Content uploaded by JC Ang
Author content
All content in this area was uploaded by JC Ang on Jun 11, 2015
Content may be subject to copyright.
Modeling of Two-dimensional Warranty Policy
using Artificial Neural Network (ANN) Approach
Hairudin A. Majid, Jun C. Ang, and Azurah A. Samah
Abstract— Modeling of two-dimensional warranty policy is an important but difficult task due to the uncertainty and instability of
data collection. Moreover, conventional numerical methods of modeling a two-dimensional warranty policy involves complex
distribution function and cost analysis. Therefore, this paper attempts to present an Artificial Intelligence (AI) technique, which is
the Artificial Neural Network (ANN) approach in order to improve the flexibility and effectiveness of the conventional method.
The proposed ANN is trained with historical data using multi-layer perceptron (MLP), feed forward back-propagation (BP)
learning algorithm. The Logarithmic (logsig) and Hyperbolic Tangent (tansig) sigmoid functions are chosen as transfer function.
Four popular training functions are adopted to obtain the best BP algorithm, that are, Levenberg-Marquardt (trainlm), Gradient
Descent (traingd), Gradient Descent with momentum (traingdm), and Gradient descent with momentum and adaptive learning
(traingdx) back propagation algorithm. This ANN model demonstrated a good statistical performance with the mean square error
(MSE) values in this four training function, especially traingd. Finally, the adopted sensitivity analysis has revealed that the
proposed model had successfully implemented.
Index Terms— Artificial Intelligence, Artificial Neural Network, Two-dimensional Warranty.
—————————— ——————————
1 INTRODUCTION
two‐dimensionalwarrantyiseitheranimpliedoran
expresscontractbetweenthemanufacturerandcon‐
sumer.Underthiscontract,manufacturersagreetopro‐
videasatisfactoryserviceeitherrepairorreplaceitems
thatfailduringthespecifiedperiodorusage(whichever
comesfirst).Nowadays,consumersalwayscomparethe
productperformance,characteristicsofcomparablemod‐
elsofcompetingbrandsbeforepurchaseaproduct.So,
warrantybecomesamajornewdirectioninmanufactur‐
ingindustrysinceitplaysanimportantroleinproviding
aguidelinetocustomers.
Intheautomobileindustry,accuratepredictionofop‐
timalwarrantyperiodandwarrantycostsisoftensought
bythemanufacturer.LeBlanc[1]mentionedthatitisdif‐
ficulttoquantifytherisksandrewardsofofferingawar‐
ranty.Itisbecausethewarrantyperiodistooshort,as
wellastoolongwhichmaybeunprofitableforthemanu‐
facturers[2].Averyshortwarrantyperiodwillinterfere
withsales,whileaverylongonewillleadtolossesfrom
compensationofconsumerclaims.Hence,applicationof
ArtificialIntelligence(AI)inwarrantymarketismuch
moreinterestingrequisitetoaffirmtherationalityand
accuracyofwarrantypolicyprediction.
Vastresearchefforthasbeendevotedtotheuseof
ANNasapracticalforecastingtool[3]. Accordingto
KhasheiandBijari[4],ANNisoneofthemostaccurate
andwidelyusedforecastingmodelsthathaveenjoyed
fruitfulapplicationsinforecastingsocial,economic,engi‐
neering,foreignexchangeandstockproblems.Apart
fromthat,theusedofhistoricaldatainANNforpredic‐
tionorforecastingisverypopularanditsefficiencyis
provenbymanyresearcherssuchas[5]‐[8].Forexample,
asurveythatwasconductedbyMarzietal.[5]hadused
twenty‐yeardatafromS&P500Europeanindexcallop‐
tionpricestoforecastthefinancialmarket,andXuand
Lim[8]hadusedrawandhistoricaldataintheirstudyto
forecastthenetflowofacarsharingsystem.
Inthispaper,wepresenttheapplicationofanANN
techniquetopredicttheminimumwarrantycostandop‐
timalinspectionintervalduringawarrantyperiod.This
paperstartswithsection1,whichintroducetwo‐
dimensionalwarrantyandresearchproblems.Thisisfol‐
lowedbySection2whichdescribestherelatedworkand
motivationofthisresearch.InSection3,theassumptions
fortheeasinessoftheimplementationoftheresearchand
resultwerepresented.Section4brieflydescribesthe
ANNapproach.Theframeworkisthoroughlydescribed
inSection5andisfollowedbySection6whichdiscussed
theeffectsofANNstructurestowardstheMSEvalues.
Thediscussionofthispaperendswiththeconclusion
whichispresentedinSection7.
————————————————
H.A.MajidiswithFacultyofComputerScienceandInformation
System,UniversitiTeknologiMalaysia,81310Skudai,Johor,Malaysia.
J.C.AngiswithFacultyofComputerScienceandInformationSystem,
UniversitiTeknologiMalaysia,81310Skudai,Johor,Malaysia.
A.A.SamahiswithFacultyofComputerScienceandInformation
System,UniversitiTeknologiMalaysia,81310Skudai,Johor,Malaysia.
A
© 2011 Journal of Computing Press, NY, USA, ISSN 2151-9617
http://sites.google.com/site/journalofcomputing/
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
48
2 RELATED WORK AND MOTIVATION
Acompletehistoricaldataispivotalduringthe
developmentofwarrantypolicy.However,itisalwaysan
exhaustinganddifficultjobfortheautomobileworkshop
tohaveaccurateandcompletehistoricaldata.Historical
warrantyclaimandservicedatausuallycontainpartial
informationasitmayberecordedincorrectlyanddueto
uncertaintyandinstabilityofdatacollection.Hence,
predictionofanaccuratewarrantypolicybecomesavery
difficulttask.ThisviewissupportedbyYang [9]who
statedthatwarrantymodelingiscomplicateddueto
warrantycensoringespeciallyfortwo‐dimensional
warranty.Thesevaguenessrealworldproblemsare
typicallytoocomplexforaformalmathematicalmodel
[10].
