lbl-camera / fvgp Goto Github PK

Although the scipy.optimize.differential_evolution supports multiprocessing, and at least some of the acquisition functions are vectorized, neither are exploited during GPOptimizer.ask(...). The design of differential_evolution doesn't seem to allow the function to be evaluated in a vectorized way. While specifying additional workers could provide some benefit, a decision could also be made to implement a strategy where calls to acquisition functions actually receive multiple points, allowing vectorization to be leveraged.

Options for improving on this design may be a topic of investigation. @MarcusMNoack

In gp.py, np.fill_diagonal is used incorrectly. This function operates in-place with no return value.

The above will only ever assign None to S.

I would submit a PR myself, but I'm also unclear on the meaning of S(x). Could this also be documented?