Add step parameter to float hyperparameters

- if set to None, no discretization performed (default behaviour)
- if specified in suggest_with_optuna(), it overrides the step attribute
- if not specified, we use the attribute (which can be None or a float)

discrete_optimization/generic_tools/hyperparameters/hyperparameter.py

@@ -181,6 +181,9 @@ class FloatHyperparameter(Hyperparameter):
 
     """
 
+    step: Optional[float] = None
+    """step to discretize if not None."""
+
     log: bool = False
     """Whether to sample the value in a logarithmic scale."""
 
@@ -226,6 +229,9 @@ def suggest_with_optuna(
             low: can be used to restrict lower bound
             high: can be used to restrict upper bound
             log: whether to sample the value in a logarithmic scale
+            step: step of discretization if specified.
+                If explicitely set to None, no discretization performed.
+                By default, use self.step (and thus default discretization only if self.step not None)
             prefix: prefix to add to optuna corresponding parameter name
               (useful for disambiguating hyperparameters from subsolvers in case of meta-solvers)
             **kwargs: passed to `trial.suggest_float()`
@@ -239,8 +245,12 @@ def suggest_with_optuna(
             high = self.high
         if log is None:
             log = self.log
+        if "step" in kwargs:
+            step = kwargs.pop("step")
+        else:
+            step = self.step
         return trial.suggest_float(
-            name=prefix + self.name, low=low, high=high, log=log, **kwargs
+            name=prefix + self.name, low=low, high=high, log=log, step=step, **kwargs  # type: ignore
         )
 
 

tests/generic_tools/hyperparameters/test_hyperparameter.py

@@ -33,7 +33,7 @@ class BigMethod(Enum):
 class DummySolver(SolverDO):
     hyperparameters = [
         IntegerHyperparameter("nb", low=0, high=2, default=1),
-        FloatHyperparameter("coeff", low=-1.0, high=1.0, default=1.0),
+        FloatHyperparameter("coeff", low=-1.0, high=1.0, default=1.0, step=0.25),
         CategoricalHyperparameter("use_it", choices=[True, False], default=True),
         EnumHyperparameter("method", enum=Method, default=Method.GREEDY),
     ]
@@ -255,6 +255,36 @@ def objective(trial: optuna.Trial) -> float:
     assert len(study.trials) == 2 * 2 * 5 * 1
 
 
+def test_suggest_with_optuna_default_float_step():
+    def objective(trial: optuna.Trial) -> float:
+        # hyperparameters for the chosen solver
+        suggested_hyperparameters_kwargs = (
+            DummySolver.suggest_hyperparameters_with_optuna(
+                trial=trial,
+                kwargs_by_name={
+                    "nb": dict(high=1),
+                    "use_it": dict(choices=[True]),
+                },
+            )
+        )
+        assert len(suggested_hyperparameters_kwargs) == 4
+        assert isinstance(suggested_hyperparameters_kwargs["method"], Method)
+        assert 0 <= suggested_hyperparameters_kwargs["nb"]
+        assert 1 >= suggested_hyperparameters_kwargs["nb"]
+        assert -1.0 <= suggested_hyperparameters_kwargs["coeff"]
+        assert 1.0 >= suggested_hyperparameters_kwargs["coeff"]
+        assert suggested_hyperparameters_kwargs["use_it"] is True
+
+        return 0.0
+
+    study = optuna.create_study(
+        sampler=optuna.samplers.BruteForceSampler(),
+    )
+    study.optimize(objective)
+
+    assert len(study.trials) == 2 * 2 * 9 * 1
+
+
 def test_suggest_with_optuna_with_choices_dict():
     def objective(trial: optuna.Trial) -> float:
         # hyperparameters for the chosen solver