Reproduction and fixes to the issue #6

carbs/carbs.py

@@ -293,6 +293,15 @@ def _get_mask_for_invalid_points_in_basic(self, input_in_basic: Tensor) -> Tenso
             input_in_basic < self.max_bounds_in_basic.unsqueeze(0)
         )
         mask = torch.logical_and(is_above_min_bounds, is_below_max_bounds)
+
+        # Also remove the samples that are in the success and failure observations
+        # Seeding in other places can lead to repeated samples, so we need to remove them here
+        observed_samples = [x.real_number_input for x in self.success_observations + self.failure_observations]
+        if len(observed_samples) > 0:
+            observed_samples = torch.stack(observed_samples)
+            new_sample_mask = ~torch.any(torch.isclose(input_in_basic[:, None], observed_samples[None, :]).all(dim=2), dim=1)
+            mask = torch.logical_and(mask, new_sample_mask)
+
         return mask
 
     def _round_integer_values_in_basic(self, input_in_basic: Tensor) -> Tensor:

carbs/utils.py

@@ -566,6 +566,15 @@ def get_pareto_groups_conservative(
         for group in grouped_observations
         if observation_group_cost(group) <= min_pareto_cost
     ]
+    # When there happens to be no observations below the min threshold, inspect observations' cost individually
+    if len(observations_below_min_threshold) == 0:
+        logger.warning(
+            "There is no grouped obs below min_pareto_cost. So inspect observations' cost individually."
+        )
+        observations_below_min_threshold = [
+            (x,) for group in grouped_observations for x in group if x.cost <= min_pareto_cost
+        ]
+
     resampled_observations_below_min_threshold = [
         group for group in observations_below_min_threshold if len(group) > 1
     ]