online_pilot
Specify a solution mode that includes the pilot cost within the sample allocation logic
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
---|---|---|---|
Optional |
For an online pilot mode, apply under-relaxation to the shared sample increments |
||
Optional |
Indicate the type of final statistics to be returned by a UQ method |
Description
Multilevel / multifidelity sampling methods are adaptive UQ methods
that utilize a pilot sample to estimate an initial set of correlations
or variances, and then augment the pilot with additional sample
increments to optimally allocate resources. In this mode, the cost of
pilot sampling is treated as online cost and is included within the
optimal sample allocation logic. It is the only solution mode that is
iterative, seeking to converge from the initial pilot to a shared set
of samples that is neither over- nor under-estimated. The former
over-estimation, while supporting more robust estimation of
correlations and variances (refer to the offline_pilot
option if a
reference solution is needed), is inefficient due to non-optimality of
the aggregate sample sets, obscuring the underlying optimal profile.
The latter under-estimation results in unnecessary inaccuracy due to
reliance on the initial correlation / covariance approximations – if
shared sample increments are indicated in the subsequent allocation
process, then these should also be used to make better estimates of
the shared covariances (through online iteration).
Default Behavior
This iterative online mode (online_pilot
) is the default.
Usage Tips
It is typically advantageous to start from a smaller pilot sample and then rely on the iterative updates to increase the shared sample levels to match the optimal sample profile (as dictated by an accuracy target or prescribed budget). Under-relaxation can also be helpful to avoid a shared sample update that is too aggressive based on inaccurate initial information.