Foraconventionalmathematicalmodeling,thefirst
steptopredictawarrantypolicyistomodeltheitems’
failuresandthecostsofrectificationactionsoverthewar‐
rantyperiod[11].Numerousfailureprobabilitymodels
suchasWeibull,ExponentialandLog‐normalhavebeen
developedforautomobilewarrantyclaimsdata[12]‐[14].
Probabilitydistributionisamathematicalfunctionused
tomodelthefrequenciesandprobabilityofoccurrences
overatime.Thosemathematicalmodelinginvolvesever‐
alstages,soitmaytakealongtimetobetrulyproficient.
Thus,inthisresearch,anattemptismadetoapplyanAI
techniqueinwarrantyanalysistoreducetheuncertainty
problemsandspeedupthecomputingtime.
AIapproacheswerebroadlyadoptedinmanyareas
whereconventionalmathematicalmodelwerereplaced
withexpertsystemstoimproveflexibilityandeffective‐
nessofthecorrespondingsystem.However,inwarranty
research,theapplicationofAItechniqueisveryfewsuch
as[15]‐[22].AmongtheAIsubjectarea,softcomputing
techniquewasfoundtobethemostpopulartechnique
thathasbeenintegratedwithwarrantyarea[23]‐[27].The
applicationsofANNinwarrantydomainfromprevious
workarediscussedinvariousresearches[28]‐[30].Hrycej
andGrabert[28]usedfailureprobabilityasgeneralfunc‐
tional.Inparticular,theauthorusedmulti‐layerperceptron
asafunctionalapproximation.Theparametersofmulti‐
layerperceptronweretrainedwithhelpofminimumcross
entropyruleinforecastingthewarrantycostofalterna‐
tivewarrantyconditionscenarios.Inotherwork,Leeat
al.[29]haddesignedanearlyclaimwarningsystemusing
neuralnetworklearning.Precisely,thesystemprotects
bothmanufacturersandconsumersbygiving“prior
warning”aboutabnormalincreaseofclaimsrateatacer‐
tainpointbasedontrendandestimationbymonitoring
variousclaimdata.Leeetal.[30]hadalsosuggested
neuralnetworklearningmodelindeterminingearly
warninggradeofwarrantyclaimsdata,whichincludes
AnalyticHierarchyProcess(AHP)analysisandknow‐
ledgeofqualityexperts.Inyear2008,Leeetal.[23]had
proposedadifferenttoolwhichisappropriateformodel‐
ingatwo‐dimensionalwarrantyplan.
3 ASSUMPTION
Afewassumptionshavebeenmadetosimplifytheim‐
plementationoftheproposedalgorithm.First,theusage
conditionsareassumedtobestatisticallysimilarandthe
warrantyclaimswerereportedimmediately,withnode‐
lay.Second,theproposedANNapproachisconsideredto
besuitableforanymodelandmakeofautomobile.Third,
theinputandtargetedoutputdataoftheANNprocess
areassumedtobecompletelyknown.Fourth,although
thenumberofdatausedinthedevelopmentofANNis
small,itisassumedthatthedataissufficientenoughto
achievetheperformancegoalinthisresearch.
4 ARTIFICIAL NEURAL NETWORK (ANN): AN
INTRODUCTION
ANNorcommonlyneuralnetwork(NN)isanintercon‐
nectedgroupofartificialneuronsthatuseamathematical
orcomputationalmodelforinformationprocessingbased
onaconnectionistapproachtocomputation[31],[32].
AccordingtoPrincipleetal.[33],oneofthemostsignifi‐
cantstrengthofANNisitsabilitytolearnfromalimited
setofexamples.ANNhasbeensuccessfullyusedinsolv‐
ingcomplicatedproblemsindifferentdomainsuchas
patternrecognition,identification,classification,speech,
vision,andcontrolsystems[34].
AnANN,whichimitatesthehumanbraininproblem
solving,iscapableinmodelingthecomplexrelationship
betweeninputandoutputtofindpatternsindata.Typi‐
cally,anANNconsistsofasetofinterconnected
processingelementsornodescalledperceptrons.The
nodesareorganizedindifferentwaystoformanetwork
structurewhereeachANNiscomposedofacollectionof
perceptrongroupedinlayers.Eachperceptronisdesigned
tomimicitsbiologicalcounterpart,theneuronandto
acceptaweightedsetofinputandrespondwithanout‐
put[35].
AsophisticatedANNmayhaveseveralhiddenlayers,
feedbackloopsandtime‐delayelements,whichdesigned
tomakethenetworksaseffectiveaspossibleindiscrimi‐
natingrelevantfeaturesorpatterns[35].Themostwell
knownANNisfeed‐forwardANN.Afeed‐forwardANN
consistsofasetofnonlinearneuronsconnectedtogether,
inwhichtheinformationflowsintheforwarddirection
[31].Amongthefeed‐forwardnetwork,multilayerPer‐
ceptron(MLP)isthemostwidelyandcommonlyused
model‐freeestimators.AMLPconsistsofatleastthree
layers,whicharetheinput,hiddenandoutputlayer.The
inputandoutputlayercontainacollectionofneurons
representinginputandoutputvariables.
TherearethreelearningtypesofANNmodels,which
aresupervised,unsupervisedandreinforcementlearning.
Thenetworkmodifiestheweightbasedonasequenceof
trainingvectorwithanassociatedtargetoutputnode,
knownassupervisedtraining.Ontheotherhand,unsu‐
pervisedtrainingreferstoanetworkthatmodifiesweight
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
49
byassigningthemostsimilarinputvectorstoanoutput
unit.Andthethirdlearningtypeisreinforcementtrain‐
ing,whichliesbetweensupervisedandunsupervised
learning.Amongthevariousneuralnetworkmodels,
backpropagationisthebestgeneralpurposemodeland
isprobablythebestatgeneralization[36],[37].Theback
propagationistheclassicalalgorithmusedforlearning.It
isaniterativegradientdescentalgorithmwhichisde‐
signedtominimizethemeansquarederrorbetweenthe
desiredoutputandthegeneratedoutputforeachinput
pattern[38].Inthisresearch,focusisgiventothefeed‐
forwardandback‐propagationmodelwithmultilayer
perceptron.
