资源描述：
《iterative learning neurocomputing - mingxuan sun》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、2009InternationalConferenceonWirelessNetworksandInformationSystemsIterativeLearningNeurocomputingMingxuanSunCollegeofInformationEngineeringZhejiangUniversityofTechnologyHangzhou310023,P.R.ChinaEmail:mxsun@zjut.edu.cnAbstractThispaperpresentsaneuralnetworkframeworkII.ARCHITECTUREOFTIME-VARYINGNE
2、URALforimplementingunknowntime-varyingmappings.AunifiedNETWORKSarchitectureoftime-varyingneuralnetworksisproposed,andthemethodologyofiterativelearningisusedforthenetworkByatime-varyingneuralnetworkwerefertothatwhichtraining.Convergenceresultsoftheiterativelearningleastweights,inputs,andoutputsar
3、eallowedtovarywithtime.squaresalgorithmarederivedunderassumptionofboundedAmultiple-inputsingle-outputtime-varyingneuralnetworkinputsignals.Periodicneuralnetworksareexploredaswelltoisdescribedasbeusedasperiodicfunctionapproximationtools.IKeywords-Neuralnetworks;time-varyingsystems;periodicy(t)=
4、wi(t)φi(x(t))(1)systems;iterativelearningi=1whereIindicatesthenumberofneurons,x(t)=[x(t),···,x(t)]Tistheinputvector,y(t)isthescalarI.INTRODUCTION1noutput,wi(t),i=1,···,I,aretheweights,andφi(x(t))Overthepastfewdecades,variousneuralnetworkmodelsarethevector-valuedactivationfunctions.havebeenpropo
5、sed,andusedinalmosteveryareaofappli-Themostcommonchoicesforfunctionφ(·)arecation[1].Artificialneuralnetworksprovideageneralcom-1)Heavisidestepfunctionputingframeworkforsolvingpracticalproblems,because1ifz≥0oftheiruniversalapproximationability.Inthepublishedφ(z)=0ifz<0literature,therehavebeeneff
6、ortsdevotedforassessingthefeasibilityofapplyingneuralnetworkstoapproximate2)Logisticfunctionφ(z)=1/(1+e−z);andunknownfunctionsariseninpracticalapplications.3)Hyperbolictangentfunctionφ(z)=(ez−Feedforwardneuralnetworksaretypicalartificialneu-e−z)/(ez+e−z).ralnetworkswherenointernalfeedbacksignalp
7、athsareNotethatφ(·)presentedabovearegloballyboundedwithpresent.Theconventionalonesarewithconstantweighs.Werespecttox.refertothemasTime-InvariantNeuralNetworks(TINNs).Itisinterestingtonotethatinrecentdevelopments,Thegradient-basedb
