Toward evolving consistent, complete, and compact fuzzy rule sets for classification problems

Abstract

This paper proposes Pitts-DNF-C, a multi- objective Pittsburgh-style Learning Classifier System that evolves a set of DNF-type fuzzy rules for classification tasks. The system is explicitly designed to only explore solutions that lead to consistent, complete, and compact rule sets without redundancies and inconsistencies. The behavior of the system is analyzed on a collection of real-world data sets, showing its competitiveness in terms of performance and interpretability with respect to three other fuzzy learners.