Task Offloading in Computing Power Network: A Graph Reinforcement Learning Approach

doi:10.1109/jiot.2026.3742019

Task Offloading in Computing Power Network: A Graph Reinforcement Learning Approach

2026-10-13

0

PRE

AI

DOI：10.1109/jiot.2026.3742019

Original

Original request for help

Save

Abstract

En 中文

Computing Power Network (CPN) is a novel paradigm that moves beyond Mobile Edge Computing (MEC) by integrating and jointly managing both computing and networking resources, thereby addressing MEC’s limited adaptability to network fluctuations in massive Internet of Things (IoT) scenarios. However, existing studies on task offloading in CPN often fail to fully mine the impact of network topological structures on offloading decisions, hindering the characterization of topology-aware computing-networking features and consequently leading to suboptimal offloading schemes under complex topological structures. To address the aforementioned challenge, we propose a graph reinforcement learning approach for task offloading in the CPN environment, called TO-GRAS, which matches optimal edge nodes and plans optimal routing paths for tasks. Specifically, the TO-GRAS first introduces a Graph Neural Network (GNN) to characterize the topological structure of the CPN and employs Deep Reinforcement Learning (DRL) to determine a computation task scheme. Subsequently, the TO-GRAS derives a routing planning scheme based on real-time network states. During runtime, the computation task scheme and routing planning scheme are adjusted alternately over multiple rounds, iteratively achieving optimal task offloading. Simulation experiments validate the effectiveness and adaptability of the TO-GRAS to complex topological structures. Moreover, compared with state-of-the-art approaches, the TO-GRAS reduces the average response time by 0.82%–25.21% and the average energy consumption by 1.80%–24.79%, respectively.

Keywords:

Mobile Internet of Things

computing power network

mobile edge computing

task offloading

graph neural network

deep reinforcement learning

Task Offloading in Computing Power Network: A Graph Reinforcement Learning Approach

Abstract

Journal

Organization

Cited Papers