An Off-Policy Natural Policy Gradient Method for a Partial Observable Markov Decision Process

Abstract

There has been a problem called " exploration-exploitation problem " in the field of reinforcement learning. An agent must decide whether to explore a better action which may not necessarily exist, or to exploit many rewards by taking the current best action. In this article, we propose an off-policy reinforcement learning method based on a natural policy gradient learning, as a solution of the exploration-exploitation problem. In our method, the policy gradient is estimated based on a sequence of state-action pairs sampled by performing an arbitrary " behavior policy " ; this allows us to deal with the exploration-exploitation problem by handling the generation process of behavior policies. By applying to an autonomous control problem of a three-dimensional cart-pole, we show that our method can realize an optimal control efficiently in a partially observable domain.