Recharging Bandits: Learning to Schedule Recurring Interventions

ABSTRACT: Traditional multi-armed bandit models posit that the payoff distribution of each action (or "arm") is stationary over time, and hence that the goal of learning is to identify the arm with the highest expected payoff and choose that one forever after. However, in many applications the efficacy of an action depends on the amount of time that has elapsed since it was last performed. Examples arise in precision agriculture, online education, and music recommendations. In this talk we introduce a generalization of the multi-armed bandit problem that models such applications. In the course of analyzing algorithms for this problem, we will encounter some interesting combinatorial questions about coloring the integers subject to bounds on the sizes of subintervals that exclude a given color.

This talk is based on joint work with Nicole Immorlica.