智能控制（英文）

IntelligentControlLecturer：王学泰GroupMembers：王学泰李媛张玉磊2011-04-021太原理工大学谢刚Dr.GangXie,TUT1INTRODUCTIONWithincreasingdemandsforhighprecisionautonomouscontroloverwideoperatingenvelopes,conventionalcontrol（传统控制）

engineeringapproachesareunabletoadequatelydealwithsystemcomplexity（复杂性）,nonlinearities（非线性）,spatial（多维）andtemporalparametervariations（多参数）,andwithuncertainty（不确定）.

IntelligentControlorself-organising/learningcontrolisanewemergingdisciplinethatisdesignedtodealwithproblems.Ratherthanbeingmodelbased,itisexperientialbased（基于经验而不是模型）.

IntelligentControlistheamalgamofthedisciplinesofArtificialIntelligence人工智能,SystemsTheory系统论andOperationsResearch运筹学Intelligentcontroldescribesthedisciplinewherecontrolmethodsaredevelopedthatattempt

toemulate（模拟

）importantcharacteristicsofhumanintelligence.Thesecharacteristicsincludeadaptationandlearning,planningunderlargeuncertaintyandcopingwithlargeamountsof

data.CharacteristicsofIntelligentSystemsAdaptation自适应andLearning再学习:

Theabilitytoadapttochangingconditionsisecessaryinan

intelligentsystem.Althoughadaptationdoesnotnecessarilyrequiretheabilitytolearn,for

systemstobeabletoadapttoawidevarietyofunexpectedchangeslearningisessential.So

theabilitytolearnisanimportantcharacteristicof(highly)intelligentsystems.Autonomy自治性

andIntelligence智能性:

Autonomyinsettingandachievinggoalsisanimportan

characteristicofintelligentcontrolsystems.Whenasystemhastheabilitytoac

appropriatelyinanuncertainenvironmentforextendedperiodsoftimewithoutexternal

interventionitisconsideredtobehighlyautonomous.StructuresandHierarchies分级递阶:

Inordertocopewithcomplexity,anintelligentsystemmust

haveanappropriatefunctionalarchitectureorstructureforefficientanalysisandevaluation

ofcontrolstrategies.Fuzzy(logic)Control模糊（逻辑）控制ANN-BasedControl(ArtificialNeuralNetwork）基于人工神经网络的控制ExpertControl专家控制GeneticAlgorithm遗传算法Intelligentcontrolcanbedividedintothefollowingmajorsub-domains:summary：Thecollectionofmethodologiescomprisingsoftcomputingincludee.g.fuzzylogic,neuralnetworks(neurocomputing)andgeneticalgorithms.Theseareoftencomplementaryratherthancompetingmethodologiesandcanoftenbecombinedinordertocreateintelligentsystems.Thestrengthsofthedifferentmethodologiesaresummarizedinthetablebelow.2FuzzyControlFuzzytheorybeganwithapaperon“fuzzysets（模糊集合）”,writtenbyProf.L.A.Zadehin1965.Fuzzysetsarethosesetswhoseboundaryisnotclear.Fuzzylogicsarecalculationproceduresonfuzzysets.Atechnologyinwhichthewholesystemcanberoughlydefined,thatis“fuzzytheory”wasproposed.Afuzzycontrolsystemisacontrolsystembasedonfuzzylogic（模糊逻辑）—amathematicalsystemthatanalyzesanaloginputvaluesintermsoflogicalvariablesthattakeoncontinuousvaluesbetween0and1,incontrasttoclassicalordigitallogic,whichoperatesondiscretevaluesofeither0or1(trueorfalse).2.1FuzzyLogicNormal“Crisp”logic

whereeverythingmustbeeither

TrueorFalse。Fuzzylogicisaformofmany-valuedlogicderivedfromfuzzysettheory（模糊集合理论）todealwithreasoningthatisfluidorapproximateFuzzylogicacknowledgesandexploitsthetoleranceforuncertaintyandimprecision.Thesentenceontheotherside

