Worst-case Performance of Greedy Policies in Bandits with Imperfect Context Observations