Optimal online learning procedures for model-free policy evaluation

Abstract

In this study, we extend the framework of semiparametric statistical inference introduced recently to reinforcement learning [1] to online learning procedures for policy evaluation. This generalization enables us to investigate statistical properties of value function estimators both by batch and online procedures in a unified way in terms of estimating functions. Furthermore, we propose a novel online learning algorithm with optimal estimating functions which achieve the minimum estimation error. Our theoretical developments are confirmed using a simple chain walk problem. © 2009 Springer Berlin Heidelberg.