神經(jīng)網(wǎng)絡(luò)初始權(quán)值優(yōu)化技術(shù)在移動機(jī)器人學(xué)習(xí)中的應(yīng)用.pdf

1、一，啦U常蜥研究生學(xué)位論文神經(jīng)網(wǎng)絡(luò)初始權(quán)值優(yōu)化技術(shù)在移動機(jī)器人學(xué)習(xí)中的應(yīng)用研究生姓名肖】主指導(dǎo)教師姓名圍壺鯉l教授!申請學(xué)付級別亟爭業(yè)名稱通信芻僮皇丞縫論文答辯u期2QQ生且墊目學(xué)位授予日期2QQ生Z目。中國海洋大學(xué)神經(jīng)網(wǎng)絡(luò)初始權(quán)值優(yōu)化技術(shù)在移動機(jī)器人學(xué)習(xí)中的應(yīng)用I呷l哪ntationofNeuraINet帕rk’s0pti帕I№ightsInitiaIizati∞T∞h∞lo科in酗ileRobotLearnlngAbstractBa

2、sedonmultiplayerfeed—forwardneuralnetworksusingBPalgorithm，consideringthefluctuationofinitialweightsandsamplesets，anoveloptimalweightsinitializationtechn0109yisproposedforthesakeofⅢatchingbetweeninitialwejghtsandsamplese

3、tsInthispaperseveralconclusionsaregained：Intheconditionofadoptingmultiplayerfeed—forwardneuralnetworkwithoneoutput(calledsin91eoutputmodelbelow)usingBPalgorithm，optimalweightsinitializationtechn0109ycangainbiggerpea_kval

4、uewithinseVeralti鵬swithgre8tpossibilityIntheconditionofadoptingneuralnetworkwithmultipleoutputs(cal】edmultipleoutputmodelbelow)，andinincreasingsampledimensionandneuralnetworkneurons，optimalweightsinitializationtechnology

5、cannotachieVebetterinitialweightsgroupwithinseveralfinitestepsAsaresult，optimalweightsinitialjzationtechn0109yissuitabletosin91eoutputmodelAnalysisontheprocessofoptimalweightsinitializationtechn0109yisachieVedandtheconcl

6、usionsmadeareasf01lows：TheF—normofweightsma七rixhaslittlefluc七uationwhenthevalueofoptimalf、lnctionchangesgreatlyAsto皿【atrixesofdifferentoptimalfunctionValues，aneWmatrix，whichisthebalancebetweentwoweightsmatrixwithdifferen

7、toptimalfunctionvalues，isconstructed。Analysisonthevalueofnewmatrix’sF—normshowsthatthedifferencebetweentwoweightsⅢatrixesisevidentOptimalweightsinitializationtechnolOgychoosesthebetterweightsgroupfarcurrentsamplesetsAsto

8、thepoorperformanceinachievingbigpeakvalueusingneuralnetworkswithmultipleoutput，aresearchismadeandtheconclusionsareasf0110ws：Tomultipleoutputmodel，asin91enetworkmustadapttoa11s繃巾1esets，andthereexistsweights’influenceinthe

9、trainingprocess：Howeverinsin91eoutputmodel，networkgroupstructureisadopted，whichconquerstheinteractionbetweencategoriesaswellasintheweights’changeprocessTothesin91eoutputmodel，thenumberofvariablesisseveraltimesto叫l(wèi)tipleou

10、tputmodel，andthereisnoinfluencebetweenvariablesConsequently，theoptimalprocessbecomeseffOrtless：TothernultipleoLJtputIilodelbecauseoftheinfluenceinweights’changeprocess，theappearancepossibiljcyofDeakvalueofoptimalfunction