Temporal-difference methods and Markov models

Abstract

The relation between temporal-difference training methods and Markov models is explored. This relation is derived from a new perspective, and in this way the particular association between conventional temporal-difference methods and first-order Markov models is explained. The authors then derive a generalization of temporal-difference methods that is suitable for Markov models of higher order. Several issues related to the performance of mismatched temporal-difference methods (i.e., the performance when the temporal-difference method is not specifically designed to match the order of the Markov model) are investigated numerically. >