5 ANN FRAMEWORK
Fourmainstageswereincludedintheframework.The
stagesaredatacollection,datadesignandpre‐processing,
back‐propagationnetworkdesignandnetworkimple‐
mentation.Figure1showsthemainflowofthisresearch.
5.1 Data Collection
Sevenhundredhistoricaldatasetswerecollectedinthis
studyfromaMalaysiaautomobilecompany,knownas
MalaysianTruckandBus(MTB).Thedatasetscomefrom
thesameautomobileproductknownasHICOMPerkasa.
Eachdatasetscomprisedoftheelementsofdatewhen
theclaimwasmade,dateclaimreceived,enginenumber,
drivingmileageandagewhentheclaimwasmade,pro‐
ductiondateandfailureordefectdata.Thehistoricaldata
isoffive‐yearsperiod,rangingfrom1998to2002.Precice‐
ly,thehistoricaldataconsistofinformationofa100ve‐
hicles.Fromthe100samples,onceamaintenanceservice
isdone,thevehiclestatusandinformationwillberecord‐
edasthehistoricaldata.
AccordingtoGeorgilakisetal.[39],thetaskofdeciding
whichoftheelementstobeselectedasinputvariableis
anarduoustask.Theselectedelementsmustcorrespond
toparameters,whichmeanthatitwilldirectlyorindirect‐
lyaffectthepredictionresult.Inthisresearch,fourmain
inputpatternswereproduced,whicharemileageandage
ofavehicleduringtheservicing,thedefectorfailurerate
forfortycomponentsofavehicle,andthedatarecorded
whichiseitherservicemaintenanceorrepairimmediate‐
ly.Thesehistoricaldataarepassedtothenextstage,the
datapre‐processing.
5.2 Data design and pre-processing
Datapre‐processingordatanormalizationhastobedone
beforeitcanbeusedintrainingthenetwork[40].Accord‐
ingtoBirbiretal.[41]andWang[42],normalizationof
datareferstoaprocessofscalingthenumbersinadata
settoimprovetheaccuracyofthesubsequentnumeric
computations.Theauthorsalsomentionedthatnormali‐
zationhelpsinshapingtheactivationfunctionduringa
trainingprocess.Basedonthisstatement,theelements
withhugedifferentiaamongthedatasuchasmileageand
ageintheinputandoptimalinspectionintervalandmin‐
imumcostinthetargetoutputarenormalizedinto[0,1]
bytheexpression:
(1)
where
xiisanobservationvalueofthefactori;
xmaxisthemaximumvalueofthefactori;
xministheminimumvalueofthefactori;
Xiisthenormalizedvalueofxi.
Thenormalizeddatawerethendividedinto70groups
(10each)andthetargetoutputsfromeachgroupare
computedusingmathematicalmodel.Finally,thedata
aredividedintothreepartsfortraining,validatingand
testingprocess.
5.3 Back Propagation (BP) Network Design
ThestructureofaBPNetworkdesignincludeselements
asillustratedinFigure2.Eachofthestructureelementsis
thoroughlydiscussedinthissection.
Fig. 2. Structure of Network Design
5.3.1 Network Architecture Determination
Inthestageofanetworkarchitecturedetermination,the
numberoflayersandthenumberofprocessingelements
perlayerareofimportantconsideration[41].Multi‐layer
backpropagationneuralnetworkcompriseofinput,
hiddenandoutputlayer.Itsarchitectureisillustratedin
figure3.Inthisresearch,inputvariablesaretheageand
mileageduringtheservicing,defectorfaultycomponents
Fig. 1. Main flow of this research
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
50
andinformationofmaintenancecheckingorimmediate
repair.Theminimumwarrantycostandoptimalinspec‐
tionintervalwereidentifiedtobetheoutputdatainthe
ANNprocess.
Fig. 3. Structure of back propagation ANN for two-dimensional war-
ranty
Inthisstage,thesetofdataweregroupedintotwo
setsi.e.inputandoutput.Thedatasetwerearrangedina
matrixformofx(input),andy(output)inacolumn.Fig‐
ure4showsthegenericformofinputandtargetoutput
design.Thedatawerekeyed‐inintoMs.Excelandsaveas
atextfilesothatitcanbeusedinMatlabtools.
(a) Inputdesign(b) Outputdesign
Fig. 4. Input and Output design
5.3.2 Hidden Neuron Number and Transfer Func-
tion Optimization
Thenumberofhiddenlayerandnodesineachofthe
hiddenlayeraffecttheperformanceofanANN.Ifthe
numberofhiddennodesistoosmall,itisnotenoughto
generalizetherulesoftrainingsample.Otherwise,the
ANNwilltakenoisydataintomemory.Todate,thereis
nospecificmethodtochoosetheoptimalnumberofhid‐
denlayerandthenumberofnodesinhiddenlayer[43].
Fauset[44]introducedaruletodeterminethenumberof
neuronnodesinhiddenlayerasillustratedinTable1.
Amongstthetransferfunctionsusedforhiddenand
outputlayerareLogarithmicSigmoidfunction(logsig)
andHyperbolicTangentSigmoidfunction(tansig).Both
Sigmoidfunctionsareoftenusedinhiddenlayerdueto
theirpowerfulnonlinearapproachcapability[45].Forthe
outputofactivationfunction,therangeoftansigis(‐1,1)
andlogsigis(0,1).Themathematicalequationsforboth
functionsare:
(2)
(3)
5.3.3 Training function optimization
Fourpopulartrainingfunctionsusedinthisstudyare
trainlm,traingd,traingdm,traingdx[46]‐[48].
