Optimal speed limit control for network mobility and safety: a twin-delayed deep deterministic policy gradient approach

Variable speed limit control (VSLC) has emerged as a promising approach for improving traffic safety and reducing congestion. However, local adjustment of VSLC may have broader impacts on the transportation network performance due to driver rerouting. This study proposes a deep reinforcement learning (DRL) controller based on twin-delayed deep deterministic policy gradient (TD3) algorithm to improve mobility and safety over a small-scale interconnected network considering rerouting behavior. The proposed DRL-based VSLC controller is designed to handle a large number of possible speed limits at each time step by utilizing a deep actor-critic framework. The study also experiments with different reward functions to characterize network mobility, safety, and traffic oscillation. Additionally, we investigate the sensitivity of the control algorithm across different traffic patterns, driving behavior, and VSLC locations, where the proposed TD3 algorithm demonstrated robustness and generalizability. Our findings indicate that implementing network-specific reward functions leads to improvements in traffic safety and mobility. Specifically, it results in a 3.84% enhancement in overall safety, as measured by time-to-collision metrics, and a 33.2% improvement in mobility by reducing total travel time compared to the scenario without VSL control. While comparable in safety performance, TD3 outperforms deep deterministic policy gradient (DDPG) algorithm by 15.1% in terms of mobility. This study contributes to the understanding of the impacts of VSLC on transportation networks and provides insights into effective ways of implementing VSLC to improve network mobility and safety.