Self-Configuring Radial Basis Function Neural Networks for Chemical Pattern Recognition

Abstract

Construction of radial basis function neural networks (RBFN) involves selection of radial basis function centroid, radius (width or scale), and number of radial basis function (RBF) units in the hidden layer. The K-means clustering algorithm is frequently used for selection of centroids and radii. However, with the K-means clustering algorithm, the number of RBF units is usually arbitrarily selected, which may lead to suboptimal performance of the neural network model. Besides, class membership and the related probability distribution are not considered. Linear averaging (L-A) was devised for selection of centroids and radii for the RBFs and computing the number of RBF units. The proposed method considers the class membership and localized probability density distribution of each class in the training sets. The parameters related to the network construction were investigated. The network was trained with the QuickProp algorithm (QP) or Singular Value Decomposition (SVD) algorithm and evaluated with the poly(chlorobiphenyl) (PCB) mass spectra and an Italian olive oil reference data set. The prediction accuracy of PCB data sets was better than 94%, and the prediction accuracy with Italian olive oil data sets achieved 100% with RMSEP as low as 2.4 × 10-3. The training times were usually about a second on a personal computer. The performance of neural networks constructed from the linear-averaging method was observed to be better than that with the K-means algorithm with these data sets.