Levenberg‐Marquardtbackpropagationalgorithm
(trainlm)isanetworktrainingfunctionthatupdates
weightandbiasvaluesaccordingtoLevenberg‐
Marquardtoptimization.Itwasdesignedtoapproach
second‐ordertrainingwithouthavingtocomputethe
Hessianmatrix.Trainlmisthefastestbackpropagation
algorithminMatlabtoolbox,andishighlyrecommended
asafirst‐choicesupervisedalgorithm,althoughitdoes
requiremorememorythanotheralgorithm.Thetraining
parametersfortrainlmareepochs,show,goal,time,
min_grad,max_fail,mu,mu_dec,mu_inc,mu_max,
mem_reduc.Thetrainingstatusisdisplayedforeveryshow
iterationofthealgorithm.Thetrainingwillterminatesin
fourconditionsi.e.ifthenumberofiterationsexceeds
epochs,iftheperformancefunctiondropsbelowgoal,ifthe
trainingtimelongerthantimeseconds,orifthemagni‐
tudeofthegradientislessthanmin_grad.Theparameter
muistheinitialvalueforμ.Iftheperformancefunctionis
reducedbyastep,themuvalueismultipliedbymu_dec.
Ontheotherhand,iftheperformanceisincreasedbya
step,thenthemuvalueismultipliedbymu_inc.However,
ifthevalueofmuisbiggerthanmu_max,thealgorithmis
terminated.
TABLE 1
NUMBER OF NEURONS OF HIDDEN LAYER
Formula Proposedby
h=nTangandFishwick
h=n/2Kang
h=2nWong
h=2n+1Lippmann
n=numberofneuronsintheinputlayer
h=numberofneuronsinthehiddenlayer
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
51
Gradientdescentbackpropagationalgorithm(traingd)
isanetworktrainingfunctionthatupdatesweightand
biasvaluesinthedirectionofnegativegradientdescent
ofperformancefunction.Thetrainingparametersfor
traingdareepochs,show,goal,time,min_grad,max_fail,and
lr.Theparameter,learningrate(lr)inTraingdalgorithm,
ismultipliedwiththenegativeofgradienttodetermine
thechangesofweightandbiases.Thelargerthelearning
rate,thebiggerthestep.Ifthevalueoflearningrateistoo
large,thealgorithmbecomesunstable.Otherwise,ifthe
valueoflearningrateistoosmall,thealgorithmtakesa
longtimetoconverge.
Gradientdescentwithmomentumbackpropagation
algorithm(traingdm)allowsthenetworktorespondnot
onlytothelocalgradient,butalsotorecenttrendsinthe
errorsurface.Thetrainingparametersfortraingdmare
epochs,show,goal,time,min_grad,max_fail,lrandmc.The
momentum,mcactslikealowpassfilter,whichallowsthe
networktoignoresmallfeaturesinerrorsurface.Anet‐
workwithoutmomentumcangetstuckinashallowlocal
minimum,whileanetworkwithamomentum,canslide
throughsuchminimum.Theinteractionoflearningrate
andmomentumleadstoanacceleratedlearning[49].
Gradientdescentwithmomentumandadaptivelearn‐
ingbackpropagationalgorithm(traingdx)isanetwork
trainingfunctionthatupdatesweightandbiasesvalues
accordingtogradientdescentmomentumandanadap‐
tivelearningrate.Thetrainingparametersfortraingdxare
epochs,show,goal,time,min_grad,max_failmax_perf_inc,lr,
lr_inc,lr_dec,andmc.Adaptivelearningrate,lr_incand
lr_decattemptstokeepthelearningstepsizeaslargeas
possiblewhilekeepinglearningstable.Ifthenewerror
exceedsthepreviouserrorbymorethanapredefined
ratioi.e..max_perf_incinthenetworkwithmomentum,
thenthenewweightsandbiasesareabandoned.
5.4 ANN Implementation
Thisimplementationstageinvolvesthreeprocesseswhich
arenetworktraining,validatingandtesting.Figure5
showstheflowtodeterminethebestANNmodel.
Fig. 5. The flow to find the best ANN model
5.4.1 ANN Model Development
TheANNdevelopmentprocessstartswithnetworktrain‐
ing,followedbynetworkvalidating.Thetrainingprocess
mayconsumealotoftime.Atthebeginning,thenetwork
modelsaretrainedwithasetofinputandtargetoutput
data.Theparameterssettinginthetrainingfunctionare
changedorderlytofindoutthebestANNmodel.The
networkadjuststheweightcoefficients,whichusually
beginwitharandomset,sothenextiterationwillpro‐
duceaclosermatchbetweenthetargetoutputandactual
outputofANN.Thetrainingmethodtriestominimize
thecurrenterrorsfromallprocessingelements,andthe
globalerroriscalculatedbyaperformanceindex.
Iftheperformanceindexintrainingprocessachieves
thetargetedgoalwhichis0.01thenthevalidatingprocess
willbecontinuedbyanothersetofinputdataandtarget
outputdata.Similartothetrainingprocess,aglobalerror
iscalculatedbyperformanceindex.Thetargetedgoalset
inthevalidationstageisintherangeof0to0.07tobe
accepted.Thenetworkmodelhadundergonetraining
andvalidatingprocesscalleddevelopedsystem.Thede‐
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
52
velopedsystemwithitsweightwillpassedtothenext
processwhichistestingprocess.
5.4.2 Model Performance Evaluation
Thenetworkperformancewasquantifiedbycalculating
MeanSquareError(MSE)forthedifferencebetweenthe
expectedandtargetedoutputdataset.Haganetal.[50]
highlightedthattheperformanceevaluationoftheback
propagationalgorithmformulti‐layerANNisMSE.The
learningalgorithmadjuststhenetworkparametersin
ordertominimizetheMSE.TheexpressionofMSEis:
(4)
where
zk isoutputpredicted,
yk isactualtargetoutputand
istheerrorofithoutputnumber1,2,…,n.
