def test_preferences(self):
d = UniformPreferenceDict(3)
a, b = (0, 1, 2), (3, 4, 5)
d[a, b] = 1
e, f = (9, 9, 9), (1, 1, 1)
d[e, f] = -1
self.assertEqual(sorted(d.preferences()), [(0, 1, 2), (1, 1, 1),
(3, 4, 5), (9, 9, 9)])
def should_return_correct_preference_key_items():
# given
a = (0, 1, 2)
b = (3, 4, 5)
e = (9, 9, 9)
f = (1, 1, 1)
preferences = UniformPreferenceDict(len(a))
# when
preferences[a, b] = 1
preferences[e, f] = -1
# then
assert sorted(preferences.preferences()) == [a, f, b, e]
def test_interface_with_direct_optimizer_scalar_lengthscale(self):
with self.test_session():
# set RNG seed
np.random.seed(RNG_SEED)
tf.set_random_seed(RNG_SEED)
# set up
bounds = [
(-3, 6),
(-3, 6),
]
optimizer = DirectOptimizer(bounds)
model = BinaryPreferenceModel(
lengthscale=1.0,
**CONFIG
)
data = UniformPreferenceDict(2)
a = (0, 0)
b = (1, 1)
c = (2, 2)
d = (3, 3)
data[a, b] = 1
data[a, c] = 1
data[b, c] = 1
data[b, d] = 1
data[c, d] = 1
# construct acquirer
acquirer = ExpectedImprovementAcquirer(data, model, optimizer)
# test that `next` needs to be called before `best`
with self.assertRaises(ValueError):
acquirer.best
# test that `next` needs to be called before `valuations`
with self.assertRaises(ValueError):
acquirer.valuations
# test `next`
a1, a2 = a
b1, b2 = b
xn = xn1, xn2 = acquirer.next
eps = 0.5
self.assertTrue(
(a1 - eps < xn1 < b1) and
(a2 - eps < xn2 < b2)
)
self.assertTrue(np.allclose(xn, acquirer.next))
# test `best`
best = acquirer.best
self.assertTrue(np.allclose(a, best))
self.assertTrue(np.allclose(best, acquirer.best))
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))
# test `update`
e = (-1, -1)
acquirer.update(e, a, 1)
acquirer.update(e, b, 1)
# test `next`
l1, l2 = (x[0] for x in bounds)
xn1, xn2 = acquirer.next
self.assertTrue(
(l1 < xn1 < a1) and
(l2 < xn2 < a2),
)
# test `best`
best = acquirer.best
self.assertTrue(np.allclose(e, best))
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))
def test_interface_with_grid_search_optimizer_ard_lengthscale(self):
with self.test_session():
# set RNG seed
np.random.seed(RNG_SEED)
tf.set_random_seed(RNG_SEED)
# set up
bounds = [
(-3, 6),
(-3, 6),
]
optimizer = GridSearchOptimizer(bounds)
model = BinaryPreferenceModel(
ard=True,
**CONFIG
)
data = UniformPreferenceDict(2)
a = (0, 0)
b = (1, 0)
c = (2, 0)
d = (3, 0)
data[a, b] = 1
data[a, c] = 1
data[b, c] = 1
data[b, d] = 1
data[c, d] = 1
ab = (0.5, 0)
bc = (1.5, 0)
cd = (2.5, 0)
data[a, ab] = 1
data[ab, b] = 1
data[b, bc] = 1
data[bc, c] = 1
data[c, cd] = 1
data[cd, d] = 1
data[ab, bc] = 1
data[ab, cd] = 1
data[bc, cd] = 1
# construct acquirer
acquirer = ExpectedImprovementAcquirer(data, model, optimizer)
# test that `next` needs to be called before `best`
with self.assertRaises(ValueError):
acquirer.best
# test that `next` needs to be called before `valuations`
with self.assertRaises(ValueError):
acquirer.valuations
# test `next`
a1, a2 = a
b1, b2 = b
xn = xn1, xn2 = acquirer.next
eps = 1.0
self.assertTrue(
(a1 - eps < xn1 < b1),
)
self.assertTrue(np.allclose(xn, acquirer.next))
# test `best`
best = acquirer.best
self.assertTrue(np.allclose(a, best))
self.assertTrue(np.allclose(best, acquirer.best))
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))
# test `lengthscale`
s1, s2 = model.lengthscale
self.assertTrue(s1 < s2)
# test `update`
e = (-1, 0)
acquirer.update(e, a, 1)
acquirer.update(e, ab, 1)
acquirer.update(e, b, 1)
acquirer.update(e, bc, 1)
# test `next`
l1, l2 = (x[0] for x in bounds)
xn1, xn2 = acquirer.next
self.assertTrue(
(l1 <= xn1 < a1),
)
# test `best`
best = acquirer.best
self.assertTrue(best < b)
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))
# test `lengthscale`
s1, s2 = model.lengthscale
self.assertTrue(s1 < s2)
def test_interface_with_grid_search_optimizer_vector_lengthscale(self):
with self.test_session():
# set RNG seed
np.random.seed(RNG_SEED)
tf.set_random_seed(RNG_SEED)
# set up
bounds = [
(-3, 6),
(-3, 6),
]
optimizer = GridSearchOptimizer(bounds)
lengthscale = np.array([1, 10], np.float32)
model = BinaryPreferenceModel(lengthscale=lengthscale, **CONFIG)
data = UniformPreferenceDict(2)
a = (0, 0)
b = (1, 0)
c = (2, 0)
d = (3, 0)
data[a, b] = 1
data[a, c] = 1
data[b, c] = 1
data[b, d] = 1
data[c, d] = 1
# construct acquirer
acquirer = ExpectedImprovementAcquirer(data, model, optimizer)
# test that `next` needs to be called before `best`
with self.assertRaises(ValueError):
acquirer.best # pylint: disable=pointless-statement
# test that `next` needs to be called before `valuations`
with self.assertRaises(ValueError):
acquirer.valuations # pylint: disable=pointless-statement
# test `next`
a1, _ = a
b1, _ = b
xn = xn1, _ = acquirer.next
eps = 0.5
self.assertTrue((a1 - eps < xn1 < b1))
self.assertTrue(np.allclose(xn, acquirer.next))
# test `best`
best = acquirer.best
self.assertTrue(np.allclose(a, best))
self.assertTrue(np.allclose(best, acquirer.best))
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))
# test `lengthscale`
self.assertAllClose(model.lengthscale, lengthscale)
# test `update`
e = (-1, 0)
acquirer.update(e, a, 1)
acquirer.update(e, b, 1)
# test `next`
l1, _ = (x[0] for x in bounds)
xn1, _ = acquirer.next
self.assertTrue((l1 < xn1 < a1))
# test `best`
best = acquirer.best
self.assertTrue(np.allclose(e, best))
# test `valuations`
valuations = acquirer.valuations
x, f = zip(*valuations)
self.assertEqual(len(x), len(data.preferences()))
self.assertTrue(all(a < b for a, b in zip(x, x[1:])))
self.assertTrue(all(a > b for a, b in zip(f, f[1:])))