The Learning Vector Quantization (LVQ) [Kohonen, 1986b,Kohonen et al., 1988] aims at defining the decision surfaces between the competing classes. The decision surfaces obtained by a supervised stochastic learning process of the training data are piecewise-linear hyperplanes that approximate the Bayesian minimum classification error (MCE) probability.
There exist several versions of LVQ described in [Kohonen, 1995,McDermott, 1990,Makino et al., 1992] which have slightly different properties. The ones used here are the LVQ2 [Kohonen et al., 1996b], which is efficient for fine tuning the decision borders between the competing classes and the LVQ3 [Kohonen et al., 1996b] which adds a stabilizing term to LVQ2 for improved long-run behavior.