A Human-Like Agent Based on a Hybrid of Reinforcement and Imitation Learning

摘要

Reinforcement learning (RL) makes it possible to build an efficient agent that handles tasks in complex and uncertain environments by maximizing future reward. However, for applications in some areas like game AI and autonomous driving, efficiency only cannot satisfy the practical use, and a human-like agent is preferable. On the other hand, in imitation learning (IL) tasks, which trains the agent to mimic actions of expert behavior provided as training data and thereby learns relatively complex tasks while achieving human-like behavior. Unfortunately, the performance of such an agent is generally limited by the expert behavior. Thus, with the aim of training an agent which achieves high performance while retaining a human-like behavior, we propose a method for mixing RL and IL, applicable to both discrete and continuous problems. We used state-of-the-art RL and IL algorithms and trained their respective models independently, before mixing them into the proposed hybrid model.