|
Abstract:
|
A new method is presented to identify "active" inputs in a computer experiment. To judge if an input is active, low-impact inputs are added as benchmarks to create a null distribution of sensitivity index activities. A two-stage group screening procedure is proposed which first partitions individual inputs into groups and fits a model to the output regarded as a function of group inputs. Active groups are identified and all inputs from these groups proceed to the second stage. Additional data are collected at Stage 2, a model is fitted using the individual inputs, and those inputs with a large total effect sensitivity index are selected as active. The grouping strategy, the construction of two-stage designs, and the effect of the magnitude of the low-impact input on the specificity and sensitivity of the procedure are described. The method is illustrated through simulated examples.
|
