pilot_samples
Initial set of samples for groups in the multilevel BLUE sampling method
Specification
Alias: initial_samples
Arguments: INTEGERLIST
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
|---|---|---|---|
Optional |
Independent pilot sampling for groups in multilevel BLUE |
||
Description
The pilot sample provides initial estimates of group covariances during the first iteration of the multilevel best linear unbiased estimator (ML BLUE) sampling approach. These initial estimates then guide the algorithm toward group sample increments in order to achieve a prescribed error at minimum total cost or minimum error for a prescribed budget.
The formulation of ML BLUE assumes independent samples per group, and
the pilot sampling can either start in this same vein (see
independent) or start
from group covariances inferred from a single set of shared samples.
The benefit in the latter shared pilot approach is one can avoid
investing any pilot samples for groups that not determined to be
active in the resource allocation process. Groups that are determined
to be active, however, must start from new sets of independent
samples, not reusing the shared samples (except for the group that
contains all models, to which the shared pilot samples are assigned).
Default Behavior
The default number of pilot samples is 100. The default pilot sampling approach is shared, where a single set of pilot samples is evaluated across all models. The group covariances are initially formed from this shared set of pilot samples for purposes of solving the initial resource allocation problem. Group covariances are then updated using new independent sample sets for all groups with new allocations.
Usage Tips
The number of specified values can be none (default values used for
all groups), one (all groups use the same specified value), or the
number of sample groups. The number of groups is determined by
combinations of the model indices, including the truth model and all
approximations. Larger counts in the number of combinations may be
throttled as described in group_throttle.
