Citation

Shi Y, Liu J, Liu C, Gu Z. Transp. Res. A Policy Pract. 2024; 183: e104069.

Copyright

(Copyright © 2024, Elsevier Publishing)

DOI

10.1016/j.tra.2024.104069

PMID

unavailable

Abstract

Autonomous vehicles have the potential to revolutionize intelligent transportation by improving traffic safety, increasing energy efficiency, and reducing congestion. In this study, a novel framework termed DeepAD was proposed and validated for decision making in intelligent autonomous driving via deep reinforcement learning. This framework incorporates multiple driving objectives such as efficiency, safety, and comfort to make informed decisions regarding autonomous vehicles (AVs). The decision-making process utilizes the origin-destination information for macrolevel routing and determines microlevel car-following and lane-changing behaviors. The lane-changing behavior is discretized and learned through a deep Q-network, and the continuous car-following behavior is learned through a deep deterministic policy gradient. Comprehensive simulation experiments on a real-world network demonstrated that DeepAD outperformed human driving while maintaining a desirable level of efficiency, safety, and comfort. In the real-world road networks experiment, multiple indexes of vehicles in the high AVs penetration rate group significantly outperformed that of the group with lower AVs penetration rate. Overall, the proposed framework provides insights into intelligent autonomous driving to improve urban mobility.

Keywords

Autonomous vehicle; Car following; Decision making; Deep reinforcement learning; Lane changing; Traffic simulation