Learning algorithms and the inversion of fuzzy neural networks

Abstract

The learning algorithm of fuzzy neural networks deserves much attention because it determines the learning capability and the performance of the networks. Several learning algorithms have been proposed, but there is still no algorithm that is accepted as the standard solution. We make an analysis of those algorithms in a comprehensive perspective and incorporate them into one learning procedure. To achieve a particular learning algorithm, we choose an adjusting scheme for weight factors, which specifies how the weight factors are represented internally. We also formulate an inversion algorithm for fuzzy neural networks. The inversion yields an estimate inverse of a given target in a fuzzy neural network. It is based on the gradient descent search and employs the strategy used by the learning procedure in adjusting weight factors. We conduct experiments on the parity-3 problem to show how the learning procedure and the inversion algorithm work in reliably-trained networks. We also demonstrate that the technique of inversion can be used for better examination of fuzzy neural networks.