Multi-Agent Reinforcement Learning Framework for Intelligent Traffic Signal Optimization in Urban Transportation Systems

Abstract

Efficient and dynamic traffic light control for intelligent transportation systems has become a pressing issue as cities have grown in recent years. A multi-agent reinforcement learning (MARL) system for dynamic signal optimization in extensive metropolitan road networks will be presented in this study. A novel approach for signal-timing plan optimization at connected intersections has been put forth that combines deep policy learning, distributed agent coordination, and real-time traffic-condition sensing. Construct a 25–49 intersection simulated city grid and conduct experiments with various inflow rates and realistic accident scenarios. In comparison to the fixed-time and actuated signal baselines, the results indicate that the MARL system has reduced the average vehicle delay by up to 30% and the average queue length by roughly 23%; network throughput has increased by approximately 15%. The approach exhibits good resilience to such issues, is comparatively steady in the face of heavy traffic and other disruptions, and continues to function normally for a considerable amount of time. According to the study, multi-agent learning can use context-sensitive policy frameworks and robust communication capabilities to solve the issue of urban mobility's lack of scale. This research provides a foundation for the further use of distributed AI in intelligent traffic control.