How to Learn on the Fly? On Improving the Uplink Throughput Performance of UAV-Assisted Sensor Networks

In UAV-assisted sensor networks, the random and unknown duty cycling of battery-powered ground Sensor Nodes (SNs) (such as wireless IoT devices) can significantly affect both energy consumption and communication performance of the data-collecting UAVs. The state of the art typically assumes a unified system with a coherent design of UAV and SNs, wherein the SNs' behavior is considered to be known, and accordingly the mobility of UAV is planned. In practical settings, however, SNs and UAVs can possibly be administered by independent stakeholders. In such an independent system, the UAV needs to learn (in online) its trajectory by understanding the random behavior of SNs in order to improve the uplink communication performance. In the independently administered UAV-assisted sensor networks, the task of maximizing the uplink throughput performance can be formulated as a sequential stochastic decision process. The key challenge of this formulation lies in choosing suitable online learning methods and selecting network parameters that help to learn and improve the throughput performance effectively. To this end, this work gives a principled approach to studying the effects of various learning methods and ways to incorporate network parameters to improve the throughput performance. The resultant online learning algorithm is demonstrated to exhibit superior performance in the extensive simulation-based evaluation study.