def process(i: int):
# synthetic data generator
dataset = SyntheticBanditDataset(
n_actions=n_actions,
dim_context=dim_context,
reward_function=logistic_reward_function,
behavior_policy_function=linear_behavior_policy,
random_state=i,
)
# define evaluation policy using IPWLearner
evaluation_policy = IPWLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]
),
)
# sample new training and test sets of synthetic logged bandit feedback
bandit_feedback_train = dataset.obtain_batch_bandit_feedback(n_rounds=n_rounds)
bandit_feedback_test = dataset.obtain_batch_bandit_feedback(n_rounds=n_rounds)
# train the evaluation policy on the training set of the synthetic logged bandit feedback
evaluation_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
# predict the action decisions for the test set of the synthetic logged bandit feedback
action_dist = evaluation_policy.predict(
context=bandit_feedback_test["context"],
)
# estimate the mean reward function of the test set of synthetic bandit feedback with ML model
regression_model = RegressionModel(
n_actions=dataset.n_actions,
action_context=dataset.action_context,
base_model=base_model_dict[base_model_for_reg_model](
**hyperparams[base_model_for_reg_model]
),
)
estimated_rewards_by_reg_model = regression_model.fit_predict(
context=bandit_feedback_test["context"],
action=bandit_feedback_test["action"],
reward=bandit_feedback_test["reward"],
n_folds=3, # 3-fold cross-fitting
random_state=random_state,
)
# evaluate estimators' performances using relative estimation error (relative-ee)
ope = OffPolicyEvaluation(
bandit_feedback=bandit_feedback_test,
ope_estimators=ope_estimators,
)
relative_ee_i = ope.evaluate_performance_of_estimators(
ground_truth_policy_value=dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=action_dist,
),
action_dist=action_dist,
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
return relative_ee_i
def test_synthetic_calc_policy_value_using_invalid_inputs(
expected_reward,
action_dist,
description,
):
n_actions = 10
dataset = SyntheticBanditDataset(n_actions=n_actions)
with pytest.raises(ValueError, match=f"{description}*"):
_ = dataset.calc_ground_truth_policy_value(
expected_reward=expected_reward, action_dist=action_dist)
def test_synthetic_calc_policy_value_using_valid_inputs(
expected_reward,
action_dist,
description,
):
n_actions = 10
dataset = SyntheticBanditDataset(n_actions=n_actions)
policy_value = dataset.calc_ground_truth_policy_value(
expected_reward=expected_reward, action_dist=action_dist)
assert isinstance(
policy_value,
float), "Invalid response of calc_ground_truth_policy_value"
def process(i: int):
# synthetic data generator
dataset = SyntheticBanditDataset(
n_actions=n_actions,
dim_context=dim_context,
reward_function=logistic_reward_function,
behavior_policy_function=linear_behavior_policy,
random_state=i,
)
# define evaluation policy using IPWLearner
ipw_policy = IPWLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
# baseline method 1. RandomPolicy
random_policy = RandomPolicy(n_actions=dataset.n_actions)
# baseline method 2. UniformSampleWeightLearner
uniform_sample_weight_policy = UniformSampleWeightLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
# sample new training and test sets of synthetic logged bandit feedback
bandit_feedback_train = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
bandit_feedback_test = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
# train the evaluation policy on the training set of the synthetic logged bandit feedback
ipw_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
uniform_sample_weight_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
# predict the action decisions for the test set of the synthetic logged bandit feedback
ipw_action_dist = ipw_policy.predict(
context=bandit_feedback_test["context"], )
random_action_dist = random_policy.predict(
context=bandit_feedback_test["context"], )
uniform_sample_weight_action_dist = uniform_sample_weight_policy.predict(
context=bandit_feedback_test["context"], )
# get the ground truth policy value for each learner
gt_ipw_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=ipw_action_dist,
)
gt_random_policy = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=random_action_dist,
)
gt_uniform_sample_weight_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=uniform_sample_weight_action_dist,
)
return gt_ipw_learner, gt_random_policy, gt_uniform_sample_weight_learner
def process(i: int):
# synthetic data generator
dataset = SyntheticBanditDataset(
n_actions=n_actions,
dim_context=dim_context,
reward_function=logistic_reward_function,
behavior_policy_function=linear_behavior_policy,
random_state=i,
)
# estimate the mean reward function of the train set of synthetic bandit feedback with ML model
regression_model = RegressionModel(
n_actions=dataset.n_actions,
action_context=dataset.action_context,
base_model=base_model_dict[base_model_for_reg_model](
**hyperparams[base_model_for_reg_model]),
)
ope_estimator = DoublyRobust()
# define evaluation policy using NNPolicyLearner
nn_policy = NNPolicyLearner(
n_actions=dataset.n_actions,
dim_context=dim_context,
off_policy_objective=ope_estimator.estimate_policy_value_tensor,
)
# baseline method 1. RandomPolicy
random_policy = RandomPolicy(n_actions=dataset.n_actions)
