dakota
Select the built in Gaussian Process surrogate
Specification
Alias: None
Arguments: None
Description
A second version of GP surrogates was available in prior versions of Dakota. For now, both versions are supported but the ``dakota`` version is deprecated and intended to be removed in a future release.
Historically these models were drastically different,
but in Dakota 5.1, they became quite similar. They now differ in that
the Surfpack GP has a richer set of features/options and tends to be
more accurate than the Dakota version. Due to how the Surfpack GP
handles ill-conditioned correlation matrices (which significantly
contributes to its greater accuracy), the Surfpack
GP can be a
factor of two or three slower than Dakota’s. As of Dakota 5.2, the
Surfpack implementation is the default in all contexts except Bayesian
calibration.
More details on the gaussian_process
dakota
model can be found in [McF08].
Dakota’s GP
deals with ill-conditioning in two ways. First, when it encounters a
non-invertible correlation matrix it iteratively increases the size of
a “nugget,” but in such cases the resulting approximation smooths
rather than interpolates the data. Second, it has a point_selection
option (default off) that uses a greedy algorithm to select a
well-spaced subset of points prior to the construction of the GP. In
this case, the GP will only interpolate the selected
subset. Typically, one should not need point selection in trust-region
methods because a small number of points are used to develop a
surrogate within each trust region. Point selection is most
beneficial when constructing with a large number of points, typically
more than order one hundred, though this depends on the number of
variables and spacing of the sample points.
This
differs from the point_selection
option of the Dakota GP which
initially chooses a well-spaced subset of points and finds the
correlation parameters that are most likely for that one subset.