Classification of apple beverages using artificial neural networks with previous variable selection

Parallel to increased consumption of fruit juices over the last years (thanks to their unrivalled nutritional benefits), fraudulent fruit juices can be found sometimes on the food supply chain. Infrared spectrometry (IR) is a fast and convenient technique to perform screening analyses to assess the quantity of pure juice in commercial beverages. The IR information has some "fuzzy" characteristics (random noise, unclear chemical assignment, etc.) and, therefore, advanced computation techniques (e.g., artificial neural networks (ANNs)) are needed to develop ad-hoc classification models. Dissapointingly, the large number of variables derived from IR spectrometry makes ANNs to take too much time to train. This work studies two different genetic algorithms (GAs) intended to select a small number of wave numbers which are to be used to develop classification models, "pruned search" and "fixed search" (in both cases a steady state GA algorithm with uniform crossover where one of the parents is chosen following the "roulette-wheel" approach, was used). In order to compare results of different assays, the number of selected variables was fixed using an external criterion (a parametric model based on Procrustes rotation). Usefulness of the GAs is evaluated by developing PLS, potential functions, SIMCA and ANNs models to classify apple juice-based commercial beverages. (C) 2004 Elsevier B.V. All rights reserved.