通過交叉驗(yàn)證準(zhǔn)則選擇線性模型.pdf

1、東南大學(xué)碩士學(xué)位論文通過交叉驗(yàn)證準(zhǔn)則選擇線性模型姓名：聞斌申請學(xué)位級別：碩士專業(yè)：概率論與數(shù)理統(tǒng)計(jì)指導(dǎo)教師：江其保20060101AbstractInthefieldofmodel(orvariable)selectionintilemultipleregressionmodel，manymethodsareappliedtoaddressthisproblem；forexample，theqcriterion，Akaikeinform

2、ationcriterion(Akaiket1974，AIC)，Bayesinformationcriterion(BIC)，the≯criterionandCrossvalidationcriterionbecausethesecriteriau∞afixedchoiceofthepenaltyJunctionHowever，afixedchoicemaybegoodinsolnesituationsandmaynotperformw

3、ellinsomeothersituations；sothechoiceofthepenaltyfunctionwillaffecttheperformanceofamodelselectioncriterionHencethereisaneedtofindadateorientedpenaltyBothataprocedurewithitsusewillperformwell田地丑r8tattempttoprovideadateori

4、entedpenaltyfunctionismadeinRaoandWu(19S9)，whichw88appliedtomodelselectionproblemsinthemultipleregre％ionmodelInthispaperourobjectistopursuetheinvestigationstartedinRaoandwh(1989)柚dmake∞merefinements，Weconsidertheproblemo

5、fmodelselectionintheclassicalregre目ionmodeIbasedoncroesvalidationwith∞addpenaltytermforpenalizingoverfittingUnderSonleweakconsidersthenewcriterionj8shownt0bestronglyconsistentinthesensethatwithprobabilityoneforalllarge’1

6、AsshowninollrMonteCarlosimulationthecriterionwith8da㈣entedpenaltyprovidesimprovedperformanceovertheer／terionwith8fixedchoiceofthepenaltyfunctionanditworkswellinsinaisamplesizesKeywords：AIC，BIC，GIC，ConsistencyCrossvalidat