IJCAI-03 Tutorial: Multiagent Learning: A Game Theoretic Perspective


Sunday, August 10, 2003, 9am-1pm

Michael Bowling and Michael L. Littman

With the explosion of networking and affordable robotics, multiagent learning has become a hot research topic, combining the mature fields of machine learning and multiagent systems. The problem of learning a goal-oriented course of action in the presence of other goal-oriented agents raises new problems for both learning and multiagent systems. Complicating the learning problem is that when other agents are adapting, traditional machine learning assumptions like stationarity are violated. This tutorial will explain the unique challenges that are being addressed and the importance of this still-emerging field. A large focus will be on the introduction of critical game-theoretic concepts that underlie much of the recent work. This will be followed by an overview of the current progress, detailing the varied approaches and algorithms. A common set of problems and examples will be carried throughout the tutorial to provide continuity and a uniform understanding of the various techniques' strengths and weaknesses.

There will also be a discussion of the future issues and open problems that still remain. No background in game theoretic concepts will be assumed. A basic understanding of Markov decision processes and reinforcement learning would be helpful, although the most relevant concepts will be briefly reviewed.

Michael Bowling

Michael Bowling, Carnegie Mellon University, is finishing his Ph.D. on "Multiagent Learning in the Presence of Limited Agents". His research examines learning and planning in multiagent systems, specifically solutions inspired by game theoretic concepts. He has also been extensively involved in robot soccer as a testbed for his research.

Michael Littman
Michael L. Littman, Rutgers University, studies decision making and optimization under uncertainty. He is a recipient of an undergraduate teaching award from Duke and a best paper award from AAAI. His work on reinforcement learning in Markov games helped introduce game theory to the rapidly expanding field of multiagent learning.