Adaptive Auto-Scaling in Mobile Edge Computing: A Deep Reinforcement Learning Approach

In recent years, Mobile Edge Computing (MEC) in 5G network has emerged as a promising paradigm with the development of the 5G. By setting up serving nodes at the edge of the network, it can be enabled for mobile devices to offload their computing tasks to the edge nodes, to achieve a lower response time and save more device power. However, due to the limited capacity of the edge nodes, it is important to implement a reasonable resource allocation method. Auto-scaling is a technology which can automatically adjusts computing resources between the containers at edge nodes, according to the demands and the strategies, to achieve a balance between the system costs and Quality of Service (QoS). To find this balance, many researchers have proposed different auto-scaling algorithms, e.g. rule-based algorithms, reinforcement learning-based algorithms, etc. For the algorithms based on reinforcement learning, most of the researchers have applied Q-Learning in auto-scaling and achieved good results, proving the feasibility of the reinforcement learning. However, the Q-Learning algorithm can only select actions in a limited action space, thus it cannot achieve a more precise control. Therefore, in this paper, based on the Markov Decision Process (MDP), we propose an auto-scaling algorithm based on the Proximal Policy Optimization (PPO), replacing the discrete action space with a continuous action space, to achieve a more precise control. Experiment shows that compared to the no-scaling method and the Q-Learning algorithm, our proposed algorithm has achieved an overall improvement of 96.9% and 86.6%, respectively.