1、原文PERFORMANCEMANAGEMENTOFINTEGRATEDMANUFACTURINGSYSTEMNETWORKSSUKLEE,ASOKRAY,KEUMSHIKHONG,JOONGSUNYOON,ANDMYUNGCHUHAN1STRUCTUREOFTHEPERFORMANCEMANAGEMENTPROCEDURENETWORKPERFORMANCE,MORESPECIFICALLYDELAYSEXPERIENCEDBYMESSAGEPACKETS,COULDBECRITICALFORDYNAMICPERFORMANCEANDSTABILITYOFREALTIMEMANUFACTURI
2、NGPROCESSESTHISISESPECIALLYTRUEWHENMULTIPLEMACHINESAREPERFORMINGATASKWITHOUTDIRECTCONNECTIONSAMONGTHEMONEOFTHEEXAMPLESISANINTELLIGENTWELDINGSYSTEMNAYAK,RAYANDVAVRECK,1987WHEREAPOSITIONINGTABLEANDAROBOTHAVETOCOMMUNICATETHROUGHANETWORKTOEXCHANGETHEDESIREDTABLEPOSITIONCOORDINATESANDVARIOUSMESSAGESFORST
3、ATUSREPORTANOTHEREXAMPLEISMANIPULATIONOFABULKYANDFLEXIBLEWORKPIECEBYMORETHANONEINDEPENDENTROBOTSWHICHINITIATETHEIROWNPRESCRIBEDTRAJECTORIESUPONRECEIVINGASIGNALFROMACONTROLLERANDREPORTTHECOMPLETIONOFTHETRAJECTORIESBACKTOTHECONTROLLERVIAANETWORKTHETIMELINESSOFTHETRANSMITTEDDATAISESSENTIALBECAUSEADELAY
4、COULDDAMAGETHEWORKPIECEORTHEROBOTSWRISTSANDARMSINORDERTOMAINTAINANACCEPTABLELEVELOFTHEDYNAMICPERFORMANCEANDSTABILITYOFVARIOUSMANUFACTURINGPROCESSES,PERFORMANCEMANAGEMENTISREQUIREDTOMANIPULATEADJUSTABLEPROTOCOLPARAMETERSINREALTIMESOTHATTHENETWORKCANADAPTITSELFTOTHEDYNAMICENVIRONMENTTHISCANBEACCOMPLIS
5、HEDINTWOSTEPSIPERFORMANCEEVALUATIONTOFINDHOWPERTURBATIONSINPROTOCOLPARAMETERSAFFECTASELECTEDPERFORMANCEMEASURE,IE,TODETERMINETHERELATIONSHIPBETWEENTHEPERFORMANCEMEASUREANDTHEPROTOCOLPARAMETERS2DECISIONMAKINGTODECIDEONHOWTOADJUSTPROTOCOLPARAMETERS,IE,TOIDENTIFYTHEDIRECTIONANDMAGNITUDEOFTHEPARAMETERAD
6、JUSTMENTVECTOR,UTILIZINGPIECESOFINFORMATIONPROVIDEDINTHEFIRSTSTEPANDTHEHISTORYOFPERFORMANCETHEANALYTICALTECHNIQUESFORPERFORMANCEEVALUATIONSUCHASQUEUEINGTHEORYVISWANADHAMANDNARAHARI,1992OFTENREQUIRESUNREALISTICASSUMPTIONSLIKEPOISSONARRIVAL,ANDTENDTOBEMATHEMATICALLYUNTRACTABLEASTHESTRUCTUREOFTHEPERFOR