5.4.3 Network Testing
Tes ti ngprocesscanonlystartoffifthetrainingandva‐
lidatingprocesseshadbeencarriedthroughbyachieving
theirgoal.Atthisstep,theneuralnetworkwiththesmal‐
lestMSEvalueisdefinedasthebestneuralnetwork
model.Inthisresearchaninterval(0–0.03)forMSEvalue
oftheoutputaredefinedasthesmallestvalue.
6 EXPERIMENTAL RESULT
AnoptimizedANNstructureisusedtoillustratetheper‐
formanceoftheproposedmodel.Sincetheneuralnet‐
workisanonlinearprocedureandthenetworkparame‐
terswillaffecteachotherthentheadjustmentofeachpa‐
rametertooptimizethewholenetworkisnotaneasytask
[51].ThissectiondiscussestheoptimizedANNstructure
whichproducedtheminimumMSEvalueineachtraining
function.Ananalysisispresentedtoverifytheaccuracy
oftheresult.
Table2and3displaythestructureofthebestANN
modelanditscorrespondingparameterssettingineach
trainingfunctionwhichachievedthesmallestMSEvalue
intestingprocess.
TABLE 2
STRUCTURE OF THE BEST ANN MODEL IN EACH TRAINING
FUNCTION
ItemValue/Character
TraingdTrai ng dmTraingdxTrainlm
Number
ofhidden
layer
ThreeThreeThreeTwo
Input
node
43
nodes43nodes43nodes43
nodes
Output2nodes2nodes2nodes2nodes
nodelogsiglogsiglogsiglogsig
Hidden
node
1x86
nodes
tansig
1x86
nodes
logsig
1x22
nodes
tansig
1x86
nodes
tansig
1x43
nodes
logsig
1x43
nodes
tansig
1x43
nodes
tansig
1x22
nodes
logsig
1x22
nodes
tansig
1x22
nodes
tansig
1x22
nodes
logsig
MSEvalue(goal=0.01)
Training0.01000.01000.01000.0096
Validating0.04240.03690.04040.0580
Tes ti ng0.00980.01720.01940.0113
TheBestANNmodelamongthetrainingfunction
TABLE 3
THE PARAMETERS USED IN EACH TRAINING FUNCTION
ParameterValu e
TraingdTrai ng dmTraingdxTrainlm
epochs
300000
(goal
met=
206128)
300000
(goalmet
=106643)
300000
(goal
met=
174933)
300000
(goal
met=
811)
goal0.010.010.010.01
learningrate
(lr)0.30.60.9‐‐
learningdec
(lr_dec)‐‐ ‐‐ 0.7‐‐
learninginc
(lr_inc)‐‐ ‐‐ 1.05‐‐
max_fail5(de‐
fault)
5(de‐
fault)
5(de‐
fault)
5(de‐
fault)
max_perf_inc‐‐ ‐‐ 1.04‐‐
momentum
constant(mc)‐‐ 0.10.9‐‐
InitialMu
(mu)‐‐ ‐‐ ‐‐ 0.006
Mudecrease
factor
(mu_dec)
‐‐
‐‐ ‐‐ 0.1
Muincrease
factor
(mu_inc)
‐‐
‐‐ ‐‐ 10
Maximum
Mu
(mu_max)
‐‐
‐‐ ‐‐ 1.0000e‐
010
min_grad1.0000e‐
010
1.0000e‐
010
1.0000e‐
010
1.0000e‐
010
show100100100100
timeInfinityInfinityInfinityInfinity
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
53
ThebestparameterssettingtoobtainthebestANN
model
6.1 Accuracy Analysis upon Experimental Result
Thederivedoutputandexpectedoutputshouldundergo
de‐normalizationbeforeanalysisprocess.Theoutputsare
de‐normalizedbytheexpression:
where
X isthede‐normalizedvalueofx
xisthederivedoutputorexpectedoutput
xmaxisthemaximumvaluederivedinnormalization
xministheminimumvaluederivedinnormalization
Thede‐normalizedderivedoutputswhichweregen‐
eratedbythebestANNmodelarecomparedwiththede‐
normalizedexpectedoutputintermsofaccuracyusing
thefollowingequation:
wherezfisthederivedoutputandzeistheexpectedout‐
put.
Table4showsthede‐normalizedoutputandtheaccuracy
oftheproposedANNmodelinderivingtheoptimum
warrantycostandinspectioninterval.
TABLE 4
THE ACCURACY OF THE PROPOSED MODEL
Expected
output
Derived
output
Accuracy
(%)
Inspection
interval
(Year)
0.58330.672384.74
1.41671.268889.56
1.08331.041996.17
War ra nty
Cost
(RM)
178.09169.1895.00
75.6489.1082.21
101.50107.9893.62
Fromthisaccuracyanalysis,itisrevealedthatanaver‐
ageof90percentofaccuracyisachievedbyusingthis
proposedmodel.Itcanbesummarizedthattheproposed
ANNmodelissuccessfullyappliedinderivingtheopti‐
mumwarrantycostandoptimuminspectioninterval.
6.2 Sensitivity Analysis: Statistic T-test
T‐testisahypothesistesttoinvestigatethesignificanceof
twosamplesfromanormallydistributedpopulation.The
T‐testisprobablythebestknowntechniqueandthemost
frequentlyusedstatisticaldataanalysismethodforhypo‐
thesistesting[52]‐[53].
Inthisstudy,T‐testisconductedtoassesstheaccuracy
oftheresultswhichwereobtainedbytheproposedANN
model.Thenullhypothesisusedinthiscasestudyis:
H0: Thereisnosignificancedifferencebetween
thederivedoutput(x1)andtheexpectedout‐
put(x2),thatisx1=x2.
H1: Thereisasignificancedifferencebetweenthe
derivedoutput(x1)andtheexpectedoutput
(x2),thatiseitherx1≠x2,x1<x2,orx1>x2.
Table5and6showtheT‐testprocessfromsteps2to6for
inspectionintervalandwarrantycost,respectively.
