交通信号灯控制任务的人在回路强化学习算法数据集

针对如何使路网的交通效率最大化，并避免交叉口内可能的交通冲突的问题，本团队提出了基于多智能体通信和动作修正的城市信号灯控制方法（MaCAR）来解决。该方法通过对当前交通态势的建模以及对未来交通态势的预测，智能的生成信号灯方案。具体的，该方法首先以各目标路口过去数个周期的流量数据、路网连接图、相位配时等信息输入交通流量预测器（TFN）获得各路口的车流量预测数据。然后，汇同各目标路口过去数个周期的流量数据、路网连接图、相位配时等信息，输入决策网络（CAN）来调节各目标信号灯的相位配时。实验室使用的JINAN数据集是宾夕法尼亚大学提供的基于真实城市路网和真实车流数据构建的城市信号灯控制任务数据集。相关数据开源在https://github.com/wingsweihua/colight。 该数据集模拟了一个真实的包含12（3X4）个路口的城市路网。车流方面，该数据集包含了由真实车流数据中采样得到的车辆到达率平均为250.70辆/5分钟的繁忙车流。

To maximize the traffic efficiency of road networks and avoid potential traffic conflicts at intersections, our team proposes an urban traffic signal control method named MaCAR (Multi-Agent Communication and Action Correction) to address this challenge. This method intelligently generates signal timing schemes by modeling the current traffic situation and predicting future traffic conditions. Specifically, the method first inputs traffic data from multiple past cycles, road network connectivity graph, phase timing and other relevant information of each target intersection into the Traffic Flow Predictor (TFN) to obtain traffic flow prediction data for each intersection. Then, combining the same types of historical data including traffic data from multiple past cycles, road network connectivity graph, phase timing, etc. of each target intersection, it is fed into the Decision Network (CAN) to adjust the phase timing of each target traffic signal. The JINAN dataset used in the experiments is an urban traffic signal control task dataset constructed by the University of Pennsylvania based on real urban road networks and real traffic flow data. The relevant dataset is open-sourced at https://github.com/wingsweihua/colight. This dataset simulates a real urban road network composed of 12 (3×4) intersections. In terms of traffic flow, the dataset includes busy traffic flows sampled from real traffic data, with an average vehicle arrival rate of 250.70 vehicles per 5 minutes.

搜集汇总

背景与挑战

背景概述

该数据集是针对交通信号灯控制任务设计的，用于支持人在回路强化学习算法的研究与评估。它基于JINAN数据集，模拟了一个包含12个路口的真实城市路网，并采用来自真实车流采样的繁忙交通数据，平均车辆到达率为250.70辆/5分钟，旨在通过多智能体通信和动作修正方法优化路网交通效率。数据集由浙江大学团队创建，包含36.39MB的71个文件，适用于计算机科学技术领域的相关应用。