Selection of the Number of Layers in Neuro-Fuzzy GMDH

Abstract

An adaptive learning network has been proposed, in which the Radial Basis Functions networks were applied to the partial descriptions of network type GMDH. In choosing the number of its layers, we can apply the statistical information criterion, such as the AIC and MDL. But the theoretical problems concerned with the local minima and nonuniqueness of parameters have been pointed out in the literature. In this paper, we apply the Differential Unbiasedness Criteria (DUC) which is based on the cross validation method whose checking samples are composed of the difference of paired outputs. The effectiveness of DUC is shown by numerical examples.