experimental_design
(Experimental) Adaptively select experimental designs for iterative Bayesian updating
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
|---|---|---|---|
Required |
Number of data points used during initial Bayesian calibration |
||
Required |
Number of candidate design points considered |
||
Optional |
Maximum number of high-fidelity model runs to be used |
||
Optional |
Number of optimal designs selected concurrently |
||
Optional |
Specify text file containing candidate design points |
||
Optional |
Use second Kraskov algorithm to compute mutual information |
||
Description
This “experimental design” algorithm uses responses produced by a high-fidelity
model as data to update the parameters of a low-fidelity model using Bayes’
Rule. It is a capability that is under active development and is currently
compatible only with queso. The user-specified high-fidelity model should
depend only on configuration variables (i.e. design conditions), such as
temperature or spatial location, while the user-specified low-fidelity model
should depend on both configuration variables and its own model parameters to
be calibrated.
The algorithm starts with a preliminary Bayesian calibration using the
number of data points specified in initial_samples. These data points can
be read in through the calibration_data_file command in the responses
block. If num_experiments is less than initial_samples or if no such data
file is provided, Latin Hypercube Samples of the design space (specified in the
variables block) will be run through the user-specified high-fidelity code
to
supplement the initial data. Once this first calibration is complete, a set of
possible experimental design conditions, specified using the configuration
variables, is proposed. The user specifies the size of this set using
num_candidates. The set of candidates itself may be explicitly given through
the import_candidate_points_file command. If the number of candidates in this
file is less than num_candidates, or if this file is omitted, the set of
candidate designs will again be supplemented with a Latin Hypercube Sample of
the design space.
For each candidate design \(\xi_i\) in the set of possible design conditions, the mutual information (MI) between the low-fidelity model parameters \(\boldsymbol{\theta}\) and the high-fidelity model response \(\textbf{y}(\xi_i)\) ,
is approximated. The high-fidelity model is replaced by the low-fidelity model
and a \(k\) -nearest neighbor approximation is used in the calculation. The
design point \(\xi^{*}\) for which MI is the largest is selected and run
through the high-fidelity model to yield a new observation \(y(\xi^{*})\) .
This new observation is added to the calibration data, and a subsequent
Bayesian calibration is performed. A new MI for each remaining candidate design
is computed, and the process repeats until one of three stopping criteria are
met. Multiple optimal designs may be selected concurrently by specifying
batch_size.
Of the three stopping criteria, two are automatically checked by Dakota. If
the relative change in the MI from one iteration to the next is sufficiently
small or if the set of candidate design conditions has been exhausted, the
algorithm teriminates. The user may specify the third stopping criteria using
max_hifi_evaluations. This limits the number of high-fidelity model
evaluations that will be performed during this algorithm. It therefore limits
the number of iterations through the algorithm that will be performed. Any
high-fidelity model runs needed to produce the data set for the initial
calibration are not included in this allocation.
In the case that the high-fidelity model must be run indepently of Dakota,
the user may set max_hifi_evaluations to zero. The optimal experimental
design point will be calculated and reported, but the high-fidelity model
will not be run. For more details, see the User’s Manual.
Expected Output
Information regarding the progress and termination condition of the experimental
design algorithm is output to the screen with varying levels of verbosity.
Further details can be found, regardless of verbosity, in the output file
experimental_design_output.txt
Usage Tips
Due to the optional file read-ins and the supplemental sampling, it is
important for the user to check consistency within the input file
specifications. For example, if num_experiments is less than the number of
experiments in the calibration_data_file, only the first lines of the file
will be used and the rest will be discarded. The same holds true for the
import_candidate_points_file and num_candidates.
