Competitive Multi-agent Inverse Reinforcement Learning with Sub-optimal Demonstrations