ofthelineisfalseThesentenceontheotherside

ofthelineisfalselinguisticvariables（语言变量）takeonlinguisticvalues（语言值）whicharewords(linguisticterms)withassociateddegreesof

membership（隶属度）intheset。

2.2FuzzyControlDesign©INFORM1990-1998 Slide17Fuzzycontrolisamethodologytorepresentandimplementa(smart)human’sknowledgeabout

howtocontrolasystem.AfuzzycontrollerisshowninFigure1.©INFORM1990-1998 Slide18Thefuzzycontrollerhasseveral

components:•Therule-baseisasetofrulesabouthowtocontrol.•Fuzzification（模糊化）istheprocessoftransformingthenumericinputsintoaformthatcanbeused

bytheinferencemechanism.•Theinferencemechanism（推理机）usesinformationaboutthecurrentinputs(formedbyfuzzification),decideswhichrulesapplyinthecurrentsituation,andformsconclusionsaboutwhattheplant

inputshouldbe.•Defuzzification（去模糊化）convertstheconclusionsreachedbytheinferencemechanismintoanumeric

inputfortheplant.MaptoFuzzySetsFuzzyRules

IFAANDBTHENL

*

*DefuzzificationInputsOutputget_inputs();fire_rules();find_output();TermDefinitions:Distance :={far,medium,close,zero,neg_close}Angle :={pos_big,pos_small,zero,neg_small,neg_big}Power :={pos_high,pos_medium,zero,neg_medium,neg_high}2.2.1Fuzzification:

-LinguisticVariables-©INFORM1990-1998 Slide21MembershipFunctionDefinition:TheLinguisticVariablesArethe“Vocabulary”ofaFuzzyLogicSystem!Computationofthe“IF-THEN”-Rules:#1:IFDistance=mediumANDAngle=pos_smallTHENPower=pos_medium#2:IFDistance=mediumANDAngle=zeroTHENPower=zero#3:IFDistance=farANDAngle=zeroTHENPower=pos_medium2.2.2Fuzzy-Inference:

-“IF-THEN”-Rules-©INFORM1990-1998 Slide22Aggregation: Computingthe“IF”-PartComposition: Computingthe“THEN”-PartTheRulesoftheFuzzyLogicSystemsArethe“Laws”ItExecutes!

2.2.3Defuzzification©INFORM1990-1998 Slide23

Defuzzificationistheprocessofproducingaquantifiableresultinfuzzylogic,givenfuzzysetsandcorrespondingmembershipdegrees。©INFORM1990-1998 Slide24summary：Fuzzycontrolistypicallyusedwhentheexplicitsystemanalyticalmodelis

notavailable.

Fuzzycontrolisintuitivetounderstandandeasytodesignfor

engineerswhoareunfamiliarwithclassicalcontroltheory.Afuzzycontroller

canbedesignedbasedone.g.ahumanoperatorsexperience.Fuzzycontrol

consistsofselectingandusing1.acollectionofrulesthatdescribethecontrolstrategy描述控制策略的规则集合2.membershipfunctionsforthelinguisticvariablesintherules在上述规则下，表示语言变量的隶属度函数3.logicalconnectionsforfuzzyrelations模糊关系之间的逻辑关系4.adefuzzicationmethod

去模糊化的方法

3Artificialneuralnetwork©INFORM1990-1998 Slide253.1BiologicalapproachtoAIArtificialneuralnetworksarecircuits,computeralgorithms（计算方法）,ormathematicalrepresentationsloosely

inspired（激励）bythemassivelyconnectedsetofneuronsthatformbiologicalneuralnetworks.

