Learning-Guided Matching Game for Decentralized Task Offloading to Multi-Functional UAVs

In this paper, we consider multiple multi-functional unmanned aerial vehicles (UAVs) with sensing, communication, and computation capabilities to simultaneously perform sensing and provide computing services to ground devices (GDs). The UAVs process both, the sensed data and the tasks offloaded by the GDs, onboard. To enable a scalable decentralized design, we formulate the task offloading problem as a matching game. In each time slot, each GD proposes to a UAV for task offloading, aiming to maximize its utility, defined as the time saved through offloading compared to local computing. Each UAV accepts to serve a subset of proposing GDs while handling its own sensing task as well. To balance resource allocation between offloaded and sensing tasks, each UAV aims to maximize its utility defined as a weighted sum of the total time saved for GDs minus the extra time incurred in processing its own sensing data when sharing partial computing resources with GDs compared to utilizing the entire computing resources for its sensing task. However, in practice, GDs may lack prior information about their channel conditions, the UAVs' communication and computing resources, and the task offloading strategies of other GDs, therefore, solving the task offloading matching game is challenging. To tackle this, we propose a novel gradient-based Multi-Armed-Bandit (GMAB) algorithm for task offloading, enabling GDs to learn and coordinate their offloading strategies in a decentralized manner. Simulation results show that the proposed GMAB algorithm outperforms several baseline task offloading schemes in terms of improving both, GD-side and UAV-side utility, by up to 34.1 % and 37.2 %, respectively.