Human Car-Following Behavior: Parametric, Machine-Learning, and Deep-Learning Perspectives

Abstract

Emerging automated vehicles and mixed traffic flow have been substantially increased demand for modeling human driving behavior in both academia and industry. As a result, many car-following (CF) models have been proposed using parametric and data-driven approaches. Considering the large number of CF models, the critical question is which CF model or category of models (e.g. machine-learning) could accurately regenerate human CF behavior. This study conducts a cross-category comparison between one parametric model (intelligent driver model (IDM)), two new machine-learning CF models based on feedforward neural network (FNN) and recurrent neural network (RNN), and one novel deep-learning CF model (Deep-RNN) with long short-term memory (LSTM). The models are developed in TensorFlow and compared at local and global levels. At the local level, Deep-RNN significantly outperformed the others, followed by RNN and FNN. At the global level, IDM demonstrated the best performance, followed closely by Deep-RNN. The result illustrates there is no one-size fit model and the model should be selected given projects’ characteristics. The result suggests a hybrid approach, which integrates parametric and deep-learning models, could precisely regenerate human car-following behavior.