Artificial

neuralnetworksareanalternativecomputingtechnologythathaveprovenusefulinavarietyof

patternrecognition（模式识别）,signalprocessing（信号处理）,estimation,andcontrolproblems.©INFORM1990-1998 Slide26©INFORM1990-1998 Slide27ArtificialneuronsNeuronsworkbyprocessinginformation.Theyreceiveandprovideinformationinformofspikes.©INFORM1990-1998 Slide28ArtificialneuralnetworksAnartificialneuralnetworkiscomposedofmanyartificialneuronsthatarelinkedtogetheraccordingtoaspecificnetworkarchitecture.Theobjectiveoftheneuralnetworkistotransformtheinputsintomeaningfuloutputs.©INFORM1990-1998 Slide293.2LearningProcessesLearning=learningbyadaptationTheyounganimallearnsthatthegreenfruitsaresour,whiletheyellowish/reddishonesaresweet.Thelearninghappensbyadaptingthefruitpickingbehavior.©INFORM1990-1998 Slide301.Learningwithateacher有教师学习

(Supervisedlearning)2.Learningwithoutateacher无教师学习（UnsupervisedLearning）3.ReinforcemengtLearning再励学习©INFORM1990-1998 Slide31a.Learningwithaperceptron（感知机）Perceptron:Data:Error:Learning:©INFORM1990-1998 Slide32Perceptron:Data:Error:Learning:©INFORM1990-1998 Slide33b.LearningwithRBF（径向基函数）neuralnetworksOnlythesynapticweightsoftheoutputneuronaremodified.AnRBFneuralnetworklearnsanonlinearfunction.©INFORM1990-1998 Slide34c.LearningwithMLP（多层感知机）neuralnetworks©INFORM1990-1998 Slide35Data:Error:xyout12…p-1p©INFORM1990-1998 Slide36d.Learningwithbackpropagation（BP反向传播神经网络）Solutionofthecomplicatedlearning:calculatefirstthechangesforthesynapticweightsoftheoutputneuron;calculatethechangesbackwardstartingfromlayerp-1,andpropagatebackwardthelocalerrorterms.©INFORM1990-1998 Slide37e.LearningwithFeed-Forward(前馈)NNx1=x2=x3=Inputsandoutputsarenumeric.©INFORM1990-1998 Slide38

Learningtasksofartificialneuralnetworkscanbereformulatedasfunctionapproximation(函数逼近)tasks.

Neuralnetworkscanbeconsideredasnonlinearfunctionapproximating（非线性逼近）tools(i.e.,linearcombinationsofnonlinearbasisfunctions),wheretheparametersofthenetworksshouldbefoundbyapplyingoptimisationmethods（最优化方法）.Theoptimisation（最优化）isdonewithrespecttotheapproximationerrormeasure.

Ingeneralitisenoughtohaveasinglehiddenlayer（单隐含层）neuralnetwork(MLP,RBForother)tolearntheapproximationofanonlinearfunction.Insuchcasesgeneraloptimisationcanbeappliedtofindthechangerulesforthesynapticweights.3.3

Training（训练）NeuralNetworks©INFORM1990-1998 Slide39Howdoweconstructaneuralnetwork?Wetrainitwithexamples.Regardlessofthetypeof

network,wewillrefertoitaswhereθisthevectorofparametersthatwetunetoshapethenonlinearityitimplements(Fcould

beafuzzysystemtoointhediscussionbelow).Foraneuralnetworkθwouldbeavectorofthe

weightsandbiases.SometimeswewillcallFan“approximatorstructure.”Supposethatwegather

input-outputtrainingdatafromafunctiony=g(x)thatwedonothaveananalyticalexpression

for(e.g.,itcouldbeaphysicalprocess).©INFORM1990-1998 Slide40TrainingoftheANN(ArtificialNeuralNet)iseffectedby:Startingwithartibrarywieghts（任意权值）Presentingthedata,instancebyinstance

adaptingtheweightsaccordingtheerrorforeachinstance.Repeatinguntilconvergence（收敛）.©INFORM1990-1998 Slide41summaryANNisamassivelyparalleldistributedprocessor（分布式并行处理）.1.Simpleprocessingunitsthatcanstoreexperience2.Alearningprocess4ExpertControlSystemsAnexpertcontrolsystemisacomputerprogramthatisdesignedtoholdtheaccumulatedknowledge（累计知识）

