RBF神经网络的结构动态优化设计

  本文关键词：RBF神经网络的结构动态优化设计，由笔耕文化传播整理发布。

图8Fig.8

D-RBF神经网络训练过程Thetrainingprocessof

D-RBF

图10Fig.10

基于D-RBF的COD建模结果

ThemodellingofCODbasedon

D-RBF

图9

Fig.9

训练过程中神经元数

Fig.11

图11基于D-RBF的COD建模误差

Theneuronsleftinthetrainingprocess

ThemodellingerrorsofCODbasedonD-RBF

推荐值.训练过程中误差变化如图8所示,训练过程中隐含层神经元的变化如图9所示,对COD的建模结果如图10和图11所示.

仿真结果表明:原水中的有机污染物(COD约300～500mg/L)得到有效去除(出水负荷COD在不同的时刻都能保持在30mg/L左右),图10和图11显示实测COD值与本模型的输出值基本吻合,相对误差小于0.02,证明该模型是有效的.1)不依赖于RBF网络的初始结构,能够根据实际对象调整RBF神经网络,获得适合对象的RBF神经网络;

2)通过神经网络输出敏感度增加和删除RBF神经网络中的神经元,最终获得的RBF神经网络结构简洁,逼近能力强;

3)本文提出的D-RBF神经网络结构稳定,为复杂系统建模提供了技术支持.

References

1YeJian,GeLin-Dong,WuYue-Xian.Anapplicationofim-provedRBFneuralnetworkinmodulationrecognition.ActaAutomaticaSinica,2007,33(6):652?654

(叶健,葛临东,吴月娴.一种优化的RBF神经网络在调制识别中的应用.自动化学报,2007,33(6):652?654)2ChenS,WangXX,BrownDJ.Sparseincrementalre-gressionmodelingusingcorrelationcriterionwithboost-ingsearch.IEEESignalProcessingLetters,2005,12(3):198?2013ChenS,WolfgangA,HarrisCJ,HanzoL.SymmetricRBF

3结论

针对RBF神经网络设计结构问题,提出了一种基于输出敏感度法的动态RBF神经网络(D-RBF),D-RBF在保证神经网络收敛性能的前提下实现结构在线调整,提高神经网络的自适应能力;通过逼近非线性函数和对非线性系统关键参数进行建模,以及与其他动态RBF神经网络进行比较,得到以下结论

:

classi?erfornonlineardetectioninmultiple-antenna-aidedsystems.IEEETransactionsonNeuralNetworks,2008,19(5):737?745

4ChaiTian-You,WangXiao-Bo.ApplicationofRBFneuralnetworksincontrolsystemoftheslabacceleratingcoolingprocess.ActaAutomaticaSinica,2000,26(2):219?225(柴天佑,王笑波.RBF神经网络在加速冷却控制系统中的应用.自动化学报,2000,26(2):219?225)5GonzalezJ,RojasI,OrtegaJ,PomaresH,FernandezFJ,DiazAF.Multi-objectiveevolutionaryoptimizationofthesize,shape,andpositionparametersofradialbasisfunctionnetworksforfunctionapproximation.IEEETransactionsonNeuralNetworks,2003,14(6):1478?14956LeungFHF,LamHK,LingSH,TamPKS.Tuningofthestructureandparametersofaneuralnetworkusinganimprovedgeneticalgorithm.IEEETransactionsonNeuralNetworks,2003,14(1):79?887BorsAG,PitasI.Medianradialbasisfunctionneuralnet-work.IEEETransactionsonNeuralNetworks,1996,7(6):1351?13648YinH,AllinsonNM.Self-organizingmixturenetworksforprobabilitydensityestimation.IEEETransactionsonNeu-ralNetworks,2001,12(2):405?4119EspositoA,MarinaroM,OricchioD,ScarpettaS.Ap-proximationofcontinuousanddiscontinuousmappingsbyagrowingneuralRBF-basedalgorithm.NeuralNetworks,2000,13(6):651?66510FuXJ,WangLP.Datadimensionalityreductionwithap-plicationtosimplifyingRBFnetworkstructureandimprov-ingclassi?cationperformance.IEEETransactionsonSys-tems,Man,andCybernetics,PartB:Cybernetics,2003,33(3):399?40911LuYW,SundararajanN,SaratchandranP.Asequential

