Optimizing the tuning parameters of least squares support vector machines regression for NIR spectra

Partial least squares (PLS) is one of the most used tools in chemometrics. Other data analysis techniques such as artificial neural networks and least squares support vector machines (LS‐SVMs) have however made their entry in the field of chemometrics. These techniques can also model nonlinear relations, but the presence of tuning parameters is a serious drawback. These parameters balance the risk of overfitting with the possibility to model the underlying nonlinear relation. In this work a methodology is proposed to initialize and optimize those tuning parameters for LS‐SVMs with radial basis function (RBF)‐kernel based on a statistical interpretation. In this way, these methods become much more appealing for new users. The presented methods are applied on manure spectra. Although this dataset is only slightly nonlinear, good results were obtained. Copyright © 2007 John Wiley & Sons, Ltd.