Sporadic "out of bounds"-error in smbo for very large search-spaces

[SimonBlanke]

There appears to be a bug in the initial sampler of the sequence-model-based optimization algorithms that only occurs in the "test_large_search_space.py"-tests. This is the error:

gradient_free_optimizers/search.py:96: in search
    self.search_step(nth_trial)
gradient_free_optimizers/search.py:185: in search_step
    self.finish_initialization()
gradient_free_optimizers/optimizers/global_opt/lipschitz_optimization.py:59: in finish_initialization
    self.all_pos_comb = self._all_possible_pos()
gradient_free_optimizers/optimizers/smb_opt/smbo.py:112: in _all_possible_pos
    if self.conv.not_in_constraint(pos):
gradient_free_optimizers/optimizers/core_optimizer/converter.py:59: in not_in_constraint
    para = self.value2para(self.position2value(position))
gradient_free_optimizers/optimizers/core_optimizer/converter.py:72: in wrapper
    return func_(self, *args)


    @returnNoneIfArgNone
    def position2value(self, position: Optional[list]) -> Optional[list]:
        value = []
    
        for n, space_dim in enumerate(self.search_space_values):
>           value.append(space_dim[position[n]])
E           IndexError: index 201 is out of bounds for axis 0 with size 100

gradient_free_optimizers/optimizers/core_optimizer/converter.py:81: IndexError

For some reason this error started to appear after switching from python 3.8 to 3.10 on my local machine.

It is strange, that this error never occurred in the CI. And I have no idea, why this was no problem in python 3.8, since the methgrid-method (which is important for the position generation in the algorithm) has not changed its behaviour (as far as I can tell).