A Neuro-Fuzzy Classifier of Patterns

Abstract

The thesis addresses the classification problem, which is an important part in the design of pattern recognition systems. The major conu•ibution of the thesis is the proposal of a new pattern recognition algorithm, a novel Neuro­ Fuzzy classi ficr. The algorithm has essential properties such as on-line adaptation, nonlinear separability, fast training, few tuning parameters and the ability to provide both sort and hard classification decisions.

The proposed model is simulated and compared with widely used neural network algorithms such as back-propagation, supervised competitive learning, and supervised fuzzy competitive learning neural networks. We describe the proposed model, the ncuro-fuzzy classifier model; the basic idea, analysis, and its application on many data sets.

We usc several benchmark data, part of them arc artificial and the others arc real to assess the proposed algorithm. The artificial databases (gaussian, concentric), are experimented to lest the efficiency of the newly developed algorithm. The real databases (iris, Monks, phoneme) are used in order to compare the results of the proposed algorithm with the previously published results of other methods. The graphs for the classification rate with changing the classifier parameters are presented, and the analysis of those results is cmTied out. The analysis shows that the neuro-fuzzy algorithm can be easily tuned to peJtorm the other neural and fuzzy techniques. Experimental results demonstrate this conclusions by achieving better .classification rate within very small training periods