Linking chemical knowledge and genetic algorithms using two populations and focused multimodal search

Many analytical problems involve measurement of a large number of experimental variables on a set of samples. Unfortunately, some of them can deteriorate the performance of classification models because not all variables yield the same quality and quantity of information. In this paper four strategies to perform variable selection in mid-infrared spectral data using genetic algorithms (GAs) are presented: fixed search, pruned search, multimodal search by hybrid two-population GA, HTP–GA, and focused HTP–GA. They can be easily tailored to favour either the reduction in the number of variables, the classification abilities of the overall model or the chemical understanding. Focused HTP–GA in particular allows introduction of chemical information in the GA fitness function and, thus, obtains mathematical solutions which are chemically-driven as well. This, therefore, simplifies the chemical interpretation of the models. To the best of our knowledge this is the first application in the Analytical Chemistry field where a GA is used to find out a solution with chemical meaning since current applications interpret the variables once they are extracted. Selected variables were used to implement a screening procedure to ascertain the amount of pure apple juice in commercial beverages employing backpropagation artificial neural networks (ANNs). The influence of each GA on the classification results was evaluated following a twofold approach. First, the GAs were coupled to an ANN with a fixed topology so that the differences on the classifications depended (mainly) on the GA-selected variables. Second, the ANNs were optimized and the overall GA–ANN models compared. © 2007 Elsevier B.V. All rights reserved.