TABLE 5
T-TEST FOR THE INSPECTION INTERVAL
Derivedoutput
(x1)
Expectedoutput
(x2)
Replicate10.67230.5833
Replicate21.26881.4167
Replicate31.04191.0833
Σx2.98303.0833
Observation(n)33
Mean( )0.99431.0278
Σd2=Σx2‐((Σ
x)2/n)0.18130.3519
Variance,σ2=Σ
d2/(n‐1)0.09070.1760
pooledstan‐
darddeviation,
σ
0.3651
T‐value0.1122
TABLE 6
T-TEST FOR WARRANTY COST
Derivedoutput
(x1)
Expectedoutput
(x2)
Replicate1169.18178.09
Replicate289.1075.64
Replicate3107.98101.50
Σx366.26355.23
Observation(n)33
Mean( )122.0867118.4100
Σd2=Σx2‐((Σ
x)2/n)3504.90035676.9234
Variance,σ
2=
Σd2/(n‐1)1752.45012838.4617
pooledstan‐
darddeviation,47.9109
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
54
σ
T‐value0.0940
Inthisstudy,95percentofconfidenceintervalisadopted.
Thesignificancelevelandcriticalregionarestatedasbe‐
low.
Significancelevel,α=0.05
CriticalRegion:Z<‐1.96orZ>1.96
BasedonthecalculationsinTable5and6,bothT‐value
forinspectionintervalandwarrantycostlieoutsidethe
criticalregion.Hence,thenullhypothesis,H0isaccepted
at5percentsignificancelevel.Itcanbeconcludedthat
thereisnosignificancedifferencebetweenthederived
outputandtheexpectedoutput.
7 CONCLUSIONS
Thispaperhaspresentedafasterandintelligentwayto
predictminimumwarrantycostandoptimalinspection
intervalduringawarrantyperiodbyusingartificialneur‐
alnetworks.Differentnetworkstructuresweretrained
andvalidatedwithanalyticalresultofmathematical
model.Themodelwasthentestedwithaseriesofhistori‐
caldata.Itwasfoundthatthemostefficientalgorithmfor
modelingthetwo‐dimensionalwarrantypolicyisback‐
propagationlearningalgorithmwithGradientDescent.In
thisresearch,althoughtheamountofexperimentaldata
islimited,significantresultprovesthattheproposedal‐
gorithmiscapabletopredictthetwo‐dimensionalwar‐
rantypolicy.Forfurtherresearch,itisrecommendedthat
otherAItechniquesisusedinmodelingthetwo‐
dimensionalwarrantypolicyinordertoreducethecom‐
plexityandtimeconsumingofconventionalmathemati‐
calmodel.
ACKNOWLEDGEMENT
TheauthorshonorablyappreciateMinistryofScience,
TechnologyandInnovation(MOSTI)forthefundingofE‐
SciencegrantandResearchManagementCenter(RMC),
UniversitiTeknologiMalaysia(UTM)forthesupportin
makingthisprojectasuccess.
REFERENCES
[1] LeBlancB.,“A n a l y s i s ofdecisionsinvolvedinofferingaprod‐
uctwarranty,”AnnualReliabilityandMaintainabilitySymposium,
2008.
[2] ChukovaS.S.,DimltrovB.N.,andRykovV. V. , “Warranty
Analysis(Review),”JournalofMathematicalSciences,vol67,no.
6,1993,34863508,doi:10.1007/BF01096273.
[3] Weigend,A.S.,D.E.Rumelhart,andB.A.Huberman,“Genera‐
lizationbyWei ght‐EliminationwithApplicationtoForecast‐
ing,”AdvancesinNeuralInformationProcessingSystems3
(NIPS*90),1991.
[4] KhasheiM.andBijariM.,“A n artificialneuralnetwork(p,d,q)
modelfortimeseriesforecasting,”ExpertSystemswithApplica‐
tions37(2010),pp.479–489,2009.
[5] MarziH.andMarziE.,“UseofNeuralNetworksinForecasting
FinancialMarket,”IEEEConferenceonSoftComputinginIndus‐
trialApplication,SMCiaʹ08,pp.240‐245,2008.
[6] Tay lo rJ.W.andBuizzaR.,“NeuralNetworkLoadForecasting
WithWe atherEnsemblePredictions,”IEEETran sac tio nOnPow‐
erSystems,vol.17,no.3,2002.
[7] Wan gQ.,YuB.andZhuJ.,“ExtractRulesfromSoftwareQuali‐
tyPredictionModelBasedonNeuralNetwork,”Proceedingof
the16thIEEEInternationalConferenceonToolswithArtificialIntel‐
ligence,ICTAI1004,2004.
[8] XuJ.‐X.andLimJ.S.,“Anewevolutionaryneuralnetworkfor
forecastingnetflowofacarsharingsystem,”IEEECongresson
EvolutionaryComputation,CEC2007,pp.1670‐1676,2007.
[9] YangG.andZaghatiZ.,“Two‐DimensionalReliabilityModel‐
ingFromWarrantyData,”IEEEProceedingAnnualReliabilityand
MaintainabilitySymposium,pp.272‐278,2002.
[10] Kastner,J.K.,“Areviewofexpertsystems,”EuropeanJournalof
OperationalResearch,vol.18,pp.285‐292,1984.
[11] MurthyD.N.P.,andDjamaludinI.,“Newproductwarranty:A
literaturereview,”InternationalJournalofProductionEconomics
79(2002),pp.231–260,2002.
[12] Chattopadhyay,G.andA.Rahman,“Developmentoflifetime
warrantypoliciesandmodelsforestimatingcosts,”Reliability
Engineering&SystemSafety,vol93,no.4,pp.522‐529,2008.
[13] MajeskeK.D.,“Amixturemodelforautomobilewarrantyda‐
ta,”ProceedingsofReliabilityEngineeringandSystemSafety,pp.
71‐77,2003.
