def test_dr_tuning_init_using_valid_input_data(lambdas, tuning_method,
description):
_ = DoublyRobustTuning(lambdas=lambdas, tuning_method=tuning_method)
_ = DoublyRobustWithShrinkageTuning(
lambdas=lambdas,
tuning_method=tuning_method,
)
_ = SwitchDoublyRobustTuning(
lambdas=lambdas,
tuning_method=tuning_method,
)
_ = SubGaussianDoublyRobustTuning(
lambdas=lambdas,
tuning_method=tuning_method,
)
def test_dr_tuning_init_using_invalid_inputs(
lambdas,
tuning_method,
use_bias_upper_bound,
delta,
use_estimated_pscore,
err,
description,
):
with pytest.raises(err, match=f"{description}*"):
_ = DoublyRobustTuning(
use_bias_upper_bound=use_bias_upper_bound,
delta=delta,
lambdas=lambdas,
tuning_method=tuning_method,
use_estimated_pscore=use_estimated_pscore,
)
with pytest.raises(err, match=f"{description}*"):
_ = SwitchDoublyRobustTuning(
use_bias_upper_bound=use_bias_upper_bound,
delta=delta,
lambdas=lambdas,
tuning_method=tuning_method,
use_estimated_pscore=use_estimated_pscore,
)
with pytest.raises(err, match=f"{description}*"):
_ = DoublyRobustWithShrinkageTuning(
use_bias_upper_bound=use_bias_upper_bound,
delta=delta,
lambdas=lambdas,
tuning_method=tuning_method,
use_estimated_pscore=use_estimated_pscore,
)
with pytest.raises(err, match=f"{description}*"):
_ = SubGaussianDoublyRobustTuning(
use_bias_upper_bound=use_bias_upper_bound,
delta=delta,
lambdas=lambdas,
tuning_method=tuning_method,
use_estimated_pscore=use_estimated_pscore,
)
),
SelfNormalizedInverseProbabilityWeighting(),
DoublyRobustTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="mse",
estimator_name="dr (tuning-mse)",
),
DoublyRobustTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="slope",
estimator_name="dr (tuning-slope)",
),
SelfNormalizedDoublyRobust(),
SwitchDoublyRobustTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="mse",
estimator_name="switch-dr (tuning-mse)",
),
SwitchDoublyRobustTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="slope",
estimator_name="switch-dr (tuning-slope)",
),
DoublyRobustWithShrinkageTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="mse",
estimator_name="dr-os (tuning-mse)",
),
DoublyRobustWithShrinkageTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, np.inf],
tuning_method="slope",
hyperparams = yaml.safe_load(f)
base_model_dict = dict(
logistic_regression=LogisticRegression,
lightgbm=GradientBoostingClassifier,
random_forest=RandomForestClassifier,
)
# compared OPE estimators
ope_estimators = [
DirectMethod(),
InverseProbabilityWeighting(),
SelfNormalizedInverseProbabilityWeighting(),
DoublyRobust(),
SelfNormalizedDoublyRobust(),
SwitchDoublyRobustTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, 10000, np.inf]),
DoublyRobustWithShrinkageTuning(
lambdas=[10, 50, 100, 500, 1000, 5000, 10000, np.inf]),
]
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description=
"evaluate the accuracy of OPE estimators on synthetic bandit data.")
parser.add_argument("--n_runs",
type=int,
default=1,
help="number of simulations in the experiment.")
