Determination of optimal support vector regression parameters by genetic algorithms and simplex optimization

Traditionally, the partial least squares (PLS) regression technique is most commonly used for quantitative analysis of near-infrared spectroscopic data. However, the use of support vector regression (SVR), a recently introduced alternative regression technique, for quantitative spectral analysis has increased over the past few years especially due to its high generalization performance and its ability to model non-linear relationships as well. Unfortunately, the practical use of SVR is limited because of its set of parameters to be defined by the user. For this reason, it was necessary to find an automated reliable, accurate and robust optimization approach to select the optimal SVR parameter settings. This paper presents a SVR parameter optimization approach based on genetic algorithms and simplex optimization, which satisfies all of the above-mentioned points. Furthermore, a comparison is made between the performance of SVR and PLS on various (noisy) data sets. From these results, it can be concluded that SVR is less sensitive to spectral noise, and hence, more robust with respect to spectral variations due to experimental circumstances. Generally, in the context of performance and robustness, the results demonstrate that SVR is a good well-performing alternative for the analysis and modelling of NIR data than the commonly applied PLS technique.