stilt-package: Separable Gaussian Process Emulator
In scrim-network/Stilt-Rpkge: Separable Gaussian Process Interpolation (Emulation)

Description Details Disclaimer Summary of main functions Summary of datasets Limitations and Caveats Author(s) References

'Stilt' is used for interpolating multivariate data in multivariate space. It is tailored to the case of interpolating ("emulating") regularly spaced time-series of computer model output between the model input parameters, but can be also used more broadly (e.g., for output in the form of regularly-spaced spatial 1D transects, or for interpolating time-series between locations on the Earth's surface). The interpolation technique relies on a separable Gaussian Process emulator. Package consists of functions to fit the Gaussian Process emulator, to interpolate to (predict at) a given model input setting, and to validate the emulator using cross-validation. In addition, a function is provided to plot model response surface as generated by the emulator as a function of two model parameters. The package works both on one and multi-parameter ensembles, with an exception of the 'rsurface.plot' function, which by definition is restricted to 2+ parameter ensembles. The emulator is restricted to multivariate model output and is not designed to work on interpolating a single scalar.

Package:	stilt
Type:	Package
Version:	1.3.0
Date:	2017-08-15
License:	GPL-3
LazyLoad:	no

This package is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. In addition, the authors maintain no responsibility for any possible code errors or bugs.

emulator: Fits a separable Gaussian Process emulator to ensemble model output
predict.emul: Predicts using a Gaussian Process emulator
rsurface.plot: Plots a 2D model response surface
test.csv: Cross-validates an emulator by removing one or many model runs, training the emulator on the remaining runs, and then predicting at the withheld model runs
test.all: Tests an emulator using leave-one-out cross validation
sep.cov: Constructs time and parameter covariance matrices

Data.1D.model and Data.1D.par: Model output and parameter settings for a simple 1-parameter ensemble example
emul.1D: Gaussian Process emulator that has been fit to the aforementioned model ensemble output
Data.AR1.Korea.model and Data.AR1.Korea.par: Temperature variability and future change for 29 CMIP5 global climate models
Data.UVic.model and Data.UVic.par: Model output and parameter settings for temperature anomaly output from 3-parameter ensemble of climate model UVic ESCM
Data.Sicopolis.model and Data.Sicopolis.par: Model output and parameter settings for ice mass loss output from a 5-parameter ensemble of ice sheet model SICOPOLIS
emul.Sicopolis: A Gaussian Process emulator that has been fit to the SICOPOLIS ensemble output

The emulator (like any other available software packages) will not work well for 'jagged' model response surfaces (high nugget): predictive uncertainty will be too high.
The emulator is restricted to output at regular intervals in time and space
The code has not been tested under conditions of extreme high / extreme low input parameter range, output time(space) coordinates range, and output range (an example would be model output ranging from -1E20 to 1E20, etc.). In such cases it is recommended to re-scale the time(space) coordinates vector, the input parameters, and/or the model output.
The emulator will not work on scalar model output – it requires multivariate data
The emulator assumes a separable covariance function, and stationarity of the covariance part of the Gaussian process.
The optimization of the emulator parameters degrades dramatically (and increases in time) as a function of number of free parameters. Hence, the emulator might be of limited use for large parameter ensembles
The emulator authors are not responsible for any code errors and/or bugs