Modeling Merging Acceleration and Deceleration Behavior Based on Gradient-Boosting Decision Tree

This paper aims to model the behavior of merging acceleration/deceleration when cars are running in a congested weaving section on a freeway during the merging implementation period by using a data-driven method called gradient-boosting decision tree (GBDT). Different from other black-box machine learning techniques, GBDT can provide abundant information about the nonlinear effects for independent variables by drawing the partial effects. Noise-filtered vehicle trajectory data collected on US Highway 101 are investigated in this study. The partial dependence plots show that the influence of independent variables on merging acceleration/deceleration is nonlinear and complicated and thus is different from the car-following behavior, which indicates that the adoption of traditional car-following models to merging execution behavior cannot reflect the distinctive behavior of merging vehicles. Evaluation of the performances in comparison with other state-of-the-art methods indicates that the proposed method can obtain more accurate results and thus is practical for simulating the merging execution behavior.