A Study on Task Offloading Based on Multi-Agent Advantage Actor Critic Algorithms in IIoT

With the rapid development of Industry 4.0, resource-constrained industrial devices are often unable to meet different demands of compute-intensive and latency-sensitive tasks in modern factories. To cope with this challenge, offloading tasks from industrial devices to edge servers or a cloud computing center to obtain lower delay is a good solution. Most of the traditional task offloading approaches rely on centralized decision-making systems, which may not be able to efficiently cope with device heterogeneity and the specific needs of different tasks in industrial environments. To address this problem, this study proposes a more comprehensive edge computing network model for Industrial Internet of Things (IIoT). By integrating an action masking technique, we establish a Markov decision model with a constrained action space. Additionally, we propose a task offloading framework based on the Multi-Agent Advantage Actor Critic (MAA2C) algorithm, adopting the Centralized Training Decentralized Execution (CTDE) architecture, aiming to achieve task offloading decision-making in modern factories and reduce the overall system delay. Validated through numerical simulation experiments, our proposed framework not only performs well in reducing system delay but also shows good adaptability in different industrial environments.