What is it?
OctGPR is an Octave package for Gaussian Process Regression, which is a robust regression technique (also known as Kriging in geostatistics). OctGPR can currently do Gaussian Process Regression with constant or linear mean (univ. Kriging), several correlation functions, and automatic estimation of inverse length scales and relative noise via concentrated maximum likelihood. Mean and variance are estimated analytically, the rest of the hyperparameters is optimized using a custom mixed-norm trust-region variable metric method. Support for RBF models with GCV estimation (and possibly more) is on the way.
The package is still in heavy development, and will likely stay so during 2008. It has become part of OctaveForge. You can download the package, documentation and track all the changes from there. I will also maintain a bundle here; however, consider the Octave Forge version as more authoritative.
Demo pictures
The following pictures are snapshots from the OctGPR demo function
(click to move to next frame):
Installation
Simply download the package, run
octave in the same directory, and type
pkg install octgpr-1.1.4.tar.gz. If there is no error,
you can load the package using pkg load octgpr and
the use the gpr_train and gpr_predict
functions. You must load the package each time you start Octave; if you
wish Octave to do this automatically, install the package using
pkg install -auto ... instead. Alternatively,
you can use pkg rebuild - see help.
Requirements
The package is coded in mixed Fortran/C/C++. Any Fortran 90 compiler as well
as some F77 ones with relevant extensions (e.g. g77) should accept the Fortran
sources. Typically, the mkoctfile provided by your Octave installation
will accept all the sources. The software relies on a number of BLAS and LAPACK
routines - these two Fortran libraries should be available on your system.
