5G-Advanced AI/ML Beam Management: Performance Evaluation with Integrated ML Models

TL;DR

This paper evaluates AI/ML-based beam management techniques in 5G, focusing on prediction accuracy, model generalization, and system-level throughput improvements, providing design guidelines for practical deployment.

Contribution

It offers a comprehensive analysis of AI/ML models for beam prediction in 5G, including accuracy, generalization, and system-level performance, with deployment guidelines.

Findings

AI/ML models improve beam prediction accuracy.

Integrated AI/ML models enhance user throughput.

Guidelines for deploying lightweight AI/ML beam management models.

Abstract

The legacy beam management (BM) procedure in 5G introduces higher measurement and reporting overheads for larger beam codebooks resulting in higher power consumption of user equipment (UEs). Hence, the 3rd generation partnership project (3GPP) studied the use of artificial intelligence (AI) and machine learning (ML) in the air interface to reduce the overhead associated with the legacy BM procedure. The usage of AI/ML in BM is mainly discussed with regard to spatial-domain beam prediction (SBP) and time-domain beam prediction (TBP). In this study, we discuss different sub-use cases of SBP and TBP and evaluate the beam prediction accuracy of AI/ML models designed for each sub-use case along with AI/ML model generalization aspects. Moreover, a comprehensive system-level performance evaluation is presented in terms of user throughput with integrated AI/ML models to a 3GPP-compliant system-level simulator. Based on user throughput evaluations, we present AI/ML BM design guidelines for the deployment of lightweight, low-complexity AI/ML models discussed in this study.