Compared with an ecological driving strategy based on an optimization method, the intelligent network connection automobile signal intersection passing strategy based on safety reinforcement learning has the advantages that the calculation time is shorter, and the real-time performance is higher; compared with a traditional ecological driving strategy based on reinforcement learning, the method is designed through a safety layer error correction mechanism, violation of safety constraints in the training process is ingeniously avoided, the safety performance is better, and the actual application value is higher. According to the method, the nonlinear traffic light constraint is converted into the time-varying linear state constraint aiming at the upper and lower reference trajectories of the vehicle position, so that the problem of reward sparseness during reinforcement learning agent trial-and-error training is effectively solved.
本发明提供了一种基于安全强化学习的智能网联汽车信号交叉口处通行策略,其相比基于优化方法的生态驾驶策略,计算时间更少,实时性更高;相比传统基于强化学习的生态驾驶策略,本发明通过安全层纠错机制设计,巧妙地避免了在训练过程中安全约束的违反,安全性能更好,实际应用价值更高。本发明针对车辆位置的上下参考轨迹将非线性交通灯约束转化为时变线性状态约束,从而有效解决了强化学习智能体试错训练时面临的奖励稀疏问题。
Intelligent network connection automobile signal intersection passing strategy based on safety reinforcement learning
基于安全强化学习的智能网联汽车信号交叉口处通行策略
2024-07-05
Patent
Electronic Resource
Chinese
Intersection passing method and system for intelligent network connection vehicles
European Patent Office | 2023
|Intersection intelligent network connection vehicle control method based on reinforcement learning
European Patent Office | 2024
|European Patent Office | 2025
|European Patent Office | 2023
|European Patent Office | 2022
|