改进的A2C算法在交通信号控制中的应用  

Application of improved A2C algorithm in traffic signal control

作　　者：曹桐 黄德启[1] 赵军 CAO Tong;HUANG De-qi;ZHAO Jun(School of Electrical Engineering,Xinjiang University,Urumqi 830047,China)

机构地区：[1]新疆大学电气工程学院,新疆乌鲁木齐830047

出　　处：《计算机工程与设计》2024年第6期1713-1719,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(51468062)。

摘　　要：针对目前以数据为驱动的交通控制算法在处理交通数据时容易忽略道路本身的空间信息的问题,提出一种结合道路拓扑结构信息的A2C(advantage actor-critic,A2C)算法。以A2C算法为基础,提取路网中车流量的信息,经过MLP(multilayer perceptron,MLP)对路口观测到的交通状态特征进行编码;结合图卷积神经网络提取道路之间的空间信息,引入多头注意力机制关注智能体之间的影响,在SUMO仿真环境中进行仿真验证。实验结果表明,改进的A2C算法相较于基线算法在等待时间、平均行驶速度上性能分别提升9.84%、7.57%,可以更好提高车辆通行效率。Aiming at the problem that the current data-driven traffic control algorithms tend to ignore the spatial information of the road itself when processing traffic data,an A2C algorithm combined with road topology information was proposed.The algorithm was based on the advantage actor-critic algorithm,the information of traffic flow in the road network was extracted.Through MLP,the observed traffic state characteristics were encoded at the intersection.Combined with the graph convolutional network,the spatial information between roads was extracted.The multi-head attention mechanism was introduced to focus on the influence between agents,and the simulation verification was carried out in the SUMO simulation environment.Experimental results show that compared with the baseline algorithm,the improved A2C algorithm improves the performance of 9.84%and 7.57%in terms of waiting time and average driving speed,respectively,which can better improve the efficiency of vehicle traffic.

关 键 词：强化学习 图卷积神经网络 优势行动者-评论家 多层感知机 多头注意力机制 交通信号控制 多智能体 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]