Reinforcement Learning in Markov Games
A classic goal of an AI agent is to learn optimize in uncertain environments, often modelled as Markov decision processes. At Brown, we are designing agents that learn to strategize in uncertain environments that are inhabited by other agents, modelled as Markov games. In the single-agent case, dynamic programming (DP) and reinforcement learning (RL) algorithms converge to optimal stationary policies. However, multiagent generalizations of DP and RL techniques need not converge to stationary equilibrium policies. On the contrary, we are investigating the space of nonstationary (e.g., cyclic) equilibrium policies that can arise in the multiagent case.
Project status: Active
Project Home Page: http://www.cs.brown.edu/people/amy/multiagent.html
Research Areas
| Artificial Intelligence |
| Machine Learning |
Research Themes
| Machine Decision and Game Theory |
People
| David Gondek |
| Amy Greenwald |
| Keith Hall |
| Michael Littman |
| Martin Zinkevich |
Funding
Computational Social Choice Theory, National Science Foundation, $375,000, 3/1/2002 - 2/28/2007
Publications
Zinkevich, M., Greenwald, A., and Littman, M. Cyclic Equilibria in Markov Games. In Advances in Neural Information Processing Systems (2006), MIT Press. To Appear. [ postscript | pdf ]
Greenwald, A., and Zinkevich, M. A Direct Proof of the Existence of Correlated Equilibrium Policies in General-Sum Markov Games. Tech. Rep. CS-05-07, Brown University, Department of Computer Science, Jun 2005.
Greenwald, A., and Hall, K. Correlated Q Learning. In Proceedings of the Twentieth International Conference on Machine Learning (Aug 2003), pp. 242-249. [ pdf ]
| Page Owner: Webmaster | Last Modified: Mon Oct 23 14:57:09 2006 |