Greedy confidence bound techniques for restless multi-armed bandit based Cognitive Radio

Abstract

In this paper, we deal with Bayesian restless multi-armed bandit (RMAB) techniques which are appliced to Cognitive Radio. We assume there are multiple arms, each of which evolves as a Markov chain with known parameters. A player seeks to activate more than one arms at each time in order to maximize the expected total reward with multiple plays. We consider non-Bayesian RMAB where the parameters of the Markov chain are unknown. We propose a simple but effective algorithm called two-slot greedy confidence bound algorithm (Two-slot GCB), which perform better than existing upper confidence bound (UCB) algorithms.