Multi-agent Imitation Learning with Copulas

Multi-agent imitation learning aims to train multiple agents perform tasks from demonstrations by a mapping between observations and actions, which is essential for understanding physical, social, team-play systems. However, most existing works on modeling multi-agent interactions typically assume that make independent decisions based their observations, ignoring the complex dependence among agents. In this paper, we propose use copula, powerful statistical tool capturing random variables, explicitly model correlation coordination in Our proposed able separately learn marginals capture local behavioral patterns of each individual agent, as well copula function solely fully captures structure Extensive experiments synthetic real-world datasets show our outperforms state-of-the-art baselines across various scenarios action prediction task, generate new trajectories close expert demonstrations.