Fuzzy rough granular neural networks for pattern analysis

Abstract

Granular computing is a computational paradigm in which a granule represents a structure of patterns evolved by performing operations on the individual patterns. Two granular neural networks are described for performing the pattern analysis tasks like classification and clustering. The granular neural networks are designed by integrating fuzzy sets and fuzzy rough sets with artificial neural networks in a soft computing paradigm. The fuzzy rough granular neural network (FRGNN) for classification is based on multi-layer neural networks. On the other hand, the fuzzy rough granular self-organizing map (FRGSOM) for clustering is based on self-organizing map. While the FRGNN uses the concepts of fuzzy rough set for defining its initial connection weights in supervised mode, the FRGSOM, as its same implies, exploits the same in unsupervised manner. Further, the input vector of FRGNN & FRGSOM and the target vector of FRGNN are determined using the concepts of fuzzy sets. The performance of FRGNN and FRGSOM is compared with some of the related methods using several life data sets.