Abstract
Four car-following models with artificial neural network (ANN) structure were developed with various input variables in the car-following behavior. A four-layer ANN structure was set up and a genetic algorithm (GA) and back-propagation methodology were utilized for determining the synaptic weights in the models, however the models sometimes had a difficulty in learning such enormous number of raw data points. Therefore, a methodology of data point conversion was developed with, Kohonen Feature Map (KFM), a self-organization neural network model. In order to evaluate the ANN models, the General Motors’ (GM) model was also calibrated. This paper concluded that the ANN models were successfully developed with KFM data conversion without deteriorating the original data quality. In comparing the results among the four ANN models, it was implied that the accelerations of the following vehicle and leading vehicle can also become key input variables for improving the modeling of car-following behavior.
