Perkins' Monte Carlo exploring starts for partially observable Markov decision processes (MCES-P) integrates Monte Carlo exploring starts into a local search of policy space to offer a template for reinforcement learning that operates under partial observability of the state. In this paper, we generalize the reinforcement learning under partial observability to the self-interested multiagent setting. We present a new template, MCES-IP, which extends MCES-P by maintaining predictions of the other agent's actions based on dynamic beliefs over models. MCES-IP is instantiated to be approximately locally optimal with some probability by deriving a theoretical bound on the sample size that in part depends on the allowed error from the sampling; we refer to this algorithm as MCESIP+PAC. Our experiments demonstrate that MCESIP+PAC learns policies whose values are comparable or better than those from MCESP+PAC in multiagent domains while utilizing much less samples for each transformation.
展开▼
