def test_meta_estimate_intervals_using_invalid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description_1: str,
alpha,
n_bootstrap_samples,
random_state,
err,
description_2: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of estimate_intervals using invalid data
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[dm])
with pytest.raises(err, match=f"{description_2}*"):
_ = ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
)
# estimate_intervals function is called in summarize_off_policy_estimates
with pytest.raises(err, match=f"{description_2}*"):
_ = ope_.summarize_off_policy_estimates(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
)
def test_meta_evaluate_performance_of_estimators_using_invalid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description_1: str,
metric,
ground_truth_policy_value,
err,
description_2: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of evaluate_performance_of_estimators using invalid data
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[dm])
with pytest.raises(err, match=f"{description_2}*"):
_ = ope_.evaluate_performance_of_estimators(
ground_truth_policy_value=ground_truth_policy_value,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
action_dist=action_dist,
metric=metric,
)
# estimate_intervals function is called in summarize_off_policy_estimates
with pytest.raises(err, match=f"{description_2}*"):
_ = ope_.summarize_estimators_comparison(
ground_truth_policy_value=ground_truth_policy_value,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
action_dist=action_dist,
metric=metric,
)
def test_meta_post_init(
synthetic_multi_bandit_feedback: BanditFeedback) -> None:
"""
Test the __post_init__ function
"""
# __post_init__ saves the latter estimator when the same estimator name is used
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[ipw, ipw2])
assert ope_.ope_estimators_ == {
"ipw": ipw2
}, "__post_init__ returns a wrong value"
# __post_init__ can handle the same estimator if the estimator names are different
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[ipw, ipw3])
assert ope_.ope_estimators_ == {
"ipw": ipw,
"ipw3": ipw3,
}, "__post_init__ returns a wrong value"
# __post__init__ raises RuntimeError when necessary_keys are not included in the bandit_feedback
necessary_keys = ["action", "position", "reward"]
for i in range(len(necessary_keys)):
for deleted_keys in itertools.combinations(necessary_keys, i + 1):
invalid_bandit_feedback_dict = {key: "_" for key in necessary_keys}
# delete
for k in deleted_keys:
del invalid_bandit_feedback_dict[k]
with pytest.raises(RuntimeError, match=r"Missing key*"):
_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=invalid_bandit_feedback_dict,
ope_estimators=[ipw])
def test_meta_evaluate_performance_of_estimators_using_valid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description_1: str,
metric,
ground_truth_policy_value,
description_2: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of evaluate_performance_of_estimators using valid data
"""
if metric == "relative-ee":
# calculate relative-ee
eval_metric_ope_dict = {
"ipw":
np.abs((mock_policy_value + ipw.eps - ground_truth_policy_value) /
ground_truth_policy_value),
"ipw3":
np.abs((mock_policy_value + ipw3.eps - ground_truth_policy_value) /
ground_truth_policy_value),
}
else:
# calculate se
eval_metric_ope_dict = {
"ipw":
(mock_policy_value + ipw.eps - ground_truth_policy_value)**2,
"ipw3":
(mock_policy_value + ipw3.eps - ground_truth_policy_value)**2,
}
# check performance estimators
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[ipw, ipw3])
performance = ope_.evaluate_performance_of_estimators(
ground_truth_policy_value=ground_truth_policy_value,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
action_dist=action_dist,
metric=metric,
)
for k, v in performance.items():
assert k in eval_metric_ope_dict, "Invalid key of performance response"
assert v == eval_metric_ope_dict[
k], "Invalid value of performance response"
performance_df = ope_.summarize_estimators_comparison(
ground_truth_policy_value=ground_truth_policy_value,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
action_dist=action_dist,
metric=metric,
)
assert_frame_equal(
performance_df,
pd.DataFrame(eval_metric_ope_dict,
index=[metric]).T), "Invalid summarization (performance)"
def test_meta_estimate_policy_values_using_valid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of estimate_policy_values using valid data
"""
# single ope estimator
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[dm])
ope_.is_model_dependent = True
assert ope_.estimate_policy_values(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
) == {
"dm": mock_policy_value
}, "OffPolicyEvaluation.estimate_policy_values ([DirectMethod]) returns a wrong value"
# multiple ope estimators
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[dm, ipw])
ope_.is_model_dependent = True
assert ope_.estimate_policy_values(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
) == {
"dm": mock_policy_value,
"ipw": mock_policy_value + ipw.eps,
}, "OffPolicyEvaluation.estimate_policy_values ([DirectMethod, IPW]) returns a wrong value"
def test_meta_create_estimator_inputs_using_valid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the _create_estimator_inputs using invalid data
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[ipw])
estimator_inputs = ope_._create_estimator_inputs(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
