基于AutoML的实验设计工作流程（DOE）选择和用仿真模型对数据采集策略进行基准测试.pdfVIP

OPENAutoMLbasedworkflowfordesign

ofexperiments(DOE)selection

andbenchmarkingdataacquisition

strategieswithsimulationmodels

1121

XukuanXu,DonghuiLi,JinghouBiMichaelMoeckel

Designofexperiments(DOE)isanestablishedmethodtoallocateresourcesforefficientparameter

spaceexploration.Modelbasedactivelearning(AL)datasamplingstrategieshaveshownpotential

forfurtheroptimization.ThispaperintroducesaworkflowforconductingDOEcomparativestudies

usingautomatedmachinelearning.Basedonapracticaldefinitionofmodelcomplexityinthecontext

ofmachinelearning,theinterplayofsystematicdatagenerationandmodelperformanceisexamined

consideringvarioussourcesofuncertainty:thisincludesuncertaintiescausedbystochasticsampling

strategies,imprecisedata,suboptimalmodeling,andmodelevaluation.Resultsobtainedfrom

electricalcircuitmodelswithvaryingcomplexityshowthatnotallALsamplingstrategiesoutperform

conventionalDOEstrategies,dependingontheavailabledatavolume,thecomplexityofthedataset,

anddatauncertainties.Trade-offsinresourceallocationstrategies,inparticularbetweenidentical

replicationofdatapointsforstatisticalnoisereductionandbroadsamplingformaximumparameter