def test_nudge_individual_correct_nudge_size(self):
input_dist = np.array([[
[0.2, 0.05, 0.1],
[0.05, 0.1, 0.025],
], [[0.1, 0.08, 0.02], [0.05, 0.025, 0.2]]])
nudge_size = 0.01
for _ in range(100):
new_dist = nudge.nudge_individual(input_dist, 0.01)
self.assertAlmostEqual(nudge_size,
np.sum(np.absolute(new_dist - input_dist)))
for _ in range(100):
input_dist = np.random.dirichlet(np.array([6, 5, 3, 4, 8]))
new_dist = nudge.nudge_individual(input_dist, 0.01)
self.assertAlmostEqual(nudge_size,
np.sum(np.absolute(new_dist - input_dist)))
def test_nudge_individual_create_new_probability_dist(self):
input_dist = np.array([[
[0.2, 0.05, 0.1],
[0.05, 0.1, 0.025],
], [[0.1, 0.08, 0.02], [0.05, 0.025, 0.2]]])
nudge_size = 0.01
for _ in range(100):
new_dist = nudge.nudge_individual(input_dist, 0.01)
self.assertAlmostEqual(1, np.sum(new_dist))
self.assertTrue(np.all(new_dist >= 0))
nudge_size = 0.01
for _ in range(100):
input_dist = np.random.dirichlet(np.array([5, 5, 5, 5, 5]))
new_dist = nudge.nudge_individual(input_dist, 0.01)
self.assertAlmostEqual(1, np.sum(new_dist))
self.assertTrue(np.all(new_dist >= 0))
def test_nudge_individual_makes_copy(self):
input_dist = np.array([[
[0.2, 0.05, 0.1],
[0.05, 0.1, 0.025],
], [[0.1, 0.08, 0.02], [0.05, 0.025, 0.2]]])
input_dist_copy = np.copy(input_dist)
new_dist = nudge.nudge_individual(input_dist, 0.01)
self.assertTrue(np.all(input_dist == input_dist_copy))
print(new_dist)
print(np.sum(np.absolute(new_dist - input_dist)))
def test_nudge_individual_correct_nudge_with_noise_vector(self):
input_dist = np.array([[
[0.2, 0.05, 0.1],
[0.05, 0.1, 0.025],
], [[0.1, 0.09, 0.01], [0.05, 0.025, 0.2]]])
noise_vector = np.array([0.1, -0.025, -0.075])
nudge_size = 0.2
new_dist = nudge.nudge_individual(input_dist, 0.01, noise_vector)
expected_new_input_dist = np.array([[
[0.235, 0.04125, 0.07375],
[0.0675, 0.095625, 0.011875],
], [[0.115, 0.085, 0], [0.0775, 0.018125, 0.179375]]])
self.assertTrue(np.allclose(new_dist, expected_new_input_dist))
shape = [5, 6, 6, 4, 3]
total_number_of_states = reduce(lambda x, y: x * y, shape)
for _ in range(100):
input_dist = np.random.dirichlet(
np.array(total_number_of_states * [1]))
input_dist = np.reshape(input_dist, shape)
new_dist = nudge.nudge_individual(input_dist, 0.2, noise_vector)
self.assertAlmostEqual(np.sum(new_dist), 1)
self.assertTrue(np.all(new_dist >= 0))
def evaluate(self, start_distribution, cond_output):
"""
Find the impact of the nudge on the output distribution and
set the score.
Parameters:
----------
start_distribution: an nd-array
Representing a probability distribution
cond_output: an nd-array
The output conditioned on all input variables
"""
new_input_dist = nudge.nudge_individual(start_distribution,
self.nudge_size, self.genes)
nudge_impact = nudge.find_nudge_impact(start_distribution,
new_input_dist,
cond_output,
measure="absolute")
#to make it a minimization rather than a maximization problem
self.score = -nudge_impact