Recently, the focus of research in car-following theory has been on the stimulus-reaction approach. Researchers often use linear regression to calibrate the stimulus-reaction function after processing the original data with nature logarithm. This procedure tends to be less precise when some zero-value data are deleted, although it is computationally convenient. In this study, the composition of artificial neural networks and genetic algorithms are employed to resolve this problem and to improve the performance of artificial neural networks. A prototype freeway driving simulator developed with the virtual reality technique is employed for data collection. These data are used for model calibration and validation. The results show that the combined model is more likely to acquire lower errors than ordinary back propagation neural networks. Networks with one hidden layer tend to perform better. The proposed models may suitably identify different behaviors between different drivers.