Comparative analysis of interpretable data-driven and analytical car-following models on highways: Robustness, sensitivity and transferability

Abstract

Driving behaviour has traditionally been explained using analytical car-following (CF) models. Nowadays, the vast availability of naturalistic driving data facilitates the development of models driven by their enhanced predictive performance rather than their ability to physically explain a phenomenon. In this study, we develop simple but powerful data-driven CF models with up to two leaders for highways and investigate their robustness. The models’ sensitivity to changes in the sampling interval of the dataset is tested and compared with calibrated versions of Gipps’, Krauss’ and Intelligent Driver models in order to simulate the effects of a varying reaction time parameter on their performance. Thereafter, the models’ predictive capabilities are evaluated using unseen data from different locations. The results indicate that both analytical and data-driven models are relatively susceptible to reaction time parameter changes. Data-driven models perform better when transferred to different locations with similar traffic patterns, while analytical models are more suitable for traffic management scenarios, with some data-driven methods providing comparable results. For the first time, a comprehensive comparison between the sensitivity of data-driven and analytical models to changes in the data is conducted.

Shahriar Iqbal Zame

Doctoral Candidate & Research Associate

Research interests include agent-based simulations, optimization, mode choice modeling, freight electrification and automation.