Prior distributions for Gaussian-process hyperparameters

Constructors for the prior families accepted by gp_prior(). Each returns a small kagu_prior object (a family name plus its parameters) that Kagu can both sample from (prior predictive simulation) and evaluate (MAP fitting).

• prior_loguniform(lower, upper) - uniform on log(x) over [lower, upper].

• prior_lognormal(meanlog, sdlog) - log(x) ~ Normal(meanlog, sdlog).

• prior_invgamma(shape, scale) - inverse-gamma; a common lengthscale prior because it puts little mass near zero (very wiggly functions).

• prior_halfnormal(scale) - |Normal(0, scale)|.

• prior_normal(mean, sd) - unbounded; only valid for root_mean.

Usage

prior_loguniform(lower, upper)

prior_lognormal(meanlog = 0, sdlog = 1)

prior_invgamma(shape, scale)

prior_halfnormal(scale = 1)

prior_normal(mean = 0, sd = 1)

Arguments

lower, upper

Positive bounds of the log-uniform prior.

meanlog, sdlog

Mean and sd of log(x) for the log-normal prior.

shape

Shape of the inverse-gamma prior.

scale

Scale of the inverse-gamma or half-normal prior.

mean, sd

Mean and sd of the normal prior.

Value

A kagu_prior object.

See also

gp_prior()

Examples

prior_lognormal(0, 0.5)
#> <kagu_prior> lognormal(meanlog = 0, sdlog = 0.5) 
prior_invgamma(5, 5)
#> <kagu_prior> invgamma(shape = 5, scale = 5)