Back-Propagation Learning Algorithm with Dynamic Learning Coefficient by Introducing Fuzzy Inference

概要

論文の詳細を見る
In this paper, we propose a new learning algorithm for the multi layered neural network. It determines dynamic learning coefficient by introducing fuzzy inference. In the previous paper, we proposed a fast learning algorithm which changed the learning coefficient for the hidden units (used to a fixed parameter in general) by the sum of the square error for the output units. This learning algorithm was able to accelerate learning by selecting appropriately the ratio of the learning acceleration coefficient to the learning continuation one defined in the paper. Since both learning coefficients were determined by trial and error, the learning algorithm had very troublesome problems whenever it was applied for complicated systems. In this paper, we propose a new learning algorithm for determination of dynamic learning coefficient that is not based on trial and error by introducing fuzzy inference using 6 rules. In the proposed learning algorithm, we also consider the parts where the sum of the square error is not decreased considerably in the medium parts of it. Therefor, the learning algorithm makes more acceleration possible on convergence of the sum of the square error for the output units. The effectiveness of this method (on convergence of the training errors) has been verified by computer simulation for 14 characters pattern recognition problem using the proposed learning algorithm.
東海大学の論文