On effectiveness of the Mirror Decent Algorithm for a stochastic multi-armed bandit governed by a stationary finite Markov chain

Abstract

In this article, we study the effectiveness of the Mirror Descent Randomized Control Algorithm recently developed to a class of homogeneous finite Markov chains governed by the stochastic multi-armed bandit with unknown mean losses. We prove the explicit, non-asymptotic both upper and lower bounds for the mean losses at a given (finite) time horizon. These bounds are very similar as functions of problem parameters and time horizon, but with different logarithmic term and absolute constant. Numerical example illustrates theoretical results.