简介:Aprimarychallengeofagent-basedpolicylearningincomplexanduncertainenvironmentsisescalatingcomputationalcomplexitywiththesizeofthetaskspace(actionchoicesandworldstates)andthenumberofagents.Nonetheless,thereisampleevidenceinthenaturalworldthathigh-functioningsocialmammalslearntosolvecomplexproblemswithease,bothindividuallyandcooperatively.Thisabilitytosolvecomputationallyintractableproblemsstemsfrombothbraincircuitsforhierarchicalrepresentation...
简介:Inthispaper,wepresentatechniqueforensuringthestabilityofalargeclassofadaptivelycontrolledsystems.WecombineIQCmodelsofboththecontrolledsystemandthecontrollerwithamethodoffilteringcontrolparameterupdatestoensurestablebehaviorofthecontrolledsystemunderadaptationofthecontroller.Wepresentaspecificapplicationtoasystemthatusesrecurrentneuralnetworksadaptedviareinforcementlearningtechniques.Theworkpresentedextendsearlierworksonstablereinforce...
简介:Inthispaper,weformulateandexplorethecharacteristicsofiterativelearninginballisticcontrolproblems.Theiterativelearningcontrol(ILC)theoryprovidesasuitableframeworkforderivationsandanalysisofballisticcontrolunderlearningprocess.Toovercometheobstaclescausedbyuncertaingradientandredundantcontrolinput,weincorporateextratrialsintoiterativelearning.Withthehelpoftrialresults,propercontrolandupdatingdirectioncanbedetermined.Then,iterativelearningcanbeappliedtoballisticcontrolproblem.Severalinitialstatelearningalgorithmsarestudiedforinitialspeedcontrol,forcecontrol,aswellascombinedspeedandanglecontrol.Intheend,shootinganglelearninginthebasketballshotprocessissimulatedtoverifytheeffectivenessofiterativelearningmethodsinballisticcontrolproblems.
简介:模式经由确定的学习理论当模特儿和识别在这篇论文被介绍的为心电图(ECG)的一个方法。而不是认识到ECG表明beat-to-beat,包含很多心跳的每个ECG信号被认出。方法完全基于时间的特征(即,动力学)ECG模式,它包含ECG模式的完全的信息。一个动态模特儿被雇用表明方法,它能够产生合成ECG信号。基于动态模型,方法在下列二个阶段被显示出:鉴定(训练)阶段和识别(测试)分阶段执行。在鉴定阶段,ECG模式的动力学精确地被建模并且通过确定的学习表示了为经常的RBF神经重量。在识别阶段,当模特儿的结果被用于ECG模式识别。建议方法的主要特征是ECG模式的动力学精确地被当模特儿并且被用于ECG模式识别。用Physikalisch-TechnischeBundesanstalt(PTB)数据库的试验性的研究被包括表明途径的有效性。
简介:Adaptivecritic(AC)basedcontrollersaretypicallydiscreteand/oryieldauniformlyultimatelyboundedstabilityresultbecauseofthepresenceofdisturbancesandunknownapproximationerrors.Acontinuous-timeACcontrollerisdevelopedthatyieldsasymptotictrackingofaclassofuncertainnonlinearsystemswithboundeddisturbances.TheproposedAC-basedcontrollerconsistsoftwoneuralnetworks(NNs)-anactionNN,alsocalledtheactor,whichapproximatestheplantdynamicsandgeneratesappropriatecon...
简介:Amultiresolutionstate-spacediscretizationmethodwithpseudorandomgriddingisdevelopedfortheepisodicunsupervisedlearningmethodofQ-learning.Itisusedasthelearningagentforclosed-loopcontrolofmorphingorhighlyreconfigurablesystems.Thispaperdevelopsamethodwherebyastate-spaceisadaptivelydiscretizedbyprogressivelyfinerpseudorandomgridsaroundtheregionsofinterestwithinthestateorlearningspaceinanefforttobreaktheCurseofDimensionality.Utilityofthemethodi...
