are studied by what is named the BDI approach [26], [11], proposing and operational model of agents, founded on a logical formalism. The notion of artificial agent presented in this research is also driven by beliefs (concepts and links between those concepts) and intentions (schemes to attain goals) – e.g., agents have an intentional attitude. Nevertheless, this proposition has the advantage of representing mental states using cognitive maps, a non-formal tool really used in organizations and by psychologists to represent knowledge. The use of cognitive maps to represent knowledge can be put side by side with an artificial intelligence approach the uses a graphical notation, named semantic networks [27]. Such as cognitive maps, semantic networks represent knowledge through nodes connected by arcs. Nevertheless, in those networks, nodes are hierarchically typed, with derivation, according to the generality level of the nodes. Those systems are mainly used to classify or to group knowledge in natural language systems. On another side, cognitive mapping concerns less restraint notions, which do not need some particular typing – it is a general methodology, and one of its strengths is precisely the adaptability to a large variety of domains. The same argument can be used when comparing cognitive maps with another graphical knowledge representation: Bayesian networks. Actually, those two tools have already been associated to define the qualitative probabilistic networks [29], a sort of cognitive mapping with causal probabilistic links, allowing bayesian reasoning in cognitive maps. However, the use of the original version of cognitive maps has the advantage of simplicity – cognitive maps can represent a larger domain of situations, it is a tool used by psychologists and allows qualitative reasoning. The POOL2 system, proposed by [32], composes collective maps through the aggregation of individual cognitive maps. This system is, as STRAGENT, based on the NPN Logic. POOL2 doesn’t incorporate the notion of interaction between artificial agents. In A-POOL – Agent-Oriented Open System Shell [33] the same authors use cognitive maps to represent artificial agents knowledge. The communication is done through the exchange of partial cognitive maps and the purpose of interactions is to compose an organisational map. The most recent evolution of this system includes the propagation of numerical values [34]. However, the use of quantitative inference is far from the qualitative spirit of cognitive mapping. In the line of thought of APOOL, [7] proposes a method of causal reasoning adapted to multi-agent negotiation. Chaib-draa introduces the notion of interaction matrix to represent several points of view concerning the same subjects. Nevertheless, the conflict detection is not dynamic along interactions, it’s done at a given moment – this model isn’t adapted to

Modeling Context-Aware Distributed Knowledge

211

artificial agents that dynamically and continuously adjust their knowledge to a changing environment.

6 Conclusion I propose a model of multi-dimensional reasoning in a multi-agent system. In this model, cognitive maps are used as instruments to represent agent’s mental states, and to compose a collective solution to a goal through a distributed and incremental process, based on agent’s interactions. Rational relations between agent’s mental states are mapped during its internal reasoning processes. Finally, the emergence of collective knowledge, where interactions give rise to some kind of organizational culture, are represented in agent’s cognitive.

References 1. Allen, C.: Lectures on the nature of mind, http://www-phil.tamu.edu/~colin/Phil251/, (2002) 2. Axelrod, R.: Structure of decision : the cognitive maps of political elites, Princeton University Press, (1976) 3. Baars, B.: A Cognitive Theory of Consciousness, Cambridge: Cambridge University Press, (1988) 4. Block, N.: The Mind as the Software of the Brain, in An Invitation to Cognitive Science, Osher, D. and Sternberg, S., ed., MIT Press, (1995) 5. Bougon, M., Komocar, J.: Façonner et diriger la stratégie. Approche holistique et dynamique, in Cossette, P., ed., Cartes Cognitives et Organisations, Laval, Canada: Les presses de l'Université Laval (1994), 57-81 6. Carlsson, C., Walden, P.: Cognitive maps and hyperknowledge support system in strategic management, Group Decision and Negociation, nº 6, (1996), 7-36 7. Chaib-draa, B.: A relational modeling of cognitive maps, in Proceedings of the Canadian Society for Computational Studies of Intelligence on Advances in Artificial Intelligence (AI-98), LNAI, Vol.1418, Springer-Verlag, (1998), 322-333 8. Cossette, P.: La carte cognitive idiosyncrasique. Etude exploratoire des schèmes personels de propriétaires-dirigeants de PME, in Cossette, P., ed., Cartes Cognitives et Organisations, Laval, Canada: Les Presses de l'Université Laval, (1994), 113-115 9.. Dennett, D.: The intentional stance, MIT Press, (1987) 10. Franklin, S.: Conscious Software: A Computational View of Mind, in Soft Computing Agents: New Trends for Designing Autonomous Systems, Loia, V. and Sessa, S. eds., Berlin: Springer (Physica-Verlag), (2001), 1-46 11. Georgeff, M., Pell, B., Pollack, M., Tambe, M., Wooldridge, M.: The belief-desireintention model of agency, in Proceedings of Agents, Theories, Architectures and Languages – ATAL’99, (1999) 12. Goldman, A.: Consciousness, Folk Psychology, and Cognitive Science, in Consciousness and Cognition, nº2, (1993), 364-382 13. Lawless, W.F.: A Quantum Approach to Knowledge Fusion on Organizational Mergers, 2003 Spring Symposium on Agent-Mediated Knowledge Management, Technical Report SS-03-01, Stanford University, (2003), 71-73 14. Miguens, S.: Dennett e a IA (in portuguese), in Intelectu, nº 9, Lisboa, (1999)

