befelix
Effect of search space boundaries on exploration
(Only boundaries for GP changed in the 2 plots) Do you have any hints how a change in the boundaries effect exploration that much? Our understanding was, that boundaries shouldn't effect exploration at all but only act as a limit.
It shouldn't all inputs being equal (the code is deterministic). There are three main sources for behavior change that come to mind:
- Random observation noise can affect exploration.
- If you do hyperparameter optimization that can affect exploration
- If you're discretizing the space, the behavior can change depending on the discretization. This could be implicit: e.g., if you have 100 points per dimension, then changing the domain affects the discretization.
On Mon, 26 Oct 2020 at 17:00, Webbah [email protected] wrote:
[image: grafik] https://user-images.githubusercontent.com/1822491/97196157-6fa79700-17ac-11eb-9126-940ab900535a.png
(Only boundaries for GP changed in the 2 plots) Do you have any hints how a change in the boundaries effect exploration that much? Our understanding was, that boundaries shouldn't effect exploration at all but only act as a limit.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/befelix/SafeOpt/issues/7, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKGGKFWJ4DVXV3KEHWEX3TSMWMJ7ANCNFSM4S7SMU6Q .
We have random noise, but it is sampled from a seeded distribution. Hence, we have a deterministic behavior, i.e. with the same bounds we can reproduce each of the above shown plots exactly.