Enhancing RL Generalizability in Robotics through SHAP Analysis of Algorithms and Hyperparameters

This article presents a novel framework utilizing SHapley Additive exPlanations (SHAP) to analyze the impact of algorithms and hyperparameters on the generalization performance of Reinforcement Learning (RL) in robotics. It establishes a theoretical link between Shapley values and RL generalizability, revealing consistent configuration impacts across various tasks, which leads to improved generalization through SHAP-guided configuration selection. This approach offers practitioners a systematic method for optimizing RL configurations, potentially enhancing deployment efficacy in real-world scenarios.