Documentation and general usage - Optimizer - differential_evolution about pypesto HOT 7 CLOSED

So if you have defined x_guessesin the problem, use pypesto.optimize.ScipyDifferentialEvolutionOptimizer and call pypesto.optimize.minimze, the following will happen:

• in pypesto/optimize/optimize.py line 89-93, if you have not defined a startpointmethod, it will use uniform as start point method, which of the class FunctionStartpoints, a subclass of CheckedStartpoints.
• In lines 105-109 of optimize, the start points are assigned, which uses as a call argument the problem.
• This call function, given you use default options will return x_guesses, which will then be supplied to the optimizer as argument of x_0 here

I think that is the argument line here. I hope this helps in clarifying it :)

from pypesto.

So if you have defined x_guessesin the problem, use pypesto.optimize.ScipyDifferentialEvolutionOptimizer and call pypesto.optimize.minimze, the following will happen:

• in pypesto/optimize/optimize.py line 89-93, if you have not defined a startpointmethod, it will use uniform as start point method, which of the class FunctionStartpoints, a subclass of CheckedStartpoints.
• In lines 105-109 of optimize, the start points are assigned, which uses as a call argument the problem.
• This call function, given you use default options will return x_guesses, which will then be supplied to the optimizer as argument of x_0 here

I think that is the argument line here. I hope this helps in clarifying it :)

Though I might have to correct myself, I think it is not passed on, as it is lost in the minimise function. Thanks a lot for the question, will fix that and also clarify in the documentation.

from pypesto.

will also have a look at the PySwarms optimizer 👍🏼

from pypesto.

Many thanks 🙏🏼 A fix for the ScipyDifferentialEvolutionOptimizer as well as at least a warning for the pyswarms optimizer will be released within this week :)

from pypesto.

merged at least warning for optimizers in #1027.

from pypesto.

This does help, but I have been doing the same thing, walking through code to figure out how potentially conflicting options are handled. For instance, I thought that it would send the x_guesses in the PyswarmsOptimizer to act as init_pos for the particles in the swarm. But instead first x0 is set by startpoint.

Then x0 is used only to define dimension, not init_pos.

I think some documentation describing options could be useful, but if it's too much work we can always figure things out by looking at src.

from pypesto.

@PaulJonasJost I've taken some notes on my exploration. They might be useful to you. Excuse them for their brevity. I think the last example showing generate_swarm will be informative.

• Say we set pypesto.problem.x_guesses ($n \times d$ array). Then the following loop will occur $n$ times, with random starts when using the PySwarms optimizer. This works well as we will have particles starting in new locations.

for startpoint, id in zip(startpoints, ids):
        task = OptimizerTask(
            optimizer=optimizer,
            problem=problem,
            x0=startpoint,
            id=id,
            history_options=history_options,
            optimize_options=options,
        )
        tasks.append(task)

• If the PySwarms option init_pos=x_guesses then there will still be $n$ loops but then the number of particles par_popsize will be ignored, and be len(x_guesses), and the initialization will be the same for each new set of particles. This would not be desireable for properly running the optimizer. Because of this is probably best to advise against using init_pos.

• side note: x0 is evaluated for both init_pos=x_guesses and the default init_pos=None.

• I've tested how particles are generated to show this. If we specify the init_pos, n_particles is overwritten and we only get an initialization of the size of init_pos.

>>> from pyswarms.backend.generators import generate_swarm
>>> init_pos = np.ones((2,3))
>>> generate_swarm(n_particles=4, dimensions=init_pos.shape[1],init_pos=init_pos)
array([[1., 1., 1.],
       [1., 1., 1.]])

The dimension arg needs to be specified, but is overwritten by kwarg init_pos.

>>> generate_swarm(n_particles=4, dimensions=5,init_pos=init_pos)
array([[1., 1., 1.],
       [1., 1., 1.]])

init_pos=None is normal default behavior, with random initialization of particles. Doing this we actually get 4 particles.

>>> generate_swarm(n_particles=4, dimensions=init_pos.shape[1],init_pos=None)
array([[0.17696656, 0.03424968, 0.55395097],
       [0.85859682, 0.5250245 , 0.41581642],
       [0.64526319, 0.04922589, 0.31136249],
       [0.53644966, 0.48004648, 0.32024561]])

from pypesto.