212

Jorge Louçã

15. Novak, J., Wurst, M., Fleischmann, M., Strauss, W.: Discovering, Visualizing and Sharing Knowledge through Personalized Learning Knowledge Maps, 2003 Spring Symposium on Agent-Mediated Knowledge Management, Technical Report SS-03-01, Stanford University, (2003), 101-108 16. Kampis, G.: The Natural History of Agent, in Agents Everywhere, Gulyás, L., Tatai, G. and Váncza, J. eds., Springer, Budapest, (1999), 24-48 17. Krippendorf, K.: On the Essential Contexts, of Artifacts or on the Proposition that Design Is Making Sense (of Things), in Design Issues, vol.5, nº2, (1989), 9-39 18. Langfield-Smith, K.: Exploring the need for a shared cognitive map, in Journal of Management Studies, Vol.29, No.3, (1992), 349-368 19. Lissack, M.: Complexity: the Science, its Vocabulary, and its Relation to Organizations, in Emergence, Vol.1, No.1, (1999) 20. Lissack, M., Gunz, H.: Managing Complexity in Organizations: A View in Many Directions Quorum Books, Westport, (1999) 21. Lissack, M., Ross, M.: The Next Common Sense, London: Nicholas Brealey Publishers, (1999) 22. Louçã, J. : Cartographie Cognitive, Réflexion Stratégique et Interaction Distribuée : une Approche Multi-Agent (in french), PhD Thesis, Université Paris IX Dauphine and Faculdade de Ciências da Universidade de Lisboa, (2000) 23. Louçã, J.: Cognitive Mapping in a Multi-Agent System for Strategic Planning, in 12th Euro Conference on Decision Support Systems, Bruxels, Belgium, (2002) 24. Louçã, J.: Autonomous agents as systems of emerging consciousness, in Sixt Conference of the Association of the Scientific Study of Consciousness - Consciousness and language: reportability and representation in humans and animals, Barcelone, Spain, (2002) 25. Louçã, J.: Modeling intentional agents: contextual cognitive maps standing for mental states, in Conference on Intentionality: Past and Future, Miskolc, Hungary, (2002) 26. Rao, A., Georgeff, M.: BDI agents : from theory to practice, ICMAS-95, (1995) 27. Sowa, J.F. (Ed.): Principles of semantic networks - explorations in the representation of knowledge, Morgan Kaufmann Publishers, (1991) 28. Stanford Encyclopedia of Philosophy: Folk Psychology as a Theory, http:// plato.stanford.edu/entries/folkpsych-theory/ , (2002) 29. Wellman, M.: Inference in cognitive maps, in Mathematics and Computers in Simulation, n°36, (1994), 137-148 30. Weick, K.: The Social Psychology of Organizing, 2nd Edition, Reading, MA: AddisonWesley, (1979) 31. Weick, K., 1995, Sensemaking in Organizations, Thousand Oaks, CA: Sage Publications, (1995) 32. Zhang, W., Chen, S., Wang, W., King, R.: A cognitive-map-based approach to the coordination of distributed cooperative agents, in IEEE Transactions on Systems, Man, and Cybernetics, Vol.22, Num.1, January/February, (1992) 33. Zhang, W., Wang, W., King, R.: A-POOL : an agent-oriented open system shell for distributed decision process modelling, in Journal of organizational computing, Vol.4, Num.2, pp.127-154, (1994) 34. Zhang, W.: NPN fuzzy sets and NPN qualitative algebra : a computational framework for bipolar cognitive modeling and multiagent decision analisys, in IEEE Transactions on Systems, Man, and Cybernetics – Part B : Cybernetics, Vol.26, No.4, (1996)

Discovering, Visualizing, and Sharing Knowledge through Personalized Learning Knowledge Maps Jasminko Novak1 , Michael Wurst2 , Monika Fleischmann1 , and Wolfgang Strauss1 1

Fraunhofer Institute for Media Communication, MARS Exploratory Media Lab, Schloss Birlinghoven, D-53754 Sankt Augustin, Germany [email protected] 2 University of Dortmund, Artiﬁcial Intelligence Dept. D-44221 Dortmund, Germany [email protected]

Abstract. This paper presents an agent-based approach to semantic exploration and knowledge discovery in large information spaces by means of capturing, visualizing and making usable implicit knowledge structures of a group of users. The focus is on the developed conceptual model and system for creation and collaborative use of personalized learning knowledge maps. We use the paradigm of agents on the one hand as model for our approach, on the other hand it serves as a basis for an eﬃcient implementation of the system. We present an unobtrusive model for proﬁling personalised user agents based on two dimensional semantic maps that provide 1) a medium of implicit communication between human users and the agents, 2) form of visual representation of resulting knowledge structures. Concerning the issues of implementation we present an agent architecture, consisting of two sets of asynchronously operating agents, which enables both sophisticated processing, as well as short respond times necessary for enabling interactive use in real-time.

1

Introduction

The basic point of departure of our work can be related to the approach which argues that knowledge consists largely of a very personal, diﬃcultly articulable and partly unconscious component, usually referred to as implicit or tacit knowledge [1]. Accordingly, a key to the communication and shared use knowledge, lies in the transformation of implicit knowledge and hidden assumptions to explicit structures perceivable und usable by others. This recognition leads us to the following question: How can existing, but not yet explicitly formulated knowledge structures, of a given community or a group L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 213–228, 2003. c Springer-Verlag Berlin Heidelberg 2003

214

Jasminko Novak et al.

of experts be discovered, visualized and made usable for cooperative discovery of knowledge in heterogeneous information pools? In formulating a practical approach to addressing these issues we introduce the following constraints and deﬁnitions. We relate the notion of knowledge discovery to supporting the discovery of semantic contexts and relationships in an information pool which is either 1) too big or too fast growing to be scanned and categorized manually, or 2) consists of too heterogeneous content to impose one ﬁxed categorization structure, or 3) serves diﬀerent user groups with heterogeneous interests. This deﬁnition immediately reﬂects the relevance of our approach and research challenge to practical applications. On one hand these conditions apply today to a vast range of Intranet/Internet portals in their own right. On the other hand, they can also be generalized to the problem of connecting existing information sources on the Internet in a way that allows semantic exploration of information and creation of both personalized and shared structures of knowledge. The paradigm of agents is a very promising approach to overcome some of the problems connected with heterogeneity on the side of the data sources as well as on the side of the users. As agents should operate autonomously and can be loosely coupled, they are well suited for the integration of distributed heterogeneous data sources, building unifying wrappers around them. This becomes especially beneﬁcial, if agents can learn to extract information from an information source automatically (see for example [2]). On the side of the users, the paradigm of Personal Information Agents oﬀers a way to encapsulate the interests, the knowledge as well as the preferences of individual users. This is especially important in a system serving diﬀerent groups of users. While agents in some systems mainly ﬁlter and distribute information (as in [3] for distributing Knowledge Discovery results) they are also very well suited for the task of capturing the (tacit) knowledge of users, as to make it accessible to others. Therefore Personal Agents can take the role of mediators between users and information sources, as well as between users among each other (see also [4] and [5]). Based on the paradigm of “Agent Mediated Knowledge Management”, we present a model for expressing implicit knowledge structures of individuals and groups of users and for using this as a means for semantic navigation and discovery of relationships in heterogeneous information spaces. We will show, how this model enables the implicit, as well as the explicit exchange of knowledge between users through intelligent agents. In particular, we discuss a model for unobtrusive generation and proﬁling of personalized user agents based on eﬀects of user interaction with information and a related model for visualising and navigating resulting knowledge structures. Furthermore we present an agent architecture consisting of two sets of asynchronously operating agents. This architecture enables us to perform sophisticated data and interaction analysis, without loosing the property of short respond times essential for interactive work in real-time.

