Explaining Reinforcement Learning Agents via Inductive Logic Programming

Logic-based explanations can be systematically measured: The paper introduces objective metrics (activation rate, feature coverage, syntactic distance, and semantic distance) to quantify how well symbolic rules explain RL agent behavior, moving beyond subjective assessments and addressing a gap in explainable AI where "the systematic quantification of the explainability degree of logical representations remains an open problem."

Symbolic rules offer compact, human-readable alternatives to black-box policies: By using Inductive Logic Programming to extract symbolic representations of RL policies, the approach produces "compact, human-readable abstractions of decision-making" that can be understood by non-experts, which is particularly important for safety-critical applications.

Explainability metrics reveal how agents make decisions: The proposed metrics track which features matter most in decision-making (feature coverage), how often rules are applied (activation rate), and how policies evolve during training, providing transparency into agent behavior across different scenarios and multiple agents.

This addresses a critical gap in responsible AI evaluation: The paper tackles the problem that explainable RL has "mostly been based on user studies, thus targeting the needs of a specific audience and lacking shared evaluation metrics," establishing standardized ways to assess whether explanations are actually meaningful.

The approach works across different RL environments: Experiments demonstrate the metrics' effectiveness "across different RL domains," suggesting the framework could be broadly applicable to various real-world applications requiring transparent AI decision-making.