def test_boundedness_of_snipw_using_random_evaluation_policy(
synthetic_bandit_feedback: BanditFeedback, random_action_dist: np.ndarray
) -> None:
"""
Test the boundedness of snipw estimators using synthetic bandit data and random evaluation policy
"""
action_dist = random_action_dist
# prepare snipw
snipw = SelfNormalizedInverseProbabilityWeighting()
# prepare input dict
input_dict = {
k: v
for k, v in synthetic_bandit_feedback.items()
if k in ["reward", "action", "pscore", "position"]
}
input_dict["action_dist"] = action_dist
# make pscore too small (to check the boundedness of snipw)
input_dict["pscore"] = input_dict["pscore"] ** 3
estimated_policy_value = snipw.estimate_policy_value(**input_dict)
assert (
estimated_policy_value <= 1
), f"estimated policy value of snipw should be smaller than or equal to 1 (because of its 1-boundedness), but the value is: {estimated_policy_value}"
# ipw with estimated pscore
snipw_estimated_pscore = SelfNormalizedInverseProbabilityWeighting(
use_estimated_pscore=True
)
input_dict["estimated_pscore"] = input_dict["pscore"]
del input_dict["pscore"]
estimated_policy_value = snipw_estimated_pscore.estimate_policy_value(**input_dict)
assert (
estimated_policy_value <= 1
), f"estimated policy value of snipw should be smaller than or equal to 1 (because of its 1-boundedness), but the value is: {estimated_policy_value}"
def test_ipw_using_invalid_input_data(
action_dist: np.ndarray,
action: np.ndarray,
reward: np.ndarray,
pscore: np.ndarray,
position: np.ndarray,
use_estimated_pscore: bool,
estimated_pscore: np.ndarray,
description: str,
) -> None:
# prepare ipw instances
ipw = InverseProbabilityWeighting(use_estimated_pscore=use_estimated_pscore)
snipw = SelfNormalizedInverseProbabilityWeighting(
use_estimated_pscore=use_estimated_pscore
)
sgipw = SubGaussianInverseProbabilityWeighting(
use_estimated_pscore=use_estimated_pscore
)
ipw_tuning = InverseProbabilityWeightingTuning(
lambdas=[10, 1000], use_estimated_pscore=use_estimated_pscore
)
sgipw_tuning = SubGaussianInverseProbabilityWeightingTuning(
lambdas=[0.01, 0.1], use_estimated_pscore=use_estimated_pscore
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ipw.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ipw.estimate_interval(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = snipw.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = snipw.estimate_interval(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ipw_tuning.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = ipw_tuning.estimate_interval(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = sgipw.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = sgipw.estimate_interval(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = sgipw_tuning.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)
with pytest.raises(ValueError, match=f"{description}*"):
_ = sgipw_tuning.estimate_interval(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_pscore=estimated_pscore,
)