针对大规模复杂车流的PG强化交通灯控制算法

PG enhanced traffic light control algorithm for large-scale complex traffic flow

优化城市道路中的交通信号灯控制是低成本地提升城市交通路网性能的方法之一。该研究提出了一种利用策略梯度(Policy Gradient, PG)强化调优的交通灯控制算法。该算法引入了道路压力项、旅程时间项和黑名单机制项,利用统计方式预测汽车行程轨迹,并采用策略梯度估计的优化算法调整算法中的参数。在数据挖掘国际会议Knowledge Discovery and Data Mining (KDD)组织的算法竞赛KDD Cup 2021城市大脑挑战赛中,获得了冠军的成绩。在该挑战赛提供的城市路网规模复杂车流仿真平台上的实验结果表明,算法具有应用于实际场景的价值。

Optimizing traffic light control is one of the low cost methods to improve the performance of urban traffic network.This study proposes a traffic signal control algorithm using the Policy Gradient(PG)algorithm.The model of this algorithm introduces the road pressure term,the travel time term and the blacklist mechanism term,predicts the car travel trajectory statistically,and adjusts the parameters in the algorithm with the optimization algorithm of PG estimation.The algorithm won the KDD Cup 2021 City Brain Challenge organized by the Data Mining field international conference-Knowledge Discovery and Data Mining(KDD).The experiment results of the simulation platform which has city-scale road network with complex traffic flow provided by the challenge commission show that the algorithm has the value of application in practical scenarios.