def _init_optimizer(
self,
trans: _SearchSpaceTransform,
direction: StudyDirection,
population_size: Optional[int] = None,
randomize_start_point: bool = False,
) -> CmaClass:
lower_bounds = trans.bounds[:, 0]
upper_bounds = trans.bounds[:, 1]
n_dimension = len(trans.bounds)
if self._source_trials is None:
if randomize_start_point:
mean = lower_bounds + (upper_bounds - lower_bounds
) * self._cma_rng.rand(n_dimension)
elif self._x0 is None:
mean = lower_bounds + (upper_bounds - lower_bounds) / 2
else:
# `self._x0` is external representations.
mean = trans.transform(self._x0)
if self._sigma0 is None:
sigma0 = np.min((upper_bounds - lower_bounds) / 6)
else:
sigma0 = self._sigma0
cov = None
else:
expected_states = [TrialState.COMPLETE]
if self._consider_pruned_trials:
expected_states.append(TrialState.PRUNED)
# TODO(c-bata): Filter parameters by their values instead of checking search space.
sign = 1 if direction == StudyDirection.MINIMIZE else -1
source_solutions = [
(trans.transform(t.params), sign * cast(float, t.value))
for t in self._source_trials if t.state in expected_states
and _is_compatible_search_space(trans, t.distributions)
]
if len(source_solutions) == 0:
raise ValueError("No compatible source_trials")
# TODO(c-bata): Add options to change prior parameters (alpha and gamma).
mean, sigma0, cov = get_warm_start_mgd(source_solutions)
# Avoid ZeroDivisionError in cmaes.
sigma0 = max(sigma0, _EPS)
if self._use_separable_cma:
return SepCMA(
mean=mean,
sigma=sigma0,
bounds=trans.bounds,
seed=self._cma_rng.randint(1, 2**31 - 2),
n_max_resampling=10 * n_dimension,
population_size=population_size,
)
return CMA(
mean=mean,
sigma=sigma0,
cov=cov,
bounds=trans.bounds,
seed=self._cma_rng.randint(1, 2**31 - 2),
n_max_resampling=10 * n_dimension,
population_size=population_size,
)
def _get_trans_params(trials: List[FrozenTrial], trans: _SearchSpaceTransform) -> numpy.ndarray:
return numpy.array([trans.transform(trial.params) for trial in trials])