Modeling Car Following with Feed-Forward and Long-Short Term Memory Neural Networks

Abstract The paper investigates the capability of modeling the car following behavior by training shallow and deep recurrent neural networks to reproduce observed driving profiles, collected in several experiments with pairs of GPS-equipped vehicles running in typical urban traffic conditions. The input variables are relative speed, spacing, and vehicle speed. In the model, we assume that the reaction is not instantaneous. However, it may occur during a time interval of the order of a few tenth of seconds because of both the psychophysical driver’s reaction process and the mechanical activation of braking or dispensing the traction power to the wheels. Experimental results confirm the reliability of this assumption and highlight that the deep recurrent neural network outperforms the simpler feed-forward neural network.