Improvement of the Relaxation Procedure in Concurrent Q-Learning

Abstract

In this paper, we point out problems in concurrent Q-learning (CQL), which is one of the adaptation techniques to dynamic environment in reinforcement learning and propose the modification of the relaxation procedure in CQL. We apply the proposed algorithm to the problem of maze in reinforcement learning and validate what kind of behavior the original CQL and the proposed algorithm show for the changes of environment such as the change of goals and the emergence of obstacles.