# baseline method 2. UniformSampleWeightLearner
uniform_sample_weight_policy = UniformSampleWeightLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
# sample new training and test sets of synthetic logged bandit feedback
bandit_feedback_train = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
bandit_feedback_test = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
# estimate the mean reward function of the train set of synthetic bandit feedback with ML model
estimated_rewards_by_reg_model = regression_model.fit_predict(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
n_folds=3, # 3-fold cross-fitting
random_state=12345,
)
# train the evaluation policy on the training set of the synthetic logged bandit feedback
nn_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
uniform_sample_weight_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
# predict the action decisions for the test set of the synthetic logged bandit feedback
nn_policy_action_dist = nn_policy.predict(
context=bandit_feedback_test["context"], )
random_action_dist = random_policy.predict(
context=bandit_feedback_test["context"], )
uniform_sample_weight_action_dist = uniform_sample_weight_policy.predict(
context=bandit_feedback_test["context"], )
# get the ground truth policy value for each learner
gt_nn_policy_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=nn_policy_action_dist,
)
gt_random_policy = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=random_action_dist,
)
gt_uniform_sample_weight_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=uniform_sample_weight_action_dist,
)
return gt_nn_policy_learner, gt_random_policy, gt_uniform_sample_weight_learner
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
estimated_rewards_by_reg_model=estimated_rewards_by_reg_model,
)
# predict the action decisions for the test set of the synthetic logged bandit feedback
random_action_dist = random_policy.compute_batch_action_dist(
n_rounds=n_rounds)
ipw_learner_action_dist = ipw_learner.predict(
context=bandit_feedback_test["context"], )
nn_policy_learner_action_dist = nn_policy_learner.predict_proba(
context=bandit_feedback_test["context"], )
# evaluate learners' performances using ground-truth polocy values
random_policy_value = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=random_action_dist,
)
ipw_learner_policy_value = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=ipw_learner_action_dist,
)
nn_policy_learner_policy_value = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=nn_policy_learner_action_dist,
)
policy_value_df = DataFrame(
[
[random_policy_value],
[ipw_learner_policy_value],
[nn_policy_learner_policy_value],
def process(i: int):
# synthetic data generator
dataset = SyntheticBanditDataset(
n_actions=n_actions,
dim_context=dim_context,
reward_function=logistic_reward_function,
behavior_policy_function=linear_behavior_policy,
random_state=i,
)
# sample new training and test sets of synthetic logged bandit data
bandit_feedback_train = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
bandit_feedback_test = dataset.obtain_batch_bandit_feedback(
n_rounds=n_rounds)
# defining policy learners
ipw_policy = IPWLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
q_policy = QLearner(
n_actions=dataset.n_actions,
base_model=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
nn_policy = NNPolicyLearner(
n_actions=dataset.n_actions,
dim_context=dim_context,
off_policy_objective="ipw",
)
# baseline method 1. RandomPolicy
random_policy = RandomPolicy(n_actions=dataset.n_actions)
# baseline method 2. UniformSampleWeightLearner
uniform_sample_weight_policy = UniformSampleWeightLearner(
n_actions=dataset.n_actions,
base_classifier=base_model_dict[base_model_for_evaluation_policy](
**hyperparams[base_model_for_evaluation_policy]),
)
# policy training
ipw_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
q_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
nn_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
uniform_sample_weight_policy.fit(
context=bandit_feedback_train["context"],
action=bandit_feedback_train["action"],
reward=bandit_feedback_train["reward"],
pscore=bandit_feedback_train["pscore"],
)
# prediction/making decisions
ipw_action_dist = ipw_policy.predict(
context=bandit_feedback_test["context"], )
q_action_dist = q_policy.predict(
context=bandit_feedback_test["context"], )
nn_action_dist = nn_policy.predict(
context=bandit_feedback_test["context"], )
random_action_dist = random_policy.predict(
context=bandit_feedback_test["context"], )
uniform_sample_weight_action_dist = uniform_sample_weight_policy.predict(
context=bandit_feedback_test["context"], )
# evaluation
gt_ipw_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=ipw_action_dist,
)
gt_q_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=q_action_dist,
)
gt_nn_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=nn_action_dist,
)
gt_random_policy = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=random_action_dist,
)
gt_uniform_sample_weight_learner = dataset.calc_ground_truth_policy_value(
expected_reward=bandit_feedback_test["expected_reward"],
action_dist=uniform_sample_weight_action_dist,
)
return (
gt_ipw_learner,
gt_q_learner,
gt_nn_learner,
gt_random_policy,
gt_uniform_sample_weight_learner,
)