Use deterministic order of vertices in PolytopeNoise

ampyc/noise.py

@@ -9,6 +9,7 @@
 
 from abc import ABC, abstractmethod
 import numpy as np
+from polytope import extreme
 
 from ampyc.utils import Polytope, qhull
 
@@ -98,7 +99,9 @@ class PolytopeVerticesNoise(NoiseBase):
 
     def __init__(self, W: Polytope, seed: int | None = None) -> None:
         assert seed is None or seed >= 0
-        self.V = W.V
+        np.random.seed(0)  # get deterministic behavior of `extreme`
+        # create new polytope to force re-computation
+        self.V = extreme(Polytope(W.A, W.b))
         self.rng = np.random.default_rng(seed)
 
     def _generate(self, N: int | None = None) -> np.ndarray:
@@ -116,7 +119,9 @@ class PolytopeNoise(NoiseBase):
 
     def __init__(self, W: Polytope, seed: int | None = None) -> None:
         assert seed is None or seed >= 0
-        self.V = W.V
+        np.random.seed(0)  # get deterministic behavior of `extreme`
+        # create new polytope to force re-computation
+        self.V = extreme(Polytope(W.A, W.b))
         self.rng = np.random.default_rng(seed)
 
     def _generate(self, N: int | None = None) -> np.ndarray:

tests/noise/test_noise.py

@@ -0,0 +1,35 @@
+from typing import Callable
+from ampyc.noise import GaussianNoise, NoiseBase, PolytopeNoise, PolytopeVerticesNoise
+from ampyc.utils import Polytope
+import pytest
+import numpy as np
+
+
+def get_unit_cube_polytope(noise_dim):
+    A = np.zeros((2 * noise_dim, noise_dim))
+    b = np.zeros((2 * noise_dim,))
+    for i in range(noise_dim):
+        A[2 * i, i] = 1
+        b[2 * i] = 1
+        A[2 * i + 1, i] = -1
+        b[2 * i + 1] = 1
+    return Polytope(A, b)
+
+
+@pytest.mark.parametrize(
+    "noise_factory",
+    [
+        lambda noise_dim: GaussianNoise(
+            mean=np.zeros((noise_dim,)), covariance=np.eye(noise_dim)
+        ),
+        lambda noise_dim: PolytopeNoise(get_unit_cube_polytope(noise_dim)),
+        lambda noise_dim: PolytopeVerticesNoise(get_unit_cube_polytope(noise_dim)),
+    ],
+)
+def test_noise_is_deterministic_with_seed(noise_factory: Callable[[int], NoiseBase]):
+    noise_dim = 6
+    noise1 = noise_factory(noise_dim)
+    noise2 = noise_factory(noise_dim)
+    noise1.seed(5)
+    noise2.seed(5)
+    assert np.all(noise1.generate() == noise2.generate())