7、MANCEMEASUREBECOMESCOMPLEXFURTHERMORE,NETWORKTRAFFICSTATISTICSSUCHASDISTRIBUTIONSOFMESSAGEGENERATIONINTERVALANDMESSAGELENGTH,WHICHAREREQUIREDASINPUTSTOTHEANALYTICALMODEL,AREVERYDIFFICULTTOESTIMATEINREALTIMEONTHEOTHERHAND,DISCRETEEVENTSIMULATIONLAWANDKELTON,1991ISAVIABLEALTERNATIVETOANALYTICALTECHNIQ
8、UESAMAJORADVANTAGEOFSIMULATIONOVERANYANALYTICALTECHNIQUEISTHATADEDSCANBEMODELEDWITHMUCHLESSSTRINGENTASSUMPTIONS,ANDMORECOMPLEXPERFORMANCEMEASURESCANBEHANDLEDWITHRELATIVEEASEHOWEVER,DISCRETEEVENTSIMULATIONUSUALLYSUFFERSFROMSIGNIFICANTCOMPUTATIONALBURDENBECAUSEASINGLESIMULATIONRUNREPRESENTSONLYONEREAL
9、IZATIONOFASTOCHASTICPROCESSINORDERTOOBTAINANACCURATEPERFORMANCEESTIMATEUNDERAGIVENSETOFPARAMETERS,SEVERALINDEPENDENTRUNSORALENGTHYRUNIFTHEPROCESSISERGODICARENEEDED,ANDTHESERUNSSHOULDBEREPEATEDFORDIFFERENTSETSOFPERTINENTPARAMETERSINORDERTOAVOIDTHEESTIMATIONOFNETWORKTRAFFICSTATISTICSWHICHARESTILLREQUI
10、RED,ONECANRECORDTIMEOFGENERATIONANDLENGTHFOREACHMESSAGEANDFEEDTHISINFORMATIONINTOASIMULATIONMODELHOWEVER,THISREQUIRESALARGEAMOUNTOFINFORMATIONTRANSFERFROMEACHANDINDIVIDUALSTATIONTOTHECOMPUTERONWHICHTHEMODELISRUNNING,WHICHMAYDEGRADETHEOVERALLNETWORKPERFORMANCEOVERTHELASTDECADE,HOANDHISCOLLEAGUESHOAND
11、CAO,1991HOANDLI,1988ANDREFERENCESTHEREINHAVEDEVELOPEDTHETECHNIQUEOFPERTURBATIONANALYSISPATOCIRCUMVENTTHEDIFFICULTIESOFCONVENTIONALANALYSISANDSIMULATIONINDEDSPAESTIMATESTHEDEDSPERFORMANCEUNDERPERTURBEDCONDITIONSWITHDIFFERENTPARAMETERVALUESBYOBSERVINGTHESEQUENCEOFEVENTSOCCURRINGOVERAPERIODOFTIMEINTHEN
12、OMINALIE,UNPERTURBEDSYSTEMINFACT,PACONSTRUCTSPARTSOFEVENTSEQUENCEFORTHEPERTURBEDSYSTEMBASEDONTHENOMINALONETHISAPPROACHHASACOMPUTATIONALADVANTAGEOVERREPETITIVESIMULATIONESPECIALLYWHENNOANALYTICTECHNIQUEISAVAILABLEWHENTHEEFFECTSOFNPARAMETERSONAPERFORMANCEMEASUREARETOBEEVALUATED,THECONVENTIONALDISCRETE
