A new method for initializing reference vectors in LVQ

A new method for setting initial locations of reference vectors in learning vector quantization (LVQ) is proposed to obtain stably high-performance classification results. The initial locations of reference vectors are important for obtaining adequate results rapidly in the LVQ, because the initial locations affect the convergence of LVQ. On the basis of the convergence property of LVQ, this method locates reference vectors in such a manner that they match the probability distribution of training data with a self-organizing map (SOM). Then, it determines the categories of the reference vectors as representatives of respective Voronoi regions. Numerical simulations confirm better classification results with the present method than with conventional methods