Explainable Reinforcement Learning Achieved via Prolog Expert Systems

Deep reinforcement learning (RL) policies are typically opaque "black boxes," making their decisions difficult to understand or trust. This research introduces a method to transform such policies into explainable, executable logic programs using Prolog expert systems. The process involves three post-hoc stages: first, extracting a frozen Proximal Policy Optimization (PPO) teacher; second, inducing an ordered rule list from its decisions using classical relational learning techniques; and third, emitting the result as a Prolog program. This program's decisions can be executed by an off-the-shelf logic engine, and it can be edited, with changes accepted only if policy evaluation certifies a return increase. The paper provides four key guarantees, including a return-loss bound that makes the distilled program a machine-checkable certificate in finite Markov Decision Processes, and a monotonically improving, terminating expansion loop. For continuous-observation settings, it proves that conversion to arbitrary fidelity is possible, with disagreement and return gaps closing at a rate of O(1/B) as resolution B grows. Empirically, the expanded Prolog program achieved exact optimal return on a two-room key-and-door task and matched or recovered a significant portion of the neural teacher's return on continuous-control tasks like Acrobot and CartPole, demonstrating its practical viability and explainability benefits.