Key points are not available for this paper at this time.
Most freeway accidents do not cause complete traffic breakdowns but create bottlenecks that significantly compromise traffic safety and efficiency, making safe and efficient vehicle passage essential. This paper proposes a proactive emergency traffic control architecture based on deep reinforcement learning to optimize traffic flow in accident-induced bottlenecks. Based on the traffic flow mechanism analysis in bottlenecks, a combined control strategy, integrating hierarchical differential variable speed limits (HDVSL) and lane change guidance (LCG), is devised to improve both macroscopic traffic coordination and microscopic vehicle behavior. Furthermore, to overcome perceptual limitations under complex traffic dynamics, we developed the Recurrent Soft Actor-Critic (R-SAC) algorithm. The ability to capture temporal sequential dependencies is accomplished by embedding a recurrent temporal modeling module within the policy network. Additionally, the memory and state augmentation module further enhances the adaptability to dynamic traffic patterns. A simulation-based evaluation was conducted in high-accident frequency mainline scenarios, and the results demonstrate that the proposed approach significantly outperforms existing strategies and models in both safety and efficiency. Specifically, it improves safety by 38.6% (measured by speed standard deviation) and efficiency by 35.7% (measured by travel time through the bottleneck) compared to the no-control scenario, and further surpasses the soft actor-critic algorithm by 5.4% and 5.9%, respectively. Moreover, the systematic evaluations of state engineering and generalization establish a principled basis for effective state selection and verify the model’s robustness, providing theoretical insights and practical guidance for emergency traffic management.
Li et al. (Wed,) studied this question.