Fix wrong desirabilities in predictive strategies

bofire/data_models/domain/features.py

@@ -783,24 +783,39 @@
     def __call__(
         self,
         experiments: pd.DataFrame,
+        experiments_adapt: Optional[pd.DataFrame] = None,
         predictions: bool = False,
     ) -> pd.DataFrame:
         """Evaluate the objective for every feature.
 
         Args:
             experiments (pd.DataFrame): Experiments for which the objectives should be evaluated.
+            experiments_adapt (pd.DataFrame, optional): Experimental values which are used to update the objective
+                parameters on the fly. This is for example needed when a `MovingMaximizeSigmoidObjective` is used
+                as this depends on the best experimental value achieved so far. For this reason `experiments_adapt`
+                has to be provided if `predictions=True` ie. that the objectives of candidates are evaluated.
+                Defaults to None.
             predictions (bool, optional): If True use the prediction columns in the dataframe to calc the
-                desirabilities `f"{feat.key}_pred`.
+                desirabilities `f"{feat.key}_pred`, furthermore `experiments_adapt` has to be provided.
 
         Returns:
             pd.DataFrame: Objective values for the experiments of interest.
 
         """
+        if predictions and experiments_adapt is None:
+            raise ValueError(
+                "If predictions are used, `experiments_adapt` has to be provided.",
+            )
+        else:
+            experiments_adapt = (
+                experiments if experiments_adapt is None else experiments_adapt
+            )
+
         desis = pd.concat(
             [
                 feat(
                     experiments[f"{feat.key}_pred" if predictions else feat.key],
-                    experiments[f"{feat.key}_pred" if predictions else feat.key],
+                    experiments_adapt[feat.key].dropna(),
                 )
                 for feat in self.features
                 if feat.objective is not None

bofire/strategies/predictives/predictive.py

@@ -156,8 +156,10 @@ def predict(self, experiments: pd.DataFrame) -> pd.DataFrame:
             predictions=predictions,
             outputs=self.domain.outputs,
         )
-        desis = self.domain.outputs(predictions, predictions=True)
-        predictions = pd.concat((predictions, desis), axis=1)
+        objectives = self.domain.outputs(
+            predictions, experiments_adapt=self.experiments, predictions=True
+        )
+        predictions = pd.concat((predictions, objectives), axis=1)
         predictions.index = experiments.index
         return predictions
 

tests/bofire/data_models/domain/test_outputs.py

@@ -12,6 +12,7 @@
     ConstrainedObjective,
     MaximizeObjective,
     MaximizeSigmoidObjective,
+    MovingMaximizeSigmoidObjective,
     Objective,
     TargetObjective,
 )
@@ -300,6 +301,32 @@ def test_outputs_call(features, samples):
     ]
 
 
+def test_outputs_call_adapt_experiment():
+    outputs = Outputs(
+        features=[
+            ContinuousOutput(key="of1", objective=MaximizeObjective()),
+            ContinuousOutput(
+                key="of2",
+                objective=MovingMaximizeSigmoidObjective(tp=0, steepness=10, w=1.0),
+            ),
+        ],
+    )
+    candidates = pd.DataFrame(
+        columns=["of1_pred", "of2_pred"], data=[[1.0, 5.0], [2.0, 5.0]]
+    )
+
+    experiments = pd.DataFrame(columns=["of1", "of2"], data=[[1.0, 5.0], [2.0, 6.0]])
+
+    with pytest.raises(
+        ValueError,
+        match="If predictions are used, `experiments_adapt` has to be provided.",
+    ):
+        outputs(candidates, predictions=True)
+
+    outputs(experiments)
+    outputs(candidates, experiments_adapt=experiments, predictions=True)
+
+
 def test_categorical_objective_methods():
     obj = ConstrainedCategoricalObjective(
         categories=["a", "b"],