简介:Artificalneuralnetworks(ANN)arenowwidelyusedsuccessfullyastoolsforhithenergyphysics.Thepapercoverstwoaspects.First,mappingANNsontotheproposedringandlinearsystolicarrayprovidesanefficientimplementationofVLSI-basedarchitecturessinceinthiscaseallconnectionsamongprocessingelementsarelocalandregular,Second.itisdiscussedalgorthmicorganizingofsuchstructuresonthebaseofmodularalgebrawhoseusecanprovideanessentialincreaseofsystemthroughput.
简介:WeproposeanapplicationoftheelasticneuralnetforringrecognitioninRICHdetectors.Themethodhasbeendevelopedtofindringsdistortedduetomisalignmentofdetectorsandcontaminatedbynoise.ThealgorithmwastestedonsimulateddataofCOMPASSRICH-1detector.Reconstructionefficiencyis99.95%fortripleLEPTOeventstaking5msperevent.
简介:TheneuralnetworkrealtimeeventselectionfortheDIRACewperimentatCERNispresented.Itcomprisesoftwoindependentparts.OneusesplasticscintillatorsandtheothertheverticalScintillatingFibres,Theglobaleventdecisionistakeninlessthan250ns.Signaleventsareselectedwithanefficiencyosmorethan0.99withabackgroundratereductionofabout2.
简介:IntheresearchofusingRadialBasisFunctionNeuralNetwork(RBFNN)forecastingnonlineartimeseries,weinvestigatehowthedifferentclusteringsaffecttheprocessoflearningandforecasting.Wefindthatk-meansclusteringisverysuitable.Inordertoincreasetheprecisionweintroduceanonlinearfeedbacktermtoescapefromthelocalminimaofenergy,thenweusethemodeltoforecastthenonlineartimeserieswhichareproducedbyMackey-Glassequationandstocks.Byselectingthek-meansclusteringandthesuitablefeedbackterm,muchbetterforecastingresultsareobtained.
简介:Therobustexponentialstabilityofalargerclassofdiscrete-timerecurrentneuralnetworks(RNNs)isexploredinthispaper.Anovelneuralnetworkmodel,namedstandardneuralnetworkmodel(SNNM),isintroducedtoprovideageneralframeworkforstabilityanalysisofRNNs.MostoftheexistingRNNscanbetransformedintoSNNMstobeanalyzedinaunifiedway.ApplyingLyapunovstabilitytheorymethodandS-Proceduretechnique,twousefulcriteriaofrobustexponentialstabilityforthediscrete-timeSNNMsarederived.Theconditionspresentedareformulatedaslinearmatrixinequalities(LMIs)tobeeasilysolvedusingexistingefficientconvexoptimizationtechniques.Anexampleispresentedtodemonstratethetransformationprocedureandtheeffectivenessoftheresults.
简介:Inthispaper,weproposeanovelapproachtoachievespectrumprediction,parameterfitting,inversedesign,andperformanceoptimizationfortheplasmonicwaveguide-coupledwithcavitiesstructure(PWCCS)basedonartificialneuralnetworks(ANNs).TheFanoresonanceandplasmon-inducedtransparencyeffectoriginatedfromthePWCCShavebeenselectedasillustrationstoverifytheeffectivenessofANNs.WeusethegeneticalgorithmtodesignthenetworkarchitectureandselectthehyperparametersforANNs.OnceANNsaretrainedbyusingasmallsamplingofthedatageneratedbytheMonteCarlomethod,thetransmissionspectrapredictedbytheANNsarequiteapproximatetothesimulatedresults.Thephysicalmechanismsbehindthephenomenaarediscussedtheoretically,andtheuncertainparametersinthetheoreticalmodelsarefittedbyutilizingthetrainedANNs.Moreimportantly,ourresultsdemonstratethatthismodel-drivenmethodnotonlyrealizestheinversedesignofthePWCCSwithhighprecisionbutalsooptimizessomecriticalperformancemetricsforthetransmissionspectrum.Comparedwithpreviousworks,weconstructanovelmodel-drivenanalysismethodforthePWCCSthatisexpectedtohavesignificantapplicationsinthedevicedesign,performanceoptimization,variabilityanalysis,defectdetection,theoreticalmodeling,opticalinterconnects,andsoon.
