资源描述：
《Stochastic Modelling Hints for Neural Network Prediction》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。
1、StochasticModellingHintsforNeural1NetworkBasedTimeSeriesPredictions2RaduDrossuZoranObradovicrdrossu@eecs.wsu.eduzoran@eecs.wsu.eduSchoolofElectricalEngineeringandComputerScienceWashingtonStateUniversity,Pullman,Washington,99164-2752AbstractTheobjectiveofthisstudyistoinvestigatethere
2、lationshipbetweenstochasticandneuralnetworkapproachestotimeseriesmodelling.Experimentsonbothacomplexreallifepredictionproblem(entertainmentvideotracseries)aswellasonanarti ciallygeneratednonlineartimeseriesonthevergeofchaoticbehavior(Mackey-Glassseries)indicatethattheinitialknowledg
3、eobtainedthroughstochasticanalysisprovidesareasonablygoodhintfortheselectionofanappropriateneuralnetworkarchitecture.Althoughnotnecessarilytheoptimal,sucharapidlydesignedneuralnetworkarchitectureperformedcomparableorbetterthanmoreelaboratelydesignedneuralnetworksobtainedthroughexpens
4、ivetrialanderrorprocedures.Keywords:timeseries,non-stationaryprocess,ARMAmodelling,neuralnetworkmod-elling,predictionhorizon.1Correspondence:Z.Obradovic,phone:(509)335-6601,Fax:(509)335-38182ResearchsponsoredinpartbytheNSFresearchgrantNSF-IRI-9308523.1INTRODUCTIONAtimeseriesxcanbede
5、 nedasarandom(ornondeterministic)functionxofantindependentvariablet[6].Itsmaincharacteristicisthatitsfuturebehaviorcannotbepredictedexactlyasinthecaseofadeterministicfunctionoft.However,thebehaviorofatimeseriescansometimesbeanticipatedbydescribingtheseriesthroughprobabilisticlaws.Com
6、monly,timeseriespredictionproblemsareapproachedeitherfromastochasticperspective[1]or,morerecentlyfromaneuralnetworkperspective[9,12].Eachoftheseapproacheshasadvantagesanddisadvantages:thestochasticmethodsareusuallyfast,butoflimitedapplicabilitysincetheycommonlyemploylinearmodels,wher
7、eastheneuralnet-workmethodsarepowerfulenough,buttheselectionofanappropriatearchitectureandparametersisatimeconsumingtrialanderrorprocedure.At rstglanceitmightseemthatthereisn'tanydirectrelationshipbetweentimeseriesandneuralnetworks,butthereareatleasttworeasonsthatmightmakeneuralnetwo
8、rksveryattra
