Partial Simulation for Imitation Learning

Model-based imitation learning methods require full knowledge of the transition kernel for policy evaluation. In this work, we introduce the Expert Induced Markov Decision Process (eMDP) model as a formulation of solving imitation problems using Reinforcement Learning (RL), when only partial knowledge about the transition kernel is available. The idea of eMDP is to replace the unknown transition kernel with a synthetic kernel that: a) simulate the transition of state components for which the transition kernel is known (s_r), and b) extract from demonstrations the state components for which the kernel is unknown (s_u). The next state is then stitched from the two components: s={s_r,s_u}. We describe in detail the recipe for building an eMDP and analyze the errors caused by its synthetic kernel. Our experiments include imitation tasks in multiplayer games, where the agent has to imitate one expert in the presence of other experts for whom we cannot provide a transition model. We show that combining a policy gradient algorithm with our model achieves superior performance compared to the simulation-free alternative.