Hierarchical joint optimization of traffic signals and vehicle trajectories at intersections via deep reinforcement learning

Existing methods typically focus on single-agent optimization, which struggles to achieve global optimality, or only consider longitudinal trajectory control while neglecting the traffic benefits of lane-changing. This study proposes a hierarchical joint optimization framework based on multi-agent deep reinforcement learning. At the macroscopic level, a traffic light agent is constructed using Proximal Policy Optimization (PPO) to dynamically adjust signal phases and green durations. At the microscopic level, a vehicle agent is developed based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to jointly optimize both lateral and longitudinal trajectories of Connected and Automated Vehicles (CAVs), facilitating rapid platooning and efficient intersection passage. Simulation results demonstrate that, compared with baseline schemes and comparable methods, the proposed framework achieves significant advantages in traffic efficiency: average speed improves by 7.54%-25.5%, queue length and loss time are reduced by 5.37%-19.3% and 11.45%-31.55%, respectively. Multi-intersection extension experiments validate the generalization capability of the framework, and safety analysis shows that the risk factor decreases most significantly with increasing penetration rate. Further analysis indicates that the framework achieves the most pronounced effectiveness under medium-to-high CAV penetration rates.