Probabilities of Feasibility for Classifier based constraints in Acquisition Functions

botorch/acquisition/monte_carlo.py

@@ -47,7 +47,10 @@
 from botorch.exceptions.warnings import legacy_ei_numerics_warning
 from botorch.models.model import Model
 from botorch.sampling.base import MCSampler
-from botorch.utils.objective import compute_smoothed_feasibility_indicator
+from botorch.utils.objective import (
+    compute_probabilities_of_feasibility_indicator,
+    compute_smoothed_feasibility_indicator,
+)
 from botorch.utils.transforms import (
     concatenate_pending_points,
     match_batch_shape,
@@ -188,6 +191,7 @@ def __init__(
         sample_reduction: SampleReductionProtocol = torch.mean,
         q_reduction: SampleReductionProtocol = torch.amax,
         constraints: list[Callable[[Tensor], Tensor]] | None = None,
+        probabilities_of_feasibility: list[Callable[[Tensor], Tensor]] | None = None,
         eta: Tensor | float = 1e-3,
         fat: bool = False,
     ):
@@ -248,6 +252,7 @@ def __init__(
         self._sample_reduction = partial(sample_reduction, dim=sample_dim)
         self._q_reduction = partial(q_reduction, dim=-1)
         self._constraints = constraints
+        self._probabilities_of_feasibility = probabilities_of_feasibility
         self._eta = eta
         self._fat = fat
 
@@ -328,6 +333,18 @@ def _apply_constraints(self, acqval: Tensor, samples: Tensor) -> Tensor:
                 fat=self._fat,
             )
             acqval = acqval.add(ind) if self._log else acqval.mul(ind)
+        if self._probabilities_of_feasibility is not None:
+            if not self._log and (acqval < 0).any():
+                raise ValueError(
+                    "Constraint-weighting requires unconstrained "
+                    "acquisition values to be non-negative."
+                )
+            ind = compute_probabilities_of_feasibility_indicator(
+                probabilities_of_feasibility=self._probabilities_of_feasibility,
+                samples=samples,
+                log=self._log,
+            )
+            acqval = acqval.add(ind) if self._log else acqval.mul(ind)
         return acqval
 
 

botorch/utils/objective.py

@@ -180,6 +180,18 @@ def compute_smoothed_feasibility_indicator(
     return is_feasible if log else is_feasible.exp()
 
 
+def compute_probabilities_of_feasibility_indicator(
+    probabilties_of_feasibility: list[Callable[[Tensor], Tensor]],
+    samples: Tensor,
+    log: bool = False,
+) -> Tensor:
+    is_feasible = torch.zeros_like(samples[..., 0])
+    for constraint in probabilties_of_feasibility:
+        # 1 means feasible, zero means not feasible
+        is_feasible = is_feasible + constraint(samples).log()
+    return is_feasible if log else is_feasible.exp()
+
+
 def apply_constraints(
     obj: Tensor,
     constraints: list[Callable[[Tensor], Tensor]],