deodorant.covar-functions
Covariance functions for Deodorant.
matern32-grad-K
(matern32-grad-K x-diff-squared log-sig-f log-rho)
Gradient for matern32. Syntax as per matern32
except returns a DxNxN array giving derivatives
in the different directions. The first entry
of the first dimension corresponds to the derivative
with respect to log-sig-f, with the others wrt
log-rho
matern32-grad-z
(matern32-grad-z xs-z-diff log-sig-f log-rho)
Jacobian of side kernel matrix w.r.t. new data point z for Matern 32.
If using a gradient based solver for the acquisition funciton, then
needed for calculating derivative of Expected Improvement, EI(z), as outlined
on page 3 of
http://homepages.mcs.vuw.ac.nz/~marcus/manuscripts/FreanBoyle-GPO-2008.pdf.
Accepts:
xs-z-diff - NxD matrix whose (i, j)th entry is x_ij - z_j
log-sig-f - scalar; parameter of kernel function
log-rho - D-dimensional parameter of kernel function
Returns:
[NxD] Jacobian of side kernel matrix w.r.t. new data point where
(i, j)th entry is d(kernel(x_i, z)) / d(z_j).
matern32-K
(matern32-K x-diff-squared log-sig-f log-rho)
Covariance function for matern-32.
Accepts: x-diff-squared - a NxDxN matrix of squared distances
or NxD matrix of squared distances of old points
and new point
log-sig-f - a scalar
log-rho - a vector
Returns: A matrix K
matern32-plus-matern52-grad-K
(matern32-plus-matern52-grad-K x-diff-squared log-sig-f-32 log-rho-32 log-sig-f-52 log-rho-52)
Gradient of compound covariance function for matern-32 and matern-52.
Accepts: x-diff-squared - a NxDxN matrix of squared distances
log-sig-f-32 - a scalar
log-rho-32 - a vector
log-sig-f-52 - a scalar
log-rho-52 - a vector
Returns: An DxNxN array grad-K giving derivatives
in the different directions. The first entry
of the first dimension corresponds to the derivative
with respect to log-sig-f, with the others wrt
log-rho
matern32-plus-matern52-grad-z
(matern32-plus-matern52-grad-z x-z-diff log-sig-f-32 log-rho-32 log-sig-f-52 log-rho-52)
Jacobian of side kernel matrix w.r.t. new data point z for Matern 32 + Matern 52.
If using a gradient based solver for the acquisition funciton, then
needed for calculating derivative of Expected Improvement, EI(z), as outlined
on page 3 of
http://homepages.mcs.vuw.ac.nz/~marcus/manuscripts/FreanBoyle-GPO-2008.pdf.
Accepts:
xs-z-diff - NxD matrix whose (i, j)th entry is x_ij - z_j
log-sig-f - scalar; parameter of kernel function
log-rho - D-dimensional parameter of kernel function
Returns:
[NxD] Jacobian of side kernel matrix w.r.t. new data point where
(i, j)th entry is d(kernel(x_i, z)) / d(z_j).
matern32-plus-matern52-K
(matern32-plus-matern52-K x-diff-squared log-sig-f-32 log-rho-32 log-sig-f-52 log-rho-52)
Compound covariance function for matern-32 and matern-52.
Accepts: x-diff-squared - a NxDxN matrix of squared distances
log-sig-f-32 - a scalar
log-rho-32 - a vector
log-sig-f-52 - a scalar
log-rho-52 - a vector
Returns: A matrix K
matern32-xs-z
(matern32-xs-z xs z log-sig-f log-rho)
Side covariance matrix for matern-32, i.e. vector k where
k_i = kernel(x_i, z).
Accepts:
xs - a NxD vector of vectors of xs
z - [Dx1] vector of new data point
log-sig-f - a scalar
log-rho - a vector
Returns: A vector k sized N.
matern52-grad-K
(matern52-grad-K x-diff-squared log-sig-f log-rho)
Gradient for matern52. Syntax as per matern52
except returns a DxNxN array giving derivatives
in the different directions. The first entry
of the first dimension corresponds to the derivative
with respect to log-sig-f, with the others wrt
log-rho
matern52-grad-z
(matern52-grad-z xs-z-diff log-sig-f log-rho)
Jacobian of side kernel matrix w.r.t. new data point z for Matern 52.
If using a gradient based solver for the acquisition funciton, then
needed for calculating derivative of Expected Improvement, EI(z), as outlined
on page 3 of
http://homepages.mcs.vuw.ac.nz/~marcus/manuscripts/FreanBoyle-GPO-2008.pdf.
Accepts:
xs-z-diff - NxD matrix whose (i, j)th entry is x_ij - z_j
log-sig-f - scalar; parameter of kernel function
log-rho - D-dimensional parameter of kernel function
Returns:
[NxD] Jacobian of side kernel matrix w.r.t. new data point where
(i, j)th entry is d(kernel(x_i, z)) / d(z_j).
matern52-K
(matern52-K x-diff-squared log-sig-f log-rho)
Covariance function for matern-52.
Accepts: x-diff-squared - a NxDxN matrix of squared distances
or NxD matrix of squared distances of old points
and new point
log-sig-f - a scalar
log-rho - a vector
Returns: A matrix K
matern52-xs-z
(matern52-xs-z xs z log-sig-f log-rho)
Side covariance matrix for matern-52, i.e. vector k where
k_i = kernel(x_i, z).
Accepts:
xs - a NxD vector of vectors of xs
z - [Dx1] vector of new data point
log-sig-f - a scalar
log-rho - a vector
Returns: A vector k sized N.