Multi-agent Deep Reinforcement Learning-Based Decentralized Adaptive Traffic Signal Control for Corridor-Level Safety Optimization

Adaptive traffic signal management (ATSM) improves traffic flow by adjusting signal timings in real time. However, most ATSM systems overlook safety optimization, partly due to limited tools for real-time safety evaluation. With the rise of advanced sensing and vehicle technologies, there is increasing interest in integrating safety into ATSM. This study introduces a multi-agent decentralized deep reinforcement learning ATSM algorithm using proximal policy optimization (PPO) for corridor-level safety improvement. The algorithm was trained in a simulated environment using a safety reward based on real-time crash risk predictions from extreme value theory (EVT) models. A safety measure called return level (RL) quantified crash risk per signal cycle. Evaluation on a corridor with three signalized intersections showed that the proposed method reduced average RL by 27%, expected crashes by 35%, and delay by 21% compared to existing controllers. These results highlight the algorithm's potential to enhance safety without sacrificing traffic efficiency.