Overload protection in mobile edge computing using multi-agent reinforcement learning

近年来，实时在线游戏、视频点播、增强和虚拟现实应用等计算密集型智能手机应用激增。由于高计算需求，这些服务经常被移动到使用移动边缘计算(MEC)的所谓边缘服务器。为了确保MEC集群的可靠性，防止服务器过载是很重要的。这种服务器过载可以通过先发制人地卸载请求并将其路由到其他服务器或完全丢弃它们来防止。在这个项目中，我们建议开发新的理论和算法来利用多智能体强化学习(MAIL)来改进作业卸载和路由策略。具体地说，我们将：(I)为多服务器MEC集群建立一个基于多智能体多阶段优化问题的数学模型；(Ii)识别最优接纳控制和路由策略的相关定性属性；(Iii)开发低复杂度的MAIL算法，该算法利用前面步骤中得到的定性属性；以及(Iv)进行详细的基于模拟的研究，以评估所提出的算法在真实环境中的性能。预计该项目将导致今后在顶级会议和期刊上发表文章。

In recent years, there has been a proliferation of computationally intensive smartphone applications such as real-time online gaming, video-on-demand, augmented and virtual reality applications, etc. Due to the high computational demand, these services are often moved to what are called Edge Servers using a paradigm known as Mobile Edge Computing (MEC).To ensure the reliability of a MEC cluster, it is important to prevent overloading of the servers. Such server overloading can be prevented by pre-emptively offloading the requests and routing them to other servers or completely dropping them. In this project, we propose to develop new theories and algorithms to improve job offloading and routing policies using multi-agent reinforcement learning (MARL).Specifically, we will: (i) develop a multi-agent multi-stage optimization problem based mathematical model for multi-server MEC clusters; (ii) identify pertinent qualitative properties of optimal admission control and routing policies; (iii) develop low-complexity MARL algorithms which exploit the qualitative properties derived in the previous step; and (iv) perform a detailed simulation-based study to evaluate the performance of the proposed algorithms in real-world settings. It is anticipated that the project will lead to future publications in top-tier conferences and journals.