Dynamic Programming for Partially Observable Stochastic Games

We develop an exact dynamic programming algorithm for partially observable stochastic games (POSGs). The algorithm is a synthesis of dynamic programming for partially observable Markov decision processes (POMDPs) and iterative elimination of dominated strategies in normal form games. We prove that it iteratively eliminates very weakly dominated strategies without first forming the normal form representation of a finite-horizon POSG. This is the first dynamic programming algorithm for iterative strategy elimination in these types of games. For the special case in which agents share the same payoffs, the algorithm can be used to find an optimal solution. We present preliminary empirical results and discuss ways to further exploit POMDP theory in solving POSGs.

• Type: Conference
• Year: 2004
• Where: AAAI
• Authors: Eric A. Hansen, Daniel S. Bernstein, Shlomo Zilberstein
• Keywords: AAAI 2004, Dynamic Programming Algorithm, Intelligent Agents, Normal Form, Observable Stochastic Games