Otimização de modelos de regressão multivariados empregando métodos de seleção de variáveis.

Abstract

It has a great increase in the use of infrared spectroscopy techniques for chemical analyses in the industry, due to rapidity, low cost and preservation of the samples in this kind of analysis. Thus it is desirable a study of infrared spectra acquisition techniques, multivariate regression and variable selection methods. This dissertation has as objective the study and implementation of a genetic algorithm, jointly with interval partial least-squares multivariate regression technique (iPLS), capable to select the variables most pertinent the property that if it desires to measure and to create more robust multivariate regression models. In this research was made the determination of hydroxyl value of hydroxylated soybean oils, where the gotten results had been 14.97% betters in relation to the prediction error and 15.63% betters in relation to the average percentile error of the values calculated for the prediction samples in comparison with the results found through the iPLS method. Also it was made the propranolol hydrochloride determination in tablets, where the gotten results had been 76.1% betters in relation to the prediction error and 73.99% betters in relation to the average percentile error of the values calculated for the prediction samples in comparison with the results found through the iPLS method. Observing such values, it can be concluded that the use of genetic algorithms jointly with the iPLS method was capable to optimize the solutions, selecting of efficient way the involved spectral variables, finding more predictive and robust models.