摘要
强化学习(reinforcement learning,RL)由于其解决高度动态环境中复杂决策问题的能力,成为信号灯控制中一种具有前景的解决方案。大多数基于强化学习的方法独立生成智能体的动作,它们可能导致交叉口的动作冲突、道路资源浪费。因此,本文提出了基于图卷积网络和长短期记忆(graph convolution network-long short-term memory,GCNLSTM)的多交叉口信号灯控制方法。首先,基于二进制权重网络对多交叉口进行构图。其次,通过图卷积网络聚合周围交叉口的空间状态信息,利用长短期记忆(long short-term memory,LSTM)获得交叉口的历史状态信息。最后,通过基于竞争网络框架的Q值网络进行动作的选择,实现对交叉口相位的控制。实验结果表明,与其他强化学习方法相比,本文方法在多交叉口的信号灯控制中能够减少交叉口的队列长度,并使道路网络中的车辆获得更少的等待时间。
Reinforcement learning(RL)has become a promising solution in signal control because of its ability to solve complex decision-making problems in a highly dynamic environment.Most methods based on reinforcement learning generate agent actions independently,which may lead to action conflicts at intersections and waste of road resources.Therefore,a multi intersection signal control method based on graph convolution network-long short-term memory(GCN-LSTM)is proposed.Firstly,multi intersections are mapped based on binary weight network.Secondly,long short-term memory(LSTM)obtains the historical state information of intersections by aggregating the spatial state information of surrounding intersections through graph convolution network.Finally,the Q value network based on the competitive network framework is used to select actions to control the intersection phase.The experimental results show that compared with some reinforcement learning methods,the queue length at intersections and the waiting time of vehicles in the road network can be reduced in the signal light control at multiple intersections.
作者
徐东伟
朱宏俊
郭海锋
周晓刚
汤立新
XU Dongwei;ZHU Hongjun;GUO Haifeng;ZHOU Xiaogang;TANG Lixin(Institute of Cyberspace Security,Zhejiang University of Technology,Hangzhou 310023;College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023;Ningbo Ninggong Transportation Engineering Design Consulting Co.,Ltd.,Ningbo 315010;Zhejiang Expressway Company Limited,Hangzhou 311500)
出处
《高技术通讯》
北大核心
2025年第5期472-479,共8页
Chinese High Technology Letters
基金
国家自然科学基金(62373325,6190334,52072343)
浙江省自然科学基金(LY21F030016,LY20E080023,LQ16E080011)资助项目。
关键词
智能交通系统
交通信号灯控制
多智能体强化学习
长短期记忆
图卷积网络
intelligent transportation system
traffic light control
multi-agent reinforcement learning
long short-term memory
graph convolution network
