Energy-efficient UAV-aided computation offloading on THz band: a MADRL solution

• Main Authors: Li, Yuanjian, Madhukumar, A. S., Tan, Ernest Zheng Hui, Zheng, Gan, Saad, Walid, Aghvami, A. Hamid
• Other Authors: College of Computing and Data Science
• Format: Conference or Workshop Item
• Language: English
• Published: 2025
• Subjects: Computer and Information Science; Terahertz communication; Computation offloading; Multi-agent deep reinforcement learning
• Online Access: https://hdl.handle.net/10356/183054; https://edas.info/p31420

In this paper, the problem of energy-efficient unmanned aerial vehicle (UAV)-assisted computation offloading over the Terahertz (THz) spectrum is investigated. In the studied system, several UAVs are deployed as edge servers to aid task executions for multiple energy-limited computation-scarce terrestrial user equipments (UEs). Then, an expected energy efficiency maximization problem is formulated, aiming to jointly optimize UAVs’ trajectories, UEs’ local central processing unit (CPU) clock speeds, UAV-UE associations, time slot slicing, and UEs’ offloading powers. To tackle the considered multi-dimensional optimization problem, the duo-staggered perturbed actor-critic with modular networks (DSPAC-MN) solution in a multi-agent deep reinforcement learning (MADRL) setup, is proposed and tailored, after mapping the original problem into a stochastic (Markov) game. Compared to representative benchmarks in simulations, e.g., multi-agent deep deterministic policy gradient (MADDPG) and multi-agent twin-delayed DDPG (MATD3), the proposed DSPAC-MN can achieve the optimal performance of average energy efficiency, while ensuring 100% safe flights.