ABC with a GP surrogate, four methods — abc_gp

Wilkinson (2014) history matching + GP-emulated log-posterior on a grid, or one of the Meeds & Welling (2014) samplers.

Usage

abc_gp_emulator(
  sim,
  obs,
  X_grid = NULL,
  kernel = "sqexp",
  method = "wilkinson",
  prior_ppf = NULL,
  log_prior = NULL,
  theta0 = NULL,
  n_sim = 50L,
  epsilon = 1,
  summary = NULL,
  n_waves = 3L,
  n_design = 32L,
  threshold = 10,
  n_sd = 3,
  accept_kernel = "gaussian",
  n_iter = 200L,
  proposal_sd = 0.5,
  xi = 0.05,
  delta_s = 10L,
  n_alpha = 64L,
  seed = 0L
)

Arguments

sim: See Usage.
obs: See Usage.
X_grid: See Usage.
kernel: See Usage.
method: See Usage.
prior_ppf: See Usage.
log_prior: See Usage.
theta0: See Usage.
n_sim: See Usage.
epsilon: See Usage.
summary: See Usage.
n_waves: See Usage.
n_design: See Usage.
threshold: See Usage.
n_sd: See Usage.
accept_kernel: See Usage.
n_iter: See Usage.
proposal_sd: See Usage.
xi: See Usage.
delta_s: See Usage.
n_alpha: See Usage.
seed: See Usage.

Examples

# \donttest{
sim <- function(theta, e) theta + c(-0.5, 0, 0.5)
obs <- c(0.5, 1, 1.5)
fit <- abc_gp_emulator(sim, obs, prior_ppf = list(-2, 4), n_waves = 2L, n_design = 12L, n_sim = 5L)
names(fit)
#>  [1] "estimate"          "grid"              "log_likelihood"   
#>  [4] "log_likelihood_sd" "posterior"         "waves"            
#>  [7] "ensemble_size"     "lengthscale"       "tau2"             
#> [10] "beta"              "n_simulations"     "method"           
# }