Título: The Effective Use of Diverse Rule Bases in Fuzzy Classification
Autores: Coelho, Guilherme P.; Castro, Pablo A. D.; Von Zuben, Fernando J.
Resumo: This paper is concerned with the synthesis of diverse rule bases for fuzzy classification. An immune-inspired approach for combinatorial optimization, capable of controlling the size and diversity of the population along the search, is applied to generate multiple high-quality solutions. A preliminary comparison of the obtained rule bases indicates the existence of inconsistency, mainly characterized by the presence of rules with the same antecedent part and distinct consequent parts. Based on a winnertakes-all reasoning method, the effective portion of the input space allocated to each rule will depend on a competitive procedure. So rules with the same antecedent part in distinct rule bases may fire at distinct portions of the input space, and possibly with distinct consequent parts. This presumed disadvantage, when interpretability issues are concerned, can be assertively explored to produce an ensemble of fuzzy classifiers, with increment in performance precisely for the same reason. High-quality and diverse solutions are essentially the basic requisites for successful implementation of ensembles. A qualitative disadvantage may then become a quantitative advantage.
Código DOI: 10.21528/CBRN2005-220
Artigo em PDF: CBRN2005_220.pdf
Arquivo BibTex: CBRN2005_220.bib