lcs研究概述ppt课件PPT文档格式.pptx
《lcs研究概述ppt课件PPT文档格式.pptx》由会员分享,可在线阅读,更多相关《lcs研究概述ppt课件PPT文档格式.pptx(47页珍藏版)》请在冰豆网上搜索。
,Outline,IntroductionDefinition,HistoryBasicIdeaofLCSTypes,ApproachesOurCurrentProgressWhatwehavedone?
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?
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?
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:
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,Handling