learningschemeforfunctionapproximationusingminimalradialbasisfunction(RBF)neuralnetworks.NeuralCom-putation,1997,9(2):461?47812PanchapakesanC,PalaniswamiM,RalphD,ManzieC.Ef-fectsofmovingthecentersinanRBFnetwork.IEEETrans-actionsonNeuralNetworks,2002,13(6):1299?130713HuangGB,SaratchandranP,SundararajanN.Ageneral-izedgrowingandpruningRBF(GGAP-RBF)neuralnet-workforfunctionapproximation.IEEETransactionsonNeuralNetworks,2005,16(1):57?6714FengHM.Self-generationRBFNsusingevolutionalPSO

learning.Neurocomputing,2006,70(1-3):241?25115HuangDS,DuJX.Aconstructivehybridstructureop-timizationmethodologyforradialbasisprobabilisticneu-ralnetworks.IEEETransactionsonNeuralNetworks,2008,19(12):2099?2115

16ChenS,HongX,LukBL,HarrisCJ.Constructionoftun-ableradialbasisfunctionnetworksusingorthogonalforwardselection.IEEETransactionsonSystems,Man,andCyber-netics,PartB:Cybernetics,2009,39(2):457?46617SaltelliA,RattoM,TarantolaS,CampolongoF.Sensitivity

analysispractices:strategiesformodel-basedinference.Re-liabilityEngineeringandSystemsSafety,2006,91(10-11):1109?1125

18RattoM,TarantolaS,SaltelliA.Sensitivityanalysis

inmodelcalibration:GSA-GLUEapproach.ComputerPhysicsCommunications,2001,136(3):212?22419CariboniJ,GatelliD,LiskaR,SaltelliA.Theroleofsensi-tivityanalysisinecologicalmodeling.EcologicalModelling,2007,203(1-2):167?18220CampolongoF,CariboniJ,SaltelliA.Ane?ectivescreening

designforsensitivityanalysisoflargemodels.Environmen-talModellingandSoftware,2007,22(10):1509?151821BuzziC,GrippoL,SciandroneM.Convergentdecomposi-tiontechniquesfortrainingRBFneuralnetworks.NeuralComputation,2001,13(8):1891?192022LiuYY,StarzykJA,ZhuZ.Optimizedapproximational-gorithminneuralnetworkswithoutover?tting.IEEETrans-actionsonNeuralNetworks,2008,19(6):983?99523HenzeM,GradyCPL,GujerW,MaraisGVR,Matsuo

T.Ageneralmodelforsingle-sludgewastewatertreatmentsystems.WaterResearch,1987,21(5):505?515

乔俊飞北京工业大学教授.主要研究方向为智能控制,神经网络分析与设计.E-mail:junfeq@bjut.edu.cn

(QIAOJun-FeiProfessoratBeijingUniversityofTechnology.Hisresearchinterestcoversintelligentcontrol,andanalysisanddesignofneuralnetworks.)

韩红桂北京工业大学博士研究生.主要研究方向为复杂过程建模与控制,神经网络分析与设计.本文通信作者.E-mail:isibox@sina.com

(HANHong-GuiPh.D.candidateatBeijingUniversityofTechnology.Hisresearchinterestcoversmodelingandcontrolincomplexprocess,and

analysisanddesignofneuralnetworks.Correspondingau-thorofthispaper.)

博泰典藏网btdcw.com包含总结汇报、表格模板、计划方案、初中教育、外语学习、农林牧渔、教学研究、人文社科、党团工作以及RBF神经网络的结构动态优化设计_乔俊飞_图文等内容。

本文共2页12

  本文关键词：RBF神经网络的结构动态优化设计，由笔耕文化传播整理发布。