적대적 생성 모방학습 기반 종방향 운전자 모델에 관한 연구

Abstract

With recent improvements in AI technology, the application of artificial intelligence is being attempted in various research area. It is being used in the development of driver model or control design of autonomous vehicle. Especially, study on reinforcement learning or imitation learning algorithm is being actively researched. Imitation Learning is algorithm for mimicking given expert’s trajectory. Behavioral Cloning(BC), Dataset Aggregation(DAgger) and Inverse Reinforcement Learning(IRL) are kind of most known imitation learning method. In this paper, we propose an algorithm to develop human-like longitudinal driver model by using Generative Adversarial Imitation Learning(GAIL), which is type of Inverse Reinforcement Learning algorithm. Soft Actor Critic(SAC) RL algorithm is applied for interaction with longitudinal driving environment. Human driver’s driving data is obtained from Driver In the Loop Simlation environment by using expert trajectory for GAIL agent. Train result is compared between PI controller based model and Intelligent Driver Model(IDM) result. GAIL-based longitudinal driver model can generate more human-like velocity profile better than other methods.