13、EVENTSIMULATIONNEEDSNIRUNSONEWITHTHENOMINALPARAMETERSANDNRUNS,EACHWITHONEPERTURBEDPARAMETERANDTHEREMAININGNOMINALVALUESONTHEOTHERAND,PANEEDSONLYONERUNBECAUSEITCALCULATESTHEPERFORMANCEMEASUREOFTHEPERTURBEDSYSTEMBASEDONTHEINHERENTINFORMATIONFROMTHESIMULATIONWITHTHENOMINALPARAMETERTHEREFORE,THERATIOOFC
14、OMPUTATIONTIMECANBEAPPROXIMATELYITONIIFTHEPROCESSINGTIMEFORPAALGORITHMSISNEGLIGIBLECOMPAREDTOTHATFORDISCRETEEVENTSIMULATIONFORPERFORMANCEMANAGEMENT,PAISVERYSUITABLEBECAUSETHISTECHNIQUECANDIRECTLYUTILIZEONLINEOBSERVATIONOFEVENTSTHISDOESNOTREQUIREANYIDENTIFICATIONOFSTATISTICALPARAMETERSOFTHENETWORKTRA
15、FFICANDISCOMPUTATIONALLYMOREEFFICIENTTHANDISCRETEEVENTSIMULATIONFURTHERMORE,PASTILLRETAINSTHEINHERENTADVANTAGESOFSIMULATIONOVERANALYTICALTECHNIQUESDECISIONMAKINGREQUIRESPARAMETEROPTIMIZATION,ANDCANBEACCOMPLISHEDNUMERICALLYBYSTOCHASTICAPPROXIMATIONSAWHICHUTILIZESRANDOMMEASUREMENTSOVERAFINITEPERIODOFT
16、IMETOESTIMATETHEFINITEDIFFERENCEQUOTIENTOFTHEPERFORMANCEMEASUREWITHRESPECTTODECISIONVARIABLESRUBINSTEIN,1986SINCETHEMEASUREMENTFORPERFORMANCEMEASUREISARANDOMVARIABLEWITHANUNKNOWNDISTRIBUTION,THEESTIMATEDQUOTIENTSHAVEANONZEROVARIANCEATEVERYPOINTCONSEQUENTLY,THESATECHNIQUEHASTOREDUCEITSSTEPSIZEASTHEEX
17、TREMALPOINTISAPPROACHEDINMANYSITUATIONS,HOWEVER,MORETHANONEOPTIMIZATIONALGORITHMMAYBEREQUIREDBECAUSEANYINDIVIDUALALGORITHMISLIKELYTOBEEFFICIENTONLYINSOMEREGIONOFTHEPROTOCOLPARAMETERSETTINGSUNDERGIVENSTATISTICSOFNETWORKTRAFFICANADDITIONALLEVELOFDECISIONMAKINGISDESIRABLETOSELECTTHEMOSTAPPROPRIATEOPTIM
18、IZATIONALGORITHMACCORDINGTOTHECURRENTPARAMETERSETTINGSANDTRAFFICSTATISTICSTHISAPPROACHISLIKELYTOENHANCETHEEFFICIENCYANDCREDIBILITYOFPERFORMANCEMANAGEMENTINADYNAMICOPERATINGENVIRONMENTWHOSECHARACTERISTICSAREUNKNOWNORPARTIALLYKNOWNTECHNIQUESLIKELEARNINGAUTOMATALACANBEUSEDFORDECISIONMAKINGATTHEUPPERLEV
