Fuzzy learning vector quantization generation of codebooks

A modification of the learning vector quantization (LVQ) method used for the generation of codebooks is presented. It retains the simplicity of the LVQ method while eliminating the non uniform spatial distribution of the prototype vectors which could result from an inadequate choice of the input signal sequence and/or from the initial choice of the prototype vectors. The method is based upon the segmentation of the input vector space in fuzzy partitions. A fuzzy objective function is defined. An algorithm for its minimization is presented. Simulation results are given