Multi-agent Adversarial Inverse Reinforcement Learning with Latent Variables

We introduce an algorithm for inferring reward functions from expert human trajectories in multiagent environments. Current techniques exhibit poor sample-efficiency, lack stability in training, or scale poorly to large numbers of agents. We focus on settings with a large, variable number of agents and attempt to resolve these settings by exploiting similarities between agent behaviors. In particular, we learn a shared reward function using adversarial inverse reinforcement learning and a continuous latent variable. We demonstrate our algorithm on two real-world settings: traffic on highways and in terminal airspace.