简介:最近,为小摆动差错的快速的察觉的一条途径为连续时间的系统基于确定的学习理论被建议。在这份报纸,一个差错察觉计划经由确定的学习为非线性的分离时间的系统的一个班被建议。由使用确定的学习算法的分离时间的扩展,一般差错工作(即,内部动力学)内在的正常和非线性的分离时间的系统的差错模式被分离时间的动态光线的基础功能(RBF)局部地精确地接近网络。然后,有嵌入的系统动力学的获得的知识的评估者的一个银行被构造,并且一套剩余被获得并且过去常测量监视系统的动力学和训练系统的动力学之间的差别。一个差错察觉决定计划根据最小的剩余原则被介绍,即,一个差错的出现能被比较剩余的大小在一个分离时间的背景检测。差错detectability分析被执行,察觉时间的上面的界限被导出。一个模拟例子被给说明建议计划的有效性。
简介:Inthispaper,anoptimalhigherorderlearningadaptivecontrolapproachisdevelopedforaclassofSISOnonlinearsystems.Thisdesignismodel-freeanddependsdirectlyonpseudo-partial-derivativesderivedon-linefromtheinputandoutputinformationofthesystem.Anovelweightedone-step-aheadcontrolcriterionfunctionisproposedforthecontrollaw.Theconvergenceanalysisshowsthattheproposedcontrollawcanguaranteetheconvergenceundertheassumptionthatthedesiredoutputisasetpoint.Simulationexamplesareprovidedfornonlinearsystemstoillustratethebetterperformanceofthehigherorderlearningadaptivecontrol.
简介:Thispaperaddressesthedesignproblemofrobustiterativelearningcontrollersforaclassoflineardiscrete-timesystemswithnorm-boundedparameteruncertainties.Aniterativelearningalgorithmwithcurrentcyclefeedbackisproposedtoachievebothrobustconvergenceandrobuststability.Thesynthesisproblemoftheproposediterativelearningcontrol(ILC)systemisreformulatedasaγ-suboptimalH-infinitycontrolproblemviathelinearfractionaltransformation(LFT).AsufficientconditionfortheconvergenceoftheILCalgorithmispresentedintermsoflinearmatrixinequalities(LMIs).Furthermore,thelineartransferoperatorsoftheILCalgorithmwithhighconvergencespeedareobtainedbyusingexistingconvexoptimizationtechniques.Thesimulationresultsdemonstratetheeffectivenessoftheproposedmethod.
简介:这份报纸论述OCPA(发生作用之人或物调节概率的自动机)仿生的自治学习系统基于为解决一个二车轮的灵活机器人的平衡控制问题的Skinners发生作用之人或物调节理论。学习系统的OCPA由二个阶段组成:在第一个阶段,一个发生作用之人或物行动从一套发生作用之人或物行动随机地被选择然后用作控制系统的输入;在第二个阶段,学习系统收集系统的取向信息并且为优化使用它直到完成控制目标。同时,发生作用之人或物行动集合的尺寸能自动地为避免小概率事件在学习过程期间被减少。理论分析为在纸学习系统的设计OCPA被做,它理论上证明在学习系统的OCPA学习机制的发生作用之人或物调节的集中,也就是,发生作用之人或物行动熵将与学习进程收敛到最小。然后OCPA学习系统被用于姿势二车轮的灵活自我平衡的机器人的平衡控制。机器人不把posutre平衡处于起始的状态和在发生作用之人或物集合的每发生作用之人或物的选择概率的技巧是相等的。与学习继续,最佳的发生作用之人或物的选择可能性逐渐地逐渐地趋于到一个和发生作用之人或物行动熵趋于到最小,并且机器人逐渐地因此学习了姿势平衡技巧。
简介:Intheprocedureofthesteady-statehierarchicaloptimizationwithfeedbackforlarge-scaleindustrialprocesses,asequenceofset-pointchangeswithdifferentmagnitudesiscarriedoutontheoptimizationlayer.Toimprovethedynamicperformanceoftransientresponsedrivenbytheset-pointchanges,afilter-basediterativelearningcontrolstrategyisproposed.Intheproposedupdatinglaw,alocal-symmetric-integraloperatorisadoptedforeliminatingthemeasurementnoiseofoutputinformation,asetofdesiredtrajectoriesarespecifiedaccordingtotheset-pointchangessequence,thecurrentcontrolinputisiterativelyachievedbyutilizingsmoothedoutputerrortomodifyitscontrolinputatpreviousiteration,towhichtheamplifiedcoefficientsrelatedtothedifferentmagnitudesofset-pointchangesareintroduced.Theconvergenceofthealgorithmisconductedbyincorporatingfrequency-domaintechniqueintotime-domainanalysis.Numnericalsimulationdemonstratestheeffectivenessoftheproposedstrategy.
