基于DQN算法的单点交叉口信号控制被引量：1

Isolated Intersection Traffic Signal Control Based on Deep Reinforcement Learning

在线阅读下载PDF

导出

摘要我国城市交通信号控制系统是一个复杂的系统,交通流具有非线性、时变性、不确定性等特点,传统的建模和控制方式难以解决多变的交通流变化。为此,将深度强化学习方法引入交通控制系统中,构建一个单点路口的Agent模型,对交叉口进行智能化控制。首先,为了更加精确地描述交通状态,以车辆的排队数和车辆的平均速度进行交通状态空间设计。其次,采用灵活的相位组合方案,扩大强化学习的动作空间,奖励函数主要是采用交叉口累计延误、一段时间内交叉口通过的车辆数、交叉口排队长度和等指标系数的加权和。最后,以深圳市宝安区新湖路—金田路为研究对象,利用SUMO(Simulation of Urban MObility)仿真平台对算法进行验证。实验结果表明,与定时控制相对比,所提出方法的交叉口总延误,排队长度和车辆等待时间有不同程度的提升。 China's urban traffic signal control system is a complex system,with traffic flow characteristics such as nonlinearity,time-varying,and uncertainty.Traditional modeling and control methods are difficult to solve the variable traffic flow changes.Therefore,the deep reinforcement learning method is introduced into the traffic control system,and an agent model of a single point intersection is constructed to intelligently control the intersection.Firstly,in order to describe the traffic state more accurately,the traffic state space is designed based on the number of vehicles queuing and the average speed of vehicles.Secondly,the flexible phase combination scheme is adopted to expand the action space for strengthening learning,and the reward function mainly uses the weighted sum of index coefficients such as the cumulative delay at the intersection,the number of vehicles passing through the intersection within a period of time,and the sum of queue lengths at the intersection.Finally,taking Xinhu Road—Jintian Road,Bao'an District,Shenzhen City as the research object,the algorithm was validated using the SUMO(Simulation of Urban MOBILITY)simulation platform.The experimental results show that compared with timing control,the total delay of the cross port of the proposed method,the length of queuing length and the waiting time of the vehicle have increased to varying degrees.

作者陈元洁 CHEN Yuanjie(Shenzhen Urban Transport Planning Center Co.,Ltd,Shenzhen 518000,China)

机构地区深圳市城市交通规划设计研究中心股份有限公司

出处《交通与运输》 2023年第S01期89-94,共6页 Traffic & Transportation

关键词交通工程交通信号控制深度强化学习神经网络 SUMO Traffic engineering Traffic signal control Deep reinforcement learning Neural Networks SUMO

分类号 U491 [交通运输工程—交通运输规划与管理]

作者简介陈元洁(1994-),女,汉族,广东汕头人,学士,助理工程师,主要研究方向:交通控制。