Scholarlink

In UAV-assisted Mobile Edge Computing (MEC) systems, the joint optimization of UAV trajectory and resource allocation is a major challenge due to high environmental dynamics, the complex coupling between continuous trajectory control and resource management, and an NP-hard resource allocation subproblem that demands computationally efficient solutions. To tackle this, we propose a novel hierarchical framework that decouples long-term trajectory planning from instantaneous resource allocation. Our framework features a high-level Deep Reinforcement Learning (DRL) agent, based on an attention-enhanced Soft Actor-Critic (SAC) algorithm, which learns an adaptive flight policy. For any given UAV position, a low-level greedy solver then efficiently manages partial computation offloading and resource distribution based on marginal utility. Extensive simulations demonstrate that our proposed approach significantly outperforms several DRL and heuristic baselines by achieving a higher task completion rate, lower average task latency, and more stable training convergence, thus validating the effectiveness of a hierarchical framework for complex, real-time decision-making in dynamic UAV-MEC environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hierarchical Deep Reinforcement Learning Framework for Joint Trajectory and Resource Allocation in UAV-MEC Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tianrong Wu,
  • Haifeng Sun

摘要

In UAV-assisted Mobile Edge Computing (MEC) systems, the joint optimization of UAV trajectory and resource allocation is a major challenge due to high environmental dynamics, the complex coupling between continuous trajectory control and resource management, and an NP-hard resource allocation subproblem that demands computationally efficient solutions. To tackle this, we propose a novel hierarchical framework that decouples long-term trajectory planning from instantaneous resource allocation. Our framework features a high-level Deep Reinforcement Learning (DRL) agent, based on an attention-enhanced Soft Actor-Critic (SAC) algorithm, which learns an adaptive flight policy. For any given UAV position, a low-level greedy solver then efficiently manages partial computation offloading and resource distribution based on marginal utility. Extensive simulations demonstrate that our proposed approach significantly outperforms several DRL and heuristic baselines by achieving a higher task completion rate, lower average task latency, and more stable training convergence, thus validating the effectiveness of a hierarchical framework for complex, real-time decision-making in dynamic UAV-MEC environments.