Reducing Traffic Congestion with Reinforcement Learning-Driven Signal Control Systems

The road network is a critical component of modern urban mobility systems, with signalized traffic intersections playing a pivotal role. Traditionally, traffic light phase timings and duration at intersections are designed by transportation engineers using historical traffic data. Some modern intersections employ trigger-based mechanisms to improve traffic flow; however, these systems often lack global awareness of traffic conditions across multiple intersections within a network. With the increasing availability of traffic data and advancements in machine learning, traffic light systems can be enhanced by modeling them as agents operating in an environment. This paper proposes a Reinforcement Learning (RL)-based approach for multi-agent traffic light systems within a simulation environment. The simulation is calibrated using real-world traffic data, enabling RL agents to learn effective control strategies based on realistic scenarios. A key advantage of using a calibrated simulation is that RL agents can experiment with different control actions without compromising safety in real-world traffic. The proposed system demonstrates that RL agents can coordinate and learn optimal policies, effectively reducing overall vehicle wait times in heavy urban traffic scenarios. The reward function is carefully designed to minimize traffic congestion by reducing vehicle wait times. A comparative study between static phase timings, currently used by conventional controllers, and the RL-based policy highlights significant reductions in overall wait times. Additionally, this simulation-based approach allows RL agents to be deployed in real-time, continuously learning and adapting to live traffic data. By implementing updated control policies for traffic light phase timings, the system can effectively reduce congestion and improve traffic flow across the network.