A Study of Explainability Inquiry Based on Reinforcement Learning

With the application of Reinforcement Learning (RL) in security-sensitive fields such as healthcare, autonomous driving, and finance, the lack of Explainable Reinforcement Learning (XRL) restricts the technology's grounding and social trust. This paper reviews XRL's progress, constructing a framework covering core challenges, methods, and applications. Core challenges include black-box decision-making, reward bias, and complex multi-intelligence interaction. Among the existing methods, the intrinsic interpretability of the model is limited by the generalization ability, the ex-post interpretation method faces the problem of poor local-global consistency, and the hybrid method is difficult to dynamically adapt due to the high complexity of the system. XRL shows potential in medical decision-making, autonomous driving safety, and financial risk control through causal reasoning and multimodal interpretation, but further optimization is needed. This paper emphasizes building a standardized evaluation system and explores cutting-edge directions like cross-domain migration and multimodal human-computer collaborative interpretation to deepen XRL's theoretical framework and promote its safe application in high-risk domains.