切比雪夫神经网络权值与结构确定及其分类应用  

Weights and structure determination of Chebyshev-polynomial neural networks for pattern classification

作　　者：殷勇华[1] 张雨浓[1] 

机构地区：[1]中山大学信息科学与技术学院,广州510006

出　　处：《软件》2012年第11期172-180,共9页Software

基　　金：教育部高等学校博士学科点专项科研基金博导类课题(20100171110045);国家大学生创新训练项目(201210558042)

摘　　要：本文提出一种可用于模式分类的新型的单输出切比雪夫(Single-Output Chebyshev-Polynomial)神经网络(简称SOCP网络)。基于SOCP网络,本文进而提出另一种可用于模式分类的多输出切比雪夫(Multi-Output Chebyshev-Polynomial)神经网络(简称MOCP网络)。为避免冗长的迭代学习过程,本文采用权值直接确定法一步获得网络的最优权值。另外,为使网络同时具备较好的拟合和泛化能力,本文提出四折交叉验证法用于确定网络适当的隐层神经元数目。结合权值直接确定法和四折交叉验证法,本文最终提出SOCP网络和MOCP网络相对应的权值与结构确定法。数值实验结果验证了所提出的SOCP网络、MOCP网络以及相对应权值与结构确定法的有效性,并且由该算法所确定的网络在模式分类的应用中具有很高的准确率和很强的鲁棒性。This paper firstly proposes anewtype ofsingle-outputChebyshev-polynomial feed-forwardneuralnetwork（SOCPNN） for pattern classification. Then, based on such a SOCPNN, another new type of multi-output Chebyshev-polynomial feed-forward neural network （MOCPNN） is proposed for pattern classification in this paper. In order to avoid lengthy iterative-learning procedure, the weights- direct-determination （WDD） method is applied to obtaining the optimal weights of the SOCPNN and MOCPNN. In addition, the 4-fold cross-validation （4FCV） method is used to determine the appropriate numbers of the SOCPNN and MOCPNN hidden-layer neurons such that they can achieve good performances （i.e., approximation and generalization）. By combining the presented WDD with 4FCV methods, two weights-and-structure-determination （WASD） algorithms, one for the SOCPNN and the other for the MOCPNN, are thus proposed for pattern classification. Furthermore, experiment results substantiate the high accuracy and strong robustness of the proposed SOCPNN and MOCPNN equipped with the WASD algorithms for pattern classification.

关 键 词：切比雪夫多项式 神经网络 权值与结构确定 模式分类 鲁棒性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]