Echo State Networks for Adaptive Filtering

《Echo State Networks for Adaptive Filtering》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、UniversityofAppliedSciencesDepartmentofComputerScienceMasterThesisEchoStateNetworksforAdaptiveFilteringAliUygarKüçükemreAthesissubmittedtotheUniversityofAppliedSciencesBonn-Rhein-SiegforthedegreeofMasterofScienceinAutonomousSystemsstRefereeandTutor:1Prof.Dr.Paul-GerhardPlöge

2、rndReferee:2................................................rdexternalReferee:3................................................Submitted:30.04.2006ECHOSTATENETWORKSFORADAPTIVEFILTERINGI,theundersigned,declarethatthisworkhasnotpreviouslybeensubmittedtothisoranyotherUniversity

3、,andthatunlessotherwisestated,itisentirelymyownwork.........................................................DateAliUygarKüçükemreA.U.KÜÇÜKEMREIIECHOSTATENETWORKSFORADAPTIVEFILTERINGACKNOWLEDGEMENTSDevotedtothemydearfriendÖzgünDoğan....SpecialthankstomyprofessorPaul-GerhardPl

4、ögerforhisvaluablesupport,teachingandunderstandingduringmystudies.ForthefirsttimeinmylifeIlikedMathematicsafterhisshortdiscoursesduringthelectures...Nevertheless,Ithanktomydearparentsforalltheirsupportduringthisthesiswork.A.U.KÜÇÜKEMREIIIECHOSTATENETWORKSFORADAPTIVEFILTERING

5、ABSTRACTAnAdaptiveFilterisaselfdesigningsystemthatcanworkinnon-stationaryenvironments.Itreliesonarecursivealgorithmwhichcontinuouslyupdateitsparametersinordertotrackthechangesitssurrounding.Ifthenon-stationarityoftheenvironmentisgovernedbynonlineardynamicalprocessesaRecurren

6、tNeuralNetworks(RNNs)canbeusedsuccessfullylikeanadaptivefilter,providedtheycanlearninanonlinemanner.Thisisanecessaryconditionforadaptivefiltering.Inthisthesis,wewilldiscussusageofaspecialtypeofaRNNforthosetasks,theyarecalledtheEchoStateNetworks(ESNs).Theyconstituteanovelappr

7、oachthatcanbeusedforRNNtraining.Countertotheconventionalmethods,whichusuallyrequireahugeoftimetolearnagivensystem,anESNcanbetrainedinlineartimeinaveryquickway.Mostofthetime,anofflinelearningalgorithmisusedtoteachtheESN.Butthisdoesnotmeanthattrainingisonlylimitedtothisalgorit

8、hm.ThestructureoftheESN,letstheonlineadaptationalgorithmsfromadaptivefilter