A-New-Active-Queue-Management-Algorithm-based-on-S

ANewActiveQueueManagementAlgorithmbasedonSelf-adaptiveFuzzyNeural-networkPIDcontrollerYanQiaoCollegeofComputerandSoftwareShenzhenUniversityShenzhen,[e-mail]ormationEngineeringShenzhenUniversityShenzhen,[e-mail]—Activequeuemanagement(AQM)isaveryimportantresearchareaincongestioncontrol.ButthecomplexityanddynamiccharacteristicofthecomputernetworkcausethetraditionalPIDcontrolalgorithmlowadaptabilitytodynamicenvironmentduetoitsfixedparameters.Inordertoovercomesuchshortcomings,intelligentcontroltheorywasintroducedtocongestioncontrolresearch,andanewAQMalgorithmcalledFAPIDNNwasproposed.Fuzzycontrollerautomaticallycomputersthelearningrateηaccordingtothecurrentnetworkstate,andtheneuralnetworkPIDcontrollercalculatethepacketdroppingprobabilitybasedonthelearningrateprovidedbythefuzzycontroller.SimulationresultsshowthatFAPIDNNalgorithmissuperiortothepresentedPIDcontrolleronthequeuestability,convergencespeedandtimedelay.Keywords-congestioncontrol;activequeuemanagement;fuzzycontrol;neuralnetworks;self-adaptiveI.INTRODUCTIONInrecentyears,withtherapiddevelopmentofcomputerandnetworktechnologyandtherapidincreaseinmultimediaapplications,peoplehaveahigherdemandforthequalityofnetworkservice.Activequeuemanagement(AQM)hasbeenplayingasignificantroleinnetworkcongestioncontrolresearch.TraditionalPI[1]andPID[2]controllerforAQMbasedonclassicalcontroltheoryaretypicalrepresentatives,whichwereappliedtosolvecongestioncontrolproblemoftheintermediatenodesontheInternet,andthequalityofnetworkservicewasimproved.Hollotetal.[1]proposedanactivequeuemanagementbasedonPIcontroller,andtheperformanceofnetworkwasimproved,suchasqueuestability.Howeveritcannotachievebothsatisfyingtransientresponseandsmalldeviationfromsteady-statebehavioroverawiderangeofnetworkdynamics.Renetal.[2]proposedPIDAQMalgorithmbasedongainandphasemargins.PIDcontrollercanreducetheregulationtimeofAQMsystemandimprovethetransientperformance.However,thecomplexityanddynamiccharacteristicofthecomputernetworkcausethelowadaptabilityoftraditionalPIDcontrolalgorithmduetoitsfixedparameters.Inordertoachieveastableandexpectedqueuelength,lowpacketlossandhighlinkutilizationinatime-delayedTCP/IPnetwork,intelligentcontroltheorywasintroducedtocongestioncontrolresearch.Inthispaper,anewAQMalgorithmbasedonself-adaptivefuzzyneural-networkPIDcontrollercalledFAPIDNNwasproposed.ThenewcontrollerintegratesthemeritsofbothPIDcontrollerbasedonneuralnetworkandfuzzycontroller,whichcanimprovetheperformanceandrobustnessofAQMcontrolsystem.Thefollowingofthepaperisorganizedinsuchaway:SectiontwointroducestheanalysisanddesignoftheFAPIDNNalgorithmindetails.SectionthreeshowsthesimulationresultsinordertovalidatethedesignandcomparestheperformanceofFPIDNNwithotherAQMalgorithm.Finally,aconclusionisgiveninSectionfour.II.THEFAPIDNNCONTROLLERSTRUCTUREAnewPIDcontrollerbasedonneuralnetworkdefinedin[3],whichnamedImprovedPIDNN,cantunethelearningrateautomatically.However,becauseoftheroughcontrolofit,whendisturbancesincreased,therobustnessoftheAQMcontrolsystemwillbeworse,andtherateofsystemresponsewillreduce.Inordertoavoidtheweightsoscillateinlearningprocessandtheslowconvergencespeed,weproposedanewself-adaptiveAQMalgorithmcalledFAPIDNNbasedonPIDNNdesignedin[4].Thenewalgorithmconsideredafuzzycontrollertojoininthecontrolsystem,whichcanrealizethedynamicadjustmentoflearningrate,speeduptheconvergencerateofthequeueandimprovethestabilityofAQMcontrolsystem.FAPIDNNwhichisshowninFig1consistsoftwomainparts:thePIDcontrollerbasedonneuralnetworkandafuzzycontroller.FAPIDNNintegratesthemeritsofthetwoparts,whichinheritsnotonlythestabilityofPIDcontrol,butalsothequickresponseandadaptabilityoftheneuralnetworks.Therefore,thenewAQMalgorithmisrobusttothevariationsinnetworkparametersandmodelingerrors.TheresultsofsimulationshowthatFAPIDNNissuperiortotheImprovedPIDNN[3]andthetypicalPIDcontroller[2]ontheperformanceofcontrolsystem.TheblockdiagramoftheAQM978-1-4244-7255-0/11/$26.00©2011IEEEcontrolsystembasedonself-adaptivefuzzyneural-networkPIDcontrollerisillustratedinFig.1.Figure1.FAPIDNNcontrollerInFig.1,qrefisexpectedqueuelength,qisactualqueuelength.FAPIDNNconsidersqrefandqastheinputs,thefuzzycontrollerdynamicallytunethelearningrateofneuralnetworksaccordingtothelinkconditions,thentheneural-networkPIDcontrollerwilltunetheweightsandcalculatethedropprobabilityastheoutputofAQMcontrolsystem.A.TheNeural-networkPIDControllerTheneural-networkPIDcontrollershowninFig.2isformedbyathreelayersfeed-forwardneuralnetwork.Theinputlayerhastwoinputneurons,oneisexpectedqueuelengthoftheintermediatenodesandtheotheristheactualqueuelength.Therearethreenodesinthehiddenlayer,andtheyareproportional,integralanddifferentialneuronrespectively.ThereisonlyoneneuronintheoutputlayeranditsoutputisthedropprobabilityoftheAQMcontrolsystem.Theblockdiagramoftheneural-networkPIDcontrollerisillustratedinFig.2.Figure2.Neural-networkPIDcontroller[4]FromFig.2,qrefisexpectedqueuelength,qisactualqueuelengthandpisdropprobability.Atthesamplingtime,theinp