5G Network Slice Admission Control Using Optimization and Reinforcement Learning
Md Ariful Haque, Vassilka Kirova
Abstract
Network slicing can enable operators to maximize revenue by optimizing the utilization of network resources. As a result, slice admission control and resource allocation decisions become critical tasks for 5G network orchestration. In this work, we examine two complimentary approaches that address the challenge of slice admission control: (1) dynamically assessing and accepting a set of slices that generate greater revenue using a queue-based optimization algorithm, and (2) automating the admission decision using reinforcement learning (RL). The simulation results from the first approach show that the proposed optimization algorithm always delivers higher revenue than accepting the slices in any other order. In the second approach, we utilize several reinforcement learning algorithms to train an agent (orchestrator) to dynamically make decisions based on the environment. Our results show that the Deep Q Network (DQN) algorithm provides higher cumulative rewards (i.e., cumulative revenue) than the other tested algorithms (SARSA, Expected SARSA, and Q-Learning). The two approaches are complementary, and can be deployed either separately or simultaneously to maximize network operator’s revenue.