A multi-agent-based joint optimization method for intersection signal control and vehicle trajectories

Abstract

The development of Cellular Vehicle-to-Everything (C-V2X) communication technology has accelerated the evolution of urban transportation systems toward connectivity and intelligence. Traditional signal control methods exhibit limited adaptability to time-varying traffic and are incapable of effective coordination with connected and automated vehicles (CAVs). To address mixed traffic conditions involving CAVs and human-driven vehicles (HDVs), this paper proposes a multi-agent-based joint optimization method for intersection signal control and vehicle trajectories, termed Signal-Trajectory Optimizing Collaborative Agents (STOCA). A two-layer collaborative optimization architecture is developed by designing a signal control agent and a CAV trajectory control agent, which enhances intersection traffic efficiency while allowing CAVs to jointly optimize driving comfort and safety. To improve the robustness of trajectory control under non-ideal C-V2X conditions, delayed observations are further introduced into the CAV trajectory-control training and evaluation process. To balance overall traffic efficiency and individual benefits of controlled vehicles, a joint reward function integrating signal control and CAV trajectory control is designed. A Tree-Structured Parzen Estimator (TPE)-based offline hyperparameter optimization strategy is adopted to calibrate the weight of the CAV trajectory-cost term in the joint reward. Simulation results based on a digital intersection in Guilin, China, show that the proposed method reduces average delay by 29.71 %, and decreases average queue length by 22.43 % compared with baseline methods. Sensitivity analysis indicates improved performance with higher CAV penetration and good applicability under low-penetration peak traffic conditions.