lcs研究概述ppt课件.pptx
- 文档编号:11476785
- 上传时间:2023-06-01
- 格式:PPTX
- 页数:47
- 大小:1.25MB
lcs研究概述ppt课件.pptx
《lcs研究概述ppt课件.pptx》由会员分享,可在线阅读,更多相关《lcs研究概述ppt课件.pptx(47页珍藏版)》请在冰点文库上搜索。
ResearchonLearningClassifierSystem,学习分类系统研究概述,Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
HotIssuesandFutureDirectionWhatisapromisingfutureresearchdirection?
Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
HotIssuesandFutureDirectionWhatisapromisingfutureresearchdirection?
Adaptiverule-basedproductionsystemSetofrulesTrialanderror:
强化学习通过环境的反馈来调整自身的行为Survivalofthefittest:
遗传算法以遗传算法来探索、发现规则,IntroductiontoLCS,RuleA,RuleB,ReinforcementLearning,Environment,RLAgent,action,Reward&State,GeneticAlgorithm,Population,
(2),Population,Population,NewIndividual,GoodIndividual,
(1),Individual,Selection,Reproduction,ComponentsinLCS,ComponentsinLCS,ComponentsinLCS,BriefViewofLCSHistory,1971Holland首次提出分类系统概念,1978Holland正式确立学习分类系统名称,并提出大概框架,1988Holland定义标准框架(太复杂)LCS研究停滞,1995Wilson进一步提出XCS,从此LCS的研究进入新的阶段,1994Wilson简化了标准LCS,提出更易实现的ZCS,1998Stolzmann提出不同于传统LCS的A-LCS,新的方向,Relative,ConferencesandMagazinesGeneticandEvolutionaryComputationConference(GECCO)InternationalWorkshoponLearningClassifierSystems(IWLCS)SEAL,EvolutionaryComputation/IEEETransactiononECPapersandApplicationsH.Ishibuchi.FuzzyGenetics-BasedMachineLearningSEAL2012PierLucaLanzi.XCSwithAdaptiveActionMappingSEAL2012R.Urbanowicz.Instance-LinkedAttributeTrackingandFeedbackforMichigan-StyleSupervisedLearningClassierSystemsGECCO2012M.Iqbal.ExtractingandUsingBuildingBlocksofKnowledgeinLearningClassierSystemsGECCO2012,Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
HotIssuesandFutureDirectionWhatisapromisingfutureresearchdirection?
BriefViewofLCSHistory,1971Holland首次提出分类系统概念,1978Holland正式确立学习分类系统名称,并提出大概框架,1988Holland定义标准框架(太复杂)LCS研究停滞,1995Wilson进一步提出XCS,从此LCS的研究进入新的阶段,1994Wilson简化了标准LCS,提出更易实现的ZCS,1998Stolzmann提出不同于传统LCS的A-LCS,新的方向,HollandsLCS,缺陷:
1.无节制使用遗传算法2.桶队列算法的依赖性,规则条件/动作/预测匹配集M动作选择动作集A,BriefViewofLCSHistory,1971Holland首次提出分类系统概念,1978Holland正式确立学习分类系统名称,并提出大概框架,1988Holland定义标准框架(太复杂)LCS研究停滞,1995Wilson进一步提出XCS,从此LCS的研究进入新的阶段,1994Wilson简化了标准LCS,提出更易实现的ZCS,1998Stolzmann提出不同于传统LCS的A-LCS,新的方向,WilsonsXCS,最重要的改进部分:
重新定义了适应度计算方法HollandsLCS:
规则的权值WilsonsXCS:
引入了新的参数通过计算精确度来度量遗传算法,BriefViewofLCSHistory,1971Holland首次提出分类系统概念,1978Holland正式确立学习分类系统名称,并提出大概框架,1988Holland定义标准框架(太复杂)LCS研究停滞,1995Wilson进一步提出XCS,从此LCS的研究进入新的阶段,1994Wilson简化了标准LCS,提出更易实现的ZCS,1998Stolzmann提出不同于传统LCS的A-LCS,新的方向,StolzmannsACS,Model-FreeLCSsZCS/XCSNoknowledgeaboutresultofactionsModel-BasedLCSAnticipatoryclassifiersystems(ACS,1998)Anticipatorylearningclassifiersystems(ACS2,2000)Knowledgeaboutresultofactions,TwoApproaches,
(1)MichiganApproach:
Searchforgoodrules
(2)PittsburghApproach:
Searchforagoodrulecombination,Champions=Goodplayers+Goodcooperation,MichiganApproach,FitnessEvaluationofEachRuleDirectOptimizationofRulesNewrulesaregeneratedfromgoodrulesIndirectSearchforaGoodRuleSetAsetofgoodrulesisnotnecessarilyagoodruleset,RuleA,RuleC,RuleE,RuleG,RuleB,RuleD,RuleF,RuleH,PittsburghApproach,FitnessEvaluationofEachSubRuleSetDirectOptimizationofRuleSetsNewrulesetsaregeneratedfromgoodrulesetsIndirectSearchforGoodRulesGoodrulesinapoorrulesetcannotsurvive,RuleARuleBRuleC,RuleDRuleERuleF,RuleGRuleHRuleI,RuleJRuleKRuleL,Michigan-PittsburghHybridApproach,H.Ishibuchietal.HybridizationofFuzzyGBMLApproachesforPatternClassificationProblems,IEEET-SMCPartB(2005),Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
HotIssuesandFutureDirectionWhatisapromisingfutureresearchdirection?
