A Dynamic GPF scheduler proposal backed by Differential Evolution and Neural Network on 5G HetNet simulated scenarios

Abstract

To achieve the desired 5G goals, a network must be efficient and fair simultaneously, allowing high data rates for regular users while also equally serving low-power users. Static schedulers, as the name implies, don’t change their scheduling policy regardless of the number of users, giving rise to the need for dynamic schedulers that can shift this paradigm by sensing the number of network users and other parameters like Signal-to-Interferenceplus- Noise Ratio (SINR) and changing the scheduling policy accordingly, raising network efficiency and fairness. Thus, this work presents an optimization technique that uses the Differential Evolution (DE) algorithm and, later, trains a feedforward Neural Network (NN) to mimic the DE’s decision-making to act as a dynamic Generalized Proportional Fair (GPF), adapting its two internal parameters, α and β, assuring a threshold Jain’s fairness index value while seeking a throughput maximization in a simulated 5G Heterogeneous Network (HetNet). Results show that the method achieves a minimal 0.7 Jain’s fairness index for most numbers of users, from 40 to 200 users, while having a better average throughput when compared to a related scheduler. It was also noted that the implemented DE algorithm took an impractical runtime to work as an online optimizer; in contrast, using the created dynamic NN scheduler added only 0.03% to 0.57% in simulation runtime compared to GPF.