Using Gaussian Stochastic Processes to Improve Fourier Amplitude Sensitivity Test

• Activity Number: 522
• Type: Contributed
• Date/Time: Wednesday, August 4, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Amy Hoeksema*+ and Max Morris
• Companies: Iowa State University and Iowa State University
• Keywords: computer experiments ; sensitivity analysis ; Fourier Amplitude Sensitivity Test ; Gaussian Stochastic Processes

Abstract

Fourier Amplitude Sensitivity Test (FAST) is a method of sensitivity analysis commonly used for computer models. Given a model for the data FAST uses Weyl's Theorem to express high-dimensional integrals as one-dimensional integrals allowing simple computation of moments from the joint output distributions. The moments are then used to compute sensitivity indices which determine the importance of input variables. FAST requires a large number of model evaluations to make the analyses effective. We propose an alternative hybrid method using Gaussian stochastic processes. Then MLE's can compute an estimate of the original function that will require fewer model evaluations and so is less expensive to compute. Using this surrogate model, we can increase the sample sizes and get better sensitivity estimates. We show this in comparing results from test functions to the original analyses.