提高多层前馈网络推广能力的权值控制算法

Controlled Weight Method for Improving the Generalization Ability of Layered Forward Neural Networks

摘要: 研究了多层前馈网络的“过拟合”现象,针对网络拟合效果好,而推广能力差的问题,提出了与BP算法相结合的、以权值修正量小于某一给定的较小值为收敛标准的权值控制算法.在紫坪埔水库工程洪水预报的神经网络模型中的应用表明,该方法具有提高多层前馈网络推广能力的作用.

Abstract: In this paper,the pheno menon of "over fitting" of ANN is studied.A new method called control led-weight is proposed to improve the generalization ability of layered forward neural networks.Its efficiency is given by using daturm from the field of hydrological forecasting.