Multi Objective Genetic Algorithm for Variable Selection in Multivariate Classification Problem: A Study Case in Biodiesel Adulteration Evaluation

Multi Objective Genetic Algorithm for Variable Selection in Multivariate Classification Problem: A Study Case in Biodiesel Adulteration Evaluation

Abstract

This work aims to propose a multiobjective formulation using genetic algorithms for the problem of variable selection in multivariate calibration. A problem involving the classification of biodiesel samples to detect adulteration through a classifier called linear discriminant analysis is presented. The role of the multiobjective genetic algorithm is to reduce the dimensionality of the original set of variables in order to provide a more robust classification model and consequently better generalization capacity. In particular the work makes an implementation of the fast elitism version of the genetic algorithm of ordering by non-dominance (NSGA-II). A comparison is made between a monoobjective and multiobjective implementation with respect to the model and robustness in relation to the presence of noise. The results show that on average the monoobjective genetic algorithm selects 20 variables and has an error rate of 14% and the multiobjective selects 7 variables and has an error rate of 11%. It was possible to demonstrate that the multiobjective formulation provides classification models with less sensitivity to instrumental noise when compared to the monoobjective formulation.

Publication
In Proceedings of 11th Brazilian Congress on Computational Intelligence (CBIC)
Date
Links

Text in Portuguese, Original Title: Algoritmo Genético Multi-Objetivo Para Seleção de Variáveis em Problemas de Classificação Multivariada: Um Estudo de Caso na Verificação de Adulteração de Biodiesel.