基于深度强化学习的边缘视频任务卸载策略  

Edge video offloading strategy based on deep reinforcement learning

作　　者：周陈静 骆淑云 ZHOU Chenjing;LUO Shuyun(School of Computer Science and Technology(School of Artificial Intelligence),Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学计算机科学与技术学院(人工智能学院),杭州310018

出　　处：《智能计算机与应用》2024年第2期118-123,共6页Intelligent Computer and Applications

基　　金：浙江省尖兵研发攻关计划项目(2023C01041)。

摘　　要：随着物联网的发展以及智能设备的普及,视频处理技术已广泛应用于生活中。自动驾驶、产品质检等应用场景对视频处理技术的实时性需求逐步提高,移动边缘计算为计算能力不足和能源受限的设备提供计算资源以执行时延敏感性任务,为实时视频处理提供了新的计算架构。本文搭建了一个视频计算卸载场景,并以视频检测为任务,以系统时延为优化目标,建立了计算卸载模型和马尔可夫决策模型;考虑到计算卸载场景的复杂动态因素,如带宽波动、设备数量、任务大小等,以最小化系统时延为目标,提出了一种基于深度强化学习的计算卸载策略进行求解。实验表明,与其他基线方案相比,该卸载策略能够适应较复杂卸载场景,有效降低系统时延。With the development of the Internet of Things and the proliferation of smart devices,video processing technology has been widely applied in daily life.Applications such as autonomous driving and product quality inspection increasingly demand realtime video processing.Mobile edge computing provides computing resources for devices with limited computational power and energy,supporting time-sensitive tasks and offering a new computational architecture for real-time video processing.This paper constructs a video computation offloading scenario focused on video detection tasks,with system latency as the optimization objective.It establishes a computation offloading model and a Markov decision model.Considering the complex dynamic factors of the computation offloading scenario,such as bandwidth fluctuations,number of devices,and task size,a computation offloading strategy based on deep reinforcement learning is proposed to minimize system latency.Experiments show that this offloading strategy adapts well to complex offloading scenarios and effectively reduces system latency.

关 键 词：移动边缘计算 实时视频处理 深度强化学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN929.5[自动化与计算机技术—控制科学与工程]