Operational Characteristics of Mixed-Autonomy Traffic Flow on the Freeway With On- and Off-Ramps and Weaving Sections: An RL-Based Approach

Abstract

As well as offering tremendous opportunities for existing transportation systems, autonomous driving presents challenges for traffic management, particularly in mixed traffic flow. The design of driving strategies for autonomous vehicles (AVs) considers both safety and efficiency to promote smooth traffic flow with no collisions. However, the complex traffic conditions at freeway bottlenecks mean that the traffic impacts of AVs in such scenarios remain unclear. Therefore, this study employs several driving strategies utilizing deep reinforcement learning (RL) to form a mixed-autonomy traffic flow on a one-way two-lane freeway with on- and off-ramps and weaving sections in SUMO traffic microsimulations. The study also examines the overall traffic impacts in terms of congested patterns, safety, efficiency, and passenger comfort. Multiple experiments are conducted with different AV driving strategies, penetration rates, and key traffic conditions. The results show that AVs with a deep-RL driving strategy significantly mitigate traffic-congested patterns and prevent the propagation of shockwaves, but at the expense of speed at low penetration rates. Moreover, the proposed driving strategy is conducive to a safer capability in vehicle dynamics. However, AVs under a deep-RL strategy lead to poorer passenger comfort at bottlenecks. Therefore, it is crucial to design a driving strategy with high robustness and scalability so that deep RL-based AVs can adapt to more complex scenarios and traffic events. Overall, the present research improves the understandings of AV driving mechanisms and provides insights into corresponding future traffic-management measures.