简介:Thispaperdealswiththeiterativelearningcontrol(ILC)designformultiple-inputmultiple-output(MIMO),time-delaysystems(TDS).TwofeedbackILCschemesareconsideredusingtheso-calledtwo-dimensional(2D)analysisapproach.Itshowsthatcontinuous-discrete2DRoessersystemscanbedevelopedtodescribetheentirelearningdynamicsofbothILCschemes,basedonwhichnecessaryandsufficientconditionsfortheirstabilitycanbeprovided.Anumericalexampleisincludedtovalidatethetheoreticalanalysis.
简介:Inthispaper,iterativelearningcontrol(ILC)designisstudiedforaniteration-varyingtrackingprobleminwhichreferencetrajectoriesaregeneratedbyhigh-orderinternalmodels(HOIM).AnHOIMformulatedasapolynomialoperatorbetweenconsecutiveiterationsdescribesthechangesofdesiredtrajectoriesintheiterationdomainandmakestheiterativelearningproblembecomeiterationvarying.TheclassicalILCfortrackingiteration-invariantreferencetrajectories,ontheotherhand,isaspecialcaseofHOIMwherethepolynomialrenderstoaunitycoefficientoraspecialfirst-orderinternalmodel.ByinsertingtheHOIMintoP-typeILC,thetrackingperformancealongtheiterationaxisisinvestigatedforaclassofcontinuous-timenonlinearsystems.Time-weightednormmethodisutilizedtoguaranteevalidityofproposedalgorithminasenseofdata-drivencontrol.
简介:Arobustadaptiverepetitivelearningcontrolmethodisproposedforaclassoftime-varyingnonlinearsystems.Nussbaum-gainmethodisincorporatedintothecontroldesigntocounteractthelackofaprioriknowledgeofthecontroldirectionwhichdeterminesthemotiondirectionofthesystemunderanyinput.Itisshownthatthesystemstatecouldconvergetothedesiredtrajectoryasymptoticallyalongtheiterationaxisthroughrepetitivelearning.Simulationiscarriedouttoshowthevalidityoftheproposedcontrolmethod.
简介:Somenewcriteriaforthechaoticlagsynchronizationareproposed.Atfirst,lagsynchronizationschemeforidenticalmaster-slaveLur'esystemsbyreplacingvariablescontrolandtherelevanterrorsystemaregiven,andtherelationsbetweenabsolutestabilityoftheerrorsystemandthechaoticlagsynchronizationaredescribed.Then,basedonaquadraticLyapunovfunction,twonewLur'ecriteriafortheabovechaoticlagsynchronizationareproved.FourcorrespondingfrequencydomaincriteriaarefurtherderivedbymeansofMeyer-Kalman-YacuboviaLemma.Thesefrequencydomaincriteriaareappliedtoanalyzethelagsynchronizationofgeneralmaster-slaveChua'scircuitssothatsomerangesoftheparametersinwhichthemaster-slaveChua'scircuitsachievechaoticlagsynchronizationbyreplacingsingle-variablecontrolareattained.Finally,someexamplesaregiventoverifythetheoreticalresults.
简介:NecessaryandsuffcientconditionsfortheexistenceofaLyapunovfunctionintheLur'eformtoguaranteetheabsolutestabilityofLur'econtrolsystemswithmultiplenon-linearitiesarediscussedinthispaper.Itsimplifiestheexistenceproblemtooneofsolvingasetoflinearmatrixinequalities(LMIs).IfthoseLMIsarefeasible,freeparametersintheLyapunovfunction,suchasthepositivedefinitematrixandthecoefficientsoftheintegralterms,aregivenbythesolutionoftheLMIs.Otherwise,thisLyapunovfunctiondoesnotexist.Somesufficientconditionsarealsoobtainedfortherobustabsolutestabilityofuncertainsystems.Anumericalexampleisprovidedtodemonstratetheeffectivenessoftheproposedmethod.