From Bandits to Experts: On the Value of Side-Observations