Adversarial bandit problems: the power of randomization

In this tutorial we discuss sequential prediction problems in which the forecaster has limited information about the past outcomes of the sequence. We concentrate on the so-called "adversarial" framework in which no probabilistic is available for the sequence. We describe various models of limited feedback and pay special attention to the so-called "multi-armed bandit" problem. We discuss various randomized prediction methods and analyze their behavior.