[14] AskarK.,DoughertyM.,andRocheT.,“Ag e n t rasedsystem
thatsupportreliabilitytransportengineering,”Proceedingsofthe
InternationalConferenceonApplicationsofAdvancedTech nol og i es
inTransportationEngineeringinBeijing,2004.
[15] DeepR.,etal.“Bit‐MappingClassifierExpertSystemInWar ‐
rantySelection,”ProceedingsoftheIEEEConferenceonNational
AerospaceandElectronicsinDayton,USA.
[16] Derr,J.H.andR.J.Louch,“Ad v a n c e d methodologyforproject‐
ingfieldrepairratesandmaintenancecostsforvehicleelectron‐
icsystems,”SAE(SocietyofAutomotiveEngineers)Tran sa c ti ons ,
1991.100(Sect2),pp.1‐11,1991.
[17] Hyman,W.A.(1989).“Legalliabilityinthedevelopmentand
useofmedicalexpertsystems.”JournalofClinicalEngineering,
14(2):p.157‐163.
[18] KalerJr,G.M.“Expertsystempredictsservice,”HPACHeating,
Piping,AirConditioning,vol.60,no.11,pp.99‐101,1988.
[19] KasraviK.,“Improvingtheengineeringprocesseswithtext
mining.”ProceedingsoftheConferenceonASMEDesignEngineer‐
ingTechnicalinSaltLakeCity,UT,2004.
[20] Lee,S.andL.M.Chang,“Digitalimageprocessingmethodsfor
assessingbridgepaintingrustdefectsandtheirlimitations,”
Proceedingsofthe2005ASCEInternationalConferenceonCompu‐
tinginCivilEngineeringinCancun,2005.
[21] Lin,P. C . , J.Wan g,andS.S.Chin,“Dynamicoptimizationof
price,warrantylengthandproductionrate,”InternationalJour‐
nalofSystemsScience,vol.40,no.4,pp.411‐420,2009.
[22] LeeS.H.,J.H.Lee,etal.,“AFuzzyReasoningModelofTwo ‐
dimensionalWarrantySystem,”7thInternationalConferenceon
AdvancedLanguageProcessingandWebInformationTe ch n ol ogy ,
Liaoning,PEOPLESRCHINA,IEEEComputerSoc,2008b.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
55
[23] LeeS.H.,LeeD.S.,etal.,“AFuzzyLogic‐ BasedApproachto
Two‐DimensionalWarrantySystem,”4thInternationalConfe‐
renceonIntelligentComputing,Shanghai,PEOPLESRCHINA,
Springer‐VerlagBerlin,2008c.
[24] Vujosevi eM.,Makajie‐NikolieD.,StrakM.(2004),“FuzzyPetri
netbasedreasoningforthediagnosisofbuscondition,”Seventh
SeminaronNeuralNetworkApplicationsinElectricalEngineering‐
Proceedings,NEUREL2004,Belgrade,2004.
[25] ZhouG.,CaoZ.,MengZ.,andXuQ.,“AGAbasedapproachon
arepairlogisticsnetworkdesignwithM/M/smodel,”Proceed‐
ingsInternationalConferenceonComputationalIntelligenceandSe‐
curity,CIS2008,Suzhou2008.
[26] HotzE.andNakhaeizadehG.,PetzscheB.andSpiegelberger
H.,“WAPS,aDataMiningSupportEnvironmentforthePlan‐
ningofWarrantyandGoodwillCostsintheAutomobileIndus‐
try.”ProceedingsofthefifthACMSIGKDDinternationalconference
onKnowledgediscoveryanddatamining,SanDiego,California,
UnitedStates,pp.417–419,1999b.
[27] Hrycej,T.andM.Grabert,“WarrantyCostForecastBasedon
CarFailureData.”ProceedingsofInternationalJointConferenceon
NeuralNetworks,Orlando,Florida,USA,pp.108‐113,2007.
[28] LeeS.H.,SeoS.C.,etal.,“AStudyonWarning/DetectionDe‐
greeofWar ranty ClaimsDataUsingNeuralNetworkLearn‐
ing.”SixthInternationalConferenceonAdvancedLanguage
ProcessingandWebInformationTe ch n ol ogy (ALPIT),2007.
[29] LeeS.etal.,“Ondeterminationofearlywarninggradebased
onAHPanalysisinwarrantydatabase,”LectureNotesinCom‐
puterScience(includingsubseriesLectureNotesinArtificialIntelli‐
genceandLectureNotesinBioinformatics).2008,Shanghai,pp.84‐
89,2008a.
[30] DomingoM.,AgellN.andParraX.,“Connectionisttechniques
toapproachsustainabilitymodeling,”RevistaInternacionalDe
Tecnologia,SostenibilidadyHumanismo,diciembre2006,no.1,pp.
61‐73,2006.
[31] ZabiriH.andMazukiN.,“ABlack‐BoxApproachinModeling
Val ve Stiction,”InternationalJournalofMathematical,Physicaland
EngineeringSciences4:12010,2010.
[32] Principe,J.,NeuralNetworksandAdaptiveSystems,JohnWiley
andSons:NewYor k,NY,1999.
[33] SozenA,ArcakliogluE.,“SolarpotentialinTurkey,” Applied
Energy,80(1),pp.35–45,2004.
[34] OladokunV.O.,AdebanjoA.T.andCharles‐OwabaO.E.,
“PredictingStudents’AcademicPerformanceusingArtificial
NeuralNetwork:ACaseStudyofanEngineeringCourse,”The
PacificJournalofScienceandTech nol og y ,vol.9,no.1,2008.
[35] LawrenceJ.,IntroductiontoNeuralNetwork:Design,Theoryand
Application,6thed.NevadaCity.CA:CaliforniaScientificSoft‐
ware,1994.
[36] MitchellT.,M.MachineLearning,1stedition,NewYor k:
McGraw‐HillScience/Engineering/Math,1997.