19、ELFORSELECTINGTHEMOSTEFFICIENTANDCREDIBLEOPTIMIZATIONALGORITHMBASEDONTHEPASTPERFORMANCENARENDRAANDTHATACHAR,1989ALEARNINGAUTOMATONCONSISTSOFASETOFACTIONS,ACORRESPONDINGSETOFACTIONPROBABILITIES,ANDAREINFORCEMENTSCHEMETHEACTIONPROBABILITIESAREUPDATEDBYTHEREINFORCEMENTSCHEMEACCORDINGTOTHERESPONSEFROMTH
20、EENVIRONMENTWHICHREACTSTOTHEACTIONOFTHELEARNINGAUTOMATONAPERFORMANCEEVALUATORPEASAPARTOFTHEENVIRONMENTISREQUIREDTOINTERPRETTHERESPONSEIFTHEINTERPRETATIONISFAVORABLE,THENTHEPROBABILITYOFTHECHOSENACTIONISINCREASEDANDTHOSEFOROTHERACTIONSAREDECREASEDBYREPEATINGTHISPROCESS,THELEARNINGAUTOMATONCANSELECTTH
21、EBESTACTIONUNDERTHECURRENTENVIRONMENTINADYNAMICENVIRONMENT,THEPERFORMANCEMANAGEMENTALGORITHMMUSTKNOWWHETHERTHENETWORKTRAFFICSTATISTICSHAVECHANGEDBECAUSEITHEOPTIMIZATIONALGORITHMISLIKELYTOHAVEITSOWNABILITYOFADAPTATIONSUCHASREDUCTIONOFTHESTEPSIZEINSAANDIIACHANGEINNETWORKTRAFFICMAYMISLEADTHELEARNINGAUT
22、OMATONINEVALUATINGTHEPERFORMANCEOFTHEOPTIMIZATIONALGORITHMSTHEREFORE,WHENEVERANYCHANGEINNETWORKTRAFFICSTATISTICSTAKESPLACE,THESTEPSIZEOFTHEOPTIMIZATIONALGORITHMANDTHEPEMAYHAVETOBERESETTHEPROPOSEDPERFORMANCEMANAGEMENTISBASEDONTHEPRINCIPLESOFPERTURBATIONANALYSISPA,STOCHASTICAPPROXIMATIONSA,ANDLEARNING
23、AUTOMATALATHELINEARTOKENPASSINGBUSLTPBPROTOCOLSOCIETYOFAUTOMOBILEENGINEERS,1987HASBEENSELECTEDTODEMONSTRATETHEDLICACYOFTHISPERFORMANCEMANAGEMENTPROCEDUREFORADJUSTMENTOFTOKENHOLDINGTIMERTHTANDTHREETOKENROTATIONTIMERSTRTI,TRT2,ANDTRT3INREALTIMEONTHEBASISOFMEASUREDNETWORKPERFORMANCETHEOPERATINGPRINCIPL
24、EOFLTPBANDITSPRIORITYMECHANISMAREDESCRIBEDINAPPENDIXA2IMPLEMENTATIONOFTHEPERFORMANCEMANAGEMENTTOOLTHTPERFORMANCEMANAGEMENTALGORITHMHASBEENIMPLEMENTEDONANETWORKTESTBEDWHERETHELTPBPROTOCOLUNDERCONSIDERATIONWASEMULATEDBYTWOINTERACTINGAPPLICATIONPROCESSESTHETESTBEDISOPERATEDONTHEIEEE8024TOKENBUSPROTO,OL
25、INTHEENVIRONMENTOFTHE10MBPSBROADBANDMANUFACTURINGAUTOMATIONPROTOCOLMAPTHEPHYSICALCONFIGURATIONOFTHETESTBEDCONSISTSOFALENGTHOFCOAXIALCABLE,AHEADENDREMODULATOR,ANDTHREEHOSTSEACHHOSTCOMPUTERISEQUIPPEDWITHTWONETWORKCARDSFROMINDUSTRIALNETWORKINGINCORPORATEDINDUSTRIALNETWORKINGINCORPORATED,1987ONEOFTHEHOS
