Pattern Classification using Multilayer Hybrid Fuzzy Neural Networks with Technologies of Advanced Computational Intelligence

Abstract

The objective of this research work is to study and propose the complementary newlinehybrid intelligent systems which combine the strengths of the components of computational newlineintelligence, that include artificial neural network, fuzzy logic and/or genetic newlinealgorithms, for pattern recognition and classification. Such classification systems are newlinesupposed to possess human like expertise within a specific domain, adapt themselves newlineand learn to do better in changing environments, and explain how they make decisions newlineor take actions. Also, such systems should combine information, skills, and newlineapproaches from different sources. newlineIn this work, initially two types of modifications are proposed in the prediction newlinephase of Fuzzy Hyperline Segment Neural Network (FHLSNN) to improve its classification newlineaccuracy. In the first modification (MFHLSNN1), Euclidean distance between newlinethe applied input pattern and the centroid of the patterns falling on the hyperline newlinesegments is calculated to classify the applied input pattern. In the second newlinemodification (MFHLSNN2), use of both the membership value and Euclidean distance newlineis proposed to decide the class of applied input pattern. The performance of newlineboth, MFHLSNN1 and MFHLSNN2 is evaluated using three benchmark problems newlinei.e. Wine dataset, Iris dataset and Sonar dataset obtained from UCI machine learning newlinerepository and handwritten character recognition database. Both the proposed newlinemodifications improve the prediction accuracy of the FHLSNN without affecting its newlineimportant feature of incremental learning. newlineLater, above modified versions of FHLSNN are extended further using pruning newlineapproach to reduce the network complexity. The Pruned fuzzy hyperline segment newlineneural network (PFHLSNN) and Pruned modified fuzzy hyperline segment neural newlinenetwork (PMFHLSNN) are proposed. The pruning approach used in these methods newlineis to train the network to get 100 % recognition rate for training dataset and newlinethen remove the parts (hyperline segments) that are less important and even after newlinetheir removal, acceptable classification ac