The optimal brain surgeon (OBS) pruning procedure for automatic selection of the optimal neural network architecture was applied in multivariate calibration studies of two different near infrared data sets. These spectroscopic data sets were first preprocessed by using principal component analysis (PCA), and the scores of these principal components were the input into the neural network. In the first (linear) data set, the optimized architecture converged to a linear model, and the results were similar to linear PCR and PLS. In the second (non-linear) data set, the pruning procedure improved the generalization ability, reducing the errors in a test set when compared to a non-pruned architecture, and produced better results than PCR and PLS. When using OBS in a network with both linear and non-linear transfer functions, a diagnostic for non-linearity results. In case of a linear model, the net is automatically reduced to principal component regression (PCR). (C) 1998 Elsevier Science B.V. All rights reserved.37541671187195