基于深度强化学习的自适应交通信号控制研究  

Adaptive Traffic Signal Control Based on Deep Reinforcement Learning

作　　者：徐建闽[1] 周湘鹏 首艳芳 XU Jianmin;ZHOU Xiangpeng;SHOU Yanfang(School of Civil and Transportation Engineering,South China University of Technology,Guangzhou 510640,Guangdong,China;Guangzhou Institute of Modern Industrial Technology,South China University of Technology,Guangzhou 510640,Guangdong,China)

机构地区：[1]华南理工大学土木与交通学院,广东广州510640 [2]华南理工大学广州现代产业技术研究院,广东广州510640

出　　处：《重庆交通大学学报（自然科学版）》2022年第8期24-29,共6页Journal of Chongqing Jiaotong University(Natural Science)

基　　金：国家自然科学基金面上项目(61873098);广东省自然科学基金项目(2018A030313250);广东省科技计划项目(2016A030305001)。

摘　　要：为了提高交通控制算法的适应性和鲁棒性,缓解城市交通拥堵,提出了一种改进的D3QN(dueling double deep Q-network, D3QN)自适应信号控制方法。首先对几种强化学习自适应控制模式进行分析,然后在固定步长动作模式的基础上提出了不定步长动作模式,并构造了一种基于空间占有率的奖励函数;最后使用Sumo软件,对中山市东区街道某交叉口分别在稳定流和随机流场景下进行仿真。仿真结果表明:该方法具有良好的收敛性,有效地降低了延误时间和排队长度。In order to improve the robustness and adaptability of traffic control algorithms and ease urban traffic congestion,an adaptive traffic signal control method based on improved D3QN(dueling double deep Q-network,D3QN)was proposed.Firstly,several adaptive traffic control modes based on reinforcement learning were analyzed.Subsequently,a variable step-size action mode was proposed based on the fixed step-size action mode and a reward function based on space occupancy was constructed.Finally,an intersection in East Street of Zhongshan was simulated by software Sumo in steady flow and stochastic flow.The simulation results show that the proposed method exhibits excellent convergence and effectively reduces the delay time and the queue length.

关 键 词：交通工程 交通仿真 自适应控制 交通流 深度强化学习 

分 类 号：U491.51[交通运输工程—交通运输规划与管理]