KRIging Scilab Package
Kriging based regression and optimization package for Scilab.
This toolbox implements kriging based regression (also known as Gaussian process
regression) and optimization of deterministic simulators.
The toolbox consists of two main components:
1. functions for creation of kriging model for deterministic or noisy data
(correlation kernels, hyper-parameter estimation, prediction, cross
2. methods implementing Expected Improvement based optimization for time
consuming deterministic functions.
Included trend functions: constant, linear, quadratic polynomials.
Included correlation kernels: Gaussian, Matern_5_2, power exponential (p=1.9)
and exponential (p=1);
For sub-optimization (kriging hyper-parameter learning and EI maximization) use
CMA-ES (v0.955) (http://www.lri.fr/~hansen/cmaes_inmatlab.html) or
Nelder-Mead (2.01) (http://atoms.scilab.org/toolboxes/simplex).
The main functions are:
macros/Kriging/ - functions for kriging model creation.
km - create kriging model.
findtheta - find kriging hyper-parameters.
predict - calculate prediction of the kriging mean and variance.
LOOCV - calculate Leave One Out Cross-validation given by the kriging model.
logLL - calculate negative log likelihood of the kriging model.
simGaussian - simulate Gaussian processes based on the covariance and trend
SimCondGaussian - simulate conditional Gaussian processes based on the kriging
macors/EGO/ - functions for optimization of deterministic simulator based on
EGO- create the optimizer.
EGO_ask - get the next data points.
EGO_tell - to tell the optimizer response values at the new data points.
EGO_stop - indicates if the stopping criterion has been reached.
EGO_best - returns the best point and criterion value.
- C functions for correlation matrix and correlation vector creation.
For computational efficiency correlation matrix and vector creation is
implemented in C.
CorrMatrix.sci and sci_corr_D.c - for creation of kriging correlation
CorrVectorData.sci and sci_corr_vector.c - for creation of correlation vector
between point and observed data points.
createCorr.sci and sci_corr_X.c - for creation of arbitrary correlation
corr_functions.h - macros of correlation kernels.
findTheta.sci modified: CMA restart and iteration control loop.
Correlation matrix and vector creation implemented using C gateways and Scilab
km.sci modified: input parameters, kmodel structure.
findTheta.sci modified: CMA default with restarts; optimization settings are
saved into kmodel structure.
ego.sci modified: input parameters, EGO structure does not contain kriging
parameter estimation settings any more.
EGO_ask.sci modified: CMA default with restarts.
EGO_tell.sci modified: findTheta is called using optimization settings in
some errors in help files corrected.