Explainable Deep Reinforcement Learning: State of the Art and Challenges

Interpretability, explainability, and transparency are key issues to introducing artificial intelligence methods in many critical domains. This is important due ethical concerns trust strongly connected reliability, robustness, auditability, fairness, has consequences toward keeping the human loop high levels of automation, especially cases for decision making, where both (human machine) play roles. Although research community given much attention explainability closed (or black) prediction boxes, there tremendous needs closed-box that support agents act autonomously real world. Reinforcement learning methods, their deep versions, such methods. In this article, we aim provide a review state-of-the-art explainable reinforcement taking also into account operators—that is, those who make actual decisions solving real-world problems. We formal specification problems, identify necessary components general framework. Based on these, comprehensive categorizing them classes according paradigm they follow, interpretable models use, surface representation explanations provided. The article concludes by identifying open questions challenges.