Tensor decomposition and parallelization of Markov Decision Processes

Abstract

Markov Decision Processes (MDPs) with large state spaces arise frequently when applied to real world problems. Optimal solutions to such problems exist, but may not be computationally tractable, as the required processing scales exponentially with the number of states. Unsurprisingly, investigating methods for efficiently determining optimal or near-optimal policies has generated substantial interest and remains an active area of research. A recent paper introduced an MDP representation as a tensor composition of a set of smaller component MDPs, and suggested a method for solving an MDP by decomposition into its tensor components and solving the smaller problems in parallel, combining their solutions into one for the original problem. Such an approach promises an increase in solution efficiency, since each smaller problem could be solved exponentially faster than the original. This paper develops this MDP tensor decomposition and parallelization algorithm, and analyzes both its computational performance and the optimality of its resultant solutions.