Vehicle-following has been an important topic of traffic flow research in the past 50 years. Many deterministic vehicle-following models have been proposed and studied [1] and many of them are being used in microscopic traffic simulation tools [2]. Earlier studies, for example [3], relied on limited sets of data collected from instrumented vehicles driven in test tracks. Results of the earlier studies have been Gemcitabine clinical trial developed into the well-known Gazis, Herman, and Rothery or simply the GHR model [3,

4]. Users of the GHR model or other deterministic models have assumed that the selected model, once calibrated with its fixed parameter values, was applicable to all driver-vehicles; that is, the driver-vehicle population is homogenous. Some microscopic traffic simulation tools distinguish the behavior between different driver-vehicles by using the same model but vary the parameter values between different driver-vehicles. With the large-scale vehicle trajectory data collection efforts

enabled by remote sensing techniques in the past decade, several researchers have begun studies on heterogeneous vehicle-following behavior between driver-vehicles and/or for the same driver-vehicle [5–8]. Such studies still relied on one or more prespecified vehicle-following equations. The researchers either (i) calibrated different equations to show that different driver-vehicles responded with different driving rules; (ii) calibrated the same equation but different parameter values between driver-vehicles; or (iii) calibrated the same equation but different parameter values between acceleration and deceleration. Such studies still depend on the deterministic equations, which may need to be calibrated to different segments of the driver-vehicle population.

In this paper, we use the term vehicle-following instead of the conventional term car-following, as the lead or following vehicle may be a truck instead of a car. We propose to use the self-organizing feature map (SOM) to replicate vehicle-following behavior. The SOM consists of neurons arranged systematically on a two-dimensional surface GSK-3 (known as a “map”). Each neuron has a prototype weight vector that represents the characteristic features in the input space. Such structure is capable of mapping patterns in the high dimensional input space into a two-dimensional map. According to the unsupervised learning rule, vectors that are similar to each other in the multidimensional space will be clustered in the same neighborhood in the SOM’s two-dimensional space, which makes it possible to be adopted as a tool of data classification. Conventional neural networks do not have the unsupervised clustering capability. Because of its unique structure, users of the SOM do not need to specify the function between the input features and its output variable. No equation needs to be predefined and no parameter calibration is necessary.