D-optimal design used to optimize a multi-response class-modelling method

Abstract An experimental strategy, based on a D-optimal design, to systematically study the influence of some metaparameters that affect the behaviour of a class-modelling method is described. The class-modelling method computes class-models by using neural networks trained by an evolutionary algorithm. The key is that the neural networks are trained to find a set of models that behave differently as regards sensitivity and specificity and that constitute the Pareto-optimal models for the class-modelling problem. A measure for comparing different Pareto-optimal fronts has been defined. In this way, by studying the effects from the D-optimal experimental design, the metaparameters that influence the behaviour of both neural networks and evolutionary algorithms when modelling a class are determined. As a case-study to explain the procedure, it has been applied to model the acceptance or rejection of 117 dry-cured ham samples based on their pastiness (a sensory property), using the NIR spectra (1050 variables) as predictor variables.

[1]  Rafael Cela,et al.  Multi-objective optimisation using evolutionary algorithms: its application to HPLC separations , 2003 .

[2]  Luis A. Sarabia,et al.  How to search the experimental conditions that improve a Partial Least Squares calibration model. Application to a flow system with electrochemical detection for the determination of sulfonamides in milk , 2008 .

[3]  Randall S. Sexton,et al.  Simultaneous optimization of neural network function and architecture algorithm , 2004, Decis. Support Syst..

[4]  George Cybenko,et al.  Approximation by superpositions of a sigmoidal function , 1992, Math. Control. Signals Syst..

[5]  Parag C. Pendharkar,et al.  An empirical study of impact of crossover operators on the performance of non-binary genetic algorithm based neural approaches for classification , 2004, Comput. Oper. Res..

[6]  Luis A. Sarabia,et al.  On Pareto-optimal fronts for deciding about sensitivity and specificity in class-modelling problems , 2005 .

[7]  Kalyanmoy Deb,et al.  Multi-objective optimization using evolutionary algorithms , 2001, Wiley-Interscience series in systems and optimization.

[8]  W. J. Studden,et al.  Theory Of Optimal Experiments , 1972 .

[9]  Luis A. Sarabia,et al.  Vectorial optimization as a methodogical alternative to desirability function , 2006 .

[10]  Roger Phan-Tan-Luu,et al.  Pharmaceutical Experimental Design , 1998 .

[11]  M. D. Luque de Castro,et al.  Prediction of texture and colour of dry-cured ham by visible and near infrared spectroscopy using a fiber optic probe. , 2005, Meat science.

[12]  M. C. Ortiz,et al.  Sensitivity and specificity of PLS-class modelling for five sensory characteristics of dry-cured ham using visible and near infrared spectroscopy , 2006 .

[13]  Luis A. Sarabia,et al.  GINN (Genetic Inside Neural Network): towards a non-parametric training , 1997 .

[14]  M. C. Ortiz,et al.  Pareto-optimal front as a tool to study the behaviour of experimental factors in multi-response analytical procedures. , 2008, Analytica chimica acta.

[15]  Kurt Hornik,et al.  Multilayer feedforward networks are universal approximators , 1989, Neural Networks.

[16]  Luis A. Sarabia,et al.  A stochastic trained neural network for nonparametric hypothesis testing , 2002 .