Year: 2026
Type: article
Abstract: The growing integration of AI and the Industrial Internet has intensified the need for rapid deployment of AI inference services at the edge.However, existing task offloading approaches often fail to address framework heterogeneity, computational load, and latency sensitivity, resulting in inefficient resource utilization and high delays.To tackle these challenges, this paper proposes IF-DQN, a network-aware, autonomous agent-driven task offloading framework for Model-as-a-Service (MaaS) scenarios.In IF-DQN, edge inference nodes are modeled as agents that continuously perceive bandwidth, latency, and resource load, while deep reinforcement learning (DQN) is employed to autonomously learn cross-layer offloading strategies, enabling dynamic task allocation and resource coordination across cloud-edge-end layers.An inference framework matching mechanism is further incorporated to ensure compatibility between tasks and runtime environments, enhancing stability and utilization.Experimental results on the self-developed MaaSim simulation platform show that IF-DQN achieves rapid convergence, higher cumulative rewards, reduced inference latency, improved resource utilization, and nearly 100% task completion rates under diverse conditions.These findings highlight IF-DQN as an effective and robust solution for embodied AI in collaborative cloud-edge-end environments.
Institution: Tongji University
Language: English
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7204773491