26、TSOPERATESASANETWORKMANAGEMENTCONSOLEFORINITIALDOWNLOADINGOFTHEPROTOCOLSTOTHEREMAININGTWOHOSTSFORTHESETWOHOSTS,ASOFTWAREPACKAGEHASBEENDEVELOPEDTOEMULATEANUMBEROFLTPBSTATIONSBYGENERATING,TRANSMITTING,ANDRECEIVINGMESSAGESWHICHAREESSENTIALLYPACKETSOFTHEASSOCIATIONCONTROLSERVICEELEMENTACSEOFMAP3IMPLEMEN
27、TATIONSTRATEGYSINCETHEPAALGORITHMINVOLVESONLYLOGICANDADDITIONOPERATIONS,ITHASBEENIMPLEMENTEDINADISTRIBUTEDMANNERSOTHATTHEALGORITHMISEXECUTEDATEACHSTATIONTOESTIMATETHENETWORKPERFORMANCEUNDERPERTURBATIONSTHEALGORITHMISCAPABLEOFFOLLOWINGFOURPERTURBEDPATHSINWHICHONEOFTHEFOURTIMERSETTINGSOFTHELTPBPRIORIT
28、YMECHANISMISPERTURBEDTHEEFFECTOFATIMERPERTURBATIONPERTURBATIONINTHETOKENRECEPTIONINSTANTONTHENEXTSTATIONPROPAGATESTHROUGHTHELOGICALRINGVIATHETOKENWHICHCARRIESTHISINFORMATIONINTHISDISTRIBUTEDIMPLEMENTATION,ADDITIONALTRAFFICDUETOMANAGEMENTOPERATIONSISEXPECTEDTOBESIGNIFICANTLYSMALLERTHANTHATFORACENTRAL
29、IZEDPAALGORITHMWHICHREQUIRESINFORMATIONONCONTENTSOFQUEUESANDTIMERSTATUSFROMALLSTATIONSONTHEOTHERHAND,THEDECISIONMAKINGMODULETHATINCLUDESSTOCHASTICAPPROXIMATIONANDLEARNINGAUTOMATONISCENTRALIZEDATADESIGNATEDSTATION,HEREAFTERREFERREDTOASPERFORMANCEMANAGEREVENTHOUGHTHENETWORKMAYHAVEMORETHANONESTATIONWIT
30、HPARTIALORCOMPLETECAPABILITYTOEXECUTEDECISIONMAKINGFUNCTIONS,THECENTRALIZEDSTRATEGYALLOWSONLYONEACTIVECOPYOFEACHDECISIONMAKINGFUNCTIONTHERATIONALEFORSELECTINGCENTRALIZEDDECISIONMAKINGISTHATTHEDISTRIBUTEDSTRATEGY,WHEREEVERYSTATIONCOULDMAKEDECISIONSAUTONOMOUSLYBASEDONLYONITSLOCALPERFORMANCE,WOULDRESUL
31、TININCONSISTENCYANDCONFLICTBYHAVINGDIFFERENTTIMERSETTINGSOVERTHENETWORKNETWORKOPERATIONSUNDERTHEPROPOSEDPERFORMANCEMANAGEMENTPROCEDUREINVOLVEASERIESOFITERATIONSWHICHCONSISTOFANOBSERVATIONPERIODANDSUBSEQUENTMANAGEMENTACTIONSATTHEBEGINNING,THEPERFORMANCEMANAGERBROADCASTSTHEINITIALTIMERSETTINGSANDTIMER
