基于改进的回声状态神经网络的非线性预测

《基于改进的回声状态神经网络的非线性预测》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、南京工业大学硕士学位论文基于改进的回声状态神经网络的非线性预测姓名：王瑟申请学位级别：硕士专业：计算机应用技术指导教师：蔚承建20060515摘要关键词：回声状态神经网络；小波神经网络；小波分解；混沌时间序列预测；先验性；PSO；集群智能II硕士学位论文ABSTRACTNonlinearsystempredictionusingneuralnetworksappearsgreatefficiencyandhasabundanceofapplications.Recurrentneuralnetworksshowsmoreadvancedadvantagesamongthem

2、againstthesepredictiontasks,althoughitslearningmethodshavenotimprovedmuchmoreforlongtime.Echostatenetworkisonenovelstructureofrecurrentneuralnetwork(RNN)alsoonenovellearningmethodforRNNaswell,it’ssimilarwiththosebio-neural-networksstructurally,andithastheperfectSTMcapabilityasoneRNN.Itempl

3、oysonelargescaleRNNasinformationreservoircalleddynamicalreservoir,thenminimizesthemeaningsquarederror(MSE)duringtrainingtogetthelearningusingcomputingsimpleregressionweightmatrixfrominternalstatestowardsoutputunit.However,thereisonecontradictionexistinginESN,itis:toemploynonlinearneuroncan

4、raisethenonlinearcapabilityofESNbutreducetheSTMofitsimultaneously.IthastoemployoneverylargescaleDRwhenfacethosetoughtaskwhichrequirenotonlyhighnonlinearitybutalsoniceMClikechaotictimeseriesprediction.ThiscausestherunningprocessofESNslowingdownandbecomingmoreinstableduringexploitationperiod

5、.AccordingtothetranscendentalknowledgetheoryofANN,theESNcanemployotherneuralnodetoimprovetheperformance,thewavelonusinginWNNchoseninthisthesis.Theinternalstatespaceisenlargedwheninputsometunedwavelon.TheSWHESNcanpredict46%furtherthantheoriginalESNwithouttypicaldeviationbutonlyconsumingonly

6、30%timeofwhatESNdowhenlearningsamedatasample.Wecan’tforgetthatESNhasimprovedthebestprevious[1]technologybyfactor700.Thisthesisshowstri-highlightviews:1.WeintroducewavelonintoRNNwhichappearsinforwardANNtraditionally.2.Wereducedthediversitybetweenwavelonforthereasontosmoothworkingconditionin

7、ESNratherthanaugmentingthemwhichforwardANNwhoneedlargerbasicvectorfunctionembedded.3.Theparametersinechostatenetworksinvolvedinapplicationaresetbyexpertofechostatenetworkscommonly,whichusuallywasteofcomputationresource,inthispaperwepresentonemethodthattooptimi