Learning to Allocate Resources with Censored Feedback
The study addresses allocating a budget across K arms over T rounds when rewards are censored—an arm succeeds only if it's activated and its allocation exceeds a random threshold. The authors prove a lower bound of Omega(T^(1/3)) for this problem. They propose RA-UCB, an optimistic algorithm that achieves O(sqrt(T)) regret when the budget is known, and MG-UCB for unknown budgets, showing effective strategies for balancing exploration and exploitation.