ABSTRACT: Car-following behavior plays a vital role in microscopic traffic simulation, intelligent transportation systems and advanced driver assistance systems. The key variables of the car-following behavior, such as speed, acceleration, and inter-vehicle distance, are closely related to rear-end collisions, passenger experience and traffic efficiency. How to control and optimize the car-following trajectories is an important research field in automated driving. Based on the Deep Reinforcement Learning (DRL), this study proposed a velocity control algorithm in the car-following scenario for automated driving. Field data, extracted from the 161,055 km of data collected in the Shanghai Naturalistic Driving Study (SH-NDS), are used for training and simulated data are used for testing the algorithm respectively in the automated driving simulation platform Car Learning to Act (CARLA). Through trial and error in simulation environment, the agent maximizes cumulative rewards to learn to control vehicle speed.Results show that the three parts of DRL correspond with each of the objectives: the minimal longitude distance in the Responsibility-Sensitive Safety (RSS) for ensuring safety, the jerk for keeping comfortable, and the lognormal distribution of time gaps for improving efficiency. Due to the DRL’s demonstrated high performance in car-following scenarios on freeways, it is reasonable to assume the algorithm may be used to handle car-following trajectory planning in other type of roadway.
Ping Sun, Han Chen, Xuesong Wang*. Automated Car-following Control Algorithm for Multiple Objectives on Freeways. Transportation Research Board 100th Annual Meeting, Washington D.C., USA, 2021. 1.25-29.