Improvement and Evaluation of the Policy Legibility in Reinforcement Learning

When we work with intelligent agents, such as fighting a battle with other agents in computer games, it is difficult to achieve seamless collaboration if we can't figure out what the agents are doing. Especially in a complex problem domain, the agents are well trained and their actions could be too sophisticated to be comprehended by humans. In this article, we propose a novel reward shaping mechanism to improve the legibility of reinforcement learning that is used to train agents' policies. More importantly, we develop an interactive system to seek for users' evaluation of the policy legibility and show performance of the new learning approach.