On a Class of Restless Multi-armed Bandits with Deterministic Policies

We describe and analyze a restless multi-armed bandit (RMAB) in which, in each time-slot, the instantaneous reward from the playing of an arm depends on the time since the arm was last played. This model is motivated by recommendation systems where the payoff from a recommendation on depends the recommendation history. For an RMAB with N arms, and known reward functions for each arm that have a finite support (akin to a maximum memory) of M steps, we characterize the optimal policy that maximizes the infinite horizon time-average of the reward. Specifically, using a weighted-graph representation of the system evolution, we show that a periodic policy is optimal. Further, we show that the optimal periodic policy can be obtained using an algorithm with polynomial time and space complexity. Some extensions to the basic model are also presented; several more are possible. RMABs with such large state spaces for the arms have not been previously considered.