Implementing Low Latency and High Energy Efficiency Task Scheduling in MEC Systems Using Improved DDPG Algorithm

Lihong Zhao

doi:10.1109/ACCESS.2024.3502400

2024-01-01

Lihong Zhao

X. X. Ding *

Shuqin Wang

DOI：10.1109/ACCESS.2024.3502400

原文链接

原文求助

摘要

En 中文

The Mobile Edge Computing (MEC) system is crucial in modern networks, particularly for applications requiring low latency and high energy efficiency. Efficient task scheduling in MEC environments to balance these requirements is challenging. This paper presented a method that improves the Deep Deterministic Policy Gradient (DDPG) algorithm for better performance in MEC systems. Stability is enhanced through a soft update mechanism for target network parameters and an experience replay buffer to store past experiences, reducing sample correlation and improving training efficiency. It also proposed a reward function considering both latency and energy consumption, comparing the effects of linear and nonlinear reward functions on system performance. The task scheduling process includes state representation, action selection, environment feedback, network updates, and target network soft updates. The experimental results indicate that by analyzing the performance of different reward functions, a nonlinear reward function has been determined to achieve faster convergence. The improved DDPG algorithm outperforms traditional task scheduling algorithms in terms of latency and energy consumption under different task loads, and also has significant advantages in robustness. The ablation experiment further verified the positive contribution of each improved part to the overall performance. At Task load 10, the latency of DDPG, DDPG+Soft update, DDPG+Experience reply buffer, and this paper model were 267 ms, 263 ms, 260 ms, and 250 ms, respectively, with energy consumption of 988J, 974 J, 969 J, and 965 J. The improved DDPG algorithm in this paper not only effectively solves the task scheduling problem in MEC systems, but also provides important references for future research.

Keyword:

Training

Servers

Processor scheduling

Convergence

Scheduling

Optimization

Energy consumption

Stability analysis

Correlation

System performance

Mobile edge computing