Discovering, Visualizing, and Sharing Knowledge

2

215

Personalized Learning Knowledge Maps

In order to develop a working solution for capturing and visualizing implicit knowledge structures of human users based on their interaction with information, two basic problems need to be solved: 1. a context for user actions has to be created in order to be able to interpret the meaning of user interaction with information items. The lack of a clear interaction context is the main diﬃculty of general user-tracking and interaction-mining approaches such as [6]. 2. a form of visual representation has to be found that communicates to the user both the semantics of the information space in itself (content, structure and relationships) and relates this to the meaning of his actions. As a practical context for addressing these issues we take the process of information seeking and semantic exploration of a document pool. This can be understood as a process in which the users interaction with information both reﬂects their existing knowledge and produces new knowledge structures. In the concrete solution we develop a model of agents learning personalized knowledge maps. The notion of a knowledge map in our approach refers to the representation of information spaces in which the individual information items are not isolated but structured according to possible meanings and semantic relationships. This concept serves as a point of departure for both providing an unobtrusive context for interpreting user actions as well as for visualizing the resulting knowledge structures and exchanging them between users.

2.1

Capturing User Knowledge

The basic idea is to build agents, that provide the users with a semantically structured overview of a document pool as a basis for their exploration and interaction with information. The results of their interaction can then be taken as the basis for generating user-speciﬁc templates. These templates (personal maps) are the basis for generating and proﬁling personal information agents which can then automatically generate a semantically structured map of a document pool, in a way that reﬂects a users particular point of view. In our approach the generation of user-speciﬁc templates is based on a two-stage model. First the user is presented with an agent-generated knowledge map created by means of methods for autonomous machine clustering such as in [7], [8], [9], [10]. This map serves as an initial context and navigation guide for the users exploration of the document space. As she explores the information space, the user identiﬁes relevant documents and relationships between them which she can express by selecting individual items into personal collections and by (re-)arranging them according to her personal understanding of their meaning (e.g. by moving objects between groups,

216

Jasminko Novak et al.

creating new groups, adding labels and relationships). In this way the user creates a personal map as a natural result of her exploration of information. This template can now be learned by a personal information agent by means of methods for supervised learning. Having learned a user-speciﬁc template, the agent can semantically structure arbitrary information pools or dynamically classify unknown information items. 2.2

Visualizing the Knowledge Structures

The challenge for the visual representation of the knowledge maps is to develop a visual tool for both navigating a large information space as well as for discovering possible contexts and relationships between groups of items. This applies both to relationships uncovered by the machine analysis and those stemming from interpretation and knowledge of human users. To achieve this the two main elements of the knowledge map visualization are: the Content Map and the Concept Map.

Fig. 1. The content map

The Content Map provides an overview of the information space structured according to semantic relationships between information items. In the ﬁrst realization the Content Map visualizes clusters of related documents and oﬀers insight into implicit relationships between their content. This is the main context for users exploration and interaction with information. The Concept Map visualizes a concept-network that is extracted from the document pool and redeﬁned by the users. This provides both a navigation structure and insight into the criteria that have determined the semantic structuring in the Content Map. These criteria are a kind of semantic axes that deﬁne a given structuring out of a variety of possibilities. Since the personalized map templates have been produced by a user as an eﬀect of his interaction with information and can be dynamically applied to reﬂect his

Discovering, Visualizing, and Sharing Knowledge

217

Fig. 2. The concept map

point of view, they are a form of representation of the user’s knowledge that has previously not been expressed. Visualizing the personalized maps and the related concept structures, and making them available to other users is a way of making the users knowledge perceivable and available to others. Hence, our claim that this is a way of expressing a user’s implicit knowledge resulting out of his interaction with an information space, in a way, which makes it perceivable and usable by others.

2.3

Exchanging Knowledge

In our model, there are two major ways to enable the exchange of knowledge between users. Firstly, users can explicitly exchange knowledge maps they have created, secondly, information contained in personal maps can be analyzed implicitly (without the user being involved) and then be used to support the exploration and map editing process of other users. In chapters 4.2 and 4.3, we describe, how both of these possibilities are integrated in our system, the ﬁrst through a personal assistant to enable search in the set of knowledge maps, the second through interaction analysis used for learning personal maps.

2.4

Relationship to Related Work

The basic idea of generating user-speciﬁc templates and applying them for personalized structuring and ﬁltering of information has been previously realized in several diﬀerent ways. In one class of approaches the users have to express their preferences explicitly and as their primary task, such as by voting, preference proﬁling or initial selection of items from a given information pool (see [11] for an overview). One critical issue here is the bootstrapping problem: the available orientation for users initial identiﬁcation of relevant items in an information pool (which they are not familiar with) is based solely on already available proﬁles of

218

Jasminko Novak et al.

other users (e.g. [12]). A related problem is that of communicating the intention and meaning behind user choices that contributed to the creation of a given proﬁle to other users: the proﬁles themselves are typically neither explained, nor visualised, nor put in relation to the semantic structure of the underlying information pool. Another typical class of approaches attempts to analyze the users actions in form of click streams and navigation patterns on the web (e.g. [13], [6]). The critical issue here is the lack of a clear context for interpreting the meaning of users actions. In our approach both of these problems are addressed by introducing a system generated map as 1) a clear initial context for user actions, 2) a structure for semantic navigation in an unknown information pool, 3) form of visualising users personal knowledge structures in relation to the original information space. This approach also allows us to make the expression of personal points of view unobtrusive and not distracting from the users main task: that of discovering relevant information and internalizing it into knowledge. Furthermore, the personalized maps in our approach provide an easy and understandable way for communicating and sharing knowledge between diﬀerent users both through explicit selection of diﬀerent maps by the users themselves, as well as through implicit inference mechanisms of the agents that analyze the relationships between individual maps (Chapters 4.2, 4.3)

