Non-Linear Metamodeling Extensions to the Robust Parameter Design of Computer Simulations
Abstract
Robust parameter design (RPD) is used to identify a systems control settings that offer a compromise between obtaining desired mean responses and minimizing the variability about those responses. Two popular combined-array strategies the response surface model (RSM) approach and the emulator approachare limited when applied to simulations. In the former case, the mean and variance models can be inadequate due to a high level of non-linearity within many simulations. In the latter case, precise mean and variance approximations are developed at the expense of extensive Monte Carlo sampling. This research combines the RSM approachs efficiency with the emulator approachs accuracy. Non-linear metamodeling extensions, namely through Kriging and radial basis function neural networks, are made to the RSM approach. The mean and variance of second-order Taylor series approximations of these metamodels are generated via the Multivariate Delta Method and subsequent optimization problems employing these approximations are solved. Results show that improved prediction models can be attained through the proposed approach at a reduced computational cost. Additionally, a multi-response RPD problem solving technique based on desirability functions is presented to produce a solution that is mutually robust across all responses. Lastly, quality measures are developed to provide a holistic assessment of several competing RPD strategies.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 15, 2016
- Accession Number
- AD1017985
Entities
People
- Joseph P. Bellucci
Organizations
- Air Force Institute of Technology