Fit: Fit
In italo-goncalves/geomod3D: Implicit 3D Geological Modeling and Geostatistics

Description Usage Arguments Details Value References See Also

Training of a Gaussian Process model with a genetic algorithm.

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 `ga`, such as `maxiter`, `popSize`, etc.
`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

italo-goncalves/geomod3D documentation built on May 24, 2019, 2:49 p.m.