Evolutionary modular design of rough knowledge-based network using fuzzy attributes

Abstract

This article describes a way of integrating rough set theory with a fuzzy MLP using a modular evolutionary algorithm, for classification and rule generation in soft computing paradigm. The novelty of the method lies in applying rough set theory for extracting dependency rules directly from a real-valued attribute table consisting of fuzzy membership values. This helps in preserving all the class representative points in the dependency rules by adaptively applying a threshold that automatically takes care of the shape of membership functions. An l-class classification problem is split into l two-class problems. Crude subnetwork modules are initially encoded from the dependency rules. These subnetworks are then combined and the final network is evolved using a GA with restricted mutation operator which utilizes the knowledge of the modular structure already generated, for faster convergence. The GA tunes the fuzzification parameters, and network weight and structure simultaneously, by optimising a single fitness function. This methodology helps in imposing a structure on the weights, which results in a network more suitable for rule generation. Performance of the algorithm is compared with related techniques.