層状ニューラルネットワークにおける学習アルゴリズムの初期値問題

概要

論文の詳細を見る
Multilayer perceptrons, trained with the back-propagation algorithm, have been applied successfully in a variety of diverse areas. Basically, the BP algorithm is a gradient technique and not an optimization technique. The error surface, with the weights as coordinates, has local minima in addition to global minima. Therefore, the BP algorithm has some difficulties, which are local minima, generalization (or over learning) , size of learning parameter, etc. The local-minima problem is related to the choice of initial value of weights. In this paper we discuss four algorithms which use back-propagarion error. These algorithms are characterized by three learning parameters : the learningrate parameter α; momentum constant β; and parameter γ which is used in the proposed method for removing singular points of BP errors. We tested the performace of these algorithms for the Exclusive OR (XOR) problem which is not linearly separable. From the computer simulation, parameter γ is very effective for learning acceleration and improving the rate of convergence for setting initial values of weights.
東海大学の論文
1997-03-30