New Benchmark Evaluates RL Explainability by Bug Fixing

Current methods for evaluating Explainable Reinforcement Learning (XRL) often rely on abstract metrics like faithfulness or subjective human ratings. This preliminary paper outlines a new, more practical benchmark called EvalXRL, which aims to assess XRL methods by their real-world utility: how well they help fix bugs in RL agents. EvalXRL proposes using a Large Language Model (LLM) coding agent to interactively diagnose and repair malfunctioning RL agents. The LLM agent would utilize various XRL methods, process their outputs, form hypotheses about the malfunction, and then invoke the XRL method again with adjusted parameters to test these hypotheses, mimicking a scientific debugging process. The ultimate measure of an XRL method's effectiveness in EvalXRL would be the improvement in the RL agent's reward signal after the LLM coding agent attempts to fix it. This closed-loop evaluation provides a head-to-head comparison of XRL methods based on their direct impact on agent repair, offering a more concrete and actionable assessment than existing benchmarks.