Empirical Likelihood for Contextual Bandits

Neural Information Processing Systems 

How can we find a tight confidence interval (CI) on counterfactual estimates?