Adaptive Algorithms for Multi-armed Bandit with Composite and Anonymous Feedback