The mechanistic-empirical pavement design methodology being developed under NCHRP Project 1-37A will require accurate classification of vehicles to develop axle load spectra information needed as the design input. Scheme F, used by most states to classify vehicles, can be used to develop the required load spectra. Unfortunately, the scheme is difficult to automate and is prone to errors resulting from imprecise demarcation of class thresholds. In this paper, the classification problem is viewed as a pattern recognition problem in which connectionist techniques such as probabilistic neural networks (PNN) can be used to assign vehicles to their correct classes and hence to establish optimum axle spacing thresholds. The PNN was developed, trained, and applied to field data composed of individual vehicles' axle spacing, number of axles per vehicle, and overall vehicle weight. The PNN reduced the error rate from 9.5% to 6.2% compared with an existing classification algorithm used by the Florida Department of Transportation. The inclusion of overall vehicle weight as a classification variable further reduced the error rate from 6.2% to 3.0%. The promising results from neural networks were used to set up new thresholds that reduce classification error rate.