ofoneormoredomainexperts4.1ComponentsofanExpertSystemTheknowledgebaseisthecollectionoffactsandruleswhichdescribealltheknowledgeabouttheproblemdomainTheinferenceengineisthepartofthesystemthatchooseswhichfactsandrulestoapplywhentryingtosolvetheuser’squeryTheuserinterfaceisthepartofthesystemwhichtakesintheuser’squeryinareadableformandpassesittotheinferenceengine.Itthendisplaystheresultstotheuser.InterpreterInferenceEngineKnowledgeBasedRulesDatabaseContextSetoffactsNaturalLanguageInterfaceExpertUserExpertSystemStructure（1）KnowledgeBase（知识库）Representsallthedataandinformationimputedbyexpertsinthefield.Storesthedataasasetofrulesthatthesystemmustfollowtomakedecisions.KnowledgeAcquisitionExpertSystemKnowledgeEngineerHumanExpert

（2）InferenceEngine（推理机）Askstheuserquestionsaboutwhattheyarelookingfor.Appliestheknowledgeandtherulesheldintheknowledgebase.Appropriatelyusesthisinformationtoarriveatadecision.（3）UserInterface（人机界面）Allowstheexpertsystemandtheusertocommunicate.Findsoutwhatitisthatthesystemneedstoanswer.Sendstheuserquestionsoranswersandreceivestheirresponse.4.2CharacteristicswithExpertSystemsThereisnoexpertinthefieldTheexpertisunabletocommunicatehis/herideasTheexpertisunwillingtocommunicatehis/herideasTheexpertisnotavailableMusthaveallinformationonasubjectCanallthetestingbeaccomplished?UseracceptanceAdvantagesofExpertSystemsCanbesimpletouseEfficientresultsAccurateresultsAdaptationandadjustmentstochangingconditionsCosteffectiveProblemswithExpertSystemsLimiteddomainSystemsarenotalwaysuptodate,anddon’tlearnNo“commonsense”Expertsneededtosetupandmaintainsystem5GeneticAlgorithmsAnalgorithmisasetofinstructionsthatisrepeatedtosolveaproblem.Ageneticalgorithmconceptuallyfollowsstepsinspiredbythebiologicalprocessesofevolution.GeneticAlgorithmsfollowtheideaofSURVIVALOFTHEFITTEST-Better（适者生存）andbettersolutionsevolvefrompreviousgenerationsuntilanearoptimalsolutionisobtained.Alsoknownasevolutionaryalgorithms,geneticalgorithmsdemonstrateselforganizationandadaptationsimilartothewaythatthefittestbiologicalorganismsurviveandreproduce.Ageneticalgorithmisaniterativeprocedure（迭次求近）thatrepresentsitscandidatesolutionsasstringsofgenescalledchromosomes（染色体）.5.1Individual

Representing(个体表现)