assert set(estimator_inputs.keys()) == set(["ipw"])
assert set(estimator_inputs["ipw"].keys()) == set(
[
"reward",
"action",
"pscore",
"position",
"action_dist",
"stratum_idx",
"pscore_avg",
"estimated_rewards_by_reg_model",
"estimated_pscore",
"estimated_pscore_avg",
]
), f"Invalid response of _create_estimator_inputs (test case: {description})"
# _create_estimator_inputs function is called in the following functions
_ = ope_.estimate_policy_values(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
_ = ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
_ = ope_.summarize_off_policy_estimates(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
_ = ope_.evaluate_performance_of_estimators(
ground_truth_policy_value=0.1,
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
_ = ope_.summarize_estimators_comparison(
ground_truth_policy_value=0.1,
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
def test_meta_create_estimator_inputs_using_invalid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the _create_estimator_inputs using valid data
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[ipw])
# raise ValueError when the shape of two arrays are different
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_._create_estimator_inputs(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
# _create_estimator_inputs function is called in the following functions
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_.estimate_policy_values(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_.summarize_off_policy_estimates(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_.evaluate_performance_of_estimators(
ground_truth_policy_value=0.1,
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ope_.summarize_estimators_comparison(
ground_truth_policy_value=0.1,
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
def test_meta_estimate_intervals_using_valid_input_data(
action_dist,
estimated_rewards_by_reg_model,
description_1: str,
alpha: float,
n_bootstrap_samples: int,
random_state: int,
description_2: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of estimate_intervals using valid data
"""
# single ope estimator
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback, ope_estimators=[dm])
assert ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
) == {
"dm": mock_confidence_interval
}, "OffPolicyEvaluation.estimate_intervals ([DirectMethod]) returns a wrong value"
# multiple ope estimators
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[dm, ipw])
assert ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
) == {
"dm": mock_confidence_interval,
"ipw": {k: v + ipw.eps
for k, v in mock_confidence_interval.items()},
}, "OffPolicyEvaluation.estimate_intervals ([DirectMethod, IPW]) returns a wrong value"
def test_meta_estimated_rewards_by_reg_model_inputs(
synthetic_multi_bandit_feedback: BanditFeedback, ) -> None:
"""
Test the estimate_policy_values/estimate_intervals functions wrt estimated_rewards_by_reg_model
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[DirectMethod()])
action_dist = np.zeros((
synthetic_multi_bandit_feedback["n_rounds"],
synthetic_multi_bandit_feedback["n_actions"],
))
with pytest.raises(ValueError):
ope_.estimate_policy_values(
action_dist=action_dist,
estimated_rewards_by_reg_model=None,
)
with pytest.raises(ValueError):
ope_.estimate_intervals(
action_dist=action_dist,
estimated_rewards_by_reg_model=None,
)
def test_meta_summarize_off_policy_estimates(
action_dist,
estimated_rewards_by_reg_model,
description_1: str,
alpha: float,
n_bootstrap_samples: int,
random_state: int,
description_2: str,
synthetic_multi_bandit_feedback: BanditFeedback,
) -> None:
"""
Test the response of summarize_off_policy_estimates using valid data
"""
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=synthetic_multi_bandit_feedback,
ope_estimators=[ipw, ipw3])
value, interval = ope_.summarize_off_policy_estimates(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
)
expected_value = pd.DataFrame(
{
"ipw": mock_policy_value + ipw.eps,
"ipw3": mock_policy_value + ipw3.eps,
},
index=["estimated_policy_value"],
).T
expected_value["relative_estimated_policy_value"] = (
expected_value["estimated_policy_value"] /
synthetic_multi_bandit_feedback["reward"].mean())
expected_interval = pd.DataFrame({
"ipw": {k: v + ipw.eps
for k, v in mock_confidence_interval.items()},
"ipw3": {k: v + ipw3.eps
for k, v in mock_confidence_interval.items()},
}).T
assert_frame_equal(value,
expected_value), "Invalid summarization (policy value)"
assert_frame_equal(interval,
expected_interval), "Invalid summarization (interval)"
# check relative estimated policy value when the average of bandit_feedback["reward"] is zero
zero_reward_bandit_feedback = deepcopy(synthetic_multi_bandit_feedback)
zero_reward_bandit_feedback["reward"] = np.zeros(
zero_reward_bandit_feedback["reward"].shape[0])
ope_ = MultiLoggersOffPolicyEvaluation(
bandit_feedback=zero_reward_bandit_feedback,
ope_estimators=[ipw, ipw3])
value, _ = ope_.summarize_off_policy_estimates(
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
alpha=alpha,
n_bootstrap_samples=n_bootstrap_samples,
random_state=random_state,
)
expected_value = pd.DataFrame(
{
"ipw": mock_policy_value + ipw.eps,
"ipw3": mock_policy_value + ipw3.eps,
},
index=["estimated_policy_value"],
).T
expected_value["relative_estimated_policy_value"] = np.nan
assert_frame_equal(value,
expected_value), "Invalid summarization (policy value)"