On the full bounds, in emulatorSpace(): a space-filling design (lhs::maximinLHS(), plus replicates), then batches chosen on a GP (hetGP::mleHetGP(), Matern 5/2) fitted to every evaluation so far, on a capped log scale (see emulatorDefaults()): expected improvement over the GP's best mean (greedy, with an exclusion distance), a lower confidence bound for exploration, and replicates of the best points. A global phase, until the GP's best mean stops improving, then trust-region rounds (emulatorTrustBox(), emulatorTrustStep()), each re-centred on the best of all the points so far, shrinking while they improve by more than the objective's noise, until they stop improving or the budget is spent.

emulateObjective(evaluate, lower, upper, NP, control = list(), seed = 1L)

Arguments

evaluate

a function of a matrix of parameter sets (one per row, columns names(lower)) and an integer seed, returning one objective value per row (emulatorEvaluator()).

lower, upper

the bounds, named.

NP

integer; how many points to return.

control

list; see emulatorDefaults().

seed

integer; the design, the candidates and the evaluations' seeds follow from it.

Value

a list: pop, NP distinct evaluated points, good and spread out (emulatorSpreadPopulation(); fewer if fewer were evaluated), columns names(lower); objFunVal, the means of their evaluations; evals, every evaluation (batch, kind, value and the parameters); status, one row per batch (with the trust round, its scale and its box's widest side, boxWidth); stopReason.