Contextual Bandit Learning With Predictable Rewards