Proactive Car-Following Using Deep-Reinforcement Learning

Car-following is a fundamental operation for vehicle control for both ADAS on modern vehicles and vehilce control on autonomous vehicles. Most existing car following mechanisms react to the observations of nearby vehicles in real-time. Unfortunately, lack of capability of taking into account multiple constraints and objectives, these mechanisms lead to poor efficiency, discomfort, and unsafe operations. In this paper, we design and implement a proactive car-following model to take into account safety regulation, efficiency, and comfort using deep reinforcement learning. The evaluation results show that the proactive model not only reduces the number of inefficient and unsafe headway but also eliminates the traffic jerk, compared to human drivers. The model outperformed 79% human drivers in public data set and the road efficiency is only 2% less than the optimal bound. Compared to ACC model, the DDPG model allows 4.4% more vehicles to finish the simulation than ACC model does, and increases the average speed for 28.4%.