Adversarial Inverse Reinforcement Learning to Estimate Policies from Multiple Experts

Abstract

Inverse reinforcement learning is used for complex control tasks by using experts. However, since the learning results depend on the expert, it is impossible to imitate ungiven policies from expert when there are multiple optimal polices for the same goal, or when the environment changes from the training. The problems can be solved by giving multiple experts and representing their features in the latent space. the proposed method extends information maximizing generative adversarial imitation learning with adversarial inverse reinforcement learning to deal with such environment. Experiments show that the proposed method can not only imitate multiple experts, but also estimate ungiven polices.