A fuzzy neural network for fuzzy classification

A fuzzy neural network (FNN) for fuzzy classification of patterns is presented. An extension of the backpropagation algorithm which uses interval arithmetic is employed to accept fuzzy numbers as inputs to the network using α-level set representation. The learning phase of the neural network is supervised by two fuzzy systems which control the learning rate and the slope of the activation functions. A method for fuzzification of the training data set is proposed. The FNN is compared with a traditional backpropagation network in terms of generalisation capabilities. The training and testing are carried out using data with high degree of ambiguity and overlapping among the classes. The results suggest that the FNN provides a robust classification response in the presence of uncertainty and ambiguity in the training data. Its application as a first stage classification procedure in a hybrid system is discussed