Preference-centric Bandits in Wireless Communications: Theory and Applications

Data-driven methods are increasingly preferable to model-driven approaches in wireless communications. We study resource allocation via multi-armed bandits (MAB) which balances exploration and exploitation under unknown, time-varying channels. Moving beyond average-based metrics, we adopt the preference metric-centric framework which subsumes them, and through distortion functions, captures distribution tails relevant to the problem. This framework unifies existing formulations and extends naturally to problems lacking a standard model. We design an algorithm tailored to parametric family of distributions, and under mild assumptions, establish regret guarantees and finally, present empirical evaluations.