[3K4-J-2] Machine learning: real world interaction

[3K4-J-2-04] Comfortable Driving by Deep Inverse Reinforcement Learning

〇Daiko Kishikawa¹, Sachiyo Arai¹ (1. Chiba University)

Keywords:Inverse Reinforcement Learning, Automatic Driving

For the realization of automatic driving, not only safety but also comfortability of passengers is required for its application to the real society. We define it as comfortable driving. Comfortable driving is hard to define because the expectation for comfortability varies according to the designer of system. Therefore comfortable driving is difficult to code a rule-based algorithm manually. Reinforcement learning, which learns an optimal policy from trial-and-error by the agent, is an effective method to solve this problem. However it requires a reward function for the appropriate evaluation of action taken by the agent. In this paper we propose an approach to obtain comfortable driving by training with the reward function estimated from trajectories of comfortable driving, using deep inverse reinforcement learning. Experimentally we used low lateral acceleration as the condition of comfortable driving, and we were able to estimate a reward function with satisfying it.