大数据背景下的路径选择行为建模

路径选择建模的主要任务是基于合理假设,定量分析交通参与者的路径选择行为,并估计和预测交通参与者对交通网络的使用情况。基于此,全面总结路径选择建模的研究现状,介绍各种出行数据的特点,阐释常见的选择集生成方法,对文献中提出的众多离散选择模型进行归类和讨论,对比模型估计的2类主要方法,并展望机器学习在路径选择建模中的广阔前景。研究结果表明：随着交通感知技术的全息化发展,在海量车辆轨迹数据的支撑下,路径选择研究取得了全方位的进步;路径选择模型可分为基于路径和基于路段的模型,前者以路径为基本选项,从通过确定性或随机性方法生成的选择集中选择路径,包括多项Logit （MNL）模型以及更先进的MNL修正模型、广义极值（GEV）模型、混合Logit模型和非GEV分布模型,后者以路段为基本选项,动态地求解路径选择问题,无需生成选择集,包括各种递归Logit模型;路径选择模型的参数估计可使用有标签数据或无标签数据,前者通过地图匹配在交通网络中重构出真实路径,后者则依概率考虑一系列可能的路径。近年来,基于机器学习的路径选择模型因具有更优的预测性能而受到广泛关注。在未来的路径选择研究中,应进一步结合离散选择模型和机器学习模型,使两者优势互补。

Modeling Route Choice Behavior in the Era of Big Data

The primary tasks of route choice modeling are to quantitatively analyze the route choice behavior of traffic participants based on reasonable assumptions, and to estimate and predict their usage of the transportation network. In this paper, the state of the art in route choice modeling was comprehensively reviewed. The characteristics of different kinds of trip data were introduced. The common methods of choice set generation were interpreted. The various discrete choice models proposed in the literature were sorted out and discussed. The two major approaches of model estimation were compared. Finally, the broad prospects of machine learning in route choice modeling were illustrated. The results indicate that route choice studies have made all-round progress with the holographic development of traffic perception technology and the consequent support of massive vehicle trajectory data. Route choice models can be classified into path-based and link-based models. The former takes paths as alternatives and selects a path from a choice set generated by some deterministic or stochastic method. Such models include the multinomial logit (MNL) model, as well as the more accurate MNL-modification models, generalized extreme value (GEV) models, mixed logit models, and non-GEV models. However, the latter takes links as alternatives and dynamically solves the route choice problem without a choice set. Such models include various recursive logit models. The parameters of a route choice model can be estimated using labeled or unlabeled data. The former approach reconstructs the chosen route on the transportation network through map-matching, while the latter considers a series of possible routes by probability. In recent years, machine learning-based route choice models have aroused extensive attention due to better prediction performance. In the future, the discrete choice models and machine learning models should be further combined in order to make the two complement each other.