基于安全强化学习的交叉口多车协同决策方法OA

At urban intersections with multiple vehicles in motion,effectively characterizing the dynamic interactions between autonomous vehicles,optimizing right-of-way sequences,and ensuring driving safety remain significantly challenging.This paper proposes a collaborative control method based on safe multi-agent reinforcement learning to mitigate unsafe behaviors caused by the lack of safety constraints.First,the constrained Markov game(CMG)is integrated with multi-agent reinforcement learning.The Lagrange multiplier method and multi-agent Deep Deterministic Policy Gradient algorithm are employed to maximize the reward while minimizing the safety costs to limit dangerous behaviors.Next,a safety-awareness experience replay mechanism is proposed to mitigate the adaptability degradation and the potential for suboptimal solutions caused by over-restricting individual agents decision-making.Simulation results demonstrate the proposed algorithm outperforms existing baseline methods in safety,achieving a collision rate of 8.6%,effectively improving the collaborative decision-making capabilities of self-driving vehicles.