Fuzzy versions of Kohonen's net and MLP-based classification: performance evaluation for certain nonconvex decision regions

Abstract

Classification of certain linearly nonseparable pattern classes with nonconvex decision regions is a problem that cannot be efficiently handled by the Bayes' classifier for normal distributions or other metric-based methods. An attempt is made here to demonstrate the ability of fuzzy versions of Kohonen's net and the multilayer perceptron for classification of such patterns. In these models, the uncertainties involved in the input description and output decision have been taken care of by the concept of fuzzy sets whereas the neural net theory helps to generate the required concave and/or disconnected decision regions. Superiority of these fuzzy models (over the respective conventional versions, the Bayes' classifier and seven other existing neural algorithms) has been adequately established when they are implemented on different sets of linearly nonseparable pattern classes. The effect of fuzzification at the input has been investigated for both models. The contribution of the a priori probabilities of the pattern classes in the back-propagation procedure for weight updating has also been studied.