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?
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?