New Metrics Explain Reinforcement Learning Agent Behavior.

Researchers are advancing the field of Explainable Reinforcement Learning (XRL) by proposing a novel approach that moves beyond subjective user studies. Their method leverages Inductive Logic Programming (ILP) to extract symbolic, human-readable representations of RL policies, providing a more objective way to understand agent behavior. The core contribution is a new set of explainability metrics, including activation rate, feature coverage, syntactic distance, and semantic distance. These metrics allow for a systematic quantification of how well logical rules align with agent actions, the importance of different features in decision-making, and how policies evolve during training or across multiple agents. Experiments across various RL domains demonstrate that these metrics offer detailed insights into action-specific learning dynamics, feature importance, and coordination patterns in multi-agent systems. This work provides a principled way to quantify the explainability of logical rules, offering crucial information for policy transfer, generalization, and understanding complex AI decisions.