|
|
过程工程学报 2002
DOBDAlgorithmforTrainingNeuralNetwork:PartII.ApplicationKeywords: neuralnetwork,DOBDalgorithm,Marquardtmethod,overfitting,pruning,training,application Abstract: Inthefirstpartofthearticle,anewalgorithmforpruningnetworkDynamicOptimalBrainDamage(DOBD)isintroduced.Inthispart,twocasesandanindustrialapplicationareworkedouttotestthenewalgorithm.ItisverifiedthatthealgorithmcanobtaingoodgeneralizationthroughdeletingweightparameterswithlowsensitivitiesdynamicallyandgetbetterresultthantheMarquardtalgorithmorthecross-validationmethod.Althoughtheinitialconstructionofnetworkmaybedifferent,thefinialnumberoffreeweightsprunedbytheDOBDalgorithmissimilarandthenumberisjustclosetotheoptimalnumberoffreeweights.Thealgorithmisalsohelpfultodesigntheoptimalstructureofnetwork.
|
