Adaptive Edge Resource Allocation for Maximizing the Number of Tasks Completed on Time: A Deep Q-Learning Approach

Abstract

To relieve resource-limited mobile devices from computation-intensive tasks, reduce the transmission latency and mitigate the burden of the backhaul network for the centralized cloud-based network services, mobile edge computing (MEC) has been proposed to be a promising solution and draws increasing attention from both industry and academia. Traditional task offloading approaches focus on average-based metrics, and try to minimize the average service delay. The service delay of different tasks varies from each other, resulting in a low service reliability. To attack this challenge, this paper focuses on mobile users’ computation offloading problem in wireless cellular networks for purpose of maximizing the number of tasks completed on time. Since the environment states, including available local resources, channel conditions and remaining computation resource of the edge cloud, will vary from time to time, we use the model-free reinforcement learning (RL) framework to formulate and tackle the computation offloading problem. The agent learns through interactions with the environment to decide executing task locally on the mobile device or offloading the task to the edge cloud via the wireless link for each mobile device user. Considering high-dimensional state spaces, we use deep Q-learning (DQN) which combines reinforcement learning method Q-learning and deep neural network (DNN) to obtain the optimal approach. Simulation results show that the effectiveness of the proposed approach in comparison with baseline approaches in terms of the total number of the tasks completed on time.