nategarton13/sparseRGPs: Fit Sparse Gaussian Processes

Fit sparse and full Gaussian processes with Type-II maximum likelihood assuming Gaussian, binary, or Poisson data. In the case of non-Gaussian data, the Laplace approximation is used to approximate the marginal likelihood and select covariance parameters and/or knots. Predictions in the non-Gaussian case also use the Gaussian posterior approximation implied by the Laplace approximation. In the Gaussian data case, there is an option to use the variational approximation. Otherwise, FI(T)C models are fit. In the case that sparse models are used, knots can be held fixed, simultaneously optimized alongside covariance parameters, or optimized using the OAT knot selection algorithm proposed in Garton et al. (2020).

install.packages("remotes")
remotes::install_github("nategarton13/sparseRGPs")