From the best poolFrac of the distinct points by value (at least NP), best first, each point that is at least minDist from every point already picked, the distance in GP lengthscales (theta) of the unit cube. While fewer than NP are picked, the distance is halved and the pool is gone through again for more, keeping the points already picked (so each good region gets a point before any gets a second; at 0 every pool point qualifies).

emulatorSpreadPopulation(X, value, theta, NP, poolFrac, minDist)

Arguments

X

matrix; distinct points, unit cube.

value

numeric; one per row of X (the mean of its evaluations), lower is better.

theta

numeric; the GP's lengthscales, one per column of X (or one for all).

NP

integer.

poolFrac, minDist

numeric.

Value

integer; rows of X, in the order picked, min(NP, nrow(X)) of them.