边缘计算中时延敏感的启发式任务卸载方法

Latency-sensitive heuristic task offloading method in edge computing

关键词:  边缘计算  深度强化学习  任务卸载  多智能体  启发式算法

Key words:  edge computing  deep reinforcement learning  task offloading  multi-agent  heuristic algorithm

摘要:

针对移动边缘计算在多用户环境中卸载决策设计困难导致负载失衡、总时延过高和响应延迟问题,提出一种时延敏感的启发式任务卸载方法。首先,为解决边缘设备计算资源匮乏和电量不足的问题,建立以边缘服务器为主体的卸载架构,构建系统模型与时延优化模型;其次,提出一种改进的近端策略优化算法I-PPO,通过增加离线训练阶段、设计考虑多智能体决策影响的奖励机制、在特征中融入基于特定智能体的全局信息,使算法能够适用于多用户场景;进一步,在I-PPO的基础上,在任务卸载执行过程中引入任务优先级调度决策,构造时延敏感的轻量级启发式任务卸载算法,以降低系统时延,并提升用户体验。仿真实验表明,相比同类算法,所提I-PPO算法在收敛速度、寻优能力和鲁棒性方面表现更优,且适用于多智能体环境；所提算法在系统总时延和边缘服务器负载均衡方面优于对比算法,并具有良好的稳定性。

Abstract:

In order to address the challenge of designing reasonable offloading decisions for mobile edge computing (MEC) in multi-user environments, which leads to load imbalance, excessive total latency, and response delays, this paper proposed a latency-sensitive heuristic task offloading method. Firstly, to address the issues of limited computational resources and insufficient battery power of edge devices during computation task processing, the paper introduced an edge server-centric offloading paradigm and established a system model and a latency optimization model. Subsequently, it introduced an improved proximal policy optimization algorithm (I-PPO), which extended the offline training process, designed a reward mechanism that considers the impact of multi-agent decisions, and incorporated global information based on specific agents into the features, enabling the algorithm to be suitable for multi-user environments. Furthermore, building upon I-PPO, the paper introduced task priority scheduling decisions into the task offloading execution process, resulting in the development of a latency-sensitive lightweight heuristic task offloading algorithm, denoted as HTAI. This further optimized system latency and enhanced user satisfaction. Simulation experiments demonstrate that the I-PPO algorithm proposed in this paper, compared to similar algorithms, effectively improves convergence speed, optimization capability, and robustness, and it can be applied in multi-agent environments. Moreover, the algorithm proposed herein outperforms other algorithms in terms of total system latency and edge server load balance, exhibiting strong stability.