Research on regional signal control algorithm based on multi-agent DDQN

Traditional traffic signal timing control and single-intersection control methods face challenges in addressing the complexity and dynamics of modern traffic systems. Methods such as Discrete Traffic State Code(DSTE) and image-like approaches have been employed to address difficulties in setting the state space. This paper introduces a multi-agent deep reinforcement learning algorithm based on the Double Deep Q-Network (DDQN) and proposes a state space setting approach utilizing one-hot coding to effectively reduce data dimensionality. By leveraging collaborative multi-agent reinforcement learning methods and "local centralization" strategies, issues such as dimensionality explosion and unstable training are successfully mitigated. To assess the algorithm's performance, a road network model was constructed using SUMO, demonstrating a 22% reduction in waiting time compared to the original algorithm. The results confirm that the regional signal control model and methodology presented in this paper can significantly decrease regional traffic delays and enhance the traffic efficiency of the road network.