32、PERTURBATIONVECTORSTOALLSTATIONSDURINGANOBSERVATIONPERIOD,EACHSTATIONEXECUTESITSOWNPAALGORITHM,ANDTHEPERFORMANCEMANAGERMONITORSTHEMESSAGESFLOWINGOVERTHENETWORKWHENTHEPERFORMANCEMANAGERDECIDESTHATENOUGHDATAHAVEBEENCOLLECTED,ITWAITSFORTHETOKENANDTHENBROADCASTSTHEREQUESTFORAPERFORMANCEREPORTTOALLSTATIO
33、NSTHENTHEPERFORMANCEPERFORMANAGERPASSESTHETOKENWITHOUTANYFURTHERMESSAGETRANSMISSIONONCETHISREQUESTFROMTHEPERFORMANCEMANAGERISRECEIVED,OTHERSTATIONSINTERRUPTTHEIRNORMALOPERATIONS,PREPARETHEPERFORMANCEREPORTS,ANDTRANSMITTHESEREPORTSASSOONASTHETOKENISRECEIVEDAFTERONECOMPLETETOKENCIRCULATION,THEMANAGERR
34、ECEIVESTHETOKENAGAINAND,BYTHISTIME,REPORTSFROMALLOTHERSTATIONSHAVEBEENRECEIVEDTHEN,THEMANAGERPROCESSESTHEREPORTS,COMPUTESNEWTIMERSETTINGSANDBROADCASTSTHEMWITHNEWTIMERPERTURBATIONVECTORSFORTHENEXTITERATIONUPONRECEPTIONOFTHENEWSETTINGSANDPERTURBATIONVECTORS,ALLSTATIONSSETTHEIRCORRESPONDINGVARIABLESAND
35、WAITFORTHETOKENTORESUMENORMALOPERATIONSAPERFORMANCEMANAGEMENTALGORITHMFORMULTIPLEACCESSNETWORKSHASBEENCONCEPTUALIZEQ,ANDFORMULATEDFORATOKENBUSPROTOCOLBYUSINGTHEPRINCIPLESOFIPERTURBATIONANALYSISOFDISCRTTEEVENTDYNAMICSYSTEMSIISTOCHASTICAPPROXIMATION,ANDIIILEARNINGAUTOMATATHEPROCEDUREISAIMEDTOIMPROVETH
36、EPERFORMANCEOFCIMNLTWORKSINHANDLINGVARIOUSTYPESOFMESSAGESBYONLINEADJUSTMENTOFPROTOCOLPARAMETERS,ANDHASBEENEMULATEDONANETWORKTESTBEDTHECONCEPTUALDESIGNPRESENTEDINTHISPAPEROFFERSASTEPFORWARDTOBRIDGINGTHEGAPBETWEENMANAGEMENTSTANDARDSANDUSERSDEMANDSFOREFFICIENTNETWORKOPERATIONSSINCEMOSTSTANDARDSSUCHASIS
37、OANDIEEEADDRESSONLYTHEARCHITECTURE,SERVICES,ANDINTERFACESFORNETWORKMANAGEMENTTHEFOLLOWINGMAJORCONCLUSIONSAREDERIVEDFROMTHERESULTSOFSIMULATIONANDEMULATIONOFPERFORMANCEMANAGEMENTOFLINEARTOKENPASSINGBUSLTPBPROTOCOLASDESCRIBEDINTHISPAPER,THEPERFORMANCEMANAGEMENTOFANETWORKHASTOBEEXECUTEDWITHINSUFFICIENTA
