Multiagent Learning Group

Artificial Intelligence Lab
Department of Computer Science
Iowa State University

page since 2002-03-03

Group Members

• Dr.Vasant Honavar
• Jie Bao -> the Page for Hierarchial learning in Multiagent System
• Facundo Bromberg
  
• Charles Gieseler

intragroup maillist : multiagent@iastate.edu

Current Research Focus

• Graphical Model in Multiagent Game -> See reference
• Bayesian Method in Multiagent Leanrning
• Multiagent Learning in Distributed Knowlege Understand Environment

Talk

Date	Topic
2003-01-15	General Discussion
2003-01-22	General Inrodution of Multiagent Learning (Honavar)
2003-01-29	Game, Bayesian Network and MAL
2003-02-03	(in Lab seminar) Jie Bao : Zeus Agent System The notion of heterogeneous autonomous agents collaborating to solve problems is a powerful metaphor for the engineering of distributed and interoperable software systems.  This agent-based approach introduces a new level of abstraction ¾ of knowledge level co-operation between autonomous systems ¾ that enhances distributed systems interoperability, scalability and re-configurability.   However, thus far, the promise of the agent approach has been largely unrealised in the distributed software engineering community.  This is due to a number of factors (including the current lack of standards for agent technology), but primarily because of the inherent complexity of constructing collaborative agent systems. To facilitate large-scale realisation of the collaborative agent approach to distributed software engineering we felt frameworks, methodologies and toolkits were needed that would support the rapid development of multi-agent systems.  This has led to the development of ZEUS, a toolkit for constructing collaborative multi-agent applications.  ZEUS is a culmination of a careful synthesis of established agent technologies to provide an integrated environment for the rapid development of multi-agent systems.  ZEUS defines a multi-agent system design approach and supports it with a visual environment for capturing user specification of agents that are used to generate Java source code of the agents. See http://www.labs.bt.com/projects/agents/zeus/index.htm for detail
2003-02-05	Decision Netwok and Game Network
2003-02-20	Discussion of Graphical Game  by Kearns &  Game Networks by Pierfrancesco La Mura
2003-02-24	Ontology and MAL
2003-03-03	(in Lab seminar) Facundo Bromberg: Selected topics in Multi-Agent Learning  According to Daphne Koller: "a decision is an irrevocable allocation of environment resources".  Therefore, for an agent to act in the world, that is, to make decisions, implies irrevocable allocation of resources (e.g. processor time). If the resources supplies are infinite, a single agent would take no time to solve the most complex (solvable) problem. Needless is to say that this scenario is not even science fiction, it is impossible. Limited resources is part of every single natural process, including computations. Thus, a decomposition of the task to be solved into many process is a need that computer scientist realized not so long ago, giving rise to sub-fields of computer science like parallel computation theory and distributed systems. In both of this areas, the designer impose (builds-in) a schema for inter-processes communication, coordination and resource allocation conflicts. This requires a full knowledge of the domain and processes involved in advance, i.e. off-line. But again, the world is complex and uncertain, and the whole AI enterprise aims exactly that: to build computer systems (agents) capable of dealing with an uncertain and partially observable environment. This implies a certain amount of autonomy on behalf of the agents, autonomy from the designer, the programmer. The paradigm of distributed systems no longer applies, the designer cannot know in advance the times an agent will pretend allocate resources, nor how much resources will it consume. For an efficient use of the resources, the multi-agent system needs self-coordination, for which it needs to be able to communicate, all these at run-time. And, if more than one agent needs to allocate the same (limited amount of) resources, conflict arises. Hence the need of run-time conflict resolution capabilities built-in in the agents.  In this talk, I will present a not so short (but mostly qualitative) introduction  to the field of Multi-Agent systems. Most of the talk will be based on Michael Wooldridge's book: "An Introduction to Multi-Agent Systems" and "Learning in Multiagent Systems" by Gerhard Weiss and Sandip Sen. The focus of the talk will concentrate on giving an introduction to the field, a motivation for extending the AI enterprise to multi agent systems and how does Learning fits, extends and contribute to this field. The list of topics includes Game Theory, Reinforcement Learning, Communication, Conflict Resolution, Reaching Agreement, Coordination and Decision Making in a multi-agent environment. If you would like to taste part of the talk in advance, you can do so by navigating the slides of Wooldridge's book on-line at:
2003-03-06	Jie Bao:  Graphical Game of Kearns, focused on Graphical Models for Game Theory, by M. Kearns, M. Littman, and S. Singh, in the Proceeding of the UAI2001, 253-260 http://www.cis.upenn.edu/~mkearns/papers/graphgames.ps Nash Propagation for Loopy Graphical Games. L. Ortiz. and Michael Kearns. To appear, Proceedings of NIPS 2002. http://www.cis.upenn.edu/~mkearns/papers/nashprop.pdf
2003-03-10	(in Lab seminar) Jie Bao: Selected topics in Multi-Agent Learning II - Graphical Model in Multiagent Learning Game theory and decision theory have been widely used in  multiagent system in the recent years. A combination of Game Theory, Markov Decision Process(MDP) and Bayesian Network( and its extended version, Decision Network(Influence Diagram) ) is the trend in this young field. Several graphical models have been proposed to discover the strategy and utility independence in the game and to find Nash Equilibria(NE)approximately or exactly. - Graphical Game by M. Kearns etc. Inspired by polytree algorithm, theagents find local optimal response by local message exchanging and thenconstruct NE incrementally. - Game Network and Excepted Utility Network by La Mura etc. It's focused on finding utility independence. - Multi-Agent Influence Diagrams(MAID) by Daphne Koller etc, single IDs are linked directly and can be reduced to a Bayesian Network. NEs are found using a Divide-and-Conquer approach. The talk will cover two parts Part I - A brief introduction of incomplete information game (game with uncertainty) Part II - An overview of graphical game models and their comparison Refer to http://www.cs.iastate.edu/~baojie/acad/reference/2003-02-27_gamenet.htm for detailed information Most important references include: [KLS2001] Graphical Models for Game Theory, by M. Kearns, M. Littman,and S. Singh, in the Proceeding of the UAI2001, 253-260 http://www.cis.upenn.edu/~mkearns/papers/graphgames.ps [OK2002] Nash Propagation for Loopy Graphical Games. L. Ortiz. and Michael Kearns. To appear, Proceedings of NIPS 2002. http://www.cis.upenn.edu/~mkearns/papers/nashprop.pdf [MUR2000] Game Networks by Pierfrancesco La Mura In Proceedings of the 16th Conference on Uncertainty in Artificial Intelligence (UAI), pages 335-342, 2000. http://citeseer.nj.nec.com/438709.html  [KM2001]  Multi-Agent Influence Diagrams for Representing and Solving Games by Daphne Koller, Brian Milch 2001 http://www.cs.berkeley.edu/~milch/papers/ijcai01maids.html  ; http://citeseer.nj.nec.com/koller01multiagent.html   [KM2003] Daphne Koller and Brian Milch. (2003) " Multi-Agent Influence Diagrams for Representing and Solving Games". To appear in Games and Economic Behavior special issue of selected papers from the First World Congress of the Game Theory Society. http://www.cs.berkeley.edu/~milch/papers/geb03maids.html
2003-03-13	Charles Gieseler: