Queue-aware end-edge-cloud collaborative DNN inference in 6G industrial IoT: A multi-agent learning approach

The rapid advancement of artificial intelligence has introduced Deep Neural Network (DNN)-based inference tasks into Industrial Internet of Things (IIoT) systems. These tasks are often computation-intensive, requiring substantial resources beyond the capability of sensing devices. To address this challenge, edge computing has emerged as a solution, enabling collaborative DNN inference through task offloading. In this paper, we formulate the problem of joint DNN partitioning, task offloading, and priority-aware scheduling in end-edge-cloud systems as a dynamic long-term optimization. The objective is to minimize inference task completion cost while guaranteeing queue stability for all priority levels. To this end, we first leverage queue status information to dynamically redirect low-priority tasks to the cloud server when the deadline violation probability exceeds a threshold. Afterwards, we propose a Multi-Agent Proximal Policy Optimization (MAPPO) algorithm under Centralized Training with Decentralized Execution (CTDE) to jointly optimize partitioning and edge selection. Furthermore, we demonstrate that our framework reduces the average inference delay by 18.38% and the energy consumption by 26.86% compared with baselines across three DNN models.