Multi-Agent DRL Based Multi-UAV Trajectory Optimization for Energy Efficiency Improvement
Year: 2026
Type: conference-paper
Abstract: This paper addresses a UAV trajectory optimization problem that considers energy efficiency in an environment where multiple$\mathbf{U A V s}(\mathbf{N}=\mathbf{2}, \mathbf{4}, \mathbf{6}, \mathbf{8})$move in three-dimensional space to provide uplink service to multiple ground users. User association and bandwidth allocation are determined by heuristic rules based on channel state and remaining data demand, while deep reinforcement learning (DRL) controls the three-dimensional acceleration of each UAV. We compare five DRL algorithms spanning centralized and multi-agent paradigms: DDPG and TD3 as centralized single-agent baselines, MADDPG and MATD3 under the centralized training with decentralized execution (CTDE) framework, and MAPPO as an on-policy multi-agent approach. Experiments across four UAV scales reveal that energy efficiency peaks at N = 4 under a fixed-spectrum assumption, and that MADDPG degrades at$\mathbf{N} \boldsymbol{=} \mathbf{8}$due to quadratic critic input growth. MAPPO achieves competitive performance at small N but requires significantly more training episodes to converge at large N, a limitation confirmed by extended 2000-episode experiments.
Authors: Joon-Soo Eom, Oh-Soon Shin
Institution: Soongsil University
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224759366