Remove TPESampler for MO

package/samplers/auto_sampler/sampler.py

@@ -9,14 +9,11 @@
 from optuna.samplers import BaseSampler
 from optuna.samplers import CmaEsSampler
 from optuna.samplers import GPSampler
-from optuna.samplers import NSGAIIISampler
 from optuna.samplers import NSGAIISampler
 from optuna.samplers import RandomSampler
 from optuna.samplers import TPESampler
 from optuna.samplers._base import _process_constraints_after_trial
 from optuna.samplers._lazy_random_state import LazyRandomState
-from optuna.samplers._nsgaiii._sampler import _GENERATION_KEY as NSGA3_GENERATION_KEY
-from optuna.samplers.nsgaii._sampler import _GENERATION_KEY as NSGA2_GENERATION_KEY
 from optuna.search_space import IntersectionSearchSpace
 from optuna.trial import TrialState
 
@@ -28,13 +25,11 @@
 
 
 MAXINT32 = (1 << 31) - 1
-THRESHOLD_OF_MANY_OBJECTIVES = 4
 SAMPLER_KEY = "auto:sampler"
 
 
 class AutoSampler(BaseSampler):
     _N_COMPLETE_TRIALS_FOR_CMAES = 250
-    _N_COMPLETE_TRIALS_FOR_NSGA = 1000
 
     """Sampler automatically choosing an appropriate sampler based on search space.
 
@@ -102,31 +97,11 @@ def _include_conditional_param(self, study: Study) -> bool:
     def _determine_multi_objective_sampler(
         self, study: Study, trial: FrozenTrial, search_space: dict[str, BaseDistribution]
     ) -> BaseSampler:
-        if isinstance(self._sampler, (NSGAIISampler, NSGAIIISampler)):
+        if isinstance(self._sampler, NSGAIISampler):
             return self._sampler
 
         seed = self._rng.rng.randint(MAXINT32)
-        if not isinstance(self._sampler, TPESampler):
-            return TPESampler(
-                seed=seed,
-                multivariate=True,
-                warn_independent_sampling=False,
-                constraints_func=self._constraints_func,
-                constant_liar=True,
-            )
-
-        complete_trials = study.get_trials(deepcopy=False, states=(TrialState.COMPLETE,))
-        complete_trials.sort(key=lambda trial: trial.datetime_complete)
-        if len(complete_trials) >= self._N_COMPLETE_TRIALS_FOR_NSGA:
-            nsga_sampler_cls = (
-                NSGAIISampler
-                if len(study.directions) < THRESHOLD_OF_MANY_OBJECTIVES
-                else NSGAIIISampler
-            )
-            # Use NSGA-II/III if len(complete_trials) <= _N_COMPLETE_TRIALS_FOR_NSGA.
-            return nsga_sampler_cls(constraints_func=self._constraints_func, seed=seed)
-
-        return self._sampler  # No update happens to self._sampler.
+        return NSGAIISampler(constraints_func=self._constraints_func, seed=seed)
 
     def _determine_single_objective_sampler(
         self, study: Study, trial: FrozenTrial, search_space: dict[str, BaseDistribution]
@@ -187,18 +162,6 @@ def infer_relative_search_space(
     def sample_relative(
         self, study: Study, trial: FrozenTrial, search_space: dict[str, BaseDistribution]
     ) -> dict[str, Any]:
-        n_objectives = len(study.directions)
-        if n_objectives > 1 and isinstance(self._sampler, TPESampler):
-            # NOTE(nabenabe): Set generation 0 so that NSGA-II/III can use the trial information
-            # obtained during the optimization using TPESampler.
-            # NOTE(nabenabe): Use NSGA-III for many objective problems.
-            _GENERATION_KEY = (
-                NSGA2_GENERATION_KEY
-                if n_objectives < THRESHOLD_OF_MANY_OBJECTIVES
-                else NSGA3_GENERATION_KEY
-            )
-            study._storage.set_trial_system_attr(trial._trial_id, _GENERATION_KEY, 0)
-
         return self._sampler.sample_relative(study, trial, search_space)
 
     def sample_independent(
@@ -211,8 +174,10 @@ def sample_independent(
         return self._sampler.sample_independent(study, trial, param_name, param_distribution)
 
     def before_trial(self, study: Study, trial: FrozenTrial) -> None:
-        states_of_interest = [TrialState.COMPLETE, TrialState.WAITING]
-        if len(study._get_trials(deepcopy=False, states=states_of_interest)) != 0:
+        # NOTE(nabenabe): Sampler must be updated in this method. If, for example, it is updated in
+        # infer_relative_search_space, the sampler for before_trial and that for sample_relative,
+        # after_trial might be different, meaning that the sampling routine could be incompatible.
+        if len(study._get_trials(deepcopy=False, states=(TrialState.COMPLETE,))) != 0:
             search_space = IntersectionSearchSpace().calculate(study)
             self._sampler = self._determine_sampler(study, trial, search_space)
 

package/samplers/auto_sampler/tests/test_auto_sampler.py

@@ -70,33 +70,25 @@ def _check_constraints_of_all_trials(study: optuna.Study) -> None:
 
 
 @parametrize_constraints
-def test_choose_nsga3(use_constraint: bool) -> None:
-    n_trials_of_nsga = 100
-    n_trials_before_nsga = 100
+def test_choose_for_many_objective(use_constraint: bool) -> None:
+    n_trials = 200
     auto_sampler = AutoSampler(constraints_func=constraints_func if use_constraint else None)
-    auto_sampler._N_COMPLETE_TRIALS_FOR_NSGA = n_trials_before_nsga
     study = optuna.create_study(sampler=auto_sampler, directions=["minimize"] * 4)
-    study.optimize(many_objective, n_trials=n_trials_before_nsga + n_trials_of_nsga)
+    study.optimize(many_objective, n_trials=n_trials)
     sampler_names = _get_used_sampler_names(study)
-    assert ["RandomSampler"] + ["TPESampler"] * (n_trials_before_nsga - 1) + [
-        "NSGAIIISampler"
-    ] * n_trials_of_nsga == sampler_names
+    assert ["RandomSampler"] + ["NSGAIISampler"] * (n_trials - 1) == sampler_names
     if use_constraint:
         _check_constraints_of_all_trials(study)
 
 
 @parametrize_constraints
-def test_choose_nsga2(use_constraint: bool) -> None:
-    n_trials_of_nsga = 100
-    n_trials_before_nsga = 100
+def test_choose_for_multi_objective(use_constraint: bool) -> None:
+    n_trials = 200
     auto_sampler = AutoSampler(constraints_func=constraints_func if use_constraint else None)
-    auto_sampler._N_COMPLETE_TRIALS_FOR_NSGA = n_trials_before_nsga
     study = optuna.create_study(sampler=auto_sampler, directions=["minimize"] * 2)
-    study.optimize(multi_objective, n_trials=n_trials_before_nsga + n_trials_of_nsga)
+    study.optimize(multi_objective, n_trials=n_trials)
     sampler_names = _get_used_sampler_names(study)
-    assert ["RandomSampler"] + ["TPESampler"] * (n_trials_before_nsga - 1) + [
-        "NSGAIISampler"
-    ] * n_trials_of_nsga == sampler_names
+    assert ["RandomSampler"] + ["NSGAIISampler"] * (n_trials - 1) == sampler_names
     if use_constraint:
         _check_constraints_of_all_trials(study)