A Survey of Explainable Reinforcement Learning: Targets, Methods and Needs

Reinforcement learning systems lack transparency: The paper identifies that AI agents trained through reinforcement learning operate as "black boxes," making it difficult to understand why they take specific actions—a critical concern for deploying these systems in real-world applications where decisions affect people's lives.

Explainability requires targeting different audiences: The survey's "What" and "How" taxonomy reveals that effective explanations must be tailored to different stakeholders (developers, regulators, end-users), as different groups need different types of information about agent behavior—there is no one-size-fits-all explanation approach.

The field is rapidly expanding but fragmented: With over 250 papers reviewed, explainable reinforcement learning is an active research area, yet the lack of standardized methods and terminology suggests the field still needs to establish common frameworks and best practices for practitioners implementing these systems.

Multiple explanation methods exist but lack evaluation standards: The survey identifies various approaches to explaining RL agent decisions, but highlights a critical gap: there are insufficient standardized ways to evaluate whether explanations are actually useful, accurate, or trustworthy—making it hard to know which methods work best in practice.

Bridging the gap between research and real-world deployment remains a key challenge: The paper identifies unmet needs in the field, suggesting that while XRL research is advancing, significant work remains to make these explanation methods practical and reliable for high-stakes applications like autonomous vehicles, healthcare, or criminal justice systems.