fix: respect per-dimension bins when sampling inspiration neighbors

openevolve/database.py

@@ -30,7 +30,7 @@
     ProgramState,
 )
 from openevolve.utils.code_utils import calculate_edit_distance
-from openevolve.utils.metrics_utils import safe_numeric_average, get_fitness_score
+from openevolve.utils.metrics_utils import get_fitness_score, safe_numeric_average
 
 logger = logging.getLogger(__name__)
 
@@ -1136,6 +1136,7 @@ def _llm_judge_novelty(self, program: Program, similar_program: Program) -> bool
         Use LLM to judge if a program is novel compared to a similar existing program
         """
         import asyncio
+
         from openevolve.novelty_judge import NOVELTY_SYSTEM_MSG, NOVELTY_USER_MSG
 
         user_msg = NOVELTY_USER_MSG.format(
@@ -1811,8 +1812,14 @@ def _sample_inspirations(
             for _ in range(remaining_slots * 3):  # Try more times to find nearby programs
                 # Perturb coordinates
                 perturbed_coords = [
-                    max(0, min(self.feature_bins - 1, c + random.randint(-2, 2)))
-                    for c in feature_coords
+                    max(
+                        0,
+                        min(
+                            self.feature_bins_per_dim.get(dim, self.feature_bins) - 1,
+                            c + random.randint(-2, 2),
+                        ),
+                    )
+                    for dim, c in zip(self.config.feature_dimensions, feature_coords)
                 ]
 
                 cell_key = self._feature_coords_to_key(perturbed_coords)

tests/test_inspiration_feature_bounds.py

@@ -0,0 +1,47 @@
+"""Neighborhood sampling must use the bounds of each MAP-Elites dimension."""
+
+import unittest
+from unittest.mock import patch
+
+from openevolve.config import DatabaseConfig
+from openevolve.database import Program, ProgramDatabase
+
+
+class TestInspirationFeatureBounds(unittest.TestCase):
+    def test_neighbor_in_wide_dimension_is_selected_before_random_fallback(self):
+        db = ProgramDatabase(
+            DatabaseConfig(
+                in_memory=True,
+                num_islands=1,
+                feature_dimensions=["x", "y"],
+                feature_bins={"x": 20, "y": 2},
+                elite_selection_ratio=0.2,
+            )
+        )
+        db.feature_stats = {
+            dim: {"min": 0.0, "max": 1.0, "values": [0.0, 1.0]} for dim in ("x", "y")
+        }
+        parent = Program(id="parent", code="parent", metrics={"score": 0.1, "x": 0.9, "y": 0.75})
+        neighbor = Program(
+            id="neighbor", code="neighbor", metrics={"score": 0.2, "x": 0.99, "y": 0.75}
+        )
+        best = Program(id="best", code="best", metrics={"score": 1.0, "x": 0.0, "y": 0.0})
+        far = Program(id="far", code="far", metrics={"score": 0.3, "x": 0.0, "y": 0.75})
+        db.programs = {p.id: p for p in (parent, neighbor, best, far)}
+        db.islands = [set(db.programs)]
+        db.island_best_programs = [best.id]
+        self.assertEqual(db._calculate_feature_coords(parent), [18, 1])
+        self.assertEqual(db._calculate_feature_coords(neighbor), [19, 1])
+
+        with (
+            patch("openevolve.database.random.randint", return_value=2),
+            patch("openevolve.database.random.sample", return_value=[far.id]) as fallback,
+        ):
+            inspirations = db._sample_inspirations(parent, n=2, island_id=0)
+
+        self.assertEqual([p.id for p in inspirations], [best.id, neighbor.id])
+        fallback.assert_not_called()
+
+
+if __name__ == "__main__":
+    unittest.main()