Expected improvement over the GP's best mean at an evaluated point (greedy, each pick at least exclusion lengthscales from the others), then a lower confidence bound (exploration), then replicates of the best evaluated points (fewest replicates first). In the local phase the Latin-hypercube candidates are drawn in the trust region (emulatorTrustBox()) and more replicates are taken.

selectEmulatorBatch(m, U, B, trustScale, ctl)

Arguments

m

the GP (fitEmulatorGP()).

U

every evaluated point so far, unit cube.

B

integer; batch size.

trustScale

numeric; the trust region's scale in the local phase, NULL in the global phase.

ctl

emulatorControlOf().

Value

list: U (the batch, unit cube), kind ("ei", "lcb", "rep"), bestZ (the GP's best mean at an evaluated point, response scale), best (that point), box (the trust region, or NULL), theta (the GP's lengthscales).