Evaluation of Car-Following Input Variables and Development of Three- Vehicle Car-Following Models with Artificial Neural Networks

A four-layer artificial neural network (ANN) structure was set up in the models and a genetic algorithm (GA) and back-propagation methodology were utilized to customize individual driver's behavior. A number of combinations of the input variables was examined with the R2 values representing the model fitting. This paper concluded that there are significant differences in degrees of contribution to the models among the several input variables. The most important finding was that the leading vehicle's (LV) acceleration had stronger relationship to the following vehicle's (FV) acceleration rate rather than the relative speed between the leading vehicle and the following vehicle. Additionally, the input variables related to the preceding vehicle of the leading vehicle (PLV) were added in the model. It was also found the variables of the preceding vehicle of the leading vehicle help car-following models slightly, but they were not as much as expected.