A Scalable Computation Offloading Scheme for MEC Based on Graph Neural Networks

Recently, computation offloading has become a promising technology to support computation-intensive applications on resource-limited mobile devices (MDs). However, existing offloading schemes suffer from poor generalization and scalability to different applications, since they mainly consider one or a few specific task dependency topologies of the application. Different from the existing schemes, this paper introduces a novel neural network, i.e., graph neural network (GNN), to adapt to the diversity of the applications. Therefore, we propose a GNN and deep reinforcement learning (DRL)-based scheme that optimizes offloading to minimize the long-term average application completion time and MD’s energy consumption. Specifically, considering the variation of the wireless network, a fine-grained task-level scheduling mechanism is presented. In addition, a scalable GNN-enabled preprocessing network is designed to process the state of the various applications and the wireless network. Furthermore, we obtain the optimal scheduling policy via training based on policy gradient algorithm under the stochastic application arrival of the unlimited inter arrival time. Extensive simulations demonstrate the effectiveness and superiority of the proposed scheme under different average inter arrival times. Specifically, the proposed scheme can lower at least 172% of the application completion time and 45% of the energy consumption of the MD compared with other baseline schemes.