Overflow when setting bounds for NGOpt

[nepfaff]

Steps to reproduce

import sys

import nevergrad as ng
import numpy as np

x_array = ng.p.Array(init=np.ones(10)).set_bounds(
    -sys.float_info.max, sys.float_info.max
)
optimizer = ng.optimizers.NGOpt(parametrization=x_array, budget=1000)

def cost(x):
    return np.linalg.norm(x)

optimizer.minimize(cost)

Observed Results

I'm getting the following:

RuntimeWarning: overflow encountered in subtract
return trans.Affine(self.bounds[1] - self.bounds[0], self.bounds[0]).reverted() # type: ignore
UserWarning: Failed to find bounds for Array{(10,),Cl(-179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368,179769313486231570814527423731704356798070567525844996598917476803157260780028538760589558632766878171540458953514382464234321326889464182768467546703537516986049910576551282076245490090389328944075868508455133942304583236903222948165808559332123348274797826204144723168738177180919299881250404026184124858368,b)}:[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.], quasi-random optimizer may be inefficient.
Please open an issue on Nevergrad github
warnings.warn(
python3.10/site-packages/scipy/optimize/_optimize.py:2989: RuntimeWarning: overflow encountered in scalar multiply
w = xb - ((xb - xc) * tmp2 - (xb - xa) * tmp1) / denom

Changing the bounds to something more reasonable like [-1e10, 1e10] doesn't produce the first errors but still results in the following:

python3.10/site-packages/scipy/optimize/_optimize.py:2989: RuntimeWarning: overflow encountered in scalar multiply
w = xb - ((xb - xc) * tmp2 - (xb - xa) * tmp1) / denom

Reducing the budget to 100 gives the following error instead:

python3.10/site-packages/nevergrad/parametrization/data.py:309: RuntimeWarning: overflow encountered in subtract
return reference._to_reduced_space(self._value - reference._value)

Lastly, using bounds [-np.inf, np.inf] works for some optimizers but not others. For example, NGOpt with a budget of 100 works just fine, but NGOpt with a budget of 1000 returns nan without warning.

Expected Results

I want to be able to set large bounds without weird errors. In particular, I have a large vector of decision variables, and some of them have bounds while others do not. Ideally, I would want to set bounds using set_bounds for the entire vector, where the variables without bounds have some large values or np.inf.

[Basile-Terv]

Hello, I am having the same issue, did you find a solution ?