Menu
▾
▴
Tree [53d441] default tip / History
| File | Date | Author | Commit |
|---|---|---|---|
| inst | 2011-11-03 |
carandraug
|
[ea58d7] removal of deprecated functions: loadimage, spd... |
| src | 2012-06-12 |
i7tiol
|
[d52a5a] More compatible passing of linker flags to mkoc... |
| .hgignore | 2016-02-29 |
Carnë Draug
|
[53d441] maint: add hgignore file. |
| COPYING | 2011-06-24 |
carandraug
|
[06e76d] Updated FSF address |
| ChangeLog | 2009-08-06 |
highegg
|
[9b903c] OctGPR 1.2.0 |
| DESCRIPTION | 2012-06-12 |
i7tiol
|
[bad52c] Add missing NEWS entrys for fixing Makefiles fo... |
| INDEX | 2010-07-28 |
highegg
|
[97836b] update INDEX |
| Makefile | 2010-01-07 |
thomas-weber
|
[ec4e95] Add Makefile, so that the package creation proc... |
| NEWS | 2012-06-12 |
i7tiol
|
[bad52c] Add missing NEWS entrys for fixing Makefiles fo... |
| README | 2008-02-07 |
highegg
|
[4808d1] Initial commit of the OctGPR package. |
| TODO | 2008-04-18 |
highegg
|
[44882d] demo correction & cosmetic changes |
Read Me
OctGPR -------------------------------------------------------------------------------- OctGPR is a package for using Gaussian Process Regression (GPR) in Octave. GPR is a Bayesian statistical method of inference of unknown spatial data from known samples. It is also know as Kriging in geostatistics field. The method assumes that the known sample data are a result of a uniform spatial Gaussian Process, with a constant or linear mean and constant variance. Several models for the correlation function may be selected - gaussian, exponential or inverse multiquadrics (more might be added in the future). The mean (mu) and variance (var) parameters are ML-estimated analytically, while the inverse spatial scales (theta) and white noise (nu) are ML-estimated using a custom mixed-norm trust-region optimization algorithm. Derivatives w.r.t. theta and *two* derivatives w.r.t. nu are calculated analytically, hoping for a rapid convergence (as nu is typically the most sensitive parameter). In the future, the package will be extended with RBF models trained via GCV.
