Multiagent Adversarial Inverse Reinforcement Learning

Learning to coordinate is a hard task for reinforcement learning due to a game-theoretic pathology known as relative overgeneralization. To help deal with this, we propose two methods which apply forms of imitation learning to the problem of learning coordinated behaviors. The proposed methods have a close connection to multiagent actor-critic models, and will avoid relative overgeneralization if the right demonstrations are given. We compare our algorithms with MADDPG, a state-of-the-art approach, and show that our methods achieve better coordination in multiagent cooperative tasks.