网联环境下基于深度强化学习的单路口交通信号控制优化  

Based on Deep Reinforcement Learning Optimization of Traffic Signal Control at Single Intersection

作　　者：龙顺忠 王志建[1] 刘皓 

机构地区：[1]北方工业大学电气与控制工程学院交通设备与控制工程系,北京100144 [2]中国科学院自动化研究所复杂系统管理与控制国家重点实验室,北京100190 [3]中国科学院大学人工智能学院,北京100049

出　　处：《工业控制计算机》2020年第10期16-19,22,共5页Industrial Control Computer

基　　金：2019年北京高等学校高水平人才交叉培养计划;北京市基本科研业务费(110052971921/023);校内人才(213051360020XN173/013)。

摘　　要：深度强化学习能为交通信号控制研究带来诸多优化空间,它能够实现Agent与道路交通环境之间的交互,根据获得的惩罚或奖励不断地学习知识,从而更加适应环境。研究主要是基于深度强化学习和SUMO仿真的方法,以从仿真的方式建立的智能网联车辆环境中得到单路口交通实时动态信息,并从中输出深度Q网络(DQN)算法所需要的重要参数。基于搭建的深度神经网络,通过不断优化它的权重参数来提高性能,进一步实现对单路口交通信号控制方案的优化,并最终探索一个能够通过自主学习来高效控制单路口交通的信号控制方案。Deep reinforcement learning can bring a lot of optimization space for traffic signal control research. It can realize the interaction between agent and road traffic environment,and continuously learn knowledge according to the punishment or reward obtained,so as to adapt to the environment more.The research of this paper is mainly based on deep reinforcement learning and SUMO simulation method to obtain the real-time traffic dynamic information of single intersection from the intelligent network connected vehicle environment established by simulation,and output the important parameters required by the DQN algorithm. Based on the constructed deep neural network,the performance is improved by continuously optimizing its weight parameters,further realizing the optimization of single intersection traffic signal control scheme,and finally exploring a signal control scheme which can effectively control the traffic of single intersection through autonomous learning.

关 键 词：深度强化学习 DQN算法 SUMO仿真 信号控制策略 

分 类 号：U491.54[交通运输工程—交通运输规划与管理]