We built this repository on Open Bandit Pipeline library to implement our DR-IC estimator and benchmark it with the other estimators. We worked with this commit-version of the Open Bandit Pipeline library.
Create a virtual environment. For example
conda create -n ope python=3.7
In the virtual environment (do source activate ope), install this package:
cd zr-obp-code
python setup.py install
python -m examples.multiclass.all_est_evaluate --n_runs 500 --dataset_name letter --eval_size 0.3 --base_model_for_behavior_policy logistic_regression --base_model_for_evaluation_policy logistic_regression --base_model_for_reg_model logistic_regression --n_jobs -1 --save_logs True --stoc_reward True --random_state 12345
where
$n_runsspecifies the number of simulation runs in the experiment to estimate standard deviations of the performance of OPE estimators.$dataset_namespecifies the name of the multi-class classification dataset and should be one of the 10 UCI datasets.$eval_sizespecifies the proportion of the dataset to calculate the ground truth.$base_model_for_behavior_policyspecifies the base ML model for defining behavior policy and should be one of "logistic_regression", "random_forest", or "lightgbm".$base_model_for_evaluation_policyspecifies the base ML model for defining evaluation policy and should be one of "logistic_regression", "random_forest", or "lightgbm".$base_model_for_reg_modelspecifies the base ML model for defining regression model and should be one of "logistic_regression", "random_forest", or "lightgbm".$n_jobsis the maximum number of concurrently running jobs.save_logsis a flag to save the logs in thelogs/dataset-namepath.stoc_rewardis a flag that enables stochastic reward whenTrue.random_stateis the seed value used for all randomness in the code (numpy.random, sklearn, etc).
Use the bash script run.sh to execute more seeds by changing the dataset names and seeds if necessary.