摘要
如何合理地对车辆上的计算任务进行卸载,以降低任务的响应时间、RSU的平均负载率,是车辆边缘计算(VehicularEdgeComputing,VEC)中的重要问题。深度强化学习是解决计算卸载问题的有力工具,但当前深度强化学习方法往往未考虑具有依赖关系的卸载任务之间是否可以合并以及路侧单元间的负载均衡。对此,文章提出一种基于任务依赖和负载均衡的深度确定性策略梯度(Deep Deterministic Policy Gradient,DDPG)计算卸载方案;通过任务分类合并,减小DDPG的动作空间及批量处理任务,降低卸载任务的平均响应时间;综合考虑任务群组的优先级、紧迫程度和关键路径长度来对任务群组进行排序,以提高任务的卸载成功率;设计一种RSU动态权重分配机制,辅助DDPG实现RSU之间的负载均衡。实验结果表明,相比已有基于DQN(DeepQ-Network)和DDPG的计算卸载方法,本文所提方案的平均响应时间分别降低了21.4%和22.5%,RSU平均负载率也显著低于DQN和DDPG。
Efficiently offloading computational tasks from vehicles to reduce task response times and improve the average load rate of Roadside Units(RSUs)is a critical challenge in Vehicular Edge Computing(VEC).Deep Reinforcement Learning(DRL)has emerged as a powerful tool for addressing computational of-floading problems.However,existing DRL-based approaches often fail to consider whether dependent offloading tasks can be merged and overlook the load balancing among roadside units.To address these limitations,this paper proposes a Deep Deterministic Policy Gradient(DDPG)-based computa-tional offloading scheme that incorporates task dependency and load balancing.By categorizing and consolidating tasks,the action space of DDPG is reduced,and tasks are processed in batches,thereby decreasing the average response time for offloaded tasks.It further prioritizes task groups based on their priority,urgency,and critical path length to enhance offloading success rates.Additionally,a dy-namic weight allocation mechanism for RSUs is designed to assist DDPG in achieving load balancing among RSUs.Experimental results demonstrate that,compared to existing Deep Q-Network(DQN)and DDPG-based offloading methods,the proposed scheme reduces the average response time by 21.4%and 22.5%,respectively,while significantly lowering the average RSU load rate.
作者
钱振
Zhen Qian(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)
出处
《建模与仿真》
2025年第5期336-352,共17页
Modeling and Simulation
关键词
计算卸载
路侧单元
深度强化学习
任务依赖
负载均衡
Computation Offloading
Roadside Unit
Deep Reinforcement Learning
Task Dependency
Load Balancing
