Replace `Problem.x_guesses`?

[dweindl]

There is pypesto.problem.base.Problem.x_guesses which is used to pre-select startpoints for optimization. By default, Problem.startpoint_method will return points from x_guesses if that's set. Repeated calls will always return the same points. While this may be convenient in a few situations, it's a bit of a footgun elsewhere.

Are there any uses cases for x_guesses outside multi-start optimization (pypesto.optimize.optimize.minimize)? If not, I would propose removing Problem.x_guesses and instead adding an x0 argument to minimize().

[PaulJonasJost]

If you don't want the x_guesses I think one can just turn them off with startpoint_method.use_guesses=False, no?

Not sure on the advantages/problems here. Sure, in reapted calls, this might not be nice, but it might also be exactly what we want if we changed optimizers, etc, right?

Where I do agree, though, is that it is unnatural if I explicitly call the startpoint_method to return x_guesses. So maybe instead move the use_guesses from the startpoint method to the minimize()?

[dweindl]

My problem is that it is not very obvious (and neither well documented) how x_guesses interacts with the rest of pypesto.

If the only meaningful use case is supplying startpoints for multi-start optimization, then I think it would be way less surprising and error-prone to just supply those directly to the respective minimize call.

For any other applications, I am not convinced that it is currently well defined or understood what x_guesses is doing or supposed to do.

[PaulJonasJost]

Agreed. I am not aware of any other interaction of pypesto with x_guesses. Would vote for integrating it explicitely into the minimize() function and remove it from the startpoint method.