An Optimal Radial Basis Function (RBF) Neural Network for Hyper-surface Reconstruction

Data acquisition of chemical engineering processes is expensive and the collected data are always contaminated with inevitable measurement errors. Efficient algorithms are required to filter out the noise and capture the true underlying trend hidden in the training data sets. Regularization networks, which are the exact solution of multivariate linear regularization problem, provide appropriate facility to perform such a demanding task. These networks can be represented as a single hidden layer neural network with one neuron for each distinct exemplar. Efficient training of Regularization network requires calculation of linear synaptic weights, selection of isotropic spread (σ ) and computation of optimum level of regularization ( * λ ). The latter parameters (σ and * λ ) are highly correlated with each other. A novel method is presented in this article for development of a convenient procedure for decorrelating the above parameters and selecting the optimal values of * λ and * σ . The plot of