Comitês de classificadores baseados nas Redes SOM e Fuzzy ART com sintonia de parâmetros e seleção de atributos via metaheurísticas evolucionárias

Abstract

The ensemble-based classification paradigm has received considerable attention in scientific literature in recent years. In this context, supervised neural networks have been the most common choice for ensembles’ base classifiers. This dissertation has the intention of projecting and evaluating ensembles of classifiers built through modifications on non-supervised learning algorithms, such as the Fuzzy ART and SOM networks, originating, respectively, the ARTIE (ART in Ensembles) and MUSCLE (Multiple SOM Classifiers in Ensembles) models. The parameters’ tunning and the feature selection of the neural networks which compose the ARTIE and MUSCLE models were tackled by metaheuristic optimization, with the proposal of the I-HPSO (Improved Hybrid Particles Swarm Optimization) algorithm. The ARTIE and MUSCLE models were evaluated and compared with ensembles based on Fuzzy ARTMAP, LVQ and ELM networks in 12 real world datasets. The obtained results indicate that the proposed models present performance superior to the ensembles of supervised neural networks