Model-Based Least-Squares Policy Evaluation

Abstract

A popular form of policy evaluation for large Markov Decision Processes (MDPs) is the least-squares temporal differencing (TD) method. Least-squares TD methods handle large MDPs by requiring prior knowledge feature vectors which form a set of basis vectors that compress the system down to tractable levels. Model-based methods have largely been ignored in favour of model-free TD algorithms due to two perceived drawbacks: slower computation time and larger storage requirements. This paper challenges the perceived advantage of the temporal difference method over a model-based method in three distinct ways. First, it provides a new model-based approximate policy estimation method which produces solutions in a faster computation time than Boyan’s least-squares TD method. Second, it introduces a new algorithm to derive basis vectors without any prior knowledge of the system. Third, we introduce an iteratively improvingmo del-based value estimator that can run faster than standard TD methods. All algorithms require model storage but remain computationally competitive in terms of accuracy with model-free temporal differencing methods.