Explaining Reinforcement Learning Agents via Inductive Logic Programming

This paper uses Inductive Logic Programming (ILP) to extract symbolic representations of reinforcement-learning policies and proposes planning-oriented explainability metrics: activation rate, feature coverage, syntactic distance, and semantic distance. The authors report that these measures reveal action-specific learning dynamics beyond aggregate return, provide finer-grained feature analysis than conventional global feature-importance methods, and expose coordination, specialization, and adaptation patterns in multi-agent reinforcement learning. The framework is also applied to studying transfer and generalization of action-specific policies across agents and training stages.