Think about computation of gaps between best point so far and known global optimum in noisy case and implement an heuristic. about mlrmbo HOT 6 CLOSED

I have no ingenious idea about how to solve this problem. Maybe simply take the best point so far, evaluate it n times and compute the mean gap, or to be more robust, the median gap?

from mlrmbo.

Lets ignore this for now and do it later.

from mlrmbo.

Do it like this:

we have the true y function (approximated with noisy.evals). call that g.

gap = g(x_current_best) - g(x_opt)

you can construct g as an R function by using "interpolate" (or so) on the averaged y-values

from mlrmbo.

Actually this

http://stat.ethz.ch/R-manual/R-patched/library/stats/html/approxfun.html

from mlrmbo.

Done.

from mlrmbo.

Now it is possible to pass the true objective function (not noisy) as another argument fun.mean. Only if this is provided by the user, gaps are computed in the noisy case: we extract the parameter values of the entire opt.path and apply fun.mean to these to get the correct values. The gap is than computed as in the deterministic case.

from mlrmbo.