Feasibility Classification of New Design Points Using Support Vector Machine Trained by Reduced Dataset

Abstract

Inthispaper,weproposetouseasupportvectormachine(SVM)fortheclassificationofdesigndata.AlthoughtheSVMisaverypopulartechniqueindatamining,itisrarelyappliedtoanindustrialdesignprocessthatmayrequireinformationregardingthefeasibilityofthedesignpointofinterest.Tocheckthefeasibility,thedesignermustconductexperimentsorcomputersimulations,whichmayincurconsiderablecost.Therefore,theSVMcanbeaneffectivetoolforpredictingfeasibleandinfeasibleregionsbecauseitonlyusesthecumulativedesigndata.Inthispaper,weusedtheSVMtoclassifysampledatasetsdrawnfrommathematicaltestproblemsandfromanair-conditionerpipedesignexample.OurresultsindicatethattheSVMiscapableofveryaccuratelyidentifyingfeasibleandinfeasibleregionsinthedesignspace.Further, wewereabletoreducethetrainingtimeoftheSVMbyusingthek-meansclusteringalgorithmtoreducetheamountoftrainingdata,takingadvantageofthepowerfulgeneralizationabilitiesoftheSVM.Consequently,weconcludethattheSVM canbeaneffectivetooltoassessfeasibilityatcertaindesignpoints,avoidingsomeofthehighcomputationalcostsoftheanalysis.