2010-2016精選論文2016-shahriari-bayesopt-ieee-2016

1、INVITEDPAPERTakingtheHumanOutoftheLoop:AReviewofBayesianOptimizationThepaperintroducesthereadertoBayesianoptimizationhighlightingitsmethodicalaspectsshowcasingitsapplications.ByBobakShahriariKevinSwerskyZiyuWangRyanP.Ada

2、msNodeFreitasABSTRACT|BigDataapplicationsaretypicallyassociatedwithsystemsinvolvinglargenumbersofusersmassivecomplexsoftwaresystemslargescaleheterogeneouscomputingstagearchitectures.Theconstructionofsuchsystemsinvolvesma

3、nydistributeddesignchoices.Theendproducts(e.g.recommendationsystemsmedicalanalysistoolsrealtimegameenginesspeechrecognizers)thusinvolvemanytunableconfigurationparameters.Theseparametersareoftenspecifiedhardcodedintotheso

4、ftwarebyvariousdevelopersteams.Ifoptimizedjointlytheseparameterscanresultinsignificantimprovements.Bayesianoptimizationisapowerfultoolfthejointoptimizationofdesignchoicesthatisgaininggreatpopularityinrecentyears.Itpromis

5、esgreaterautomationsoastoincreasebothproductqualityhumanproductivity.ThisreviewpaperintroducesBayesianoptimizationhighlightssomeofitsmethodologicalaspectsshowcasesawiderangeofapplications.KEYWDS|Decisionmakingdesignofexp

6、erimentsoptimizationresponsesurfacemethodologystatisticallearningI.INTRODUCTIONDesignproblemsarepervasiveinscientificindustrialendeavours:scientistsdesignexperimentstogaininsightsintophysicalsocialphenomenaengineersdesig

7、nmachinestoexecutetasksmeefficientlypharmaceuticalresearchersdesignnewdrugstofightdiseasecompaniesdesignwebsitestoenhanceuserexperienceincreaseadvertisingrevenuegeologistsdesignexplationstrategiestoharnessnaturalresource

8、senvironmentalistsdesignsenswkstomonitecologicalsystemsdevelopersdesignsoftwaretodrivecomputerselectronicdevices.Allthesedesignproblemsarefraughtwithchoiceschoicesthatareoftencomplexhighdimensionalwithinteractionsthatmak

9、ethemdifficultfindividualstoreasonabout.FexamplemanyganizationsroutinelyusethepopularmixedintegerprogrammingsolverIBMILOGCPLEX1fschedulingplanning.Thissolverhas76freeparameterswhichthedesignersmusttunemanuallyVanoverwhel

10、mingnumbertodealwithbyh.Thissearchspaceistoovastfanyonetoeffectivelynavigate.Megenerallyconsiderteamsinlargecompaniesthatdevelopsoftwarelibrariesfotherteamstouse.Theselibrarieshavehundredsthoussoffreechoicesparameterstha

11、tinteractincomplexways.Infactthelevelofcomplexityisoftensohighthatitbecomesimpossibletofinddomainexpertscapableoftuningtheselibrariestogenerateanewproduct.Asasecondexampleconsidermassiveonlinegamesinvolvingthefollowingth

12、reeparties:contentprovidersuserstheanalyticscompanythatsitsbetweenthem.Theanalyticscompanymustdevelopprocedurestoautomaticallydesigngamevariantsacrossmillionsofuserstheobjectiveistoenhanceuserexperiencemaximizethecontent

13、provider’srevenue.ManureceivedMay12015revisedJuly62015acceptedJuly202015.DateofpublicationDecember102015dateofcurrentversionDecember182015.B.ShahriariiswiththeUniversityofBritishColumbiaVancouverBCV6T1Z4Canada(email:).K.

14、SwerskyiswiththeUniversityofTontoTontoONM5S1A1CanadaalsowithTwitterBostonCambridgeMA02139USA(email:).Z.WangiswithOxfdUniversityOxfdOX12JDU.K.alsowithGoogleDeepmindLondonN1C4AGU.K.(email:ziyu@).R.P.AdamsiswithHarvardUnive

15、rsityCambridgeMA02138USAalsowithTwitterUSA(email:).N.deFreitasiswithOxfdUniversityOxfdOX12JDU.K.withGoogleDeepMindLondonN1C4AGU.K.alsowiththeCanadianInstitutefAdvancedResearchTontoONM5G1Z8Canada(email:nodefreitas@).Digit

16、alObjectIdentifier:10.1109JPROC.2015.24942181softwarecommerceoptimizationcplexoptimizer00189219?2015IEEE.Translationscontentminingarepermittedfacademicresearchonly.Personaluseisalsopermittedbutrepublicationredistribution

17、requiresIEEEpermission.See:www.ieee.gpublications_stardspublicationsrightsindex.htmlfmeinfmation.148ProceedingsoftheIEEE|Vol.104No.1January2016Algithm1:Bayesianoptimization1:fn12...do2:newxn1byoptimizingacquisitionfuncti

18、on?xn1argmaxx?xDn3:queryobjectivefunctiontoobtainyn14:augmentdataDn1fDnxn1yn1g5:updatestatisticalmodel6:endfOneproblemwiththisminimumexpectedriskframewkisthatthetruesequentialriskuptothefullevaluationbudgetistypicallycom

19、putationallyintractable.Thishasledtotheintroductionofmanymyopicheuristicsknownasacquisitionfunctionse.g.Thompsonsampling(TS)probabilityofimprovementexpectedimprovement(EI)upperconfidenceboundsentropysearch(ES).Theseacqui

20、sitionfunctionstradeoffexplationexploitationtheiroptimaarelocatedwheretheuncertaintyinthesurrogatemodelislarge(explation)wherethemodelpredictionishigh(exploitation).Bayesianoptimizationalgithmsthenthenextquerypointbymaxi

21、mizingsuchacquisitionfunctions.Naturallytheseacquisitionfunctionsareoftenevenmemultimodaldifficulttooptimizeintermsofqueryingefficiencythantheiginalblackboxfunctionf.Therefeitiscriticalthattheacquisitionfunctionsbecheapt

22、oevaluateapproximate:cheapinrelationtotheexpenseofevaluatingtheblackboxf.Sinceacquisitionfunctionshaveanalyticalfmsthatareeasytoevaluateatleastapproximateitisusuallymucheasiertooptimizethemthantheiginalobjectivefunction.

23、A.PaperOverviewInthispaperweintroducetheingredientsofBayesianoptimizationindepth.OurpresentationisuniqueinthatweaimtodisentanglethemultiplecomponentsthatdeterminethesuccessofBayesianoptimizationimplementations.Inparticul

24、arwefocusonstatisticalmodelingFig.1.IllustrationoftheBayesianoptimizationprocedureoverthreeiterations.Theplotsshowthemeanconfidenceintervalsestimatedwithaprobabilisticmodeloftheobjectivefunction.Althoughtheobjectivefunct

25、ionisshowninpracticeitisunknown.Theplotsalsoshowtheacquisitionfunctionsinthelowershadedplots.Theacquisitionishighwherethemodelpredictsahighobjective(exploitation)wherethepredictionuncertaintyishigh(explation).Notethatthe

26、areaonthefarleftremainsunsampledaswhileithashighuncertaintyitiscrectlypredictedtoofferlittleimprovementoverthehighestobservation[27].Shahriarietal.:TakingtheHumanOutoftheLoop:AReviewofBayesianOptimization150Proceedingsof