Deep reinforcement learning in edge networks: Challenges and future directions

Abstract

Driven by the perception of IoT applications and advanced communication technologies, including beyond 5G and 6G, recent years have seen a paradigm shift from traditional cloud computing towards the local edge of the networks. Modern edge-centric networks have become autonomous and decentralized to expand IoT applications and corresponding data fusion. When edge networks are uncertain, network entities execute tasks locally to increase network performance. Over the past decade, Reinforcement Learning (RL) algorithms have been integrated into edge networks to generate optimal decisions and intelligent edge networks. However, complex edge networks with ample state and action space create several challenges in making optimal decisions with the RL technique. To address such shortcomings, Deep Reinforcement Learning (DRL) is combined with edge networks to build an intelligent edge framework. Concerning the benefits of edge intelligence, this paper summarizes the importance of traditional and advanced DRL methodologies in edge networks. Besides, we discuss different types of DRL-enabled libraries and state-of-the-art edge models for processing real-time IoT applications. Then, we review other emerging issues in edge networks regarding data offloading, caching, dynamic network access, edge information fusion, and data privacy. Moreover, we incorporate various DRL-enabled IoT applications in edge networks such as healthcare applications, industrial applications, traffic management, etc. Finally, we shed light on future trends of intelligent edge computing regarding system performance, security, and network management.