融合注意力机制LSTM网络的城市交叉口信号控制  

City intersection signal control based on fusion attention mechanism LSTM network

作　　者：陈国梁 石晴[1] 黄亚飞 曾昭汰 CHEN Guoliang;SHI Qing;HUANG Yafei;ZENG Zhaotai(Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学现代制造技术教育部重点实验室,贵阳550025

出　　处：《重庆理工大学学报（自然科学）》2024年第4期196-203,共8页Journal of Chongqing University of Technology：Natural Science

基　　金：贵州省基础研究计划项目(黔科合基础-ZK〔2023〕一般056);贵州大学引进人才科研项目(贵大人基合字〔2021〕56号);贵州大学现代制造技术教育部重点实验室开放基金项目(GZUAMT2021KF(13)号);贵州省创新人才团队项目(CXTD2022-009)。

摘　　要：随着国内机动车保有量的快速提升,城市道路交叉口场景在高密度车流情况下易导致长距离拥堵。为降低交叉口车流拥堵长度,深度强化学习逐渐应用于交叉口信号的控制。而现有交叉口信号控制策略存在对车流状态信息的权重特征欠考虑及时序特征难提取的问题,因此基于深度Q学习(deep Q-learning,DQL)算法提出了一种改进DQL算法,利用注意力机制增强长距离拥堵状态信息的权值,进一步采用长短期记忆网络(long short-term memory,LSTM)学习车流的历史数据,改进了DQL算法对数据不同部分重要性考虑不足及历史数据信息提取欠佳的问题。实验结果表明,所提改进DQL算法与原算法相比,能够降低20%的车辆累计等待时间并且减少21.2%的车辆平均排队数目,提高了交叉口的车辆通行效率。With agrowing number of vehicles in China,urban road intersections easily experience heavy congestion under high-density traffic conditions.To ease the traffic congestion at intersections,Deep Reinforcement Learning has been gradually applied to intersection signal control.However,existing intersection signal control strategies fail to consider the weighted features and have difficulties in extracting temporal features of traffic flow status information.To address these issues,this study proposes an improved DQL algorithm based on the Deep Q-learning(DQL)algorithm.The algorithm utilizes attention mechanisms to enhances the weight of long-distance congestion state information within intersections.Furthermore,a Long Short-Term Memory(LSTM)network is employed to learn the historical data of traffic flow,addressing the issues of insufficient consideration of the importance of different parts of the data and sub-optimal extraction of historical data information in DQL algorithm.Our experimental results show that the improved DQL algorithm is able to reduce the cumulative waiting time of vehicles by 20%and the average number of queued vehicles by 21.2%compared with the original algorithm,improving the efficiency of vehicle passage at intersections.

关 键 词：交叉口信号控制 深度强化学习 注意力机制 LSTM 通行效率 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]