We propose an online power allocation algorithm for optimizing energy efficiency (throughput per unit of transmit power) in multi-user, multi-carrier systems that evolve dynamically over time (e.g. due to changes in the wireless environment or the users’ load). Contrary to the static/ergodic regime, a fixed optimal power allocation profile (either static or in the mean) does not exist, so we draw on exponential learning techniques to design an algorithm that is able to adapt to system changes “on the fly”.
Specifically, the proposed transmit policy leads to no regret, i.e., it is asymptotically optimal in hindsight, irrespective of the system’s evolution over time. Importantly, despite the algorithm’s simplicity and distributed nature, users are able to track their individually optimum transmit profiles as they vary with time, even under rapidly changing network conditions.