Bandit Problems—Sequential Allocation of Experiments