Lempel-Ziv coding in reinforcement learning

Abstract

In this paper, we propose a new measure within the framework of reinforcement learning, by describing a model of an information source as a representation of a learning process. We confirm in experiments that Lempel-Ziv coding for a string of episode sequences provides a quality measure to describe the degree of complexity for learning. In addition, we discuss functions comparing expected return and its variance.