Extrapolating Beyond Suboptimal Demonstrations via Inverse...
A critical flaw of existing inverse reinforcement learning (IRL) methods is their inability to significantly outperform the demonstrator. This is because IRL typically seeks a reward function that...