Optimizer for acquisition function
Issue #31
open
Use of brute-force non-ideal for large dimensions - consider using one of the gradient-based techniques (available from scipy's 'optimize' package)
Comments (5)
-
-
- changed status to resolved
[major] Added default acquisition function optimizer
No major change to the user code.. added functionality
- Uses powell method instead of brute. should resolve #31
- general function signature: acq_optimizer(func: Callable, x0: np.array, bounds, func_deriv: Callable=None) -> np.array:
→ <<cset ae87cb682d86>>
-
- changed status to open
Still has issues with the powell optimizer. Could lead to problems in lower dimensions or with small kappa
-
One potential method is to use a default acquisition optimizator as a generator of local optima of acquisition and check if either of them is useful
-
-
assigned issue to
-
assigned issue to
- Log in to comment
This shall be addressed by allowing the user to change the
acquisition_function_optimizer
, just like how the user can change thekernel_function
The default will be changed frombrute
to a better robust method.Should be easy to change without much effect on the user code. (easy backward compatibility)