Guarantee weights_below to be finite in MOTPE

optuna/samplers/_tpe/sampler.py

@@ -516,6 +516,7 @@ def _build_parzen_estimator(
             weights_below = _calculate_weights_below_for_multi_objective(
                 study, trials, self._constraints_func
             )[param_mask_below]
+            assert np.isfinite(weights_below).all()
             mpe = self._parzen_estimator_cls(
                 observations, search_space, self._parzen_estimator_parameters, weights_below
             )
@@ -814,8 +815,13 @@ def _calculate_weights_below_for_multi_objective(
         contributions = np.asarray(
             [hv - WFG().compute(lvals[indices_mat[i]], reference_point) for i in range(n_below)]
         )
-        contributions += EPS
-        weights_below = np.clip(contributions / np.max(contributions), 0, 1)
+        contributions[np.isnan(contributions)] = np.inf
+        max_contribution = np.maximum(np.max(contributions), EPS)
+        if not np.isfinite(max_contribution):
+            weights_below = np.ones_like(contributions, dtype=float)
+            weights_below[np.isfinite(contributions)] = EPS
+        else:
+            weights_below = np.clip(contributions / max_contribution, EPS, 1)
 
     # For now, EPS weight is assigned to infeasible trials.
     weights_below_all = np.full(len(below_trials), EPS)

tests/samplers_tests/tpe_tests/test_multi_objective_sampler.py

@@ -442,7 +442,8 @@ def test_calculate_weights_below_for_multi_objective() -> None:
         None,
     )
     assert len(weights_below) == 3
-    assert all([np.isnan(w) for w in weights_below])
+    assert not any([np.isnan(w) for w in weights_below])
+    assert sum(weights_below) > 0
 
     # Three samples with two infeasible trials.
     study = optuna.create_study(directions=["minimize", "minimize"])