Trust-Based Multi-Agent Imitation Learning for Green Edge Computing in Smart Cities

Abstract

Green communications and networking technologies boost the interconnection and communication of Internet of Things (IoT) devices, so as to facilitate the task offloading. Artificial Intelligence (AI) based task offloading scheme is being widely studied. However, most of AI based task offloading schemes only reward the devices that process tasks locally, and do not consider the untrusted devices. To solve these issues, a Trust based Multi-Agent Imitation Learning (T-MAIL) scheme is proposed by us to improve task offloading for edge computing in smart cities. Firstly, we established a full task offloading incentive model, in which edge devices can get comprehensive reward from local processing and task re-offloading. Secondly, we proposed an active trust acquisition method, which can obtain the device trust efficiently and accurately. Finally, the new task offloading incentive scheme and trust acquisition method are introduced into multi-agent imitation learning. The experimental results show that, the proposed T-MAIL will effectively improve task offloading. Compared with MILP and DQN based task offloading solution, the average task completion time is reduced by 5.5% and 52.7% respectively. Compared with MILP scheme, the task offloading rate is increased by 19.2%. In addition, the trust difference ratio between trusted devices and untrusted devices can reach 56.1%.