简介:Inthispaper,amulti-objectiveparticleswarmoptimization(MOPSO)algorithmandanondominatedsortinggeneticalgorithmⅡ(NSGA-Ⅱ)areusedtooptimizetheoperatingparametersofa1.6L,sparkignition(SI)gasolineengine.Theaimofthisoptimizationistoreduceengineemissionsintermsofcarbonmonoxide(CO),hydrocarbons(HC),andnitrogenoxides(NOx),whicharethecausesofdiverseenvironmentalproblemssuchasairpollutionandglobalwarming.Stationaryenginetestswereperformedfordatageneration,covering60operatingconditions.Artificialneuralnetworks(ANNs)wereusedtopredictexhaustemissions,whoseinputswerefromsixengineoperatingparameters,andtheoutputswerethreeresultingexhaustemissions.TheoutputsofANNswereusedtoevaluateobjectivefunctionswithintheoptimizationalgorithms:NSGA-ⅡandMOPSO.Thenadecision-makingprocesswasconducted,usingafuzzymethodtoselectaParetosolutionwithwhichthebestemissionreductionscanbeachieved.TheNSGA-Ⅱalgorithmachievedreductionsofatleast9.84%,82.44%,and13.78%forCO,HC,andNOx,respectively.WithaMOPSOalgorithmthereachedreductionswereatleast13.68%,83.80%,and7.67%forCO,HC,andNOx,respectively.
简介:Ameasurementtechniquethatcanmeasuretheconcentrationofthesolidparticlesinliquidflowwasdeveloped.ThemeasurementsystemconsistsofacolorcameraandthreeLCDdisplays.ThesolidparticleswereputatthebottomofacylindricalmixingtankinwhichJetA1oilwasfilled.Transientmixingofthesolidparticleswasperformedbyrotatingapropellertypeagitatorwiththreedifferentrotationspeed(500,600,700r/min).MixingstatewasvisualizedbytheLCDdisplaysandacolorcamcorder.Thecolorintensityoftheglassparticleschangeswiththeirconcentration.ThecolorinformationwasdecodedintothreeprinciplecolorsR,G,andBsothat,thecalibrationcurveofcolor-to-concentrationwasperformedusingtheseinformation.Aneuralnetworkwasusedforthiscalibration.Thetransientconcentrationfieldofthesolidparticleswasquantitativelyvisualized.
简介:Inthispaper,weinvestigatecoherenceresonance(CR)andnoise-inducedsynchronizationinHindmarsh-Rose(HR)neuralnetworkwiththreedifferenttypesoftopologies:regular,random,andsmall-world.ItisfoundthattheadditivenoisecaninduceCRinHRneuralnetworkwithdifferenttopologiesanditscoherenceisoptimizedbyapropernoiselevel.Itisalsofoundthatascouplingstrengthincreasestheplateauinthemeasureofcoherencecurvebecomesbroadenedandtheeffectsofnetworktopologyismorepronouncedsimultaneously.Moreover,wefindthatincreasingtheprobabilitypofthenetworktopologyleadstoanenhancementofnoise-inducedsynchronizationinHRneuronsnetwork.
简介:Thealternatecombinationalapproachofgeneticalgorithmandneuralnetwork(AGANN)hasbeenpresentedtocorrectthesystematicerrorofthedensityfunctionaltheory(DFT)calculation.IttreatstheDFTasablackboxandmodelstheerrorthroughexternalstatisticalinformation.Asademonstration,theAGANNmethodhasbeenappliedinthecorrectionofthelatticeenergiesfromtheDFTcalculationfor72metalhalidesandhydrides.ThroughtheAGANNcorrection,themeanabsolutevalueoftherelativeerrorsofthecalculatedlatticeenergiestotheexperimentalvaluesdecreasesfrom4.93%to1.20%inthetestingset.Forcomparison,theneuralnetworkapproachreducesthemeanvalueto2.56%.Andforthecommoncombinationalapproachofgeneticalgorithmandneuralnetwork,thevaluedropsto2.15%.Themultiplelinearregressionmethodalmosthasnocorrectioneffecthere.