计算机工程与应用 ›› 2012, Vol. 48 ›› Issue (15): 102-106.

• 数据库、信号与信息处理 • 上一篇    下一篇

两输入幂激励前向神经网络权值与结构确定

张雨浓1,2,劳稳超1,余晓填1,李  钧1   

  1. 1.中山大学 信息科学与技术学院,广州 510006
    2.中山大学深圳研究院,广东 深圳 518057
  • 出版日期:2012-05-21 发布日期:2012-05-30

Weights and structure determination of two-input power-activation feed-forward neural network

ZHANG Yunong1,2, LAO Wenchao1, YU Xiaotian1, LI Jun1   

  1. 1.School of Information Science and Technology, Sun Yat-sen University, Guangzhou 510006, China
    2.Research Institute of Sun Yat-sen University in Shenzhen, Shenzhen, Guangdong 518057, China
  • Online:2012-05-21 Published:2012-05-30

摘要: 基于多元函数逼近与二元幂级数展开理论,构建了一个以二元幂函数序列为隐神经元激励函数的两输入幂激励前向神经网络模型。以该网络模型为基础,基于权值直接确定法以及隐神经元数目与逼近误差的关系,提出了一种网络权值与结构确定算法。计算机仿真与数值实验结果验证了所构建的网络在逼近与去噪方面具有优越的性能,所提出的权值与结构确定算法能够快速、有效地确定网络的权值与最优结构,保证网络的最佳逼近能力。

关键词: 权值与结构确定算法, 二元幂级数展开, 两输入幂激励前向神经网络, 最优结构, 权值直接确定法

Abstract: Based on the theory of multivariate function approximation and two-variable power series expansion, a Two-Input Power-Activation feed-forward Neural Network(TIPANN) model is constructed and studied, of which the hidden-layer neurons’ activation-functions are a sequence of power functions with two variables. Moreover, based on the weights-direct-determination method and the relationship between the number of hidden-layer neurons and the neural network’s approximation error, a Weights-And-Structure-Determination(WASD) algorithm is proposed to determine the optimal number of hidden-layer neurons of the TIPANN. Computer simulation and numerical verification results further substantiate the superiority of the TIPANN in terms of approximation and denoising, as well as the efficacy and accuracy of the proposed WASD algorithm to determine the weights and the optimal structure of the TIPANN.

Key words: Weights-And-Structure-Determination(WASD) algorithm, two-variable power series expansion, two-
input power-activation feed-forward neural network,
optimal structure, weights-direct-determination method