Meta-heurísticas de otimização tradicionais e híbridas utilizadas para construção de comitês de classificação

Abstract

This work deals with the construction of classification committees using traditional and hybrid meta-heuristics of optimization techniques. The problem of pattenr classification is treated as an optimization problem, trying to find the subset of attributes and classifiers of the problem that minimizes the classification error of the committee. Committees are generated by combining the techniques of k-NN, Decision Tree and Naive Bayes using the majority vote. The attributes of the base classifiers are modified by the metaheuristics of genetic algorithms, memetic algorithms, PSO, ACO, Multi Start, GRASP, Simulated Annealing, Tabu Search, ILS and VNS. We also apply algorithms from AMHM and MAGMA hybrid metaheuristics architectures. Algorithms of these metaheuristics are developed in mono and multi-objective versions. Experiments are performed in different mono and multiobjective scenarios optimizing classification error and measures of good and bad diversity. The goal is to verify that adding diversity measures as optimization goals results in more accurate committees. Thus, the contribution of this work is to determine if the measures of good and bad diversity can be used in mono and multiobjective optimization techniques as objectives of optimization for the construction of committees of classifiers more accurate than those constructed by the same process, but using only the accuracy classification as an optimization objective. We have verified that the developed metaheuristics present better results than the classic generation techniques of committees, ie, Bagging, Boosting and Random Selection. We also verified that in the majority of metaheuristics the use of diversity measures as optimization objectives does not help to generate more accurate committees than when only the classification error, obtaining in the best scenarios non statistically different results.