Deep reinforcement learning for resource allocation in wireless networks

Abstract

The advent of Internet of Things (IoT) generalizes the accessibility towards various kinds of IoT sensors, e.g., smart cameras, temperature sensors, and hence enables intelligent services to improve human quality of life. Building upon IoT, Industrial Internet of Things (IIoT) instead focuses on improving productivity in industrial processes by interconnecting industrial devices such as sensors to infrastructures such as Mobile Edge Computing (MEC) servers. This allows for the interconnection between the different devices in the IIoT network. This paper explores a novel approach that dynamically switches between policies such as Zero Wait (ZW) and Continuous Update (CU) using Deep Q-Network (DQN), Duelling Double Q-Network (D3QN) and Proximal Policy Optimization (PPO) based Deep Reinforcement Learning (DRL) method. This approach takes in real-time metrics such as server utilization rate, payload size, and acknowledgement packet size to reduce the Peak Age of Information (PAoI) violation probability. This proposed method is then evaluated against benchmark policies to observe the improvement.