MA-TREX: Mutli-agent Trajectory-Ranked Reward Extrapolation via Inverse Reinforcement Learning.

Trajectory-ranked reward extrapolation (T-REX) provides a general framework to infer users’ intentions from sub-optimal demonstrations. However, it becomes inflexible when encountering multi-agent scenarios, due to its high complexity caused by rational behaviors, e.g., cooperation and communication. In this paper, we propose a novel Multi-Agent Trajectory-ranked Reward EXtrapolation framework (MA-TREX), which adopts inverse reinforcement learning to infer demonstrators’ cooperative intention in the environment with high-dimensional state-action space. Specifically, to reduce the dependence on demonstrators, the MA-TREX uses self-generated demonstrations to iteratively extrapolate the reward function. Moreover, a knowledge transfer method is adopted in the iteration process, by which the self-generated data required subsequently is only one third of the initial demonstrations. Experimental results on several multi-agent collaborative tasks demonstrate that the MA-TREX can effectively surpass the demonstrators and obtain the same level reward as the ground truth quickly and stably.