A New Approach to Correlated Multi Armed Bandits

The paper considers a multi-armed bandit problem where the rewards obtained from different arms are correlated with one another. The source of this correlation is taken to be a hidden random variable having an unknown distribution. We propose a new approach for tackling the regret minimization problem for such a correlated bandit framework. Our approach classifies some bandit arms as non-competitive based on a competency criteria applied over a confidence set. We compare the performance of our approach to the well known UCB algorithm and to that of CUCB - an earlier algorithm designed to tackle the correlated bandit framework.