基于神经网络的教学质量评价模型研究

西安科技大学硕士学位论文基于神经网络的教学质量评价模型研究姓名：杨新佳申请学位级别：硕士专业：应用数学指导教师：龙熙华2011BPINIBP(l)(2)(3)BPBP(4)BPBPSubjectTheModelofResearchforTeachingQualityEvaluationBasedonNeuralNetworkSpecialtyAppliedMathematicsNameYangXinjia(Signature)InstructorLongXihua(Signature)ABSTRACTTheresearchofeducationqualityhasmoreandmoreattentionbypeoplewiththeunceasingdeepeninganddevelopingofcurrenthighereducationteachingreforms.Itisfocusedontheimprovementofteachingqualityandteachingevaluationiskeymeasuretoimprovethequalityofeducationandteaching.Therefore,theestablishingandperfectingofteachingqualityevaluationsystemforeducationmanagementhastheextremelyvitalsignificance.Inthispaper,theteachingqualityevaluationmodelisconstructed,whichbasedonBPneuralnetworkandINIneuralnetworkandcombinescharacteristicsofneuralnetworkwiththeresearchofteachingqualityevaluationactualityandcharacteristicandtheBPalgorithmandalgebraalgorithmwhichusedinthemodelhasthetheoreticalanalysisandthecontrastoftrainingandprovidesafeasiblesolutionresultsfortheteachingqualityevaluationmodel.Concretelythispaperdealwiththeresearchintheaspectasfollows:(l)Themainproblemsanddifficultiesarediscussedandanalyzedintheconstructingperfectsystemofteachingqualityevaluation,andthispapergivesthesolutionsandanalyzestheadvantagesanddisadvantagesofthepastteachingqualityevaluationmethod.Theteachingqualityevaluationmodelbasedonneuralnetworkisconstructedonthegroundthatsummarizingtheexistinglivermethodsofteachingqualityevaluationmodelandagainstthelimitationsoftheexistingevaluationmethod.(2)Firstly,thepaperintroducesthebasicprincipleofprincipalcomponentanalysisandfindsthatextractingprincipalcomponentfromthecorrelationcoefficientmatrixoftheindex,whichcan'treflectthedifferenceinformationofvariationdegreeofindex.Andthenitgivesthetheoreticalproofinformationofimprovedprincipalcomponentanalysiswhichcansolvethisproblem.Lastly,thepaperreducesthevaluationindexusingtheimprovedprincipalcomponentanalysisforthepromiseofretainingtheoriginalinformation.Itavoidsthatthenetworkmodelistoocomplicatedtoimpacttheresultofprediction.(3)ThepaperintroducestheknowledgeofneuralnetworkdomaincomprehensivelyandresearchesthemodelconstructingandtrainingofBPneuralnetworkbasedonthedimensionreductionsystematically,andthenthenetworkstructureandlearningalgorithmofteachingqualityevaluationmodelbasedontheBPneuralnetworkaredeterminedandthetrainingresultsareanalyzed.