An Empirical Evaluation of Thompson Sampling