ImprovementofLCS,Sub-LCS,LCSE,RuleA,RuleC,RuleE,RuleB,RuleB,RuleD,RuleF,RuleA,ability,readability,Sub-LCS,Sub-LCS,EnsembleMethod,ParallelensembleBagging,Randomsubspace,RandomforestcreatediversebaselearnersbyintroducingrandomnessSequentialensembleAdaboostcreatebaselearnersbycomplementarity,LCSE:
LCSEnsemble(Bagging),LCSE:
LCSEnsemble(Boosting),CompactRuleSet,(Supposesimplestconditions)2-DProblem:
32=9rules4-DProblem:
34=81rules6-DProblem:
36=729rules8-DProblem:
38=6,561rules10-DProblem:
310=59,049rules,lackofreadabilitytraditionalCRAistoocomplicated,CompactRuleSet,YangGao,LeiWu,JoshuaZhexueHuang.EnsembleLearningClassifierSystemandCompactRuleset.In:
Proceedingsofthe6thInternationalConferenceonSimulatedEvolutionandLearning.LNCS4247,pp:
42-49,2006.,CompactRuleSet,YangGao,LeiWu,JoshuaZhexueHuang.EnsembleLearningClassifierSystemandCompactRuleset.In:
Proceedingsofthe6thInternationalConferenceonSimulatedEvolutionandLearning.LNCS4247,pp:
42-49,2006.,LCSInLearningApplications,Preprocess:
训练数据的缺失/噪声?
LCS用于处理不同的学习情况?
L,LU,U,?
?
显然可以,效果明显,Semi-supervised,Classification,Clustering,训练数据,HandlingMissingData,D.GuandY.Gao,“Incrementalgradientdescentimputationmethodformissingdatainlearningclassiersystems,”inProceedingsofGECCO05,2005,pp.7273.,HandlingMissingData,D.GuandY.Gao,“Incrementalgradientdescentimputationmethodformissingdatainlearningclassiersystems,”inProceedingsofGECCO05,2005,pp.7273.,usetherelationshipamongvariablestoestimatethemissingvalue.,ClusteringwithLCS,Becausethefinalpopulationrulesetsuggestssomeimportantpatternsinthedataset,LCSEcanpredicttheunforeseencasescorrectly,Nocleardefinitionofwhatshouldbeinacluster,FrameworkofLCSc,LiangdongShi,YangGao,LeiWu,LinShang:
ClusteringwithXCSonComplexStructureDataset.AustralasianConferenceonArtificialIntelligence2008:
489-499.LiangdongShi,YinghuanShi,YangGao:
ClusteringwithXCSandAgglomerativeRuleMerging.IDEAL2009:
242-250.LiangdongShi,YinhuanShi,YangGao,LinShang,YubinYang,XCSc:
ANovelApproachtoClusteringwithExtendedClassifierSystem,InternationalJournalofNeuralSystem,21
(1):
79-93,2011.,Semi-supervisedLearningTask,SupervisedLearningComponentDealwithlabeleddataSimilartoUCS(sUpervisedClassifierSystem)Semi-supervisedLearningComponentProvidelabelstounlabeleddataSelf-learningTri-trainingSimilarityMeasureUnlabeledbecomelabeled,ChiSu,YangGao,ChunCao:
Learningclassifiersystemusingbothlabeledandunlabeleddata.GECCO2010:
1065-1066,Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
HotIssuesandFutureDirectionWhatisapromisingfutureresearchdirection?
Complexity-AccuracyTradeoff,KnowledgerepresentationHighInterpretabilityRulesCreditassignmentHighaccuracyDifficulty:
Theyareconflicting!
Complexity-AccuracyTradeoff,Complexity,0,Error,Testdataaccuracy,Trainingdataaccuracy,Complexity-AccuracyTradeoff,Complexity,0,Error,GoodTradeoff,Testdataaccuracy,Trainingdataaccuracy,FitnessofaSystemS,FitnessofaSystemS,w1Accuracy(S)-w2Complexity(S),Thenumberofcorrectlyclassifiedtrainingpatterns,Thenumberoffuzzyrules,1stTerm:
AccuracyMaximization2ndTerm:
ComplexityMinimization,Fitness(S)=,Complexity,Error,Testdataaccuracy,S*,0,Trainingdataaccuracy,Minimizew1Error+w2ComplexityWhentheweightforthecomplexityminimizationislarge:
Toosimple,DifficultyinWeightedSumApproach,Complexity,Error,Testdataaccuracy,S*,0,Trainingdataaccuracy,Minimizew1Error+w2ComplexityWhentheweightfortheerrorminimizationislarge:
Toocomplicated,DifficultyinWeightedSumApproach,Complexity,Error,Testdataaccuracy,S*,0,Trainingdataaccuracy,Minimizew1Error+w2ComplexityWhenthetwoweightsareappropriatelyspecified:
Bestsystem,DifficultyinWeightedSumApproach,BasicIdeaAggregationApproachMulti-objectiveApproach,SearchforParetoOptimal,Multi-ObjectiveApproach,H.IshibuchiandY.Nojima(2007)“Analysisofinterpretability-accuracytradeoffoffuzzysystemsbymultiobjectivefuzzygenetics-basedmachinelearning,”InternationalJournalofApproximateReasoning.GoogleScholarCitations:
182,FutureDirections,应用大数据(快速变化)海量流数据,音视频数据处理不平衡数据半监督学习在线学习Newtrainingpatternscomeeveryminute规则集处理方法理论LCS的框架方法改进DefinitionoffitnessLearningmethodStatisticalmethod,Thankyou!
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- lcs 研究 概述 ppt 课件