Exact Decomposition Approaches for Markov Decision Processes: A Survey

Abstract

As classical methods are intractable for solving Markov decision processes (MDPs) requiring a large state space, decomposition and aggregation techniques are very useful to cope with large problems. These techniques are in general a special case of the classic Divide-and-Conquer framework to split a large, unwieldy problem into smaller components and solving the parts in order to construct the global solution. This paper reviews most of decomposition approaches encountered in the associated literature over the past two decades, weighing their pros and cons. We consider several categories of MDPs (average, discounted, and weighted MDPs), and we present briefly a variety of methodologies to find or approximate optimal strategies.