Learning to Play Bayesian Games