资源描述:
《03 - Incremental Learning from Noisy Data》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、MachineLearning1:317-354,1986©1986KluwerAcademicPublishers,Boston-ManufacturedinTheNetherlandsIncrementalLearningfromNoisyDataJEFFREYC.SCHLIMMERRICHARDH.GRANGER,JR.(SCHLIMMER@ICS.UCI.EDU)(GRANGER@ICS.UCI.EDU)IrvineComputationalIntelligenceProject,DepartmentofInformation
2、andComputerScience,UniversityofCalifornia,Irvine,CA92717,U.S.A.(ReceivedMarch5,1986)(RevisedMay2,1986)Keywords:learningfromexamples,contingency,systematicnoise,conceptdrift,constructiveinductionAbstract.Inductionofaconceptdescriptiongivennoisyinstancesisdifficultandisfu
3、rtherexacerbatedwhentheconceptsmaychangeovertime.Thispaperpresentsasolutionwhichhasbeenguidedbypsychologicalandmathematicalresults.Themethodisbasedonadistributedconceptdescriptionwhichiscomposedofasetofweighted,symboliccharacterizations.Twolearningprocessesincrementally
4、modifythisdescription.Oneadjuststhecharacterizationweightsandanothercreatesnewcharacteriza-tions.Thelatterprocessisdescribedintermsofasearchthroughthespaceofpossibilitiesandisshowntorequirelinearspacewithrespecttothenumberofattribute-valuepairsinthedescriptionlanguage.T
5、hemethodutilizespreviouslyacquiredconceptdefinitionsinsubsequentlearningbyaddinganattributeforeachlearnedconcepttoinstancedescriptions.AprogramcalledSTAGGERfullyembodiesthismethod,andthispaperreportsonanumberofempiricalanalysesofitsperformance.Sinceunderstandingtherelat
6、ionshipsbetweenanewlearningmethodandexistingonescanbedifficult,thispaperfirstreviewsaframeworkfordiscussingmachinelearningsystemsandthendescribesSTAGGERinthatframework.1.IntroductionTheabilitytoadapttotheenvironmentisanessentialqualityforanyintelligentmechanism.Fordomai
7、nsinwhichlearnershaveextensivepreviousknowledge,suchaselectronics,itisappropriatetoviewlearningasbeingheavilyguidedbythatpriorknowledge.However,indomainsinwhichtherearenohigh-qualitytheories,suchasweatherorfinancialprediction,somefundamentalmethodsmustbeusedtoguidelearn
8、ing.Thispaperinvestigatesabottom-uplearningtechniquewhichdoesnotrelyonastrongdomaintheory.Thespecificclassofle