人工智能原理-北京大学-11--PartVLearningChapter11Paradigmsi-(

SchoolofElectronicandComputerEngineeringPekingUniversityWangWenminArtificialIntelligenceArtificialIntelligence2ArtificialIntelligencePart1.BasicsPart2.SearchingPart3.ReasoningPart4.PlanningPart5.LearningContents:ArtificialIntelligence3Part5.Learning9.PerspectivesaboutMachineLearning10.TasksinMachineLearning11.ParadigmsinMachineLearning12.ModelsinMachineLearningContents:SchoolofElectronicandComputerEngineeringPekingUniversityWangWenminSupervisedLearningParadigmArtificialIntelligence5Objectives教学目的11.ParadigmsinMachineLearningInthischapterwewilldiscussindetailabouttheparadigmsthathavebeenproposedinmachinelearning.这一章我们详细讨论针对机器学习所提出的一些范式。ArtificialIntelligence::Learning::Paradigms6Thelearningparadigmsareusedtodenotethetypicalscenariosthatarehappenedinmachinelearning.学习范式用于表示机器学习中发生的典型场景。WhatareLearningParadigms什么是学习的范式11.ParadigmsinMachineLearningWhyStudyLearningParadigms为什么要研究学习的范式Designinganalgorithmtosolvealearningtaskmaytakeadifferentparadigm,suchasbasedonitsexperienceortheinteractionwithitsenvironment.设计一种解决学习任务的算法可能会采用不同的范式，例如基于其经验、或者与其环境的交互。Whystudythelearningparadigmsisbecauseitcanforceyoutothinkaboutamostappropriateparadigmforthelearningtaskinordertogetthebestresult.研读学习范式的意义在于，它可以使你考虑一个最适合该学习任务的范式，以获得最好的结果。ArtificialIntelligence::Learning::Paradigms7HowDoesMachineLearningWork机器学习是如何工作的11.ParadigmsinMachineLearningUnseenData未知数据Output输出f(X)LearningAlgorithm学习算法h(X)Hypothesis假设TrainingData训练数据ArtificialIntelligence::Learning::Paradigms8HowDoesMachineLearningWork机器学习是如何工作的11.ParadigmsinMachineLearningUnseenData未知数据Output输出f(X)LearningAlgorithm学习算法h(X)Hypothesis假设ArtificialIntelligence::Learning::Paradigms9HowDoesMachineLearningWork机器学习是如何工作的11.ParadigmsinMachineLearningUnseenData未知数据Output输出f(X)LearningAlgorithm学习算法h(X)Hypothesis假设Feedback反馈ArtificialIntelligence::Learning::Paradigms10TypicalParadigmsinMachineLearning机器学习中的典型学习范式11.ParadigmsinMachineLearningParadigms范式BriefStatements简短描述TypicalAlgorithm典型算法Supervised有监督Thealgorithmistrainedbyasetoflabeleddata,andmakespredictionsforallunseenpoints.算法采用一组标注数据进行训练，再对所有的未知点做出预测。Supportvectormachines支撑向量机Unsupervised无监督Thealgorithmexclusivelyreceivesunlabeleddata,andmakespredictionsforallunseenpoints.算法仅接收未标注的数据，再对所有的未知点做出预测。k-meansk-均值Reinforcement强化Thealgorithminteractswithenvironment,andreceivesanrewardforeachaction.算法与外部环境交互，每个动作得到一个回报。Q-learningArtificialIntelligence1111.ParadigmsinMachineLearning11.1.SupervisedLearningParadigm11.2.UnsupervisedLearningParadigm11.3.ReinforcementLearningParadigm11.4.RelationsandOtherParadigmsContents:ArtificialIntelligence::Learning::Paradigms12Theagentreceivesasetoflabeledexamplesastrainingdata,andmakespredictionsforallunseenpoints.智能体接收一组标注的样本作为训练数据，然后对所有的未知点进行推测。Thisapproachattemptstogeneralizeafunctionormappingfrominputstooutputsbytraining,whichcanthenbeusedspeculativelytogenerateanoutputforpreviouslyunknowndata.这种方式试图生成从输入到输出的函数或映射，然后可以将其用于对预先未知的数据生成输出。WhatisSupervisedLearning什么是有监督学习11.1.SupervisedLearningParadigmItisawayof“teaching”thelearningalgorithm,likethata“teacher”givestheclasses(courses).这是一种“教”学习算法的方式，就像“老师”讲授课程那样。ArtificialIntelligence::Learning::Paradigms13Thetrainingdatainsupervisedlearning:有监督学习中的训练数据：eachtrainingdatahasaknownlabelasaninputdata,每个训练数据具有一个已知标注作为输入数据，thelabelisapairconsistingofaninputobjectandadesiredoutputvalue标注是由输入对象和预期输出值组成的对(suchasspam/not-spam,orastockpriceatatime)（例如垃圾与非垃圾邮件、或某时刻股票价格）。Anhypothesisfunctionaftertraining:训练后的假设函数：canbeusedformappingnewunseendata.可用于映射新的未知数据。WhatisSupervisedLearning什么是有监督学习11.1.SupervisedLearningParadigmArtificialIntelligence1411.1.SupervisedLearningParadigm11.1.1.OverviewofSupervisedLearning11.1.2.SuitableLearningTasks11.1.3.FormalDescription11.1.4.AlgorithmsofSupervisedLearning11.1.5.ApplicationsofSupervisedLearning11.1.6.VariantsofSupervisedLearningContents:ArtificialIntelligence::Learning::Paradigms15SixStepsbySupervisedLearning有监督学习的6个步骤11.1.1.OverviewofSupervisedLearningx1y1x2y21)Determinethetrainingtype确定训练类型2)Gatheratrainingset收集训练集3)Determinethefeatureextractionapproach确定特征提取方法4)Determinethealgorithmtothetask设计该任务的算法5)Trainingthealgorithm训练该算法6)Evaluatetheaccuracy评估其精确性LabeledData标注数据(x,y)f(X)h(X)XYHypothesis假设f(x)=yTraining训练LearningAlgorithm学习算法withsmallgeneralizationandempiricalerrors具有小的泛化和经验错误ArtificialIntelligence::Learning::Paradigms161)Determinethetrainingtype确定训练类型Youshoulddecidefirstlywhatkindofdataistobeusedasatrainingset.应该首先确定使用何种数据作为训练集。E.g.,forhandwritingrecognition,thatmaybeasinglehandwrittencharacter,anentirehandwrittenword,oranentirelineofhandwriting.例如，对于手写体识别，可以是一个手写字符、一个完整的手写单词、或是手写的一行。2)Gatheratrainingset收集训练集Thetrainingsetneedstoberepresentativeofthereal-worlduseofthefunction.训练集需要代表实际使用的功能。Thus,asetofinputobjectsisgatheredandcorrespondingoutputsarealsogathered,eitherfromhumanexpertsorfrommeasurements.因此，由人类专家或者通过测量，筛选出一组输入对象以及对应的输出。SixStepsbySupervisedLearning有监督学习的6个步骤11.1.1.OverviewofSupervisedLearningArtificialIntelligence::Learning::Paradigms173)Determinethefeatureextractionapproac