Add/Modify Dataset Class for Handling Multiple Loggers and Deficient Data

[usaito]

Overview

• Implement n_deficient_actions argument to obp.dataset.SyntheticBanditDataset. This argument controls the number of deficient actions having zero probability of being selected by the behavior policy. If there are some deficient actions, the full/common support assumption is very likely to be violated, producing some bias for the IPW-type estimators. See Sachdeva et al.(2020) for details.
• Implement obp.dataset.SyntheticBanditDatasetWithMultipleLoggers, which synthesize logged bandit data with multiple logging/behavior policies. This setting is closely related to the stratified sampling setting and specific estimators should be used for an accurate OPE. See Agarwal et al.(2018) and Kallus et al.(2021) for details. Some estimators will be implemented in another PR.

Minor

• Fix documents, comments, error messages.
• Update README

References

• Noveen Sachdeva, Yi Su, and Thorsten Joachims.
  "Off-policy Bandits with Deficient Support.", KDD2020.
  https://arxiv.org/abs/2006.09438

• Aman Agarwal, Soumya Basu, Tobias Schnabel, Thorsten Joachims.
  "Effective Evaluation using Logged Bandit Feedback from Multiple Loggers.", KDD2018.
  https://arxiv.org/abs/1703.06180

• Nathan Kallus, Yuta Saito, and Masatoshi Uehara.
  "Optimal Off-Policy Evaluation from Multiple Logging Policies.", ICML2021.
  http://proceedings.mlr.press/v139/kallus21a.html