search

uber-research / TuRBO

Other
186 stars 53 forks source link

Question about local fitting of GPs #4

Open trbedwards opened 2 years ago

trbedwards commented 2 years ago

Excellent BO algorithm by the way, but I'm a bit confused about how the GP is made local via the trust region in Turbo-1.

From reading your paper, I got the idea that GP is only trained with local X, Y data (using data that fits within the trust region). However, when looking at the code, I can see that upon each iteration of the optimize loop, a new GP is trained using the full X and Y datasets (since we use self._X, which is global).

Have I misunderstood how this works?

CharlyEmpereurmot commented 1 year ago

Looking at the code I am having doubts too. I have been timing the GP fittings per trust region for Turbo-M and I only see increasing durations of fitting times until the TR converges, suggesting all points are used for fitting the GP within a TR regardless of adapted boundaries.

@dme65 Is this not implemented as it was intended?

Cheers for the great design!