部分可观测环境中基于图强化的任务卸载与资源分配方法  

A Graph Reinforcement-Based Approach to Task Offloading and Resource Allocation in Partially Observable Environment

作　　者：代钰[1] 景宗明 杨雷 高振 DAI Yu;JING Zong-ming;YANG Lei;GAO Zhen(School of Software,Northeastern University,Shenyang 110169,China;School of Computer Science&Engineering,Northeastern University,Shenyang 110169,China)

机构地区：[1]东北大学软件学院,辽宁沈阳110169 [2]东北大学计算机科学与工程学院,辽宁沈阳110169

出　　处：《东北大学学报(自然科学版)》2025年第1期9-17,25,共10页Journal of Northeastern University(Natural Science)

基　　金：国家重点研发计划项目(2021YFF0901205).

摘　　要：为了解决部分可观测环境中由于边缘服务器之间缺乏有效通信而导致的全局信息缺失问题,构建了基于图注意力机制的边缘服务器间沟通机制,将移动边缘计算(mobile edge computing,MEC)系统构建为图结构,使边缘服务器之间可以通过图中的边进行消息传递,进而间接得到MEC系统的全局状态信息.同时引入双注意力机制,使边缘服务器更多关注对策略优化更有用的通信消息,加快模型收敛速度并提高算法性能.仿真实验结果表明,与基线算法相比,本文所提出的算法可以有效降低任务完成时延与能耗,同时具有收敛速度快的优点.To address the issue of global information loss due to ineffective communication among edge servers in partially observable environment,an inter-edge server communication mechanism based on a graph attention mechanism is constructed,where the mobile edge computing(MEC)system is represented as a graph structure,allowing message passing between edge servers through the edges in the graph to indirectly obtain the global state information of the MEC system.The dual attention mechanism is introduced to enable agents to focus more on communication messages that are more useful for policy optimization,thereby accelerating the convergence speed of the model and improving algorithm performance.Simulation experimental results demonstrate that compared to baseline algorithms,the proposed algorithm effectively reduces task completion delay and energy consumption while exhibiting faster convergence speed.

关 键 词：移动边缘计算 深度强化学习 任务卸载 资源分配 消息通信 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]