Deep Reinforcement Learning with Human Guidance for Autonomous Driving Car Following

To solve the complex modeling and difficult parameter adjustment of traditional car-following control methods, we propose the car-following framework KgPPO-CF. The KgPPOCF brings expert experience into the interactive process of Deep Reinforcement Learning, which improves the speed and robustness of training policies. We use four different reward functions and compare the effects of policies trained by different designs on car-following control. In the Carla simulator, KgPPO-CF achieves car-following control in a continuous action space. The experimental results show that, compared with the traditional control methods, the control strategy obtained by KgPPO-CF training has better adaptability.