(A) study on generating explanations for reinforcement learning policies in control tasks

Abstract

For recent years, deep reinforcement learning, which leverages deep neural networks for learning reinforcement learning agents, has achieved successful results in control tasks, and the need for understanding and explaining learning processes or outcomes of deep reinforcement learning has also been raised. In this thesis, we try to extend the notion of minimal sufficient explanation, which explains behaviors of an agent learned in tabular reinforcement learning, into continuous state spaces, in order to explain behaviors of deep reinforcement learning agent. Further, we propose a novel explanation generation algorithm for deep reinforcement learning policies with the extended notion, and show how our method is efficient for explaining behaviors of reinforcement learning agents, by comparing naive baseline algorithms.