A Closer Look at the Worst-case Behavior of Multi-armed Bandit Algorithms