berndbischl
commented
7 years ago @jakob-r why not simply add the note as daniel asked? i mean the optimizer IS configurable? and this is an issue not in our package? but in DK?
Open danielhorn opened 7 years ago
berndbischl
commented
7 years ago @jakob-r why not simply add the note as daniel asked? i mean the optimizer IS configurable? and this is an issue not in our package? but in DK?
Hi everyone,
at the moment, I am participating in the EMO real world competition, using our SMS-EGO implementation and mostly default settings. Among other things I use Kriging with rgenoud optimization. All small competition problems (2d to 10d) went really well (we will see in around 3 weeks how well they really went), but I had some rather huge problems with the 2 largest problems: One had dimension 24, one had dimension 20, and both of them started to get nasty after around 500 function evalutions (Hence, Kriging had to be fitted to a 500 times 20 matrix). After a each few iteration my R-Session crashed with the error "Gleitkomma Ausnahme (Speicherabzug geschrieben)". I solved it by switchig back from RGenoud to BFGS optimization in the Kriging learner - presumably the rgenoud is broken in this setting.
It's really hard to give a reproducable example code, since the error came a) stochastically and b) after several minutes to hours of computation time and c) I only have a limited budget in the competition, which I should not use for debugging. But it seems, that km and rgenoud makes problems in high dimension / iteration numbers. It would be nice if someone has time to further investigate sometime. For now, we could perhaps add a note, that we noticed some trouble and recommend to use BFGS in this setting?