discrepancy_emulation
Formulation for emulation of model discrepancies.
Specification
Alias: None
Arguments: None
Default: distinct
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
---|---|---|---|
Required (Choose One) |
Discrepancy Emulation Approach |
Distinct formulation for emulation of model discrepancies. |
|
Recursive formulation for emulation of model discrepancies. |
Description
In many uncertainty quantification approaches, model discrepancies are emulated using, e.g., polynomial chaos, stochastic collocation, or Gaussian process models. Two formulations are available for this emulation:
distinct
emulation (default), in which we directly approximate the difference or ratio between the evaluations of two models or solution levels.recursive
emulation (experimental option), in which we approximate a difference or ratio among the new model evaluation and the emulator approximation of the previous model.
The latter approach is a form of hierarchical emulation in which we emulate the surplus between the previous emulator and the new modeling level. This approach has a few advantages: (i) it reduces bias by correcting for emulation errors that occur at previous levels, and (ii) it does not require paired model evaluations for each discrepancy level, which reduces cost, allows for disparate sample points, and simplifies data imports.
On the other hand, its primary disadvantage is that the aggregate
emulation is only as good as its weakest link, in that a poor emulator
recovery can create difficulty in accurately resolving discrepancies
that are recursively dependent on it. Thus, the distinct
approach
may tend to be more expensive in exchange for greater robustness.
Examples
method,
multilevel_polynomial_chaos
expansion_order_sequence = 2
collocation_ratio = .9
orthogonal_matching_pursuit
discrepancy_emulation recursive