高階雙并聯(lián)神經(jīng)網(wǎng)絡(luò)批處理梯度算法收斂性

1、大連理工大學(xué)碩士學(xué)位論文高階雙并聯(lián)神經(jīng)網(wǎng)絡(luò)批處理梯度算法收斂性姓名：郭軍平申請學(xué)位級別：碩士專業(yè)：計算數(shù)學(xué)指導(dǎo)教師：吳微20090601高階雙并聯(lián)神經(jīng)網(wǎng)絡(luò)批處理梯度算法收斂性ConvergenceofBatchGradientDescentAlgorithmforHighOrderDoubleParallelNeuralNetworksAbstractBPneuralnetworkiSakindoffeedforwardneuralne

2、tworksandiswidelyusedinmanyapplicationsBPnetworkhasmanyadvantages，butitonlyusesadditiveneuronsandhencehaspoorcapabilityforsolvingcomplexnonlinearproblemsToresolvethisproblemproductneuronsareappliedinfeedforwardneuralnetw

3、orkstoenhancethenonlinearmappingcapability剛skindoffeedforwardneuralnetworkbasedonhi曲orderneuronsiscalledhi曲orderfeedforwardneuralnetworksOntheotherhand，therearemanynonlinearproblemswimlinearpartinengineeringwhichCannotbe

4、resolvedbyasimpleperceptronoracommonfeedforwardneuralnetworksDoubleparallelfeedforwardneuralnetworks(DP孫N)areproposedforthiscaseInthisthesis，hi曲orderdoubleparallelneuralnetwork(HODPNN)isproposed，basedonthenetworkstructur

5、esandfeatureoftheabovetwOnetworkstesolvethiskindofcomplexnonlinearproblemswith1inearpartNumericalexperimentsindicatethattheneuralnetworkhasabetterlearningperformanceinsolvingthiskindofproblem，comparedwithPi—Sigmaneuralne

6、tworkandDP剛GradientalgorithmisasimpleoptimizationalgorithmandisoftenappliedforneuralnetworkstrainingTherearetwodifferentwaystoinputthesamplesduringthetrainingprocess：onlineversionandbatchversion111emainworkofthisthesisis

7、tostudybatchgradientdescentalgorithmforHODPNNinconvergenceThestructureofthisthesisiSorganizedasfollows：Chapter1givesabriefintroductionofartificialneuralnetworkChapter2givesproofofweakconvergenceandstrongconvergenceofbatc

8、hgradientdescentalgorithmforHODPNNFiIlallytwonumericalexperimentsonfunctionalapproximationarecarriedoutinChapter3，wherethelearningperformanceofthenetworkisalsodiscussedKeyWords：Highorderneuralnetwork；Doubleparallelfeedfo