Deep Reinforcement Learning for Sleep Control in 5G and Beyond Radio Access Networks: An Overview

The advent of 5G and beyond networks is envisioned to support lower latency, higher data rates, and wider connectivity than previous cellular network generations. However, given the denser deployment of base stations (BSs) to accommodate such improvements, this results inevitably in a significant and unsustainable increase in the network’s energy consumption. Sleep Control (SC), which allows switching off some BS hardware components during light-traffic time, is considered a viable solution for greener and more energy-efficient Radio Access Networks (RAN). However, the optimization of SC is a highly challenging large-scale network combinatorial problem that depends on dynamic wireless channel conditions and varying traffic demands with stringent Quality-of-Service (QoS) requirements. Driven by the benefits and efficiency of Deep Reinforcement Learning (DRL), which has been successfully applied to multiple wireless network optimization problems, this paper investigates DRL approaches addressing sleep control in 5G and beyond RAN. To this end, we propose a taxonomy to classify the related literature. Then, we provide an overview of the different components of the Markov Decision Process (MDP) modeling the sequential decision-making of sleep control and the applied DRL algorithms. Finally, we highlight the main challenges in existing works and suggest novel strategies to address them.