parser.add_argument(
"--n_rounds",
type=int,
def test_dr_variants_using_valid_input_data(
action_dist: np.ndarray,
action: np.ndarray,
reward: np.ndarray,
pscore: np.ndarray,
position: np.ndarray,
estimated_rewards_by_reg_model: np.ndarray,
estimated_pscore: np.ndarray,
hyperparameter: float,
description: str,
) -> None:
# check dr variants
switch_dr = SwitchDoublyRobust(lambda_=hyperparameter)
switch_dr_tuning_mse = SwitchDoublyRobustTuning(
lambdas=[hyperparameter, hyperparameter * 10],
tuning_method="mse",
)
switch_dr_tuning_slope = SwitchDoublyRobustTuning(
lambdas=[hyperparameter, hyperparameter * 10],
tuning_method="slope",
)
dr_os = DoublyRobustWithShrinkage(lambda_=hyperparameter)
dr_os_tuning_mse = DoublyRobustWithShrinkageTuning(
lambdas=[hyperparameter, hyperparameter * 10],
tuning_method="mse",
)
dr_os_tuning_slope = DoublyRobustWithShrinkageTuning(
lambdas=[hyperparameter, hyperparameter * 10],
tuning_method="slope",
)
sg_dr = SubGaussianDoublyRobust(lambda_=hyperparameter)
sg_dr_tuning_mse = SubGaussianDoublyRobustTuning(
lambdas=[hyperparameter, hyperparameter / 10],
tuning_method="mse",
)
sg_dr_tuning_slope = SubGaussianDoublyRobustTuning(
lambdas=[hyperparameter, hyperparameter / 10],
tuning_method="slope",
)
switch_dr_estimated_pscore = SwitchDoublyRobust(lambda_=hyperparameter,
use_estimated_pscore=True)
switch_dr_tuning_estimated_pscore = SwitchDoublyRobustTuning(
lambdas=[hyperparameter, hyperparameter * 10],
use_estimated_pscore=True)
dr_os_estimated_pscore = DoublyRobustWithShrinkage(
lambda_=hyperparameter, use_estimated_pscore=True)
dr_os_tuning_estimated_pscore = DoublyRobustWithShrinkageTuning(
lambdas=[hyperparameter, hyperparameter * 10],
use_estimated_pscore=True)
for estimator in [
sg_dr,
sg_dr_tuning_mse,
sg_dr_tuning_slope,
switch_dr,
switch_dr_tuning_mse,
switch_dr_tuning_slope,
switch_dr_estimated_pscore,
switch_dr_tuning_estimated_pscore,
dr_os,
dr_os_tuning_mse,
dr_os_tuning_slope,
dr_os_estimated_pscore,
dr_os_tuning_estimated_pscore,
]:
est = estimator.estimate_policy_value(
action_dist=action_dist,
action=action,
reward=reward,
pscore=pscore,
position=position,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
estimated_pscore=estimated_pscore,
)
assert est == 0.0, f"policy value must be 0, but {est}"
estimator_name="dr_tuning_mse")
dr_tuning_slope = DoublyRobustTuning(lambdas=[1, 100],
tuning_method="slope",
estimator_name="dr_tuning_slope")
dr_os_0 = DoublyRobustWithShrinkage(lambda_=0.0)
dr_os_tuning_mse = DoublyRobustWithShrinkageTuning(
lambdas=[1, 100], tuning_method="mse", estimator_name="dr_os_tuning_mse")
dr_os_tuning_slope = DoublyRobustWithShrinkageTuning(
lambdas=[1, 100],
tuning_method="slope",
estimator_name="dr_os_tuning_slope")
dr_os_max = DoublyRobustWithShrinkage(lambda_=np.inf)
sndr = SelfNormalizedDoublyRobust()
switch_dr_0 = SwitchDoublyRobust(lambda_=0.0)
switch_dr_tuning_mse = SwitchDoublyRobustTuning(
lambdas=[1, 100],
tuning_method="mse",
estimator_name="switch_dr_tuning_mse")
switch_dr_tuning_slope = SwitchDoublyRobustTuning(
lambdas=[1, 100],
tuning_method="slope",
estimator_name="switch_dr_tuning_slope")
switch_dr_max = SwitchDoublyRobust(lambda_=np.inf)
sg_dr_0 = SubGaussianDoublyRobust(lambda_=0.0)
sg_dr_tuning_mse = SubGaussianDoublyRobustTuning(
lambdas=[0.01, 0.1],
tuning_method="mse",
estimator_name="sg_dr_tuning_mse")
sg_dr_tuning_slope = SubGaussianDoublyRobustTuning(
lambdas=[0.01, 0.1],
tuning_method="slope",
estimator_name="sg_dr_tuning_slope")
def main(cfg: DictConfig) -> None:
print(cfg)
logger.info(f"The current working directory is {Path().cwd()}")
start_time = time.time()
logger.info("initializing experimental condition..")
