MODELING OF PROPERTY PREDICTION FROM MULTICOMPONENT ANALYTICAL DATA USING DIFFERENT NEURAL NETWORKS

Two different artificial neural network (ANN) strategies for building a model for the quantitative prediction of the property called ''total color difference'' are described. The models in the study are based o n eight different complex oxide concentration measurements. The models obtained by the ANNs are compared with the multivariate linear regres sion model. Besides the two ANN strategies used for building the model s (the error backpropagation and the counterpropagation), the Kohonen learning strategy is used to make a partial experimental design, i.e., to select data most suitable for budding the models. An additional go al, building a rule or ''formal knowledge'', about the quality of the product, is achieved by overlapping eight two-dimensional maps of weig hts obtained in the counterpropagation neural network.