Wavelength selection and optimization of pattern recognition methods using the genetic algorithm

Abstract

A genetic algorithm (GA) method for wavelength selection and optimization of near-infrared (NIR) pattern recognition methods was developed to reduce misclassification errors of similar materials. Our goal was to automate completely the process of producing pattern recognition models, consequently, we felt it was important to include pre-processing options, the number of principal components and wavelength selection in the chromosomes. The SIMCA residual variance analysis and the Mahalanobis distance methods were used to classify samples of three different types of microcrystalline cellulose (Avicel PH101, PH102, and RC581) and sulfamethoxazole (SMX). Without GA optimization, approximately 15% of Avicel PH101 and PH102 test samples were misclassified since their NIR spectra are very similar. The GA was used to optimize pattern recognition performance on training sets using a figure of merit designed to maximize correct classification of acceptable samples and minimize classification of unacceptable samples or samples of dissimilar materials. After GA optimization of pattern recognition parameters, 100% correct classification of a validation set was achieved using both the residual variance analysis and the Mahalanobis distance methods .