# compared ope estimators
lambdas = list(dict(cfg.estimator_hyperparams)["lambdas"])
ope_estimators = [
InverseProbabilityWeighting(estimator_name="IPW"),
SelfNormalizedInverseProbabilityWeighting(estimator_name="SNIPW"),
DirectMethod(estimator_name="DM"),
DoublyRobust(estimator_name="DR"),
SelfNormalizedDoublyRobust(estimator_name="SNDR"),
SwitchDoublyRobustTuning(lambdas=lambdas, estimator_name="Switch-DR"),
DoublyRobustWithShrinkageTuning(lambdas=lambdas,
estimator_name="DRos"),
]
# configurations
n_seeds = cfg.setting.n_seeds
sample_size = cfg.setting.sample_size
reg_model = cfg.setting.reg_model
campaign = cfg.setting.campaign
behavior_policy = cfg.setting.behavior_policy
test_size = cfg.setting.test_size
is_timeseries_split = cfg.setting.is_timeseries_split
n_folds = cfg.setting.n_folds
obd_path = (Path().cwd().parents[5] /
"open_bandit_dataset" if cfg.setting.is_full_obd else None)
random_state = cfg.setting.random_state
np.random.seed(random_state)
# define dataset
dataset_ts = OpenBanditDataset(behavior_policy="bts",
campaign=campaign,
data_path=obd_path)
dataset_ur = OpenBanditDataset(behavior_policy="random",
campaign=campaign,
data_path=obd_path)
# prepare logged bandit feedback and evaluation policies
if behavior_policy == "random":
if is_timeseries_split:
bandit_feedback_ur = dataset_ur.obtain_batch_bandit_feedback(
test_size=test_size,
is_timeseries_split=True,
)[0]
else:
bandit_feedback_ur = dataset_ur.obtain_batch_bandit_feedback()
bandit_feedbacks = [bandit_feedback_ur]
# obtain the ground-truth policy value
ground_truth_ts = OpenBanditDataset.calc_on_policy_policy_value_estimate(
behavior_policy="bts",
campaign=campaign,
data_path=obd_path,
test_size=test_size,
is_timeseries_split=is_timeseries_split,
)
# obtain action choice probabilities and define evaluation policies
policy_ts = BernoulliTS(
n_actions=dataset_ts.n_actions,
len_list=dataset_ts.len_list,
random_state=random_state,
is_zozotown_prior=True,
campaign=campaign,
)
action_dist_ts = policy_ts.compute_batch_action_dist(n_rounds=1000000)
evaluation_policies = [(ground_truth_ts, action_dist_ts)]
else:
if is_timeseries_split:
bandit_feedback_ts = dataset_ts.obtain_batch_bandit_feedback(
test_size=test_size,
is_timeseries_split=True,
)[0]
else:
bandit_feedback_ts = dataset_ts.obtain_batch_bandit_feedback()
bandit_feedbacks = [bandit_feedback_ts]
# obtain the ground-truth policy value
ground_truth_ur = OpenBanditDataset.calc_on_policy_policy_value_estimate(
behavior_policy="random",
campaign=campaign,
data_path=obd_path,
test_size=test_size,
is_timeseries_split=is_timeseries_split,
)
# obtain action choice probabilities and define evaluation policies
policy_ur = Random(
n_actions=dataset_ur.n_actions,
len_list=dataset_ur.len_list,
random_state=random_state,
)
action_dist_ur = policy_ur.compute_batch_action_dist(n_rounds=1000000)
evaluation_policies = [(ground_truth_ur, action_dist_ur)]
# regression models used in ope estimators
hyperparams = dict(cfg.reg_model_hyperparams)[reg_model]
regression_models = [reg_model_dict[reg_model](**hyperparams)]
# define an evaluator class
evaluator = InterpretableOPEEvaluator(
random_states=np.arange(n_seeds),
bandit_feedbacks=bandit_feedbacks,
evaluation_policies=evaluation_policies,
ope_estimators=ope_estimators,
regression_models=regression_models,
)
# conduct an evaluation of OPE experiment
logger.info("experiment started")
_ = evaluator.estimate_policy_value(sample_size=sample_size,
n_folds_=n_folds)
# calculate statistics
mean = evaluator.calculate_mean(root=True)
mean_scaled = evaluator.calculate_mean(scale=True, root=True)
# save results of the evaluation of off-policy estimators
log_path = Path("./outputs")
log_path.mkdir(exist_ok=True, parents=True)
# save root mse
root_mse_df = DataFrame()
root_mse_df["estimator"] = list(mean.keys())
root_mse_df["mean"] = list(mean.values())
root_mse_df["mean(scaled)"] = list(mean_scaled.values())
root_mse_df.to_csv(log_path / "root_mse.csv")
# conduct pairwise t-tests
se_df = DataFrame(evaluator.calculate_squared_error())
se_df = DataFrame(se_df.stack()).reset_index(1)
se_df.rename(columns={"level_1": "estimators", 0: "se"}, inplace=True)
nonparam_ttests = (pg.pairwise_ttests(
data=se_df,
dv="se",
parametric=False,
between="estimators",
).round(4).drop(["Contrast", "Parametric", "Paired"], axis=1))
nonparam_ttests.to_csv(log_path / "nonparam_ttests.csv")
# save reg model metrics
DataFrame(evaluator.reg_model_metrics).describe().to_csv(
log_path / "reg_model_metrics.csv")
# print result
print(root_mse_df)
experiment = f"{campaign}-{behavior_policy}-{sample_size}"
elapsed_time = np.round((time.time() - start_time) / 60, 2)
logger.info(f"finish experiment {experiment} in {elapsed_time}min")