3

Agent-System Architecture

As already mentioned, our system consists of two diﬀerent kinds of agents (Fig 3). One group of agents is concerned with responding to user requests. These agents have to work very eﬃcient, as interactive work requires very short respond times. To achieve this, we use a second group of agents, which asynchronously preprocess data and store it in intermediate structures. These agents take much of the work load from the ﬁrst group of agents. Using this strategy we can use sophisticated and costly data and interaction analysis methods and even so have short respond times. In the following, we will roughly describe some of the systems components. 3.1

Data Preprocessing Agents

These agents allow the user to create a pool of documents by connecting heterogeneous data-sources. The user can either choose between readily available data sources or manually connect other structured data-sources (such as databases and semi-structured document repositories). This is supported by a dynamic data adapter for user-oriented semantic integration of XML-based semi-structured information. Preprocessing includes a text-analyzer for encoding semantic properties of texts into a vector space model, link&reference analysis, co-author relationships and the extraction of other properties.

Discovering, Visualizing, and Sharing Knowledge

Personal Information Agents

219

Visualization Agents

Shared Data Space

Data Analysis Agents

Online processing Offline processing

Data Preprocessing Agents Fig. 3. The Agent system structure

3.2

Data Analysis Agents

This layer contains agents for semantic processing of data and for interaction analysis. Interaction analysis processes the personal maps of all users in order to identify relations between objects (see 4.2). While preprocessing is performed only once for an object, interaction analysis is performed at regular intervals, as the set of personal maps changes.

3.3

Personal Information Agents

Personal Information Agents have three diﬀerent tasks. Firstly, they construct knowledge maps, based on unsupervised learning, allowing the user to inﬂuence this process by a set of options. We use Self Organizing Maps (SOM) for this purpose (see 4.1). Secondly, personal agents are able to learn a personal map, created by a user and to apply it to an individual object or a whole information pool. For this purpose, we use case-based reasoning, based on content and interaction analysis, as described in 4.2. The third task of a personal information agent is to provide its user with interesting maps of other users, enabling a direct exchange of knowledge between them (see 4.3).

3.4

Visualization Agents

The visualisation agents provide necessary post-processing of the data and of the interaction-analysis done by the personal information agents. They take care of collecting all necessary information from diﬀerent agents, needed to construct all the information layers of the Content Map and the Concept Map described in

220

Jasminko Novak et al.

the previous chapter. In a typical case, a personalised information agent delivers the logical map of documents grouped into clusters of related content, with basic parameters such as weight of document membership to a given cluster, typical members of each cluster etc. Based on the selected visualisation model, the visualisation agent then retrieves information stored by the data integration assistant and preprocessing agents, in order to ﬁll in additional information (e.g. titles, abstracts, term-document frequencies etc.) and compose all necessary information layers needed for a given visualisation.

3.5

Agent Communication and Coordination

We use two classical techniques for agent communication and coordination. The exchange of data between agents is realized as shared data space. The idea is, that on the one hand there are possibly several agents working on preprocessing in parallel. On the other hand, the preprocessing agents can provide data for the request processing agents asynchronously, without direct communication or coordination. Though within each group of agents, there is need for a tighter form of coordination. This is done by a simple event service based on XML and SOAP.

4

Personal Agents and Data Preprocessing

In this section, the personal agents used for automatically creating knowledge maps, for learning personal knowledge maps and for searching the set of knowledge maps from other users are described in more detail. Along with these agents themselves, the corresponding agents for preprocessing are described.

4.1

Clustering Documents Automatically Using Self Organizing Maps with Interactive Parameterisation

We use Kohonen’s self-organizing neural network ([7], [8]) to map the high dimensional word vectors onto a two dimensional map. As the vectors encode semantic properties of texts the map will position semantically correlated texts close to each other. The information describing the distribution of items and the measure of ”semantic similarity” between both individual items and groups of items (clusters) provides the basis for the visualization in form of the Content Map (Fig. 1, Fig. 4) In addition to the content map, a concept map is generated, which visualizes the relations between diﬀerent words (Fig 2, Fig 4). We employ an approach similar to that described e.g in [14] to build this map. The idea is to structure the words

Discovering, Visualizing, and Sharing Knowledge

221

by examining which other words appear in the context of a given word. The high dimensional context relations resulting from this are then mapped to a two dimensional space, again using the SOM. In this way we can create an initial set of concepts (words) that serve both as an explanation of the clustering and as a navigation structure. Our system provides the additional feature, that users can customize the aspects according to which the maps are generated by manually selecting a number of words on the concept map. The weights for these words in the vector space are increased making them the most important words. Then the mapping procedure is re-applied using these modiﬁed weights. In this way, by interactively exploring diﬀerent possible clustering variants, the users can develop an understanding of how the clustering works and what makes out the character of individual document groups. Moreover, they can develop an understanding of the overall semantic structure and relationships between groups of documents (e.g. topics, trends, representatives) and the concepts (words) that determine a particular semantic point of view. This allows semantic navigation across a document pool for identifying relevant pieces of information embedded in contexts and relationships from diﬀerent points of view. The discovered insights that are internalized by users as acquired knowledge are then reﬂected in their own personal maps.

4.2

Combining Content-Based and Collaborative Methods to Learn Personal Knowledge Maps

By creating a personal map, the user deﬁnes a set of classes. The idea of learning a personal knowledge map is to ﬁnd a function, which assign new objects to these classes automatically. After such a decision function has been found, a map can be applied to any single object or information source provided by the system. The question of whether an object can be reasonably assigned to any of the user deﬁned classes or not is to a signiﬁcant extent subject to individual preference. As a consequence, the system gives the user the possibility to interactively adjust the threshold of minimal similarity. If there is no object in the personal knowledge map to which the given document is at least as similar as deﬁned by this threshold, the object is assigned to the trash class. Otherwise the decision function is used to assign it to any of the user deﬁned classes. This allows the user to ﬁne tune the personalized classiﬁcation by exploring the inﬂuence of the threshold between two extremes: if the threshold is maximal then all objects are assigned to the trash class, if it is minimal all documents are assigned to some class and trash class is empty. As method to ﬁnd such a decision function that assigns documents to clusters we use Nearest Neighbor(e.g. [15]). This methods ﬁrst identiﬁes the most similar objects on the personal map for an object in question, and then performs a majority vote among them about the class to which to assign the object. This method oﬀers two important advantages in our context. The ﬁrst one concerns

