A shot introduction to Boosting

《A shot introduction to Boosting》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、JournalofJapaneseSocietyforArtificialIntelligence,14(5):771-780,September,1999.(InJapanese,translationbyNaokiAbe.)AShortIntroductiontoBoostingYoavFreundRobertE.SchapireAT&TLabsResearchShannonLaboratory180ParkAvenueFlorhamPark,NJ07932USAwww.research.att.com/
yoav,schapireyoav,schap

2、ire@research.att.comAbstractBoostingisageneralmethodforimprovingtheaccuracyofanygivenlearningalgorithm.ThisshortoverviewpaperintroducestheboostingalgorithmAdaBoost,andexplainstheun-derlyingtheoryofboosting,includinganexplanationofwhyboostingoftendoesnotsufferfromoverttingaswellasboo

3、sting'srelationshiptosupport-vectormachines.Someexamplesofrecentapplicationsofboostingarealsodescribed.IntroductionAhorse-racinggambler,hopingtomaximizehiswinnings,decidestocreateacomputerprogramthatwillaccuratelypredictthewinnerofahorseracebasedontheusualinformation(numberofracesrece

4、ntlywonbyeachhorse,bettingoddsforeachhorse,etc.).Tocreatesuchaprogram,heasksahighlysuccessfulexpertgamblertoexplainhisbettingstrategy.Notsurprisingly,theexpertisunabletoarticulateagrandsetofrulesforselectingahorse.Ontheotherhand,whenpresentedwiththedataforaspecicsetofraces,theexperth

5、asnotroublecomingupwitha“ruleofthumb”forthatsetofraces(suchas,“Betonthehorsethathasrecentlywonthemostraces”or“Betonthehorsewiththemostfavoredodds”).Althoughsucharuleofthumb,byitself,isobviouslyveryroughandinaccurate,itisnotunreasonabletoexpectittoprovidepredictionsthatareatleastalittl

6、ebitbetterthanrandomguessing.Furthermore,byrepeatedlyaskingtheexpert'sopinionondifferentcollectionsofraces,thegamblerisabletoextractmanyrulesofthumb.Inordertousetheserulesofthumbtomaximumadvantage,therearetwoproblemsfacedbythegambler:First,howshouldhechoosethecollectionsofracespresent

7、edtotheexpertsoastoextractrulesofthumbfromtheexpertthatwillbethemostuseful?Second,oncehehascollectedmanyrulesofthumb,howcantheybecombinedintoasingle,highlyaccuratepredictionrule?Boostingreferstoageneralandprovablyeffectivemethodofproducingaveryaccuratepre-dictionrulebycombiningroughan

8、dmode