Learning algorithm for RBF networks as features extractors

A specific learning algorithm, developed in the context of the hybrid linear-nonlinear features space filtering (FSF) system architecture, is proposed. The neural FSF system presented is based on a radial-basis functions (RBF) decomposition of the input data space. An adaptive linear combiner (ALC) is used as transversal filter. The features space is generated by the parameters of the local nonlinear function decomposition. ALC coefficients are adapted with this algorithm to minimize the distance, in the features space, between the reference features vector and the actual features vector obtained from the noisy data. The fuzzy estimation of features matching in the frame of this algorithm is also briefly discussed. Simulation results of spectrography/electrophoresis (EPK)-type data filtering are presented