Urban traffic management faces growing challenges due to increasing volumes of vehicles and pedestrians, resulting in congestion, delays, and safety concerns. This study introduces an innovative traffic signal control framework that integrates Multi-Agent Reinforcement Learning (MARL) with Visible Light Communication (VLC) and a Self-Adaptive Phase Adjustment (SAPA) module to enhance coordination across urban intersections. In the proposed architecture, each intersection is governed by an independent Deep Reinforcement Learning (DRL) agent capable of making real-time, decentralized decisions for both vehicular and pedestrian flows. Cooperative behavior emerges through the MARL framework, allowing agents to account for the dynamic states of neighboring intersections and improve network-wide efficiency. VLC provides high-resolution, low-latency data exchange between vehicles, pedestrians, and infrastructure, enabling precise sensing of position, speed, queue length, and stop duration. The introduction of the SAPA module further enhances adaptability by dynamically adjusting phase durations based on real-time queue/request/response patterns, resolving conflicts and prioritizing urgent demands. Extensive simulations and field tests demonstrate that the MARL–VLC–SAPA system significantly outperforms centralized and conventional agent-based approaches, reducing waiting and travel times while improving overall safety and responsiveness in complex urban environments.
Vieira et al. (Fri,) studied this question.