gplite (version 0.13.0)

approx: Approximations to the posterior of the latent values

Description

Functions for initializing the approximation for the latent values, which can then be passed to gp_init. The supported methods are:

approx_laplace

Laplace's method, that is, based on local second order approximation to the log likelihood. For Gaussian likelihood, this means exact inference (no approximation).

approx_ep

Expectation propagation, EP. Approximates the likelihood by introducing Gaussian pseudo-data so that the posterior marginals match to the so called tilted distributions (leave-one-out posterior times the true likelihood factor) as closely as possible. Typically more accurate than Laplace, but slower.

Usage

approx_laplace(maxiter = 30, tol = 1e-04)
approx_ep(damping = 0.9, quad_order = 11, maxiter = 100)

Value

The approximation object.

Arguments

maxiter

Maximum number of iterations in the Laplace/EP iteration.

tol

Convergence tolerance.

damping

Damping factor for EP. Should be between 0 and 1. Smaller values typically lead to more stable iterations, but also increase the number of iterations, and thus make the algorithm slower.

quad_order

Order of the Gauss-Hermite quadrature used to evaluate the required tilted moments in EP.

References

Rasmussen, C. E. and Williams, C. K. I. (2006). Gaussian processes for machine learning. MIT Press.

Examples

# Basic usage
gp <- gp_init(
  cfs = cf_sexp(),
  lik = lik_bernoulli(),
  method = method_fitc(num_inducing = 100),
  approx = approx_ep()
)