Abstract
We study the multi-armed bandit problem with arms which are Markov chains with rewards. In the finite-horizon setting, the celebrated Gittins indices do not apply, and the exact solution is intractable. We provide approximation algorithms for the general model of Markov decision processes with nonunit transition times. When preemption is allowed, we provide a (1/2 − ε)-approximation, along with an example showing this is tight. When preemption isn’t allowed, we provide a 1/12-approximation, which improves to a 4/27-approximation when transition times are unity. Our model captures the Markovian Bandits model of Gupta et al., the Stochastic Knapsack model of Dean et al., and the Budgeted Learning model of Guha and Munagala. Our algorithms improve existing results in all three areas. In our analysis, we encounter and overcome to our knowledge a new obstacle: an algorithm that provably exists via analytical argu- ments, but cannot be found in polynomial time.