QoS-Aware Scheduling for Cellular Networks Using Deep Reinforcement Learning

Jonathan Robert Malin, Gun Ko, Won Woo Ro

Abstract: This research presents a reinforcement learning (RL) approach that provides coarse-grained decisions to maximize QoS requirement satisfaction in mobile networks. Deep reinforcement learning has demonstrated agents that can capture the dynamics of impossibly complex systems. At each scheduling interval, our RL agent provides a scheduling policy that is suitable for optimal resource allocation in a mobile network. By using a deep neural network to approximate the action-value function (Q-function), scheduling decisions can be made using an optimal policy. Utilising a 4G-LTE network simulator and Pytorch, this research explores three scenarios of diverse traffic and UE density. The implementation shows stable and effective performance when compared to baseline static schedulers. Additionally, the RL agent selects the optimal scheduler for both single and mixed traffic simulations. Being both scalable and cheap to compute, the implementation offers a simple and effective method of radio resource management.