Optimal decision surfaces in LVQ1 classiffication of patterns

Kohonen's LVQ1 procedure is widely used for the classification of patterns in a multi-class distribution. This algorithm approximates the probability densities of vectors in each class, by updating reference vectors at each presentation of an input pattern. It is shown that the Bayes classification is only approximated by the decision surfaces of LVQ1 in special cases of probability densities; the authors show how to set the decision surfaces in a more general case, and how the relative number of reference vectors in each class can comply with their a priori probabilities, condition which is necessary for a good classification.