Bandit Learning Through Biased Maximum Likelihood Estimation