222

Jasminko Novak et al.

eﬃciency and respond time, the second one concerns the problem, that the user usually provides only few training data. The idea is, that the similarity between objects can be pre-computed using sophisticated algorithms based on data and interaction mining. The query processing agent needs only some few access operation to the result matrix making it very eﬃcient. An outdated similarity matrix could make the result sub-optimal, though in most cases this wont aﬀect the performance, as similarities change only slowly. In the remainder of this section, we describe how the preprocessing agents build this matrix based on content and context analysis and how this helps us to deal with the problem of few training examples. Content analysis uses properties of items (word vectors, authors, etc.) to measure the similarity of these items. The idea of context analysis is the following: If two objects appear together in many user edited clusters, then we can assume, that these objects are in some way similar. This is a very interesting feature of our system, as items are not only rated by users, like in ”collaborative ﬁltering” systems, but are put into the context of other items. This is much more powerful, as usually an item is not interesting or relevant per se, but only relevant in a given context. It helps us to deal with the problem, that the user provides only few examples, as the personal maps of all users can be used to support the learning and application of a map, not only the one of the actual user. Both the content-based similarity and the context similarity are in a ﬁrst step calculated independently of each other. Content based similarity is a linearweighted combination of individual aspects. For context similarity we use the “Dice”- coeﬃcient: sim(x, y) = 2

|X ∩ Y | |X| + |Y |

were X is the set of clusters, which contain object x and Y is the set of clusters, which contain object y. Using this measure, clusters, which do not contain any of both objects, are not counted, which seems appropriate for the given case. Also co-occurrences get double weight, as we consider them as more important than single occurrences. The membership of clusters and objects to personal maps is not taken into account at all, as it is quite unclear, how objects on the same map, but in diﬀerent clusters are related. Beside the direct use of context similarity in the combination with content similarity, there is still another possibility to take advantage of the user interactions. As mentioned above several aspects describing the content of underlying documents are combined using a weighted linear sum. Now, to ﬁnd optimal values for this function, we can take the context similarity as prototypical similarity and use it to train a linear regression model (or even more sophisticated regression models). In this point our system also diﬀers from systems that seek association

Discovering, Visualizing, and Sharing Knowledge

223

rules [16], which perform a kind of context analysis too, but which do not analyse the content of the underlying objects and put it into relation to their context. The remaining question is, how content-based and context similarity should be ﬁnally combined into a single measure, preserving the advantages of both. The advantage of content-based similarity is, that it is always applicable and does not rely on user generated data. Though content-based similarity can lead to poor results, if the underlying objects are heterogeneous, e.g. make use of diﬀerent terminology or are even written in diﬀerent languages. On the other hand, using context similarity, we avoid these problems completely. The disadvantage of context similarity is however, that if only few users add a given object to their maps or if the contexts, in which it appears, diverge, we do not get any reliable evidence on the similarity of this object to other objects. Consequently, we use a statistical test (chi-square based) to examine, whether the co-occurrences of two objects are signiﬁcant in a statistical sense. If so, only context similarity is used, as we have a very direct clue of the similarity of these objects. If not, we use only content-based similarity, as it works independent of any object occurrences. First experiments on synthetic data show that the combination of both methods is on average superior to any of the methods in isolation.

4.3

Searching the Set of Personal Maps - Matchmaking

In order for a given user to beneﬁt from the possibility of using knowledge maps of other users, there needs to be a way to eﬃciently identify knowledge maps which are relevant to him from a potentially huge set of such maps. The method we are developing is based on the following idea: on the one hand a user has preferences, long term interests and pre-knowledge. On the other hand, she has a current information need. To capture both, we are developing a search facility, which combines keyword search (current information need) with a similarity analysis between users based on their personal maps (long-term information need). Combining both aspects results in a ranked list of personal knowledge maps available in the system. As this feature is currently under development, we refer to future work for more details.

5

Visualization and Interface

The critical issue in visualizing the knowledge maps and using them as a tool for discovering new knowledge is an intuitive interface which allows the user to unobtrusively construct personalized maps as accompanying eﬀect of his exploration of an information space. On one hand, this requires that the results of the clustering and personalized classiﬁcation mechanisms need to be visualized in a way, which provides clear insight into the meaning and criteria of a given

224

Jasminko Novak et al.

Fig. 4. The Knowledge Explorer Interface

grouping. Our basic model for achieving this represents the combination of the Content Map and the Concept Map discussed in Chapter 2.2. By displaying the distribution of all items and their grouping in semantically related clusters, the Content Map gives a quick, general impression of the information pool. The semantic space of each cluster is described by a number of keywords. One kind of keywords is extracted from the data records as entered by the user, while the other is generated by the server side text-analysis. The left-hand window of the interface in Fig.4 shows one concrete implementation of the Content Map, with the corresponding Concept Map to its right. The basic mode for the user to get detailed information is by selecting documents or clusters of interests and moving them into one of the other free windows, which can also be resized at will. Creating a personal map functions in a similar way. The user can open an empty map and ﬁll it with relevant documents (or entire clusters) from the Content Map per drag&drop. The documents and clusters in the personal map can be rearranged at will, and annotated with user deﬁned labels and keywords. Also a typical object per cluster can be deﬁned. In this way a template to be learned by the personal agent is created. As this template has a clear visual represen-

Discovering, Visualizing, and Sharing Knowledge

225

tation communicating the semantics of individual elements to the human user (e.g. clusters, keywords, labels etc.) it is also a medium of (implicit) communication between the agent and the user. The result of the new, personalised maps generated by the agent is communicated to the user in the same visual way. A special issue for the visualization and interface has been the handling of navigation in large information spaces. Especially when investigating possible relationships between diﬀerent groups of documents, the user needs both to be able to keep switching between detailed views of individual groups and the views encompassing larger, global portions of the map. Furthermore, one also needs to be able to move smoothly between diﬀerent information layers such as titles, keywords (machine and human), abstracts and images. In addressing these issues we built on experiences from previous work on focus+context techniques such as in [17], [18] and [19]. As a concrete solution we have developed a model for semantic zooming with multiple zoom focuses and global and local zoom areas (Fig. 4). It allows the user to select diﬀerent zoom focuses and pin them down as ﬁxed points of interest without loosing the overview. The user can further decide whether the zooming should have only local eﬀect at the given focus area (drill-down mode) or scale through the global environment so as to always keep both focus and overview (progressive-zoom mode).

