offline_projection

Specify a solution mode that estimates performance based on projecting initial correlation/variance estimates from an offline pilot sample

Specification

• Alias: None

• Arguments: None

Description

Multilevel / multifidelity sampling methods are adaptive UQ methods that utilize a pilot sample to estimate an initial set of correlations or covariances, which is then used to initiate the process of performance estimation and resource allocation. In online and offline projection modes, evaluations are limited to the pilot sample and these initial correlation / covariance estimates are used to project the potential estimator performance, without any further sampling and without any computation of statistics. For the offline projection case, the estimator performance is computed with an online sample allocation that is unconstrained by the offline pilot.

Usage Tips

This mode is useful for comparing the projected performance of different estimators, which can inform method selection.

Note that the sample allocation for each model is lower bounded at two samples for defining nonzero variance. While not required for projecting estimator performance, this enables consistency with the corresponding non-projection mode that may follow.