Online Gaussian Process Toolbox - Reference Documentation

Welcome to the Online Gaussian Processes (OGP) online documentation pages.

The OGP software package implements general Bayesian inference using Gaussian Processes as latent variables. The package allows the usage of a large variety of likelihood functions ranging from the standard normal noise assumption for regression tasks to the use of local inverse models in data assimilation.

Additionally to flexibility, the OGP package also builds a sparse representation to the posterior process, allowing efficient inference for very datasets.

The code uses the NETLAB package ( http://www.ncrg.aston.ac.uk/netlab) and is based on the the paper Sparse Online Gaussian Processes (Csató & Opper, Neural Computation 14/3, 2002).

The online reference documentation provides direct hypertext links to specific function descriptions in the Online Gaussian Process Toolbox.

Any comments or problems: contact Lehel Csató (_csatol_ _cs_ _ubbcluj_ _ro)

Index

An alphabetic list of functions in the Online GP Toolbox.

c_class_bin: Online update coeffs. for binary classification.
c_reg_exp: Online update coeffs. for regression with positive exponential noise.
c_reg_gauss: Online update coeffs. for regression with Gaussian noise.
c_reg_lapl: Online update coeffs. for regression with Laplace noise.
cl_data: Generates two-dimensional dataset for classification.
cov_matern: returns the Matern covariance function.
covgrad_matern: Returns the gradient of the Matern kernel.
defoptions: sets the ``gpopt'' structure to default values.
demogp_class: Two-dimensional binary classification example.
demogp_class_gui: Graphical frontend for Onlinde Gaussian Process classification.
demogp_fixed: Gaussian Process regression using a fixed set of basis vectors.
demogp_matern: Example to use an external covariance function.
demogp_reg: One-dimensional regression example with different noise models.
demogp_reg_gui: Graphical frontend to Online Gaussian Process regression.
demogp_simp: Gaussian process regression using user inputs.
em_exp: Recomputes the likelihood parameter for pos.-exp. noise.
em_gauss: Recomputes the (Gaussian) noise variance.
em_lapl: Recomputes the likelihood parameter for exponential noise.
err_2class: Computes the binary classification error.
err_2logp: Computes the log-predictive probability of the labels.
err_abs: Computes the absolute error.
err_mse: Computes the mean-square error.
g_l_gauss: Recomputes the variance of the Gaussian noise using gradients
laplace: Sampling from a Laplace distribution.
logp_exp: Computes the log-predictive probability for positive exponential noise
logp_g: Computes the log-predictive probability for Gaussian likelihood
logp_l: Computes the log-predictive probability for Gaussian noise
matern: Computes the Matern kernel
ogp: Initialises the global net structure for the OGP toolbox.
ogpadjgp: Computes the GP coefficients from the TAP/EP ones.
ogpbvmin: Finds the BV that contributes the least to the GP
ogpcovarf: Calculate the covariance function for the OGP.
ogpcovarp: Calculate the prior covariance for the Sparse GP.
ogpcovdiag: Calculates the diagonal of the covariance function for the OGP.
ogpcovgrad: Evaluate gradient for kernel parameters (except bias) for the Sparse GP
ogpdelbv: Deletes the specified BVs from the BV set of the GP.
ogpemptybv: Adds input elements to the BV set without altering the GP.
ogpevid: Evaluates the evidence for Sparse OGP.
ogpevidgrad: Computes the gradient for the Sparse OGP.
ogpfwd: Forward propagation through a Sparse OGP.
ogphypcovpar: Initialises the Sparse Gaussian Process hyperparameters.
ogpinit: Likelihood initialisation for the Online Gaussian Process structure.
ogpkl: Computes the KL-distance of the GP marginals.
ogppak: Puts the Sparse OGP hyperparametrs into a vector.
ogpparadj: Adjusts GP such that the TAP/EP it. is not ill-conditioned.
ogppost: Calculation of the sparse posterior
ogpreset: Resets the Gaussian Process.
ogpsample: Generates samples from a Gaussian process.
ogpstep_ep: Updates the TAP/EP parameters after an online sweep
ogpstep_full: Performs a full online update step of the Gaussian Process
ogpstep_sp: Performs a sparse online update step of the Gaussian Process.
ogptrain: Inference for Sparse Gaussian Processes.
ogpunpak: Puts hyperparameters back into the Sparse OGP.
sinc: Sin(pi*x)/(pi*x) function. (from MATLAB)
sinc2data: Generates two-dimensional sinc test data.
sincdata: Generates one-dimensional sinc test data.