From Black Box to Executable Logic: Explainable Reinforcement Learning through Prolog Expert Systems

Short summary

This paper presents a method to convert black-box deep reinforcement learning policies into executable Prolog programs that reproduce their behavior. The three-stage pipeline extracts a frozen PPO teacher, induces ordered rule lists, and emits Prolog code verified by a logic engine, with a subsequent expansion stage that monotonically improves returns. It proves formal guarantees including a return-loss bound and shows empirical success on discrete tasks, though continuous-control conversion faces an exponential cost in observation dimensionality as theory predicts.

• •Three-stage pipeline converts PPO policies into executable, human-readable Prolog programs with formal guarantees
• •Expanded Prolog program achieves exact optimal return on a 16,944-state discrete task across all seeds
• •Continuous-control conversion recovers 97% of teacher return on CartPole but only partially on LunarLander, matching the exponential lower bound prediction

Generated with AI, which can make mistakes.