Reinforcement learning can achieve autonomous learning in dynamic and complex environments, which makes it widely used in fields such as law, medicine, and finance. However, reinforcement learning still faces many problems such as the unobservable global state space, strong dependence on the reward function, and uncertain causality, which results in its weak interpretability, seriously affecting its promotion in related fields. It will encounter limitations such as difficulty in judging whether the decision-making violates social legal and moral requirements, whether it is accurate and trustworthy, etc. In order to further understand the current status of interpretability research in reinforcement learning, this article discusses from the aspects of interpretable models, interpretable strategies, environment interaction and visualization, etc. Based on this, this article systematically discusses the research status of reinforcement learning interpretability, classifies and explains its explainable methods, and finally proposes the future development direction of reinforcement learning interpretability.

[2023-12-18] Accepted Paper /doc/2023.09.0430/v1