Description Usage Arguments Details Value References See Also
Training of a Gaussian Process model with a genetic algorithm.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | Fit(object, ...)
## S4 method for signature 'GP'
Fit(object, contribution = T, nugget = T, nugget.t = F,
maxrange = T, midrange = F, minrange = F, azimuth = F, dip = F,
rake = F, power = F, contribution.ns = F, maxrange.ns = F,
midrange.ns = F, minrange.ns = F, azimuth.ns = F, dip.ns = F,
rake.ns = F, ...)
## S4 method for signature 'GP_geomod'
Fit(object, maxrange = T, midrange = F,
minrange = F, azimuth = F, dip = F, rake = F, power = F, ...)
## S4 method for signature 'SPGP'
Fit(object, contribution = T, nugget = T, nugget.t = F,
maxrange = T, midrange = F, minrange = F, azimuth = F, dip = F,
rake = F, power = F, contribution.ns = F, maxrange.ns = F,
midrange.ns = F, minrange.ns = F, azimuth.ns = F, dip.ns = F,
rake.ns = F, metric = c("logLik", "PLPD", "NRMSE"), ...)
## S4 method for signature 'SPGP_geomod'
Fit(object, maxrange = T, midrange = F,
minrange = F, azimuth = F, dip = F, rake = F, nugget = F,
power = F, ...)
|
... |
Arguments passed on to |
contribution |
Optimize on the covariance model's amplitude? |
nugget |
Optimize on nugget? |
maxrange |
Optimize on the covariance model's range? |
midrange, minrange |
Optimize on the covariance model's anisotropy? |
azimuth, dip, rake |
Optimize on the covariance model's orientation? |
power |
Optimize on the power parameter (not relevant for all covariances)? |
metric |
Which metric to optimize? |
pseudo_inputs |
Optimize on the pseudo-inputs' locations? |
pseudo_tangents |
Optimize on the pseudo-inputs' locations for derivative data? |
This method uses a genetic algorithm to optimize the contribution
(or amplitude) and range of each covariance structure contained in the
model
slot of object
, as well as the parameters above, if
allowed.
Optimization is done with respect to the specified metric. The available metrics are the log-likelihood (default), normalized root mean square error (NRMSE), and penalized log predictive density (PLPD). The latter two are determined by leave-one-out cross validation, which is slower than the log-likelihood.
The positions of the pseudo-inputs may be constrained to match a subset of
the data (pseudo_inputs = "subset"
) or be free to lie anywhere
inside the data's bounding box (pseudo_inputs = "free"
). The
pseudo-tangents, however, must be a subset of the tangent data to avoid
overfitting, as the tangents have a greater degree of freedom (position and
direction).
See the documentation in ga
to control the optimization
process. Standard GP uses continuous optimization to fit the parameters,
using the current ones as a starting point. The sparse GP uses discrete
optimization (with binary encoding) to fit the parameters and select the
best positions for the pseudo-inputs. In both cases, it is
recommnded to set the popSize
around 20 and
pmutation
between 0.3 and 0.5. Convergence status can be visualized by
setting monitor = T
.
The variational SPGP model may pose difficulties for training. It may help to train a FIC model (or a standard GP, using a subset of the data) to obtain the best covariance parameters and nugget and use them to build a variational SPGP object.
In order to obtain reproducible results, set the seed before calling this method.
A GP
object similar to object
, with optimized
covariance parameters.
Bauer, M.S., van der Wilk, M., Rasmussen, C.E., 2016. Understanding Probabilistic Sparse Gaussian Process Approximations. Adv. Neural Inf. Process. Syst. 29.
covarianceStructure3D-class
,
GP-init
, SPGP-init
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.