6

Practical Applications

The practical test bed and ﬁrst application context of the described work is the Internet platform netzspannung.org3 [20]. Netzspannung.org aims at establishing a knowledge portal that provides insight in the intersections between digital art, culture and information technology. Typical netzspannung.org users are experts and professionals such as artists, researchers, designers, curators and journalists. The basic requirement of such an interdisciplinary knowledge portal is: a continually evolving information pool needs to be structured and made accessible according to many diﬀerent categorization schemes based on needs of diﬀerent user groups and contexts of use. By using the described system this heterogeneous user group will be able interactively compose and collaboratively structure an information pool coming from diﬀerent data sources, to visualise and explore it through personalised knowledge maps, and to construct a shared navigation structure based on the interconnection of their personal points of view. The current system prototype has been internally deployed as information access interface to the submissions of the cast01 conference4 and of the competition of student projects digital sparks. This simulates the use scenario in which users can explore possible relations between information usually isolated in separate archives of diﬀerent communities in the ﬁelds of media art, research and technology. The results can be tried out in the guided tour and partially online available 3 4

http://www.netzspannung.org http://netzspannung.org/cast01/

226

Jasminko Novak et al.

interactive demos. A very ﬁrst visualization prototype for browsing system generated maps is still being used as public information interface.

7

Summary and Ongoing Work

We have presented an approach of how to use the paradigm of knowledge maps as a central concept to integrate diﬀerent methods for interactive information search and for realising a model for collaborative discovery and sharing of knowledge. We have shown, how supervised and unsupervised learning can be used to generate knowledge maps, providing users with diﬀerent views on the content and semantic structure of an information source. We have presented an unobtrusive model for proﬁling personalised user agents based on two dimensional semantic maps that provide both a medium of implicit communication between human users and the agents, as well as a form of visual representation of the resulting knowledge structures. Furthermore, we have presented possibilities to use knowledge maps as medium for explicit and implicit exchange of knowledge between diﬀerent users. As pointed out, our system diﬀers signiﬁcantly from so called ”collaborative ﬁltering” systems, as items are not just rated by the users, but are put into context, in a way which is unobtrusively embedded into users primary activity. In this sense, our system enables ”collaborative structuring” rather than just ”collaborative ﬁltering”. Agents and Agent Mediated Knowledge Management have been used as paradigms to model and implement the system. This approach has shown to be well suited for the given problem, as it helped to structure the diﬀerent components not only in an understandable, but also in an extendable way, oﬀering the possibility of future additions and modiﬁcations. Currently we are working on diﬀerent methods, to extend and optimize the system. Firstly, we aim to add additional similarity aspects for the learning of personal maps. Secondly, editing personal knowledge maps, the user can arrange objects only in ﬂat structures, which is very intuitive and easy to handle, but not always suﬃcient. Therefore the system will contain a second editor, capable of creating hierarchical structures and other relations between objects. From the point of view of processing, the problem is to develop such methods, which fully exploit the information contained in such structures. Finally, an evaluation workshop is planned for analysing the usefulness of the system and comparing the individual contributions of the diﬀerent approaches. The evaluation will proceed in three steps: ﬁrst the basic model of capturing user knowledge through personal maps created in unobtrusive interaction with the system-generated map, will be evaluated. In the next step the exchange of knowledge between users through explicit sharing of maps, and through implicit agent inferencing as described in chapters 4.2 and 4.3 will be evaluated. Finally, the third test will evaluate the emergence of a shared navigation structure as a concept map network reﬂecting implicit knowledge of a group of users.

Discovering, Visualizing, and Sharing Knowledge

227

Acknowledgements The work described in this paper has been undertaken within the projects AWAKE - Networked Awareness for Knowledge Discovery and netzspannung.org - an Internet Media Lab, both ﬁnanced by the German Federal Ministry for Education and Research.

References [1] Nonaka, I., Takeuchi, H.: The Knowledge-Creating Company. Oxford University Press (1995) [2] Krueger, W., Nielsen, J., Oates, T., Finin, T.: Automatically generated daml markup for semistructured documents. In: Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management. (2003) [3] Haimowitz, I., Santo, N.: Agent-mediated knowledge management for tracking internet behavior. In: Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management. (2003) [4] Furtado, J.J.V., Machado, V.P.: Improving organizational memory through agents for knowledge discovery in database. In: Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management. (2003) [5] Tacla, C.A., Barthes, J.P.: A multi-agent architecture for knowledge aquisition. In: Proceedings of the AAAI Spring Symposium on Agent Mediated Knowledge Management. (2003) [6] Chalmers, M.: Paths and contextually speciﬁc recommendation. In: DELOS Workshop: Personalisation and Recommender Systems in Digital Libraries. (2001) [7] Lin, X., Soergel, D., Marchionini, G.: A self-organizing semantic map for information retrieval. In: Proc. of 14th ACM/SIGIR Conf. (1991) [8] Kohonen, T., Kaski, S., Lagus, K., Salojrvi, J., Paatero, V., Saarela, A.: Organization of a massive document collection. IEEE Transactions on Neural Networks 11 (2000) 574–585 [9] Sack, W.: Conversation map: An interface for very large-scale conversations. Journal of Management Information Systems (2000) [10] Becks, A., Sklorz, S., Jarke, M.: A modular approach for exploring the semantic structure of technical document collections. In: Proc. of AVI 2000. (2000) [11] Herlocker, J.L., Konstan, J.A., Riedl, J.: Explaining collaborative ﬁltering recommendations. In: Computer Supported Cooperative Work. (2000) 241–250 [12] Resnick, P., Iacovou, N., Suchak, M., Bergstorm, P., Riedl, J.: GroupLens: An Open Architecture for Collaborative Filtering of Netnews. In: Proceedings of ACM 1994 Conference on Computer Supported Cooperative Work, Chapel Hill, North Carolina, ACM (1994) 175–186 [13] Joachims, T.: Text categorization with support vector machines: learning with many relevant features. In: Proceedings of ECML-98, 10th European Conference on Machine Learning, Springer Verlag (1998) 137–142 [14] Honkela, T.: Self-Organizing Maps in Natural Language Processing. PhD thesis, Helsinki, Finland (1997) [15] Aha, D., Kibler, D., Albe, M.: Instance based learning algorithms. Machine Learning 6 (1991) 37–66

228

Jasminko Novak et al.

