Comparison of Machine Learning Algorithms for Priority-Based Network Slicing in 5G Systems

Abstract

Network slicing is a technique to enable multi-tenant operation in future 5G systems. Efficient implementation of slicing at the air interface requires comprehensive optimization algorithms characterized by high execution complexity. To address this issue in the paper, we first present a priority-based mechanism enabling performance isolation between slices competing for resources. Then, to speed up the resource arbitration process under high traffic conditions, when resource shares need to be re-calculated in sub-second timescales, we propose and compare several machine learning techniques: linear regression, polynomial regression, a random forest regressor, and a two-layer artificial neural network. The techniques' performance is assessed by utilizing the mean squared error. Our results show that a high order polynomial regression provides the desired balance between computational complexity and accuracy, outperforming both the simpler linear regression and the more complex random forest and neural network algorithms.