[37] RumelhartD.E.andMcClellandJ.L.,Paralleldistributed
processing:explorationsinthemicrostructureofcognition,Cam‐
bridge,Mass:MITPress,1986.
[38] GeorgilakisP. S.,HatziargyriouN.D.,DoulamisA.D.,Doula‐
misN.D.andKolliasS.D.,“ANeuralNetworkFrameworkfor
PredictingTra nsfor merCoreLosses,”Proceedingsofthe21st
1999IEEEInternationalConferenceonPowerIndustryComputer
Applications,PICAʹ99,pp.301‐308,1999.
[39] NorHaizanMohamedRadzi,HabibollahHaron,andTuan
IrdawatiTuanJohari,“LotSizingUsingNeuralNetworkAp‐
proach,”Proceedingofthe2ndIMT‐GTRegionalConferenceonMa‐
thematics,StatisticsandApplications,UniversitiSainsMalaysia,
Penang,pp.13‐15,June2006.
[40] BirbirY., NogayH.S.andTop uzV., “EstimationofTotal Har‐
monicDistortioninShortChordedInductionMotorsUsingAr‐
tificialNeuralNetwork,”Proceedingsofthe6thWSEASInterna‐
tionalConferenceonApplicationsofElectricalEngineering,Istan‐
bul,Turkey,pp.206‐210,2007.
[41] Wang,S.,NeuralNetworkApproachtoGeneratingtheLearning
Curve,INFOR.31(3),pp.136–150,1993.
[42] GaoM.,SunF.,ZhouS.,ShiY.,ZhaoY.andWangN.,“Perfor‐
mancepredictionofwetcoolingtowerusingartificialneural
networkundercross‐windconditions,”InternationalJournalof
ThermalSciences48(2009),pp.583‐589,2009.
[43] FausettL.V.,FundamentalsofNeuralNetwork:Architecture,Algo‐
rithms,andApplications,N.J.:Prentice‐Hall,1994.
[44] ZhangY., LiW.,ZengG.M.,TangL.,FengC.L.,HuangD.L.
andLiY.P. , “NovelNeuralNetwork‐BasedPredictionModel
forQuantifyingHydroquinoneinCompostwithBiosensor
Measurements,”EnvironmentalEngineeringScience,vol.26,no.
6,pp.1063‐1070,2009.
[45] MenH.,LiX.,WangJ.,andGaoJ.,“AppliesofNeuralNetwork
toidentifygasesbasedonelectronicnose,”IEEEInternational
ConferenceonControlandAutomationFrC4‐2Guangzhou,CHINA,
2007.
[46] SirajFadzilah,Yus of fNoorainiandKeeL.C.,“Emotionclassifi‐
cationusingneuralnetwork,”InternationalConferenceonCompu‐
ting&Informatics,ICOCIʹ06,pp.1‐7,6‐8June2006.
[47] ZhouM.,ZhangS.,WenJ.andWangX.,“ResearchonCVT
FaultDiagnosisSystemBasedonArtificialNeuralNetwork,”
IEEEVehi cl ePowerandPropulsionConference(VPPC),Harbin,
China,2008.
[48] KeJ.,LiuX.,andWan gG.,“TheoreticalandEmpiricalAnalysis
oftheLearningRateandMomentumFactorinNeuralNetwork
ModelingforStockPrediction,”AdvancesinComputationandIn‐
telligence.LectureNotesinComputerScience,vol.5370/2008,pp.
697‐706,2008,doi:10.1007/978‐3‐540‐92137‐0_76.
[49] HaganM.T., DemuthH.B.andBealeM.H.,NeuralNetwork
Design,PWSPublishingCompany,1996.
[50] LeeT.L.Neuralnetworkpredictionofastormsurge.OceanEngi‐
neering33,pp.483–494,2006.
[51] Marryanna,L.,SitiAisahS.andSaifulIskandarK.,“Water
qualityresponsetoclearfellingtreesforforestplantationestab‐
lishmentatBukitTerek F.R.,Selangor,”JournalofPhysical
Science,18(1),pp.33‐45,2007.
[52] NeideenT.andBraselK.,“UnderstandingStatisticalTests,”
JournalofSurgicalEducation,pp.93‐96,2007.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
56
Hairudin Abdul Majid has received Diploma and Bachelor of
Science in Computer Science-majoring Industrial Computing from
Universiti Teknologi Malaysia in 1993 and 1995 respectively. In
1998, he obtained his M.Sc. in Operational Research and Applied
Statistic from University of Salford, UK. His PhD thesis in Warranty
and Maintenance(Submited). Currently, he is a lecturer in Faculty of
Computer Science and Information System, Universiti Teknologi
Malaysia. His research interests focused on Image Processing, Op-
erations Management and Warranty and Maintenance. Mr. Hairudin
received Excellent Service Award by Universiti Teknologi Malaysia in
2004 and Excellent Staff Award by ISS Service in Manchester UK in
2006. Mr. Hairudin is the author of about 19 papers, 1 book chapter
entitled ‘Recent Operations Research Modelling and Applications
(Warranty Modelling)’ (UTM, 2009) and 1 text book entitled ‘Permo-
delan Simulasi’ (UTM, 2000). He has been a member of UK Opera-
tional Research Society and an active member of Operations and
Business Intelligence (OBI) Research Group.
Ang Jun Chin obtained her M.Sc. in Computer Science from Un-
iversiti Teknologi Malaysia in 2011 and Bachelor of Computer
Science in 2009. She is currently working as a system analyst in
Singapore.
Azurah A Samah has received the Diploma and Bachelor from Un-
iversiti Teknologi Malaysia in 1991and 1993 respectively. In 1996,
she obtained her M.Sc. from the University of Southampton, UK and
recently in 2010, she received her Ph.D from Salford university, UK.
Currently she is a lecturer in Faculty of Computer Science and In-
formation System, Universiti Teknologi Malaysia. Her research inter-
ests encompass Image Processing, Soft Computing Techniques and
Operational and Simulation Modeling.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 12, DECEMBER 2011, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG
57