Development of an Approach to Adaptive Parametric Tuning of a Car-Following Model

This paper presents the refinement of a methodology for the dynamic parametric identification of the M.G. Gorodnichev car-following model using machine learning methods. Unlike the previous approach, two simulation environments (SUMO and CARLA) are employed, along with an expanded set of regression models, including XGBoost, LSTM, PINN, and DeepONet. It has been experimentally established that DeepONet$(\mathbf{R}^{\mathbf{2}} \boldsymbol{=} \mathbf{0. 8 7})$and PINN$(\mathbf{R}^{\mathbf{2}}=\mathbf{0. 8 6})$provide the highest accuracy for predicting the dynamic gap. The parameter$\mathbf{C}_{2}$of the Gorodnichev model is calculated for various driver reaction time values. The results can be applied in autonomous driving systems and intelligent traffic flow control.