data_path=data_path)
# ground-truth policy value of a evaluation policy
# , which is estimated with factual (observed) rewards (on-policy estimation)
ground_truth_policy_value = OpenBanditDataset.calc_on_policy_policy_value_estimate(
behavior_policy=evaluation_policy,
campaign=campaign,
data_path=data_path,
test_size=test_size,
is_timeseries_split=is_timeseries_split,
)
# compute action distribution by evaluation policy
if evaluation_policy == "bts":
policy = BernoulliTS(
n_actions=obd.n_actions,
len_list=obd.len_list,
is_zozotown_prior=
True, # replicate the policy in the ZOZOTOWN production
campaign=campaign,
random_state=random_state,
)
else:
policy = Random(
n_actions=obd.n_actions,
len_list=obd.len_list,
random_state=random_state,
)
action_dist_single_round = policy.compute_batch_action_dist(
n_sim=n_sim_to_compute_action_dist)
def process(b: int):
# load the pre-trained regression model
with open(reg_model_path / f"reg_model_{b}.pkl", "rb") as f:
print(args)
# configurations
n_runs = args.n_runs
n_rounds = args.n_rounds
n_actions = args.n_actions
dim_context = args.dim_context
n_sim = args.n_sim
evaluation_policy_name = args.evaluation_policy_name
n_jobs = args.n_jobs
random_state = args.random_state
np.random.seed(random_state)
# define evaluation policy
evaluation_policy_dict = dict(
bernoulli_ts=BernoulliTS(n_actions=n_actions,
random_state=random_state),
epsilon_greedy=EpsilonGreedy(n_actions=n_actions,
epsilon=0.1,
random_state=random_state),
lin_epsilon_greedy=LinEpsilonGreedy(dim=dim_context,
n_actions=n_actions,
epsilon=0.1,
random_state=random_state),
lin_ts=LinTS(dim=dim_context,
n_actions=n_actions,
random_state=random_state),
lin_ucb=LinUCB(dim=dim_context,
n_actions=n_actions,
random_state=random_state),
logistic_epsilon_greedy=LogisticEpsilonGreedy(
dim=dim_context,
reg_model = pickle.load(f)
with open(reg_model_path / f"is_for_reg_model_{b}.pkl", "rb") as f:
is_for_reg_model = pickle.load(f)
# sample bootstrap samples from batch logged bandit feedback
boot_bandit_feedback = obd.sample_bootstrap_bandit_feedback(
test_size=test_size,
is_timeseries_split=is_timeseries_split,
random_state=b)
for key_ in ["context", "action", "reward", "pscore", "position"]:
boot_bandit_feedback[key_] = boot_bandit_feedback[key_][
~is_for_reg_model]
if evaluation_policy == "bts":
policy = BernoulliTS(
n_actions=obd.n_actions,
len_list=obd.len_list,
is_zozotown_prior=
True, # replicate the policy in the ZOZOTOWN production
campaign=campaign,
random_state=random_state,
)
action_dist = policy.compute_batch_action_dist(
n_sim=100000, n_rounds=boot_bandit_feedback["n_rounds"])
else:
policy = Random(
n_actions=obd.n_actions,
len_list=obd.len_list,
random_state=random_state,
)
action_dist = policy.compute_batch_action_dist(
n_sim=100000, n_rounds=boot_bandit_feedback["n_rounds"])
# estimate the mean reward function using the pre-trained reg_model
estimated_rewards_by_reg_model = reg_model.predict(
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 = [
DoublyRobustWithShrinkage(lambda_=lam_,
estimator_name=f"DRos ({lam_})")
for lam_ in lambdas
] + [
DoublyRobustWithShrinkageTuning(lambdas=lambdas,
estimator_name="DRos (tuning)"),
]
# 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/hypara")
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")
# 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")