二阶Newton法训练径向基函数神经网络的算法研究.pdf

《二阶Newton法训练径向基函数神经网络的算法研究.pdf》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、计算机研究与发展7544／issn1OOO～1239．2015．20140373JournalofComputerResearchandDevelopment52(7)：1477—1486，2015二阶Newton法训练径向基函数神经网络的算法研究蔡殉陈智KanishkaTyagi。于宽。李子强朱波(山东大学计算机科学与技术学院济南250101)(德克萨斯大学阿灵顿分校电子工程系美国阿灵顿76019)(山东建筑大学材料科学与工程学院济南250101)(山东大学材料科学与工程学院济南250061)(caixunzh@sdu．edu．cn)OrderNewton’SMethodforTrain

2、ingRadialBasisFunctionNeuralNetworksCaiXun，ChenZhi，KanishkaTyagi，YuKuan。，LiZiqiang．andZhuBo4(SchoolofComputerScienceandTechnology，ShandongUniversity，Jin，z250101)。(DepartmentofElectricalEngineering，UniversityofTexasatArlington，Arlingt．USA76019)。(SchoolofMaterialScienceandEngineering，ShandongJianzhu

3、University，Jin250101)(SchoolofMaterialsScienceandEngineering，ShandongUniversity。Jinan250061)AbstractAhybridtwo—stepsecond—orderbatchapproachispresentedforconstructingandtrainingradialbasisfunction(RBF)neuralnetworks．UnlikeotherRBFneuralnetworklearningalgorithms，theproposedparadigmusesNewton’smetho

4、dtotraineachsetofnetworkparameters，i．e．spreadparameters'meanvectorparametersandweighteddistancemeasure(DM)coefficientsandoutputweightsparameters．Forefficientlycalculatingthesecond—orderequationsofNewton’smethod。alltheoptimalparametersarefoundoutusingorthogonalleastsquares(OLS)withthemultiplyoptima

5、1learningfactors(MOLFs)fortrainingmeanvectorparameters．ThesimulationresuitsoftheproposedhybridtrainingalgorithmonarealdatasetarecomparedwiththoseoftherecursiveleastsauarebasedRBF(RLS—RBF)andLevenberg—MarquardtmethodbasedRBF(LM—RBF)trainingalgorithms．A1so，theanalysisofthetrainingperformanceforoptim

6、izationofeachsetofparametershasbeenDresented．TheexperimentalresultsshowthattheproposedhybridoptimalweightedDMtrainingalgorithm，whl。hISbased0ntheoptimizationofthemeanvectors·weightedDMcoefficientsandspreadparameters，hassignificantimprovementontrainingconvergencespeedcomparedwiththatofRLS-RBFandhasv

7、erylesscomputationcostcomparedwiththatofLM—RBF．ItconfirmsthatNewton，smethodsolvedbyOLSisasignificantlyvaluablemethodfortrainingtheRBFneura1network．Keywordsradialbasisfunction(RBF)neuralnetwork；Hessianmatrix；Newto