[16] Agrawal, R., Srikant, R.: Fast algorithms for mining association rules. In Bocca, J.B., Jarke, M., Zaniolo, C., eds.: Proc. 20th Int. Conf. Very Large Data Bases, VLDB, Morgan Kaufmann (1994) 487–499 [17] Robertson, G., Mackinlay, J.D.: The document lens. In: ACM UIST. (1993) [18] Sarkar, M., Snibbe, S.S., Tversky, O.J., Reiss, S.P.: Stretching the rubber sheet: A metaphor for viewing large layouts on small screens. In: ACM Symposium on User Interface Software and Technology. (1993) 81–91 [19] Bederson, B., et al.: Pad++: A zoomable graphical sketchpad for exploring alternate interface physics. J. Vis. Lang. Comput. 7:3 (1996) [20] Fleischmann, M., Strauss, W., Novak, J., Paal, S., Mueller, B., Blome, G., Peranovic, P., Seibert, C., Schneider, M.: netzspannung.org - an internet media lab for knowledge discovery in mixed realities. In: Proceedings of the Conference cast01 - living in mixed realities. (1991)

Agentized, Contextualized Filters for Information Management David A. Evans, Gregory Grefenstette, Yan Qu, James G. Shanahan, Victor M. Sheftel Clairvoyance Corporation, 5001 Baum Boulevard, Suite 700 Pittsburgh, PA 15213-1854, USA {dae, grefen, y.qu, jimi, v.sheftel}@clairvoyancecorp.com

Abstract. When people read or write documents, they spontaneously generate new information needs: for example, to understand the text they are reading; to find additional information related to the points they are making in their drafts. Simultaneously, each Information Object (IO) (i.e., word, entity, term, concept, phrase, proposition, sentence, paragraph, section, document, collection, etc.) someone reads or writes also creates context for the other IOs in the same discourse. We present a conceptual model of Agentized, Contextualized Filters (ACFs)—agents that identify an appropriate context for an information object and then actively fetch and filter relevant information concerning the information object in other information sources the user has access to. We illustrate the use of ACFs in a prototype knowledge management system called ViviDocs.

1

Information Management

Developing technology for information management (IM) is a challenge because our systems cannot be based on the perfection of any single function—such as superior information retrieval, for example—but rather must derive their usefulness from an interaction of many functions. Effective IM will depend on the integration (and exploitation) of models of (1) the user, (2) the context, and (3) the application (or information purpose) with (4) the processing of source data. Integration will be the dominant factor in making information management systems useful. To aid such integration, we seek to mobilize information in the user’s environment. IM tasks are highly contextualized, highly linked to other tasks and related information—never tasks in isolation. Every time a user engaged in work reads or writes, the user spontaneously generates new information needs: to understand the text he or she is reading or to supply more substance to the arguments he or she is creating. Simultaneously, each Information Object (IO)—word, entity, term, concept, phrase, proposition, sentence, paragraph, section, document, collection, etc.— encountered or produced creates context for the other IOs in the same discourse. An effective IM system will automatically link varieties of such IOs, dynamically preparing answers to implicit information needs. To this end, rather than focus on a system that performs a single “end-to-end” function—processing a request for information or finding “similar” documents or even “answering a question”—we have been focusing on the critical components of a system (which we call “ViviDocs”) that operates behind more ordinary user tasks, L. van Elst, V. Dignum, and A. Abecker (Eds.): AMKM 2003, LNAI 2926, pp. 229-244, 2003.  Springer-Verlag Berlin Heidelberg 2003

230

David A. Evans et al.

such as reading messages or writing reports. These tasks are not, explicitly, directed at finding information. But when performed in the workplace, these tasks continually generate new information needs; and to address these, we require a system that can ground a document in a structured web of authoritative information.

2

Agentized, Contextualized Filters

In ViviDocs, while a person reads or writes a text (an e-mail message; a report), the components of the text are continually analyzed into candidate IOs. A variety of agents are generated for each new IO. These agents identify an appropriate (typically local) context for the IO—represented by other text or information in the user’s environment—and then actively fetch and filter relevant information concerning the IO in information sources the user has access to. We call such agents “Agentized, Contextualized Filters” (ACFs). They are agents in the sense that they operate autonomously and asynchronously; they are triggered by some event; they use their own data; and they perform specific functions on the data; and they adjust to changing conditions, potentially learning from the user’s behavior [1]. They are contextualized because they are anchored to specific IOs in contexts of use. 2.1 A Conceptual Model of ACFs We define an ACF as a function that links one information object (as anchor) with another information object (output), taking into account the context of the task and the context of the user’s work environment. Formally, we define an ACF as:

ACFi ( Pi , Ri , Si ,θ i , H i ,U i , Ci , Ti , Fi )

(1)

where (for time/instance i) Pi represents the feature profile of the information object, Ri, the associated knowledge resources, Si, the target sources, θ i , the threshold, Hi, the history lists, Ui, the utility function for the user, Ci, the processing context, Ti, the triggering condition that activates the agent, and Fi, the response function and format. We elaborate on each of these factors below. Profile (Pi). The Profile is a representation of the information object based on its textual content. For example, in an information retrieval system, a profile representing an IO (e.g., a document or paragraph) might consist of a list of terms with associated weights to reflect their usages in the document or with respect to a document collection. Resource (Ri). Resource refers to language resources (e.g., stop words, grammar, lexicons, etc.), knowledge resources (e.g., abstract lexical-semantic types, taxonomies or classification schemata, semantic networks, inference rules, etc.), and statistical models (e.g., term frequency and distribution counts, language models, etc.) used for processing. Source (Si). Source refers to the target or available information sources, accessible to the user or to the agent, in which responses to information needs may be found. In a workgroup, this might include all the user’s files and the accessible files of the members of the user’s team or department. In a general business setting, this might

