Deep-Reinforcement-Learning-Based Computation Offloading for Servicing Dynamic Demand in Multi-UAV-Assisted IoT Network

• Published in: IEEE internet of things journal Vol. 11; no. 10; pp. 17249 - 17263
• Main Authors: Lin, Na, Bai, Lu, Hawbani, Ammar, Guan, Yunchong, Mao, Chaojin, Liu, Zhi, Zhao, Liang
• Format: Journal Article
• Language: English; Japanese
• Published: Piscataway IEEE 15.05.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Autonomous aerial vehicles; Computation offloading; Deep learning; deep reinforcement learning (DRL); Dynamic scheduling; Edge computing; Energy consumption; Greedy algorithms; Heuristic algorithms; Internet of Things; Mixed integer; Mobile computing; mobile edge computing (MEC); Nonlinear programming; Resource management; Servers; Task analysis; Task scheduling; unmanned aerial vehicle (UAV); Unmanned aerial vehicles; Wireless networks
• ISSN: 2327-4662; 2327-4662
• DOI: 10.1109/JIOT.2024.3356725

In wireless networks, meeting the performance requirements of all tasks solely with Internet of Things (IoT) devices is challenging due to their limited computational power and battery capacity. Given their flexibility and mobility, the application of unmanned aerial vehicles (UAVs) in the context of mobile edge computing (MEC) has garnered significant interest within the sector. However, UAVs also face constraints in terms of resources like storage and computational power. Therefore, it is vital to develop effective UAV assistance solutions to provide long-term demands of in-network services. The dynamic scheduling and computation offloading of UAVs is the subject of this article. Specifically, we propose a deep deterministic policy gradient algorithm based on a greedy (DDPGG) strategy to jointly optimize dynamic scheduling, device association, and task allocation of UAVs, with the goal of minimizing the weighted sum of total system energy consumption and time delay. The problem is formulated as a nonlinear programming problem involving mixed integers. The simulation results demonstrate that the DDPGG algorithm we have proposed exhibits a higher level of performance in comparison to its competitors.