Design of Intersection Control Algorithm Based on Deep Reinforcement Learning

Traditional traffic signal control (TSC) schemes rely on fixed timing or rule‑based adaptation, which lack real‑time responsiveness to stochastic and highly dynamic traffic flows. Deep reinforcement learning (DRL) has become a promising paradigm for adaptive signal control, yet value‑based methods such as DQN and DDQN suffer from overestimation bias, unstable training, and low sampling efficiency in complex urban intersection environments. This paper proposes a PPO-GAE control framework, which combines Proximal Policy Optimization (PPO) with Generalized Advantage Estimation (GAE) to achieve stable, efficient, and robust traffic signal timing optimization. PPO constrains policy updates within a trust region to avoid destructive updates and ensure training stability, while GAE effectively reduces gradient variance and improves the accuracy of advantage estimation for continuous traffic states. Experiments are carried out on the SUMO simulation platform using a real intersection in Xi’an. Results show that the proposed PPO-GAE method reduces cumulative delay by 41.2% and average queue length by 33.7% compared with the fixed‑time baseline, and outperforms DQN, DDQN-PER, and traditional adaptive methods.