Double deep network-based traffic signal optimization method for isolated intersections

This study addresses the limitations of existing reinforcement learning (RL)-based traffic signal control methods, which typically optimize either the signal phase sequence or phase duration independently. We propose a novel joint optimization framework based on the Double Deep Q-Network (DDQN) that simultaneously determines both the phase sequence and phase duration. To ensure stability, the base phase duration is determined using the classical Webster method. Furthermore, a hybrid state representation is developed by integrating both microscopic and macroscopic traffic features, such as queue length and vehicle delay. A Squeeze-and-Excitation (SE) attention mechanism is introduced to guide the agent's attention toward critical traffic attributes. Simulation experiments conducted on the SUMO platform demonstrate that the proposed method significantly reduces average queue length and vehicle travel time when compared to traditional fixed-time and vehicle-actuated control strategies, particularly under medium to high traffic demand. The results validate the effectiveness, robustness, and practical applicability of the method for intelligent signal control in complex urban intersections.