Update the default set of Ax transforms used in MBM (#3144)

Summary:
Pull Request resolved: #3144

This diff adds a new set of transforms (to be used by default in MBM based models) that replaces `IntToFloat` with `LogIntToFloat` and `UnitX` with `Normalize`. The new set of transforms avoid using continuous relaxation for non log-scale discrete parameter, which consistently delivers improved optimization performance on mixed integer benchmark problems.

This diff only updates single task model registry entries. I will follow it up with additional diffs to propagage the changes in multiple stages.

Reviewed By: dme65

Differential Revision: D66724547

fbshipit-source-id: 90fe3fb97048588138eb946dbe42a38e6b7481de

ax/modelbridge/registry.py

@@ -40,7 +40,7 @@
 from ax.modelbridge.transforms.derelativize import Derelativize
 from ax.modelbridge.transforms.fill_missing_parameters import FillMissingParameters
 from ax.modelbridge.transforms.int_range_to_choice import IntRangeToChoice
-from ax.modelbridge.transforms.int_to_float import IntToFloat
+from ax.modelbridge.transforms.int_to_float import IntToFloat, LogIntToFloat
 from ax.modelbridge.transforms.ivw import IVW
 from ax.modelbridge.transforms.log import Log
 from ax.modelbridge.transforms.logit import Logit
@@ -76,6 +76,10 @@
 
 logger: Logger = get_logger(__name__)
 
+# This set of transforms uses continuous relaxation to handle discrete parameters.
+# All candidate generation is done in the continuous space, and the generated
+# candidates are rounded to fit the original search space. This is can be
+# suboptimal when there are discrete parameters with a small number of options.
 Cont_X_trans: list[type[Transform]] = [
     FillMissingParameters,
     RemoveFixed,
@@ -87,8 +91,30 @@
     UnitX,
 ]
 
+# This is a modification of Cont_X_trans that aims to avoid continuous relaxation
+# where possible. It replaces IntToFloat with LogIntToFloat, which is only transforms
+# log-scale integer parameters, which still use continuous relaxation. Other discrete
+# transforms will remain discrete. When used with MBM, a Normalize input transform
+# will be added to replace the UnitX transform. This setup facilitates the use of
+# optimize_acqf_mixed_alternating, which is a more efficient acquisition function
+# optimizer for mixed discrete/continuous problems.
+MBM_X_trans: list[type[Transform]] = [
+    FillMissingParameters,
+    RemoveFixed,
+    OrderedChoiceToIntegerRange,
+    OneHot,
+    LogIntToFloat,
+    Log,
+    Logit,
+]
+
+
 Discrete_X_trans: list[type[Transform]] = [IntRangeToChoice]
 
+# This is a modification of Cont_X_trans that replaces OneHot and
+# OrderedChoiceToIntegerRange with ChoiceToNumericChoice. This results in retaining
+# all choice parameters as discrete, while using continuous relaxation for integer
+# valued RangeParameters.
 Mixed_transforms: list[type[Transform]] = [
     FillMissingParameters,
     RemoveFixed,
@@ -155,7 +181,7 @@ class ModelSetup(NamedTuple):
     "BoTorch": ModelSetup(
         bridge_class=TorchModelBridge,
         model_class=ModularBoTorchModel,
-        transforms=Cont_X_trans + Y_trans,
+        transforms=MBM_X_trans + Y_trans,
         standard_bridge_kwargs=STANDARD_TORCH_BRIDGE_KWARGS,
     ),
     "Legacy_GPEI": ModelSetup(
@@ -204,7 +230,7 @@ class ModelSetup(NamedTuple):
     "SAASBO": ModelSetup(
         bridge_class=TorchModelBridge,
         model_class=ModularBoTorchModel,
-        transforms=Cont_X_trans + Y_trans,
+        transforms=MBM_X_trans + Y_trans,
         default_model_kwargs={
             "surrogate_spec": SurrogateSpec(
                 botorch_model_class=SaasFullyBayesianSingleTaskGP

ax/modelbridge/tests/test_dispatch_utils.py

@@ -19,8 +19,7 @@
     choose_generation_strategy,
     DEFAULT_BAYESIAN_PARALLELISM,
 )
-from ax.modelbridge.factory import Cont_X_trans, Y_trans
-from ax.modelbridge.registry import Mixed_transforms, Models
+from ax.modelbridge.registry import MBM_X_trans, Mixed_transforms, Models, Y_trans
 from ax.modelbridge.transforms.log_y import LogY
 from ax.modelbridge.transforms.winsorize import Winsorize
 from ax.models.winsorization_config import WinsorizationConfig
@@ -44,7 +43,7 @@ class TestDispatchUtils(TestCase):
 
     @mock_botorch_optimize
     def test_choose_generation_strategy(self) -> None:
-        expected_transforms = [Winsorize] + Cont_X_trans + Y_trans
+        expected_transforms = [Winsorize] + MBM_X_trans + Y_trans
         expected_transform_configs = {
             "Winsorize": {"derelativize_with_raw_status_quo": False},
             "Derelativize": {"use_raw_status_quo": False},

ax/service/tests/test_ax_client.py

@@ -57,9 +57,8 @@
 )
 from ax.modelbridge.model_spec import ModelSpec
 from ax.modelbridge.random import RandomModelBridge
-from ax.modelbridge.registry import Models
+from ax.modelbridge.registry import Cont_X_trans, Models
 from ax.runners.synthetic import SyntheticRunner
-
 from ax.service.ax_client import AxClient, ObjectiveProperties
 from ax.service.utils.best_point import (
     get_best_parameters_from_model_predictions_with_trial_index,
@@ -220,7 +219,14 @@ def get_client_with_simple_discrete_moo_problem(
     gs = GenerationStrategy(
         steps=[
             GenerationStep(model=Models.SOBOL, num_trials=3),
-            GenerationStep(model=Models.BOTORCH_MODULAR, num_trials=-1),
+            GenerationStep(
+                model=Models.BOTORCH_MODULAR,
+                num_trials=-1,
+                model_kwargs={
+                    # To avoid search space exhausted errors.
+                    "transforms": Cont_X_trans,
+                },
+            ),
         ]
     )