Evolutionary Approach for Construction of the RBF Network Architecture

Feature selection (FS) and classifier design (CD) are two basic stages in the construction of a classification system. Typically, both tasks have been studied separately in literature. FS aims to remove irrelevant and redundant features whereas CD generates a prediction rule for classifying patterns whose class is unknown. Despite the relationship between FS and CD with radial basis function networks (RBFNs) is noticeable, only some works have addressed FS and CD jointly when constructing RBFNs. This paper presents a methodology for the automatic construction of the RBFN architecture by using two evolutionary algorithms (based on differential evolution, DE) for addressing FS and CD tasks simultaneously. FSDE algorithm evolves a population in order to find a reduced subset of discriminant features. After, each individual generates a subpopulation which evolves to construct the hidden layer of the net via CDDE algorithm. CDDE determines the suitable number of hidden nodes and their parameter. Two real datasets for breast lesion classification were used and the experimental results pointed out that the proposed methodology obtained high classification performance with reduced subsets of features.