Deep reinforcement learning for resource allocation and scalable numerology in NR-U enabled multi-RAT HetNets

TL;DR

This paper proposes a deep reinforcement learning framework to optimize resource allocation and numerology in NR-U enabled multi-RAT HetNets, improving user satisfaction.

Contribution

A novel DRL-based framework is introduced for joint user association, resource allocation, and numerology optimization in multi-RAT HetNets.

Findings

The proposed model achieves up to 70% user satisfaction compared to baseline approaches.

The framework efficiently allocates resources and determines optimal numerology values across network slices.

Abstract

Leveraging the new radio technology in the unlicensed band (NR-U) can alleviate traffic congestion, enhance network capacity, and help mitigate the diversity in users’ service requests. In this paper, a multiple slice multi-radio access technology (RAT) heterogeneous network (HetNet) is considered, integrating the new radio (NR) technology in the licensed and unlicensed bands. An optimization problem is proposed aiming to maximize users’ satisfaction, defined by maximizing the achievable data rate while maintaining the minimum latency slice requirement. To solve the proposed optimization problem, an iterative framework is introduced that utilizes deep reinforcement learning (DRL) algorithm jointly with the regret learning algorithm (RLA) that efficiently solves users’ association problem considering coexisting Wi-Fi users, allocates radio resources for each slice and determines the optimum scalable numerology value in each slice. The simulation results show that our proposed model improves users’ satisfaction, achieving up to 70% user satisfaction compared with other baseline approaches.

Figures