Agentized, Contextualized Filters for Information Management

231

include the contents of the company intranet, extranet, and, selectively, the internet, as well as the user’s personal files. History (Hi). History consists of lists of information objects (and perhaps “scores”) that have been generated by previous actions of ACFs. For example, in information retrieval with user feedback, the initial ranked list of documents considered as relevant by the system can be regarded as the history for the next round of retrieval with additional user feedback. Threshold ( θ i ). A threshold is used to control the cut-off points in decision making. Thresholds can be absolute numbers (e.g., the top 100 documents or passages), similarity scores, or confidence scores applied to retrieved information. Utility (Ui). Utility is used to measure and rank system outputs based on the benefits they produce for the user or on the degree to which they satisfy the user’s information needs minus the associated costs. Such measures are commonly used in information filtering and are typically calculated from an explicit or implicit statement of the tolerance for “noise” (the ratio of true-positive to false-positive responses) in the output. Context (Ci). Context provides additional information that can be associated with the profile. While this concept is inherently open-ended (and subject to overuse), we restrict it to information that can be determined operationally by the system. We distinguish at least three kinds of context: (a) global context, (b) local context, and (c) focus. In an IR-like action anchored to a specific IO (e.g., word or phrase), the global context might be the document in which the IO occurs; the local context the paragraph; the focus the sentence (essentially, the proposition expressed). Consider, for example, the following passage in a text on the German military campaign in the Soviet Union during World War II: The Battle of Stalingrad represented a major turning point for the Germany Army. The German general Paulus was out-foxed by the Russian Generals by being drawn into the city. The Russians eventually wore the Germans down, cut off their supply lines, and made retreat impossible. The simple IO corresponding to “Paulus” has several constraining contexts. The global context establishes Paulus as a German general in WWII. Local context relates specifically to his participation in the battle of Stalingrad. Focus involves his particular role in the event, namely, being “out-foxed” by the Russian generals. If we imagine stepping through the document and selecting each such IO (e.g., person-name reference) in sequence, we can see that the general context is stable, and does not need to be updated as we move from IO to IO; the local will change frequently, from passage to passage; and focus will vary from sentence to sentence. If the user were writing a text, we could imagine focus changing quite dynamically, even as the user wrote a new sentence or deleted an old one. User profiles and work-tasks can be treated as another source of context. On projects, the current set of documents that a user is working on or has access to may supply the global context, the specific document in which the information object is found can be the local context, and the immediate vicinity of the IO can be the focus. Trigger (Ti). Triggers activate the ACFs. The action associated with opening a document or beginning to compose a message could launch a battery of ACFs. Under

232

David A. Evans et al.

a GUI, triggers can take the form of highlighting, typing, clicking, etc. For example, every time the user types a full stop, an ACF can be triggered on the most recently completed sentence. Likewise ACFs could be triggered every twenty-four hours, updating the information that they associate with the IOs they are attached to. Function (Fi). Function specifies the relation that is to be established between the IO and other information by the ACF, including the format for extracting or presenting such information. The function might be as simple as “retrieval”—finding a rankordered list of documents or passages—or “answer” (a simple sentence) in response to an implicit question. But the function might also be considerably more complex, such as establishing the background facts that support the proposition that the IO asserts. Functions have associated presentation requirements or formats. Formats typically require that a set of (possibly contrastive) information be developed, such as the ranked list of responses to a query, or clusters of passages that each represents different senses of a response. More ambitious combinations of functions and formats might involve providing the user with a sense of the structure of the space of answers (via topic modeling, perhaps [2]); or the location of centers of importance (via semantic hubs and authorities, perhaps); or of related topical "regions" (via semanticspace abstractions). 2.2 ACF Parameters Generally, of course, parameters of an ACF interact with each other. For example, our model of the user affects utility. If the user is an analyst who already knows a great deal about a topic, then we probably want to maximize the novelty aspect of any information we link to the user’s work and discount the information already in the user’s background (files, past work, workgroup, etc.). On the other hand, even in the case of a user whose “normal” type is well understood, based on the user’s response to information or changing assignments, we may need to update or revise the user model and other parameters frequently. The issue of parameter interaction and calibration would seem to doom the model, especially if one considers the need to adapt to specific users over time: the “training” problem could be daunting. However, though parameters can vary quite widely in theory, we observe that, for many practical application types, the actual values of parameters may be quite limited. In short, in practical use, only a few of the parameters will vary freely and these will overwhelmingly assume only a few possible values. As an illustration, consider one of the most general functions an ACF can perform: association—finding relevant related material. Note that, while this might be implemented as a simple IR task, taking the text of a document as a query and searching available external sources, the proper association of information to a document is not a trivial matter. For example, a long document, taken as a query, will typically give high rank to documents (responses) that share terms with its dominant (high-frequency/low-distribution) terms. If the external sources are large, it is likely that virtually all the top-ranked responses will be biased to the “summary” or “centroid” sense of the document. Thus, in order to insure that all the parts of the document are properly represented, an association process should formulate many separate queries from the text of the document and merge results in a fashion that insures that all parts will be represented in “high-ranking responses.” An ACF that

Agentized, Contextualized Filters for Information Management

233

performs such a task on “start up” (when a document is opened, for example) might well follow a standard procedure to decompose the document into sequences of passages (each serving as a source of a query (P)), use default resources for term extraction (R) on each passage of approximately paragraph size, and target a default large external source (S). Such an ACF might ignore context (C) and history (H), since the document itself is term rich and the user’s session is just beginning, being triggered (T) upon opening the document. The function to be performed—in this case, multi-pass IR (F)—can be specified to establish a local cache of material that will be of high value if the user wants to explore topics or answer questions that arise in reading the text. Thus, the only open questions relate to what the operational interpretation of utility (U) and threshold (θ) should be. In this regard, a variety of heuristics may prove serviceable, e.g., (1) insure that each passage brings back at least n documents and all documents (up to a maximum, m) that score above the threshold; (2) vary the threshold for each passage based solely on the scoring potential of the passage against the data being searched; (3) aim for a final cache of documents in the range of 100 to 10,000. This might be achieved by ranking the results of each passage-query using normalized scoring—dividing the term score of each responding document by the term score of the first-ranked document—using a fixed threshold, e.g., 0.7 or 0.6 normalized score, and returning (and caching) the top n responses and any other responses (up to the mth) that score at or above threshold. Since we know how big the document is (the count of the number of passages we extract from it), we can set n and m to insure that the resulting information cache is in the target range (e.g., 100 to 10,000 documents). FindRelevantDocs Profile: Resource: Source: History: Threshold: Utility: Context: Trigger: Function: results>