An online learning algorithm that selects actions by optimizing past losses plus a regularization term to encourage exploration.