A Cybertwin-Driven Task Offloading Scheme Based on Deep Reinforcement Learning and Graph Attention Networks

With the improvement of mobile computing capability, the service demands of user terminals are also increasing. Therefore, task offloading in mobile edge computing (MEC) has always been a research focus. However, most of task offloading schemes are confined to current network paradigms. During the shaping of 6G, it is of value to explore novel task offloading architectures and methods which fit future communication system. In this paper, we present a cybertwin-based system that allows flexible resource deployment and coordination. In addition, by targeting at the graph structure of this system, we propose an energy-efficient and security-enhanced partial-offloading scheme based on a combination of deep reinforcement learning (DRL) and modified graph attention network (GAT). Numerical results demonstrate that the proposed training method converges better and the trained GAT model achieves utility improvement.