An Adaptive Sampling Algorithm for Solving Markov Decision Processes

Abstract

Based on recent results for multi-armed bandit problems, we propose an adaptive sampling algorithm that approximates the optimal value of a finite horizon Markov decision process (MDP) with infinite state space but finite action space and bounded rewards. The algorithm adaptively chooses which action to sample as the sampling process proceeds, and it is proven that the estimate produced by the algorithm is asymptotically unbiased and the worst possible bias is bounded by a quantity that converges to zero at rate $Oleft ( rac{Hln N}{N} ight)$, where $H$ is the horizon length and $N$ is the total number of samples that are used per state sampled in each stage. The worst-case running-time complexity of the algorithm is $O((|A|N)^H)$, independent of the state space size, where $|A|$ is the size of the action space. The algorithm can be used to create an approximate receding horizon control to solve infinite horizon MDPs.