Fuzzy Set Based Neural Networks: Structure, Learning and Application

Abstract

We introduce a class of neural network constructed from fuzzy set models of neurons. The network has a multilayer, feed-forward structure whose units are modeled through triangular norms and co-norms, and weights within the unit interval. The neuron models provide a wide spectrum of design choices - a desirable feature whenever real-world applications are of concern. We focus on pattern classification problems to introduce main concepts and algorithms. The learning procedure does not need any information about derivatives - a very convenient feature within fuzzy set theory that makes the procedure efficient and fast. We provide procedures to construct the network, initialize weights properly, and automatically generate classes of membership functions. Knowledge is easily extracted from the network as if-then rules. Computational examples demonstrate neuro-fuzzy network performance and efficiency. We conclude with remarks on computational complexity analysis and a prospectus for further developments.