A traffic signal control method based on deep reinforcement learning adopts near-end strategy optimization, generalized advantage estimation and other technologies to improve the overall performance of a model, then designs a new reward value function to significantly improve the convergence rate of the model, and moreover, the phase sequence of a signal timing scheme generated by the model is fixed, the timing difference between the adjacent periods is small, higher safety and practicability are achieved, finally, the trained model can be continuously optimized in the actual traffic environment, and the changing traffic requirements are better met. Experimental results based on real traffic flow data show that the method can effectively cope with two flow modes of flat peak and peak, compared with other models, common evaluation indexes such as average queuing length, average travel time and average vehicle delay are obviously reduced, and model convergence time is also obviously shortened.
一种基于深度强化学习的交通信号控制方法,首先采用了近端策略优化和广义优势估计等技术提升了模型的整体性能,其次设计了新的奖励值函数,显著提升了模型的收敛速度,而且,模型生成的信号配时方案相位顺序固定,相邻周期配时差异较小,具有更高的安全性和实用性,最后,训练好的模型可以在实际交通环境中进行持续优化,更好地满足变化的交通需求。基于真实交通流量数据的实验结果表明:该方法能够有效应对平峰和高峰两种流量模式,与其它模型相比,平均排队长度、平均旅行时间和车辆平均延误等常见评价指标明显降低,而且模型收敛的时间也明显缩短。
Traffic signal control method based on deep reinforcement learning
一种基于深度强化学习的交通信号控制方法
2021-01-12
Patent
Electronic Resource
Chinese
Deep Reinforcement Learning-Based Traffic Signal Control
ASCE | 2024
|Deep Reinforcement Learning-based Traffic Signal Control
IEEE | 2023
|Multi-intersection traffic signal control method based on deep reinforcement learning
European Patent Office | 2023
|Multi-intersection traffic signal control method based on deep reinforcement learning
European Patent Office | 2024
|Reinforcement learning-based traffic signal control
IEEE | 2024
|