User-Oriented Dynamic MEC Application Deployment in Edge Cloud Network

Multi-Access Edge Computing (MEC) have become the core technologies to meet users’ needs for 5G and beyond wireless networks. MEC applications can be flexibly created and placed at the network edge through virtual network functions (VNFs) to provide users with specific services with lower latency. In this paper, we consider a multi-user dynamic MEC network, where user trajectories follow Lévy walks, and each user has a set of requested target MEC applications. Our goal is to obtain an online deployment algorithm that is able to serve dynamic user requests within a tolerable latency while balancing the computational load among Mobile Edge Platforms (MEPs) as much as possible. This requires real-time processing of intractable NP-hard optimization problems. To tackle this problem, we propose an online deployment framework for MEC applications based on deep reinforcement learning, whose policy-based features adapt to the characteristics of the large action space in the problem. The framework learns binary deployment decisions from experience without solving NP-hard optimization problems, which greatly reduces computational complexity. Through simulations, we demonstrate the ability of our scheme to balance the computational load among the available MEPs and to satisfy the dynamic service requests of users.