Anindividualisdatastructurerepresenting

the“geneticstructure（基因组合）”ofapossiblesolution.Geneticstructureconsistsofanalphabet(usually0,1)BinaryEncoding二进制编码MostCommon–stringofbits,0or1. Chrom:A=1011001011Chrom:B=1111110000GivesyoumanypossibilitiesExampleProblem:KnapsackproblemTheproblem:therearethingswithgivenvalueandsize.Theknapsackhasgivencapacity.Selectthingstomaximizethevalues.Encoding:Eachbitsays,ifthecorrespondingthingisintheknapsackPermutationEncoding排列编码Usedin“orderingproblems”Everychromosomeisastringofnumbers,whichrepresentsnumberisasequence.ChromA:153264798ChromB:857723149Example:TravellingsalesmanproblemTheproblem:citiesthatmustbevisited.Encodingsaysorderofcitiesinwhichsalesmanwilllvisit.ValueEncoding值编码Usedforcomplicatedvalues(realnumbers)andwhenbinarycodingwouldbedifficultEachchromosomeisastringofsomevalues.ChromA:1.23235.32430.4556ChromB:abcdjeifjdhdierjfdChromC:(back),(back),(right),(forward),(left)Example:Findingweightsforneuralnets.Theproblem:findweightsfornetworkEncoding:RealvaluesthatrepresentweightsRulebasesystem规则库Givenarule(ifcolor=redandsize=smallandshape=roundthenobject=apple.Assumethateachfeaturehasfinitesetofvalues(e.g.,size=small,large)Representthevalueasasubstringoflengthequltothenumberofpossiblevalues.Forexample,small=10,large=01.Theentirerulewouldbe10010010100–setofrulesconcatenatingthevaluestogether.5.2SelectionCriteria（选择标准）Fitnessproportionateselection,rankselectionmethods.

Fitnessproportionate（适应度比例法）–eachindividual,I,hastheprobabilityfitness(I)/sum_over_all_individual_jFitness(j),whereFitness(I)isthefitnessfunctionvalueforindividualI.

Rankselection（

随机抽样法）–sortsindividualbyfitnessandtheprobabilitythatanindividualwillbeselectedisproportionaltoitsrankinthissortedlist.FitnessFunction（适应度函数）Representsarankofthe“representation”Itisusuallyarealnumber.Thefunctionusuallyhasavaluebetween0and1.Onecanhaveasubjectivejudgment(e.g.1-5forrecipe2-1-4.)Similarlythelengthoftherouteinthetravelingsalespersonproblemisagoodmeasure,becausetheshortertheroute,thebetterthesolution.

5.3Reproduction（复制）Reproduction-Throughreproduction,geneticalgorithmsproducenewgenerationsofimprovedsolutionsbyselectingparentswithhigher

fitnessratings（适应度）Crossover（交叉）-Manygeneticalgorithmsusestringsofbinarysymbolsforchromosomes,asinourKnapsackexample,torepresentsolutions.Crossovermeanschoosingarandompositioninthestring(say,after2digits)andexchangingthesegmentseithertotherightortotheleftofthispointwithanotherstringpartitionedsimilarlytoproducetwonewoffspring.CrossoverExampleParentA011011ParentB101100“bysplittingeachnumberasshownbetweenthesecondandthirddigits(positionisrandomlyselected)01*1011 10*1100Mutation（变异）-Mutationisanarbitrarychangeinasituation.Sometimesitisusedtopreventthealgorithmfromgettingstuck.Theprocedurechangesa1toa0,or0toa1.Thischangeoccurswithaverylowprobability(say1in1000)10101111100011Parent1Parent210100111100110Child1Child2MutationMutationandCrossoverCrossoverOperators

Single­pointcrossover:ParentA:10010|11101ParentB:01011|10110ChildAB:1001010110ChildBA:0101111101

Two­pointcrossover:ParentA:1001|011|101

ParentB:0101|110|110ChildAB:1001110101

ChildBA:0101011110UniformCrossoverandMutationUniformcrossover:ParentA:1001011101ParentB:0101110110ChildAB:1101111101

ChildBA:0001010110Mutation:randomlytoggleonebitIndividualA:1001011101IndividualA':1000011101Crossover–PermutationEncodingSinglepointcrossover-onecrossoverpointisselected,tillthispointthepermutationiscopiedfromthefirstparent,thenthesecondparentisscannedandifthenumberisnotyetintheoffspringitisadded

(123456789)+(453689721)=(123456897)Mutation

Orderchanging-twonumbersareselectedandexchanged(123456897)=>(18345629

7)

Crossover–ValueEncodingCrossover

AllcrossoversfrombinaryencodingcanbeusedMutation

Addingasmallnumber(forrealvalueencoding)-toselectedvaluesisadded(or