基于非高斯估計(jì)的Kernel PCA核優(yōu)化方法及應(yīng)用.pdf

1、代號(hào)分類號(hào)107IJl_●_●一一TP39l學(xué)號(hào)蜊；j；jj曼密級(jí)套五兩簧雹手抖毅競警碩士學(xué)位論文題(、英文)目一基于韭直斬估進(jìn)的一K豎墮!一膽墨援餞化直法羼廛且舅馥叫一Q刨趔；州娥劍呂0照thm_照K卿!T￡壘曼；卿—!n_M9衛(wèi)：旦；蚋啦!y一№臼啦!地Ⅱ作者姓名王毅娜指導(dǎo)教師姓名、職務(wù)張軍英教授學(xué)科門美王堂一學(xué)科、專業(yè)苴簋妞i剿攮握一一提交論文日期二oo六年一月AbstractKemel’basedPrincipleCompone

2、ntAnalysis(KemelPCA)isanewlyproposedfcatureex“矗ctionmethodwhichgeneralizesprinciplecomponentanalysis(PCA)t‘’lhcnonlinearcasebyuseofkemeltrickKemelPCAcane療’ectjvelyextractnonlinearfeaturesofdatasetMoreover，comparedtoPCA，j

3、thasnoconstraintondatadistributioninoriginalspacea11dnoincreaseincomputationalcomplexitySoKemelPCAhasbeenwidelyusedindatacompression，denoisingandreconstructionThemainproblemofKemeIPCAisthatitsperf、ormanceisscronglyinnucn

4、cedbyt1eparametelsjnthekemelfunctionusedforkernelPCA，andthemappeddala，whjchcouldnotbeobtainedexplicitly，makestheparameteroptimizationworkmoredimcultInthispaper，weproposeanovelpararneteroptimizingalgorithmbasedonthenongau

5、ssiandistributionestimationinfeaCurespaceBasedontheideatha“heoptimaIparametershouIdleadthemappeddatainfeatllrespacebeasGaussianaspossble，ourmethodanalysesthenongaussianstructureofthemappeddata，andtheninVcrselyestimatesth

6、edegreeofitsdistributionclosetotheGaussianoneinfeaturespaceTheexperiments，bothonsimulateddataa11drealworlddata，demonstrateexcellentresultswhichshowea’ectivenessofthemethodproposedinthispaperKeywords：Kernel書asedPrincipleC