DRL-Based Joint Clustering and Trajectory Optimization for UAV-Assisted Emergency Networks

Document Type

Article

Publication Date

1-1-2026

Abstract

The sixth-generation (6G) mobile network is expected to trigger an unprecedented surge in data traffic. unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) has emerged as a promising paradigm for emergency communications, yet dynamic user mobility and heterogeneous task demands pose severe challenges. Existing reinforcement learning (RL) methods (such as proximal policy optimization (PPO)), while effective in continuous control, often suffer from unstable convergence and neglect task-priority differentiation, leading to excessive delays for critical tasks. To address these issues, we propose a novel KM-PPO framework that integrates interference-aware k-means clustering with PPO. The clustering stage aggregates users by spatial distribution and task urgency, yielding an interference-suppressed initial UAV deployment. The trajectory optimization stage employs PPO with clipped probability ratios, which serve to constrain policy updates and maintain stability, and a priority-sensitive reward function, enabling UAVs to adaptively adjust motion trajectories under dynamic conditions. Compared with the baseline algorithms of k-means integrated with twin delayed deep deterministic policy gradient (KM-TD3) and k-means integrated with soft actor-critic (KM-SAC), our method reduces average system latency by 15.9% and 24.6%, respectively, while achieving an 86.7% success rate for high-priority tasks. These results demonstrate that KM-PPO not only ensures stable convergence in the environments but also guarantees quality of service (QoS) for mission-critical tasks, highlighting viability for UAV MEC deployments.

Keywords

deep reinforcement learning (DRL), emergency communication, trajectory optimization, UAV, user association

Publication Title

Iet Communications

ISSN

17518628

DOI

10.1049/cmu2.70126

Volume

20

Issue

1

Share

COinS