def test_tree_ei(self):
"""
Tests that using EI with DecisionTree works properly, by decomposing the process into each of its step.
"""
# Define a tree and fit it on the historical data
dtstdr = DecisionTreeSTDRegressor()
dtstdr.fit(fake_history["parameters"], fake_history["fitness"])
# Predict the regression values over the whole grid
combination_ranges = np.array(np.meshgrid(*ranges)).T.reshape(
-1, len(ranges))
predictions_means, predictions_std = dtstdr.predict(combination_ranges,
return_std=True)
# Compute the EI and save the max
current_optimum = np.min(fake_history["fitness"])
computed_ei = compute_expected_improvement(current_optimum,
predictions_means,
predictions_std)
max_ei_index = np.argmax(computed_ei)
expected_next_parameter = combination_ranges[max_ei_index]
# Compare it to the normal process
surrogate_model = SurrogateModel(
regression_model=DecisionTreeSTDRegressor,
next_parameter_strategy=expected_improvement,
)
next_parameter = surrogate_model.choose_next_parameter(
fake_history, ranges)
np.testing.assert_array_equal(next_parameter, expected_next_parameter)
def test_choose_next_parameter_categorical(self):
"""Tests that the selection of the next parameter behaves as expected
whenever there are any categorical variables.
"""
ranges = np.array(
[[1, 2, 3, 4, 5, 6], ["titi", "toto", "tutu"], ["popo", "jojo"]],
dtype=object)
fake_history = {
"fitness":
np.array([10, 5, 4, 2, 15]),
"parameters":
np.array(
[[1, "tutu", "popo"], [2, "toto", "popo"], [4, "toto", "popo"],
[2, "titi", "jojo"], [3, "tutu", "popo"]],
dtype=object),
"truncated":
np.array([True, True, False, False, False]),
}
expected_new_parameter = np.array([1, 'titi', 'jojo'], dtype=object)
surrogate_model = SurrogateModel(
regression_model=GaussianProcessRegressor,
next_parameter_strategy=expected_improvement,
)
real_new_parameter = surrogate_model.choose_next_parameter(
fake_history, ranges)
assert_array_equal(expected_new_parameter, real_new_parameter)
def test_fit_return_prediction_function(self):
"""
Tests that there is no error when computing the regression function.
"""
surrogate_model = SurrogateModel(
regression_model=GaussianProcessRegressor,
next_parameter_strategy=expected_improvement,
)
surrogate_model.regression_function(fake_history, ranges)
def test_evaluate_quality(self):
"""
Tests that the RMSE is properly returned.
"""
expected_score = -2.184857
surrogate_model = SurrogateModel(
regression_model=GaussianProcessRegressor,
next_parameter_strategy=expected_improvement,
)
surrogate_model.choose_next_parameter(fake_history, ranges)
real_score = surrogate_model.evaluate_quality(fake_history)
np.testing.assert_array_almost_equal(expected_score, real_score)
def test_censored_bayesian(self):
"""
Tests that censored bayesian + EI work properly
"""
expected_new_parameter = [0, 3]
surrogate_model = SurrogateModel(
regression_model=CensoredGaussianProcesses,
next_parameter_strategy=expected_improvement,
)
real_new_parameter = surrogate_model.choose_next_parameter(
fake_history, ranges)
np.testing.assert_array_equal(real_new_parameter,
expected_new_parameter)
def test_choose_next_parameter_ei(self):
"""
Checks that the selection of the next parameter works properly when using EI.
"""
expected_new_parameter = [4, 4]
surrogate_model = SurrogateModel(
regression_model=GaussianProcessRegressor,
next_parameter_strategy=expected_improvement,
)
real_new_parameter = surrogate_model.choose_next_parameter(
fake_history, ranges)
np.testing.assert_array_equal(real_new_parameter,
expected_new_parameter)
def test_hot_encoding(self):
"""Tests that the hot encoding feature works as expected by
hot encoding an array.
"""
surrogate_model = SurrogateModel(
regression_model=GaussianProcessRegressor,
next_parameter_strategy=expected_improvement,
)
ranges = np.array(
[[1, 2, 3], ["titi", "toto", "tutu"], ["popo", "jojo"]],
dtype=object)
# manually get categorical ranges
surrogate_model.get_categorical_ranges(ranges)
parameter_array = np.array(
[[1, "tutu", "popo"], [2, "toto", "popo"], [4, "toto", "popo"]],
dtype=object)
hot_encoded = surrogate_model.hot_encode(parameter_array)
assert_array_equal(
[[1, 0, 0, 1, 1, 0], [2, 0, 1, 0, 1, 0], [4, 0, 1, 0, 1, 0]],
hot_encoded)
def test_regression_model_no_predict(self):
"""
Tests that when the regression model does not have a predict method, an error is raised.
"""
# class without predict
class AintGotNoPredict:
"""
Fake class that doesn't have a predict method.
"""
def fit(self):
"""
Only a fit method.
"""
print("I don't have no predict !")
with self.assertRaises(AttributeError):
SurrogateModel(
regression_model=AintGotNoPredict,
next_parameter_strategy=expected_improvement,
)