Proactive handover for task offloading in UAVs

Abstract

Unmanned Aerial Vehicles (UAVs) are usually deployed alongside Internet of Things (IoT) devices in smart city applications, particularly for critical tasks such as disaster management that require continuous service. UAVs often handle resource-intensive and sensitive tasks through offloading, but unexpected task interruptions due to UAV dropouts can generate safety risks and increase costs. Although existing approaches in the literature have already addressed proactive handovers to mitigate such disruptions, their primary focus is on communication issues arising from UAV movement and are unable to handle offloading related issues. In this paper, we include in our model, in addition to communication, factors such as energy, computation requirements, and dynamic environmental conditions (e.g., wind speed and incentive), pushing toward a comprehensive solution for UAV task offloading and resource allocation. In fact, we formulate our problematic as a Markov game, which we solve using a Multi Agent Deep Q Network (MADQN). In our experiments, we assessed our approach using a federated learning scenario to illustrate its effectiveness in a realistic distributed application setting against several baselines from the state of the art. Results showed that our approach outperforms its peers in terms of system utility, and tradeoff between cost and dropout rates, leading to an improved handover management of computational and energy resources in UAV-IoT based systems. In fact, it reduces the dropout rate by approximately 45% compared to the second-best baseline, leading to a 2% improvement in model accuracy and a 50% reduction in deployment costs.