基于回歸型支持向量機的數(shù)據(jù)建模方法研究.pdf

1、江南大學碩士學位論文基于回歸型支持向量機的數(shù)據(jù)建模方法研究姓名：邵信光申請學位級別：碩士專業(yè)：控制理論與控制工程指導教師：楊慧中20060201浮章^‘f顧”學位論文AbstractsupponVectorMachine(SVM)isanewdatamodelingmethod，theoreticallybasedonstatisticlearningtheoryEmployingthecriteriaofstructuralriskm

2、inimization，whichminimizesthefittingerrorsofthesampledataandcontrolsimuhaneouslythecomplexityofthelearningfunction，|hustheSVM’sgeneralizationisbetterthanothersInthispaper，highlightswerefbcusedonSupportVectorRegression(SV

3、R)，andaseriesofworkonhowtoimprovetheSVR’sspeedandgeneralizationwasdiscussedAfteranalyzingtheSVRindetail，aheuristicreducedSVRalgorithmwasproposedbasedonthecharacteristicofsolutionsparsenessandthekernelsimilaritymeasuremen

4、tThroughnonlinearmappingandket。11cItrickthenewnrcthudmappcdtheoriginalspacesample～dataintoahighdimensionall色aturespace，andthenthefeaturesspacebecameclearerLotsofsimilardatawasdiscardedbasedonthekernelsimilaritymeasuremen

5、!，andthesizeofsample—dataforSVRtrainingbecamesmall；thetrainingtimewasdecreasedgreatlywithoutcompromisingthegeneralizationcapabilitySeveralnumericalsimulationsshowthatthisnewmethodwaseffectiveInordertospeeduptheSVRfurther

6、asmoothingtechniqueinmathematicalprogrammingwasintroduced，andancweinsensitivesmoothsupportvectorregressionbasedonLogisticlossfunction，shortlyL￡一SSVRwasproposedinthispaperTheconventionalconditionalquadraticprogrammingprob

7、leminSVRwastransferredintoanunconditionalminimizingproblem，thus，thenumbersofsolvingvariableswerereducedbynmIhcIe。SSVRwasimplcnrcntedwithaQuasi—NewtonMethoddirectlywithoutanyoptimizationpackagcs，andtheCPUtimewasdecreasedS

8、imulationsofartificialandrealdatademonstratedthattheL￡一SSVRwasalsoeffecliveThepreviousresearchesshowthattheSVR’Sgeneralizationcapacitywasgreatlyimpactedbyitsparameters，suchaskernelparameters，penaltyt、actorandcSincetherea

9、refewanalyticalmethodstochoosetheSVR’Sparameters，allautomaticparametersselectionstrategybasedonParticleSwarmOptimization(PSO)algorithmwasproposedInthisnewmethod，eachparticleindicatedagroupofSVRparametersandthe5foldCROSS—

10、validationerrorwasusedasthefitnessfunction()fPSOSimulationsofaaificialandrealdatashowthatthccxccIlentlv2IobaIsearchingabilityot、PSOcontributedthetaskol‘paranlclcissclccli【)114lfcaII、rConsideringtheentireextentmodelselect