Normalization approach to the stochastic gradient radial basis function network algorithm for odor sensing systems

Abstract

A method of adapting centers and weights in the radial basis function network (RBFN) is introduced using a normalization method to the stochastic gradient (RBFN-SG) algorithm for odor classification. The RBFN input data vector is from a conducting polymer sensor array. Using Taylor's expansion, a normalized form of the RBFN-SG algorithm is derived. The tracking dynamics of the normalized method appear to be less sensitive to widely varying inputs than the RBFN-SG. Experimental results of the proposed method have shown a faster learning speed, a lower mean squared error (MSE) and better classification performance.