基于并行結構的BP改進算法研究.pdf

1、山東師范大學碩士學位論文基于并行結構的BP改進算法研究姓名：郭彥偉申請學位級別：碩士專業(yè)：計算機軟件與理論指導教師：王洪國20090518山東師范人學碩十學位論文ABSTRAClTheresearchaboutartificialneuralnetworkhasgottenimportantdevelopmentinrecentyearsItinspiresmanyscientists’enthusiasmandinterestinth

2、efieldofcomputerscience，brainneuralscienceandartificialintelligenceArtificialneuralnetworkissimulationtotheinformation—processingmechanismofthebrainItisexpectedthatrealizethebrainfunctionbysimulatingthestructureandthinki

3、ngofbrainItstheorieshavebeenapplledinmanyfieldsThemultilayerperception，trainedbythebackpropagationalgorithm，iscurrentlythemostwidelyusedneuralnetworkItsolvesthequestionofhiddenlayerlearningrulethatutilizethemethodoferror

4、backpropagationTheessenceofbackpropagationnetworksisthatmakethechangeofweightsbecome1ittlebygradientdescentmethodandfinallyattaintheminimalerrorSinceitadoptsthesteepestdescentmethodinnonlinearprogramming，ithasthedrawback

5、sthateasyconvergetoalocalminimumpointoftheerrorfunction，anditmayconvergeveryslowlyThesystemsofparallelcomputernowcanbeclassifiedasSIMDandMIIⅧDAtthebeginningofthedevelopmentofparallelcomputer，thetypeofSIMDsystemsplayedani

6、mportantroleHowever，withthedevelopmentofmicro—processingchiptechnology，Single—processorperformancehasbeengreatlyimprovedsince90’S，andresultsinthedevelopmentoftheMIMDsystemsSIMDisnolongerpopularButbothofthetwohavethereadv

7、antageanddefault，SOwecandesignnewparallelstructurebycombiningtheadvantageofthetwo，SOastosolvesomespecialproblemsAbriefoverviewaboutneuralnetworksandareviewofapproximatealgorithmsforcombinatorialoptimizationsaregiven：weal

8、sointroducethelearningrulesofneuralnetworksveryclearlyThenthepaperintroducesthebasicprinciplesofBPneuralnetworkandhowtoderivethealgorithm，itsstrengthsanddrawbackAvarietyofeffectiveimprovementBPalgorithmsaregivenParallelt

9、heoriesarebeingwidespreadusedinmanyfieldsThiSpaperuseParallelstructuretooptimizeBPnetworkstructureAnewalgorithmnamedPBBD(ParallelBasedBPalgorithm)isgiven，itusesmanynetworkswhicharedifferentinlearningspeedtotrainbyparalle