![]() |
Dakota Reference Manual
Version 6.16
Explore and Predict with Confidence
|
Identify metadata by label for capturing online cost estimates
Alias: none
Argument(s): STRING
Metadata can be used for a variety of response annotation purposes, including the online recovery of simulation cost (using facilities either implemented in a user's analysis driver or codified in an embedded test driver). This information can then be used by certain algorithms that allocate resources based on relative cost, e.g. multifidelity UQ methods.
Cost estimates may be provided a priori, e.g. using solution_level_cost
, in cases where they can be assumed fixed for each model form and/or resolution. In cases such as model tuning over a set of hyper-parameters, costs will generally be a function of these hyper-parameters and need to be captured online during initialization of the algorithm (e.g. during evaluation of the pilot sample).
Default Behavior No recovery of cost estimates from the response metadata.
Usage Tips It is common to employ total CPU time, aggregated across all computing cores involved in a (parallel) simulation instance. Wall time, even when the number of cores is fixed, can be subject to additional variability due to system load.
model, id_model = 'TUNABLE_LF' simulation cost_recovery_metadata = 'core_time' responses, response_functions = 1 descriptors = 'Q' no_gradients no_hessians metadata 'core_time'
These keywords may also be of interest: