An Optimal Elimination Algorithm for Learning a Best Arm Avinatan Hassidim

We consider the classic problem of (ɛ, δ)-PAC learning a best arm where the goal is to identify with confidence 1 δ an arm whose mean is an ɛ-approximation to that of the highest mean arm in a multi-armed bandit setting. This problem is one of the most fundamental problems in statistics and learning theory, yet somewhat surprisingly its worst case sample complexity is not well understood. In this paper we propose a new approach for (ɛ, δ)-PAC learning a best arm. This approach leads to an algorithm whose sample complexity converges to exactly the optimal sample complexity of (ɛ, δ)-learning the mean of n arms separately and we complement this result with a conditional matching lower bound.