Deep Reinforcement Learning for Vehicle Swarm Navigation and Urban Traffic Optimization

ABSTRACT Traffic congestion has become a prevalent phenomenon on urban roads, leading to significant challenges, including economic losses due to travel delays, increased fuel consumption, and air pollution from vehicle emissions. As it is impractical to extensively improve city road networks, vehicle routing optimization has emerged as a viable solution for alleviating congestion. However, traditional algorithms cannot effectively process information for complex and changeable traffic environments. By contrast, deep reinforcement learning (DRL) is a powerful approach for solving navigation problems. Rather than creating smart vehicles, we propose a navigation model to guide all vehicles. The model employs a graph neural network to effectively capture dynamic traffic flow patterns. We utilize the simulation of urban mobility to generate a large quantity of traffic data for use as reinforcement learning samples. We propose a parallel simulation training strategy to accelerate DRL convergence. We verify the effectiveness of our model by performing simulations on a simplified road network and a real‐life road network under multiple traffic scenarios and compare the results to those obtained using traditional methods and dynamic user equilibrium (DUE). The experimental results demonstrate that the proposed model reduces average travel time by up to 7.54% and the number of halting vehicles by up to 14.35% compared to traditional methods in high‐congestion scenarios, maintaining stability across various traffic conditions. The overall performance of the proposed method is comparable to that of DUE, indicating that traffic flow patterns mined through the deep network can be used to effectively deduce the optimal vehicle route without performing the iterations required for DUE. In general, the proposed approach has the capability to accommodate dynamic and complex traffic information, considerably mitigating traffic congestion.