Broadcast scalar constraint bounds before constructing surrogate models (#628)

Author: aalquwayfili

bayes_opt/constraint.py

@@ -37,10 +37,12 @@ class ConstraintModel:
     lb : float or np.ndarray
         The lower bound on the constraints. Should have the same
         dimensionality as the return value of the constraint function.
+        A scalar is broadcast to match the upper bound.
 
     ub : float or np.ndarray
         The upper bound on the constraints. Should have the same
         dimensionality as the return value of the constraint function.
+        A scalar is broadcast to match the lower bound.
 
     random_state : np.random.RandomState or int or None, default=None
         Random state to use.
@@ -62,8 +64,7 @@ def __init__(
     ) -> None:
         self.fun = fun
 
-        self._lb = np.atleast_1d(lb)
-        self._ub = np.atleast_1d(ub)
+        self._lb, self._ub = np.broadcast_arrays(np.atleast_1d(lb), np.atleast_1d(ub))
 
         if np.any(self._lb >= self._ub):
             msg = "Lower bounds must be less than upper bounds."

tests/test_constraint.py

@@ -125,6 +125,36 @@ def constraint_function_2_dim(x, y):
     )
 
 
+@pytest.mark.parametrize(
+    ("lb", "ub"),
+    [
+        (-np.inf, np.array([0.5, 0.25])),
+        (np.array([-0.5, -0.25]), np.inf),
+        (-0.5, np.array([0.5, 0.25])),
+        (np.array([-0.5, -0.25]), 0.5),
+        (np.array([-np.inf, -np.inf]), np.array([0.5, 0.25])),
+        (np.array([-0.5, -0.25]), np.array([np.inf, np.inf])),
+    ],
+)
+def test_multiple_constraints_broadcast_bounds(lb, ub):
+    constraint = NonlinearConstraint(lambda x: np.array([x, -x]), lb, ub)
+    optimizer = BayesianOptimization(
+        f=lambda x: -(x**2), constraint=constraint, pbounds={"x": (-1, 1)}, verbose=0, random_state=1
+    )
+
+    for x in [-0.75, 0.0, 0.75]:
+        optimizer.probe({"x": x}, lazy=False)
+
+    np.testing.assert_array_equal(
+        optimizer.space.constraint_values, [[-0.75, 0.75], [0.0, 0.0], [0.75, -0.75]]
+    )
+    assert [result["allowed"] for result in optimizer.res] == [False, True, False]
+    assert optimizer.max["params"] == {"x": 0.0}
+    assert -1 <= optimizer.suggest()["x"] <= 1
+    probabilities = optimizer.constraint.predict(np.array([[-0.75], [0.0], [0.75]]))
+    np.testing.assert_allclose(probabilities, [0.0, 1.0, 0.0], atol=1e-6)
+
+
 def test_kwargs_not_the_same(target_function):
     def target_function(x, y):
         return np.cos(2 * x) * np.cos(y) + np.sin(x)