Decentralized smart traffic signal control using IoT-based multiagent reinforcement learning and VLC communication

Urban traffic management remains a significant challenge, especially during peak hours when high vehicle and pedestrian volumes cause congestion, delays, and safety risks. As expanding physical infrastructure is often impractical, intelligent and adaptive traffic signal control is essential. This work proposes a Multi-Agent Reinforcement Learning (MARL) framework that integrates Visible Light Communication (VLC) for real-time vehicle and pedestrian detection. VLC provides high-precision, low-latency information, enabling intersections to share local traffic states and coordinate decisions efficiently. Each intersection is managed by a dedicated agent, and the system compares the performance of Deep Q-Learning (DQN) and Multi-Agent Proximal Policy Optimization (MAPPO). Simulations in SUMO with five interconnected intersections under heterogeneous traffic patterns show that MAPPO achieves faster vehicle clearance, smoother flows, and lower pedestrian waiting times when compared to DQN. While DQN improves vehicle throughput in some cases, it compromises pedestrian performance. The addition of a Strategic Anti-Blocking Phase Adjustment (SAPA) module further enhances traffic fluidity and fairness. These results highlight the combined benefits of VLC and MAPPO, demonstrating scalable, adaptive, and safe urban traffic management, with MAPPO emerging as the most effective approach for coordinating multiple intersections in complex networks.