Reviewed by
Armin Roehrl
Department of Mathematics, Swiss Federal Institute of Technology, 1015 Lausanne, Switzerland.
The theme of the 8th European Workshop on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW'97) was multi-agent rationality. This choice of topic illustrates the challenge currently posed by moving from studies of individual rationality to studies of group rationality. Artificial agents have been in use for several years now, but the underlying principle was always the maximisation of their expected utility (PMEU). Most existing user agents, such as web crawlers and mail filters, are hardly more than rudimentary decision support systems. However, dis-satisfaction with these approaches has led to the development of rival principles. The Multi-Agent System (MAS) community believes that PMEU cannot provide an effective operational definition of rationality.
It follows that social issues must come into play wherever there is room for co-operation and co-ordination. To cope with sequences of heterogeneous and complex decision situations, rationality might require agents to synthesise from different evaluations of the world. The representation and modelling of their social context thus becomes the central challenge. The proceedings of the MAAMAW'97 workshop reviewed here cover a broad spectrum of relevant application areas from multi-agent co-ordination in anti-air defense to robot vision.
Invited SpeakersThree speakers representing different schools of thought in MAS were invited to the workshop. Abstracts of all three talks and a full paper on "Delegation Conflicts" by Cristiano Castelfranchi are included in the proceedings.
Michael P. Wellman (University of Michigan) is studying the use of economic principles for designing rational agent behaviour, calling this technique "market-oriented programming". He illustrates the success potential of economic rationality in real-life MAS applications. To underline his point, he discusses the Michigan Internet Auction Bot, a configurable auction server implementing market-based negotiation over the World Wide Web.
Yuri Ermoliev (IIASA, Vienna) is one of the founders of stochastic optimisation and has used multi-agent systems to study decisions under risk. He takes a sceptical stance and claims that similarities with natural evolutionary processes can be misleading when studying man-made systems.
Cristiano Castelfranchi (CNR, Rome) argues that the adoption of the game-theoretic paradigm by Artificial Intelligence may originally have motivated effective research, but fails in the modelling of autonomous agents and MAS. His criticism also highlights overconfidence in the concept of economic rationality.
Papers15 of the 51 papers accepted for the workshop are included in the proceedings. They range from clarification of issues in classical game theory to formal representations of autonomous agents in multi-agent systems.
"Multi-Agent Co-ordination in Anti-Air Defense: A Case Study" by Noh and Gmytrasiewicz is a classic example of a game-theoretic approach. It uses the Recursive Modelling Method (RMM), enabling agents to select their rational action by examining the expected utility of alternative behaviours, and to coordinate with other agents by modelling their decision-making in a distributed multi-agent environment. It is a very good article to introduce the issues. They also provide an online demonstration in which their system tries to outperform other computer programs or humans.
"A Service-Oriented Negotiation Model Between Autonomous Agents" by Sierra, Faratin and Jennings also follows the game-theoretic approach, but puts the emphasis on negotiations. They show that the game-theoretic assumptions can be extended to introduce a kind of subjective probability. Boman extends this idea in "Norms as Constraints on Real-Time Autonomous Agent Action" by obtaining rational behaviour allowing agent reports to be overriden by norms.
The ideas of credibility and reliability assessment are used in "Distributed Belief Revisions versus Belief Revision in a Multi-Agent Environment: First Results of a Simulation Experiment" by Dragoni, Giorgini and Baffetti. Agents meet and hold elections, though they don't necessarily change their own assessement results. All sorts of complex possibilities are considered, such as mechanisms for coping with incorrect information and contradiction resulting from some degree of incompetence. Each node is equipped with its own belief revision module. This approach looks very promising, but is computationally expensive to simulate.
De Jong's paper "Multi-Agent Co-ordination by Communication of Evaluations" is about the domain competence problem. Local experts direct other agents, whose competence is more global. Co-ordination signals are used to simulate real-life encounters.
In "Causal Reasoning in Multi-Agent Systems", Chaib-draa investigates the possibilities for using Cognitive Maps in multi-agent environments. He offers a formal model to establish the mathematical basis for the manipulation of CMs. Unfortunately that model can only give partial advice.
Glaser and Morignot's paper ("The Reorganisation of Societies of Autonomous Agents") describes a model in which agents migrate between different societies. Agent societies will establish new conventions that force reorganisations based on the principle of punishment: a society punishes or favours the behaviours of its new members. The candidate agent is rewarded after its integration through the growth of its own utility function.
In "Adaptive Selection of Reactive/Deliberate Planning for the Dynamic Environment - A Proposal and Evaluation of MRR-Planning", Kurihara, Aoyagi and Onani evaluate methodologies for multi-agent real-time reactive planning. Their paper takes concepts from parallel programming, conventional real-time planning and multi-agent methods. It follows a very useful "hands on" approach with a real world example from robot vision. Fisher and Wooldridge, in "Distributed Problem-Solving as Concurrent Theorem Proving", follow similar parallel programming ideas, but at a more theoretical level. They demonstrate that distributed problem solving may be viewed as concurrent theorem proving. To solve such problems, they use an agent-based approach and finally describe Concurrent-METATEM, a multi-agent programming language.
In "Commitments Among Autonomous Agents in Information-Rich Environments", Singh gives a definition of commitments which refers to social rather than psychological conventions. His definition is particularly useful for information-rich applications such as electronic commerce. The goal of the paper is to unify principles behind commitments for single agent and multi-agent systems.
Ygge and Akkermans ("Making a Case for Multi-Agent Systems") provide a clear real life comparison between multi-agent systems and conventional engineering solutions for climate control in large buildings.
Baerentzen, Avila and Talukdar ("Learning Network Designs for Asynchronous Teams") return to the difficult problem of learning effectively from the past. They arrange asynchronous teams, consisting of networks of agents and memories. The crucial question is to determine the structure of such networks. They develop some automata to address this problem.
The high-level concurrent object oriented programming language CORRELATE with high level synchronisation primitives was developed as a testbed and development system at the crossroads of many disciplines. Joosen, Bijens, Matthijs, Robben, Oeyen and Verbaeten provide a detailed description of CORRELATE in "Building Multi-Agent systems with CORRELATE". CORRELATE might save you a lot of time when simulating agent systems. Currently it runs on DEC Alpha Compaq 64 True Unix, Sun's Solaris 2.x and SGI's Irix 5.3.
Jonker and Treur ("Modelling an Agent's Mind and Matter") use novel techniques to represent the dynamics of models including dynamic logic, cybernetics and conceptual modelling. The authors use MAS by representing the lifespan between the birth and death of multiple agent objects. They treat the mind of the agent as genuinely embodied.
The last paper is Castelfranchi and Falcone's "Delegation Conflicts", which touches on a rather wide range of subjects leading to an analytical theory of delegation. They define delegation as changing another agent's plan and also provide insights for dealing with possible conflicts resulting from delegation.
All in all, the choice and ordering of papers is very stimulating.
Overall ConclusionIt was a real pleasure reading this proceedings volume. I found the fresh approach to a very young and expanding field especially appealing. Many different subject areas and academic disciplines are covered. This approach is not only interesting in itself but also means that the audience of potential readers is large compared to many other books on agents. These proceedings are certainly worth buying for everyone who needs to get a grasp of what is going on in the area of multi-agent rationality from multiple perspectives. It is also a good source of the most important references in the field.
Return to Contents of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 1999