Merge branch 'master' into 'fix/get_extra_info'

# Conflicts:
#   pyoptgra/optgra.py

Author: waldemar-martens

pyoptgra/optgra.py

@@ -78,6 +78,13 @@ def _get_constraint_violation(
     return violation_norm, num_violations
 
 
+def _assert_finite(arr: np.ndarray, name: str):
+    mask = ~np.isfinite(arr)  # True for NaN, Inf, -Inf
+    if np.any(mask):
+        raise ValueError(f"Encountered non-finite values in {name} at indices: {np.where(mask)[0]}")
+
+
+
 class optgra:
     """
     This class is a user defined algorithm (UDA) providing a wrapper around OPTGRA, which is written
@@ -126,6 +133,7 @@ def wrapped_fitness(x):
 
             # we are using vectorisation internally -> convert to ndarray
             x = np.asarray(x, dtype=np.float64)
+            _assert_finite(x, "decision vector")  # catch nan values
 
             if khanf:
                 # if Khan function is used, we first need to convert to pagmo parameters
@@ -149,6 +157,7 @@ def wrapped_fitness(x):
             # reorder constraint order, optgra expects the merit function last, pagmo has it first
             # equivalent to rotating in a dequeue
             result = np.concatenate([result[1:], result[0:1]])
+            _assert_finite(result, "fitness")  # catch nan values
 
             return result.tolist()  # return a list
 
@@ -173,6 +182,7 @@ def wrapped_gradient(x):
 
             # we are using vectorisation internally -> convert to ndarray
             x = np.asarray(x, dtype=np.float64)
+            _assert_finite(x, "decision vector")  # catch nan values
 
             if khanf:
                 # if Khan function is used, we first need to convert to pagmo parameters
@@ -221,6 +231,8 @@ def wrapped_gradient(x):
                 khan_grad = khanf.eval_grad(x)
                 result = result @ khan_grad
 
+            _assert_finite(result, "gradient")  # catch nan values
+
             return result.tolist()  # return as a list, not ndarray
 
         return wrapped_gradient
@@ -229,8 +241,8 @@ def __init__(
         self,
         max_iterations: int = 150,
         max_correction_iterations: int = 90,
-        max_distance_per_iteration: int = 10,
-        perturbation_for_snd_order_derivatives: int = 1,
+        max_distance_per_iteration: float = 10,
+        perturbation_for_snd_order_derivatives: float = 1,
         variable_scaling_factors: List[float] = [],  # x_dim
         variable_types: List[int] = [],  # x_dim
         constraint_priorities: List[int] = [],  # f_dim

tests/python/test.py

@@ -272,6 +272,13 @@ def set_seed(self, seed):
         algo = pygmo.algorithm(pyoptgra.optgra(constraint_priorities=[1] * 61))
         algo.evolve(pop)
 
+        # check than nan in the decision vector is caught
+        x = pop.get_x()[0]
+        x[3] = np.nan
+        pop.set_x(0, x)
+        with self.assertRaises(ValueError):
+            algo.evolve(pop)
+
     def basic_no_gradient_test(self):
         # Basic test that the call works and the result changes. No constraints, not gradients.