38、PRIORIKNOWLEDGEINTHEPERFORMANCECHARACTERISTICSOFTHENETWORKANDITSSURROUNDINGENVIRONMENTITMAYBEEFFECTIVETOEMPLOYVARIOUSKNOWLEDGEBASEDTECHNIQUESSUCHASFUZZYREASONING,EXPERTSYSTEM,NEURALNETWORK,ANDGENETICALGORITHMSOURCELECTURENOTESINCOMPUTERSCIENCE2008P928译文集成制造系统网络的绩效管理SUKLEE,ASOKRAY,KEUMSHIKHONG,JOONGS
39、UNYOON,ANDMYUNGCHUHAN1、绩效管理程序的结构网络绩效，更准确地说是通过信息包邮的延迟，可能对实时制造过程的动态性和稳定性是至关重要的。这是真实的，当多台没有直接连接的机器在执行一项任务时。其中一个例子是智能焊接系统（NAYAK,RAYANDVAVRECK,1987）,就是定位平台和机器人通过交换理想的标位置和各种信息状态报告进行交流。另一个例子是通过不仅一个机器人在自己开始的轨迹收到信号通过网络完成一种控制器和报告负离子的飞行轨迹从而操纵庞大和灵活的工件。因为延迟会损坏工件或机器人的手腕和胳膊，所以发送数据的及时性事十分必要的。为了保持各种机器制造过程的动态性和稳定性的可接
40、受水平，绩效管理要求能够实时操纵可协调参数从而使网络能够适应动态的环境。这个可以在两个步骤中完成（1）绩效评估是找到协议参数如何影响绩效测量来确定绩效评估和协议参数之间的关系；（2）决策决定如何调整协议，即参数识别的方向和大小不同的参数调整向量，利用在第一步和历史绩效中提供的资讯。绩效评价分析检测技术等比如排队理论VISWANADHAMANDNARAHARI,1992往往需要不切实际的假设就像泊松的到来和使得绩效测量变得复杂的不容易加工的数学结构。此外，网络流量统计比如信息生成区间和信息长度的分布，要求作为输入的分析模型是很难实时估计的。从另外一方面来说，离散事件模型LAWANDKELTON,
41、1991是一可行选择的替代分析技术。模拟超过任何分析技术的主要优势是在不是很严密的假设下DEDS可以被模拟，使得更加复杂的绩效测量可以轻而易举的处理。然而，离散事件模拟计量通常遭受重大因为一个只代表了一个仿真运行一个随机过程，实现。为了获得一个精确估计在特定性能参数的确定，几个独立分（或长时间的运行如果这一过程遍历性）是需要的，而这些应该重复不同的相关参数。为了避免网络流量统计的估计。仍然需要时间，可以记录每个信息的生成和长度，了解仿真模型这个情况。但是，这需要大量的信息传递从他们个人站的计算机模型运行时，可降解整个网络的性能。在过去的十年里，HO和他的同事（HO和曹，1991；HO和李，19
42、88中和参考）发展了技术摄动分析方法来克服DED常规分析和仿真的困难。PA通过观察一连串发生在一段时间内的事件系统估计在摄动条件下的DEDS绩效。事实上，PA构建了部分基于名义的事件序列的摄动系统。该算法在重复仿真上具有优势，尤其在不分析技术和可用的时候。当对影响绩效评价N个参数进行评估时，传统的离散事件模拟需要N1分（一个命名参数和N分，每一个摄动参数和剩下的票面价值）从另一方面来说，PA只需要一分，因为它计算摄动系统的绩效评价是基于从模拟和标准参数上获得的内在信息。因此，计算时间的比例如果PA算法在加工时间中相比可以忽略不计离散事件的仿真，那么大约可以算为I到NI。对于绩效管理，PA是十分
43、合适的，因为这种技术可以直接利用在线观察事件。与计算仿真相比需、效率更高，这并不需要任何统计参数识别网络流量和离散事件。此外，PA在技术分析上仍旧保留着仿真固有的优点。决策需要参数优化，可以实现随机逼近数值（SA）随机测量，绩效评价对决策变量利用有限的一段时间估计商的有限差分法（RUBINSTEIN,1986）。因为绩效评价的测量是一个未知腹部的随机变量，估计商在每一点上具有非零方差。因此，SA必须技术为极点值进行探讨减少每一步的大小。然而在很多情况下，不止要求优化一个算法，因为任何个人算法在一定的统计交通网络下的一些地区的参数设置协议下得到有效提高。增加一层决策对根据当前的参数设置和交通统计
