Cooperative Traffic Light Control with Spatio-Temporal Multi-Agent Reinforcement Learning using Neural Networks

Abstract

Traffic congestion remains a major challenge in urban areas, necessitating innovative approaches to optimize traffic flow and reduce commuting times. This research proposes a novel approach for traffic light control that leverages the power of spatio-temporal multi-agent reinforcement learning (MARL) combined with neural networks.Traditional traffic light control systems often rely on fixed-time or simple rule-based strategies, lacking adaptability to dynamic traffic conditions. In contrast, the proposed approach embraces a cooperative multi-agent framework, where each traffic light intersection acts as an independent agent capable of making autonomous decisions.The MARL architecture employs deep neural networks to encode the traffic environment's spatio-temporal data, enabling agents to perceive traffic conditions comprehensively. By sharing information with neighboring agents and collaborating towards a common goal, the agents learn to make coordinated decisions that optimize traffic flow across the entire road network.The learning process is guided by a reward function that encourages the reduction of overall travel time, minimizing congestion, and enhancing intersection efficiency. Reinforcement learning algorithms, such as Q-learning or Deep Q Networks (DQNs), are adapted to train the agents to act optimally in diverse traffic scenarios.