Explainable Reinforcement Learning: A Survey

The "Performance-Transparency Trade-off" is a fundamental challenge in AI: More complex AI models tend to be more powerful but less understandable. This is particularly concerning for reinforcement learning systems that learn autonomously, where understanding why a system made a decision is critical for responsible deployment.

Most explainability solutions work backwards rather than forwards: The survey found that the majority of explainable reinforcement learning methods attempt to explain complex models after the fact by mimicking and simplifying them, rather than designing inherently simpler and more transparent systems from the start.

Current XAI methods ignore human factors: A significant gap exists in how explainability research is conducted—most methods neglect insights from psychology and philosophy about how humans actually understand and trust explanations. This means explanations may be technically accurate but fail to meaningfully communicate with non-expert users.

Interdisciplinary collaboration is essential for progress: The survey emphasizes that advancing explainable AI requires moving beyond computer science alone to incorporate expertise from psychology, philosophy, and other fields that study human understanding and decision-making.

Autonomous learning systems demand urgent attention to explainability: Because reinforcement learning systems make decisions independently without explicit programming, understanding their reasoning is not optional—it's a prerequisite for safe and trustworthy deployment in real-world applications.