Neural network methods in hydrodynamic yield estimation

A comparison of artificial neural networks and statistical analyses

“…Our results on the test set are similarto those reported by Dowla et al (1992) when they restrictedattentionto a single analysiswindow, an F-valueof 1.40. The added value provided by a statisticaltreatmentinclude warnings if the calibrationevent was found to be inappropriateand estimates of the uncertaintiesassociatedwith the yield estimate.…”

“…Lawrence LivermoreNational Laboratoryconductedan analysisof i hydrodynamicyield estimationusing an artificialneural network (Dowla,et al, 1992). The neural network approachof Dowla et al (1992) and the standard approach of Heinle and Goldwire (1991) can provide reasonablehydrodynamic yield estimates. However, both approachesfail to provide realistic uncertaintieson the yield estimatesbecausethey lack a rigorous treatmentof the uncertaintiesin the problem.…”

“…However, both approachesfail to provide realistic uncertaintieson the yield estimatesbecausethey lack a rigorous treatmentof the uncertaintiesin the problem. Dowla, et al (1992) claim that the network provides "a level of confidence,"but they do not report these levels or examinetheir performancefor the 24 Nevada Test Site events used as their testing set. The confidencelimits producedby HYIELD "appearto be striking • indicatorsof the accuracyof the fitted solution,but the error limits can be vastly underestimatedif the deviationsare not purely instrumental(random)" ' (p.13Heinle and Goldwire,1991).…”

“…Treating the above points is far from trivial. We acquiredthe CORRTEX data used in the neural networkanalysis of Dowla et al (1992) in order to describethe "value added" by a statisticaltreatmentwhich considersthe above points. We developeda methodologyto capturethe uncertaintiesin the CORRTEXcrush length data of the unknownevent and the yield of the calibrationevent and to propagatethese uncertaintiesthrough the method of correspondingpoints to the resultingestimatedyield-versus-time curve.…”

“…Of the currentlyacceptedestimationtechniques,the state-of-the-art computer code is known as HYIELD; it was developedby Heinle and Goldwire(1991). The use of neural networkswas motivatedby an attempt to remove subjectivityfrom the estimationof hydrodynamicyields Dowla et al (1992). employed an artificialneural networkwith ten input units, one hidden layer with eleven hidden units and one output unit.…”