Reward Compatibility: A Framework for Inverse RL