With diverse environments, we can analyze, diagnose and edit deep reinforcement learning models using attribution.