Learning-Aided Multi-RAT Operation for Battery Lifetime Extension in LPWAN Systems

Abstract

End-device (ED) lifetime is considered to be a crucial design factor in radio systems for massive machine-type communications. This parameter is heavily impacted by the continuously changing propagation conditions between the ED and the base station. In this paper, to extend the ED lifetime, we consider equipping a single ED with multiple low-power wide-area network (LPWAN) technologies to dynamically select the one with lower energy consumption. To facilitate this process, we propose to employ reinforcement learning (RL) algorithms. Assessing the resultant performance, we conduct two large-scale measurement campaigns that characterize the ED power consumption and the time-dependent propagation conditions for NB-IoT, Sigfox, and LoRaWAN technologies. Our numerical results demonstrate that the designed schemes effectively reduce ED power consumption by timely reacting to the varying radio conditions. Consequently, the ED lifetime expectancy is prolonged by around 10%. For instance, the Thompson sampling technique delivers the most consistent results by outperforming its counterparts and allowing to exploit up to 99% of the theoretical gains while converging over only 25-50 samples.