Model Selection of Beta Neuro-Fuzzy Classifiers for High Dimensional Data

Tarek M. HAMDANI

Engineering of computer science systems ISI

University of Sfax

National School of Engineers of Sfax

Abstract: This thesis fits into the overall process of the design of cognitive systems focused on studying and solving problems of automatic classification. This study focuses on three fundamental items of classification, namely: data dimensions, design of Classifiers models and decision making. A more specific interest is focused on problems encountered in testing neural based methods for difficult classification applications, starting with the variation of the number of training data and the number of features, to come up with the relationship between training data and the number of used features and to study classification methods and evaluation of classification performance. A first iterative method has been developed for deciding the structure of SVM and single layer neural networks. In the same way of learning process optimization of the neuro-fuzzy classifiers, we develop a method for enhancing the structure and parameters of the centers for BBF fuzzy neural network classifier construction based on data structure. Interested in reducing the size of the data, we proposed a system of hierarchical genetic algorithms with a new evaluation function and bi-coded representation for feature selection with confidence rate. The fourth contribution of this thesis is an intelligent decision making system based on multiple classifiers using confidence rates and stress parameters. It should be interesting to mention that these contributions were validated on artificial and real world applications and have shown to be superior to commonly used methods in the field.