| 引用本文: | 卢 进.一种用于神经网络训练的隐节点校正算法[J].控制理论与应用,1998,15(1):53~60.[点击复制] |
| LU Jin.A Hidden Node Value Regulation Algorithm or Neural Network Training[J].Control Theory and Technology,1998,15(1):53~60.[点击复制] |
|
| 一种用于神经网络训练的隐节点校正算法 |
| A Hidden Node Value Regulation Algorithm or Neural Network Training |
| 摘要点击 1323 全文点击 459 投稿时间：1994-11-07 修订日期：1997-06-11 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 |
| 1998,15(1):53-60 |
| 中文关键词 神经网络 学习算法 泛化能力 |
| 英文关键词 neural network learning algorithm generalization capacity |
| 基金项目 |
|
| 中文摘要 |
| 误差反传算法被广泛用于多层前馈神经网络的训练， 但该算法的收敛性问题并没有解决， 这导致训练后的网络泛化能力一般很差. 本文研究了这一问题， 并基于神经网络映射定理提出一种用于训练网络逼近单输出函数的隐节点校正（HNR）算法. 这在神经辨识领域是有用的， 因为大多数工业对象都是多输入单输出的. 我们对HNR算法的收敛性和泛化能力作了理论上的研究， 仿真实验和在催化重整过程建模中的应用实例表明该算法在一定条件下具有很高的学习速度和较强的泛化能力. |
| 英文摘要 |
| Error back propagation algorithm is widely used for the training of multi layer feed forward neural networks. But the convergency of this method is actually an open problem,which often leads to its poor generalization capacity. In this paper,we probe the situation and propose a Hidden Node Value Regulation （HNR） algorithm for approximating single output functions based on the mapping neural network existence theorem. This is useful in neural identification,for many industrial plants are multi-input-single-output. Theoretical problems of the convergence and gener-alization capacity of the HNR algorithm are also studied. Simulation results as well as an applica-tion case of modeling a catalytic reform process show some interesting characteristics of this algo-rithm,including its relatively high speed of sample learning and strong capacity of generalization under certain circumstance. |