Survey Unifies Reinforcement Learning Policy Verification Methods

Reinforcement Learning (RL) is increasingly being applied in complex and safety-critical environments, yet a significant hurdle to its widespread adoption is the lack of robust guarantees regarding the behavior of neural network-based policies. The field of RL policy verification, which aims to provide these assurances, has grown rapidly but remains conceptually fragmented. This comprehensive survey aims to bring clarity and structure to this diverse body of work. The survey introduces a novel taxonomy that categorizes existing RL verification methods based on three key axes: the verification paradigm (formal versus probabilistic approaches), the temporal scope (step-wise versus multi-step analysis), and the strength of the guarantees provided. Beyond classification, the paper unifies the underlying theoretical foundations of these methods, explicitly detailing their implicit assumptions and limitations. It also highlights promising emerging directions for future research, offering a consolidated view for researchers and practitioners navigating this critical area.