Global Function Approximation by Fuzzy-Connection of Plural Local Neural Networks

Abstract

This paper presents new model that is called a fuzzy connected neural network. The fuzzy connected neural network is expected to be better approximation and converge faster than the conventional neural network. When input-output teaching data with non-linearity are widely over a range, it is necessary to increase the hidden-layer’s neurons and neural parameters because of the simple structure of neural networks. So We propose to divided the input data space into plural sub-area in order to approximate the divided teaching data with a neural network specially prepared on the each divided space. The plural neural networks are connected with fuzzy membership functions to construct the fuzzy connected neural network on the whole input space. Plural neural networks are called local neural networks. Because each local neural network dose not learns all teaching data, it is expected that the surface of the square error function is more monotonous than it in the conventional neural network, and the learning problem of each local neural network can be solved more easily. These points give fuzzy connected neural network advantages over the conventional neural network, which needs repeated trial and error to tune weights. Finally this paper shows the ability of the proposal model.