In this paper, an optimization framework of Autonomous Vehicles (AVs) multi-agent cooperative control algorithm driven by simulation platform is proposed, aiming at solving the problems of poor adaptability, high computational complexity and high verification cost of real vehicles in dynamic traffic environment. Firstly, a high-fidelity simulation platform is built, which integrates vehicle dynamics model, intelligent driver model and V2X communication delay model to support large-scale parallel testing and extreme scene reproduction. Second, a distributed collaborative control algorithm is designed by integrating Model Predictive Control (MPC) with Multi-Agent Deep Reinforcement Learning (MADRL), enabling hierarchical decision-making that balances short-term trajectory optimization with long-term strategy learning. Furthermore, a Bayesian optimization strategy driven by simulation data is proposed, which automatically optimizes MPC weight and MADRL superparameter, and combines with anti-sample training to enhance the robustness of the algorithm. The experimental results show that this method can achieve zero collision, shorten the average transit time by 29% and increase the system throughput by 19% in the high-density intersection scene, which is significantly superior to the traditional rule method and single MPC strategy. This study provides a low-cost, high-efficiency algorithm development and verification path for autonomous multi-vehicle cooperative system, and has good engineering migration potential.
Ni et al. (Sun,) studied this question.