Feature: Make Hyperparameter tuning method flexible

[usaito]

new feature

• The original implementation (Feature: Implementing Automatic Hyperparameter Tuning #116) follows the automatic hyperparameter tuning procedure proposed in Su et al.(2020).
• I found that the original procedure favors the bias of the estimator, and thus tends to select a large value as a hyperparameter value.
• This is because only the bias term is replaced by its bias upper bound. The variance term is estimated by the sample variance, not its upper bound.
• Thus, I added an option not to use the bias upper bound in the hyperparameter tuning process. What we have to do is just to set use_bias_upper_bound=False when initializing the estimators with hyperparameter tuning. This alternative procedure works as follows:

where \hat{Bias} is the direct bias estimation. This will often lead to a better empirical performance.

reference
Yi Su, Maria Dimakopoulou, Akshay Krishnamurthy, and Miroslav Dudik.
"Doubly Robust Off-Policy Evaluation with Shrinkage.", 2020.

fix

• changed the name of the hyperparameter of Switch-DR from tau to lambda. Now, the name of the hyperparameter is consistent across OPE estimators.