44、选择最合适的优化算法是可取的。这种方法可能使进行管理在一个动态经营环境的特征是未知的或部分为人所知的情况下能提高效率和信誉。技术比如学习自动机，可以用于上层决策在基于过去进行表现基础之上，为其选择最高效和最可靠的优化算法。NARENDRA和THATACHAR,1989。一个学习自动机制有一组行动、相应的行动和概率加固法案组成。更新概率的行动方案，加强根据环境的反应堆学习是作用自动机制。一个进行评估员作为环境的一部分需要解释反应。如果解释优惠，那么选择行动的可能性增加，而且其他行为是减少。我们重复这个过程中，学习自动机可以选择最好的行动在当前动态环境中，演出管理算法必须知道这个网络流量统计数字一
45、改变，因为，（1）该算法可能有自己的能力的适应如降低在SA的台阶尺寸；（2）改变网络流量可能误导徐诶些自动机进行评价的优化算法。那么，无论何时，当有任何变化的网络流量统计的必要条件的地方，台阶尺寸的优化算法和体育可能需要重新设置。提出该绩效管理是基于摄动分析（PA），随机逼近（SA）和学习自动机LA等原则。线性令牌通过公共汽车（LTPB）协议（社会的汽车工程师，1987）被选中来验证DLICACY绩效管理程序令牌控股计时器调整尺寸和三个表示，TRT2定时器（TRT1、TRT2和TRT3）在实时测量的基础上，提出了网络性能。LTPB的工作原理及其优先机制在附录A中描述。2、绩效管理执行工具绩效管
46、理算法应用于网络站在LTPB协议考虑到有两个相互作用进行仿真应用程序。运行的网站8024令牌总线模块款，01环境中（亿位/秒）制造自动化协议（MAP）网站的身体形态包括一段同轴电缆、一个高端的REMODULATOR和三个主机。每个主机电脑配备了从工业网络卡中成立的两个网络卡（工业网络结合，1987）。其中一台主机的运行作为其他两台初始的下载网管控制台的协议。在这两个主机上，软件包被开发出来通过产生、发送、接收信息在MAP本质上协会控制服务元素来效仿大量的LTPB网站。3、实施战略YINWIEPA算法只涉及逻辑与加法，它已经实施在一个分布式的态度，这样算法才能在每一个网站平衡种群波动中执行。该算
47、法能够消除在LTPB定时器的的优先机制下的四大路径的厌烦。定时器在下一站传播的效果通过逻辑环传输这些信息。这种分布式实施，附加流量预计将比PA算法的经营管理工作大大减少，这就要求从所有站的列队和定时器位置上的内容中获取信息。从另一方面来说，决策模块，包括随机逼近和学习自主机在制定站，一下简称绩效经理。即使一个网站有不止一个站用其部分或全部能力来执行决策。集中策略允许每一个决策功能只有一个复制活动。选择集中决策的理论就是分布式策略，在每一站基于其本地绩效的基础上可以决定其大权，通过在网络设定不同程式的事件会导致矛盾和摩擦。网络操作在此提议中的绩效管理程序涉及由一段观察期和事后管理动作组成的一系列
48、迭代。在开始，绩效经理向所有站广播最初的定时器设置和定时器扰动向量。通过一段观察期，每个站交换它们自己的PA算法，然后绩效经理监控通过网络的信息流量。当绩效经理决定已经收集了足够的数据后，它就等待记号然后请求执行报告到所有站。然后绩效经理在没有任何信息输入时通过令牌。一旦从绩效经理那收到这种要求，其他站就会打断它们的正常运行，准备绩效报告，当令牌收到这些报告后马上进行传输。在一个完整的循环周期后，经理重新收回令牌，同时，其它站的报告也被收集了。然后经理处理这些报告，重新设置定时器和利用新的定时器摄动向量进行传播。在接收新设置和扰动向量时，所有站设定相应的变量来等待令牌恢复正常操作。一套多途径的
49、网络绩效管理算法已经被定义化，利用以下原则制定了一个令牌总线协议。（1）摄动分析事件动态系统；（2）随机逼近；（3）学习自动机。程序旨在提高CIMN1TWORKS在处理在线调整协议参数的各种类型的信息的绩效,和在ANETWORK试点上的仿效。概念设计进行了介绍一个一步直接爱你的差距，双语管理标准和用户的需求，因为大多数网络操作中有效途径，如ISO和电子地址建筑、服务为网络管理和接口。下面的主要结论是来自仿真结果和仿真线性绩效管理的传球公共汽车（LTPB）协议。在本文中的描述，网络绩效管理的执行在先验知识绩效特点的网络和它的环境中。可能有效利用各种知识技术，如模糊推理、专家系统、神经网络和遗传算法的优化方法。资料来源LECTURENOTESINCOMPUTERSCIENCE2008P928