(4)AccordingtotheproblemsofBPalgorithm,thispaperusesalgebraalgorithmandgivesthebasictheoryandadvantagesofthat,andthenitwasappliedtotheteachingqualityevaluationmodelbasedonneuralnetwork.Throughtheconcreteinstances,thepaperanalyzesthetrainingresultsforusingalgebraalgorithmandexplainsthevalidityandaccuracyoftheteachingqualitymodelbasedonalgebraalgorithm.KeywordsArtificialneuralnetworkTeachingqualityevaluationmodelPrincipalcomponentanalysisBPalgorithmAlgebraicalgorithmThesisApplicationResearch1111.11999[1][2][3][4]2[5]1.21.2.113(1)(2)[6][7]4(3)[8]ABCD10%20%35%20%15%()ABCDXdXcXbXa,,,KdKcKbKa,,,0,70=≥KdKa0,70=≥+KdKbKa[9]Step1U{}nUUUU,,,21L={}isiiiUUUU,,,21L=ni,,2,1L=),,,(21isiiiaaaaL=11=∑=sjija0≥ijaiU),,,(21sAAAAL=11ii∑==nA0≥iAStep2VV{}mvvV,,1L=),,2,1(mjvjL={}mvvvV,,,21L={}1,2,3,4,5,6,7,8,9,10=Step3RiUiju),,2,1(mjvjL=),,2,1(mjrijL=iRiUiiiRaB*=BiTnBBBB),,,(21L=UTnBBBAARB),,,(21L==TnnBBBAAA),,,)(,,,(2121LL=BBVTM=15()(90)(8089)(7079)(6069)(60)()Nin5,4,3,2,1=i),,,,()1(54321NnNnNnNnNnR=(1.1)1n11n15141312,,,nnnn),,,,(115114113112111nnnnnnnnnn(1.2))5,4,3,2)(,,,,(54321=innnnnnnnnniiiiiiiiii(1.3))5,4,3,2,1,)(()(===jinnppiijij(1.4)P),,,,(54321xxxxx9080706050543215060708090xxxxxs++++=(1.5)1.2.261.31.3.1BP1.3.21234BPBP17BPMatlab5INIINIINIBP6822.11943McCullochPitts(MP)6060[10](1)1943McCullochPittsMP1949D.HebbHebb1957F.Rosenblat1960B.WidrowM.Hoff(2)1969M.MinskyPapertM.Minsky207029Boltzmann(3)1985Rumelhart(BP)M.MinskyA)B)C)D)[11]102.22.2.1(1)MPMcCullochPittsMP[12]2.12.1MPixiwiθiy)(iufiy))((1jiyfynjijiji≠-=∑=θϖ(2.1))(xf(ActivationFunction)MP(2)(Perception)2112.2MP()(HiddenUnit)2.32.3()()S2L[13]2.2.2(1)⎩⎨⎧≥==0,00,1)(puuufy(2.2)12⎩⎨⎧-≥===0,10,1)sgn()(puuuufy(2.3))sgn(⋅(2.2)(2.3)2.4(a)(b)2.4(2)yuuufy==)((2.4)2.5(3)y)11(21)(--+==uuufy(2.5)2.62.52.6(4)SigmoidalSigmoidalS213SigmoidalSigmoidalueufyλ-+==11)((2.6)uueeufyλλ--+-==11)((2.7)λSigmoidalλ(2.6)(2.7)(2.6)Sigmoidal(2.7)SigmoidalSigmoidal(BP)(5)()(RBF)22)(σueufy-==(2.8)σσσ2.2.3(1)(supervisedlearning){}iidp,Ni,,2,1L=ipid()ipid(unsupervisedlearning)(self-organizedlearning)()(2)()14[]Tnθ,,,,21L=[]TnxxxX1,,,,21-=L[]Tnθ,,,,)(21L=Xd)(kWΔ)()()1(kWkWkWΔ+=+η(2.9)η10ppηδδδA))(WJ)(kW)()()1(kWkWkWΔ+=+(2.10))(kWΔ[][])()1(kWJkWJp+(2.11))(kWΔ[])1(+kWJ[][][])()()()()()1(kWkgkWJkWkWJkWJTΔ+≈Δ+=+(2.12)[])()()(kWWkWJkg=∇=(2.13))(WJ)(kWW=)()(kgkWη-=Δ(2.14)η10ppη(2.12)2(2.11)B)δδSigmoidalueufyλ-+==11)((2.15)uueeufyλλ--+-==11)((2.16)215Sigmoidal(2.2),11XWWXxwuTTnjjj===∑+=[]Tnθ,,,,21L=[]TnxxxX1,,,,21-=LδW{}dX,22))((21)(21)(XWfdydWJT-=-=(2.17)XXWfydWJT)()()('-=∇(2.18))()(WJkW∇-=ΔηXXWfydkWT)()()('--=Δη(2.19)XXWfydkWkWT)()()()1('--=+η(2.20)δBPC)Window-HoffWindow-Hoff[14]W-H(Ada