R1-2500391 report

AI/ML for beam management

From Ericsson
Status: not treated
WI: NR_AIML_air
Agenda: 9.1.1
Release: Rel-19
Source: 3gpp.org ↗
Ericsson's prior position on 9.1.1 at RAN1#119 · AI-synthesized, paraphrased
verify sources →
Advocates for a comprehensive AI/ML framework that leverages existing CSI mechanisms with minimal specification impact while maximizing functionality. Strongly supports expanding aperiodic CSI-RS resources from 16 to 64 beams to enable practical AI/ML deployment. Proposes uncertainty quantification in UE predictions to enable trustworthy AI/ML and adaptive Top-K beam selection based on model confidence. Supports extensive overhead reduction mechanisms for NW-sided models. Opposes complex new signaling frameworks, favoring reuse of existing CSI infrastructure with targeted enhancements.

Summary

Ericsson presents 21 proposals and 6 observations for AI/ML beam management in NR, focusing on UE-sided model configuration, inference reporting, performance monitoring, and NW-sided data collection overhead reduction. The document argues for configuring associated IDs at the ResourceSet level rather than CSI-ReportConfig to support larger beam sets and flexible aperiodic triggering, while also proposing mechanisms for uncertainty reporting, adaptive Top-K selection, and joint activation of monitoring and inference configurations.

Position

Ericsson proposes configuring the associated ID at the CSI-ResourceSet level rather than the CSI-ReportConfig level to preserve fundamental CSI framework assumptions and enable predictions beyond current resourceSet size limits. They require support for separate CSI-ResourceConfigIds for Set A and Set B, and propose extending aperiodic resource sets to 64 NZP CSI-RS resources. For inference reporting, Ericsson supports adaptive Top-K values and the inclusion of probability and RSRP confidence intervals to handle model uncertainty and data drift. They argue for joint activation of monitoring and inference configurations to minimize signaling overhead and propose specific performance metrics based on Top-1/Top-K alignment. For NW-sided models, they propose mechanisms to reduce reporting overhead, including pre-processing Set B beams and omitting duplicated or unstable training data samples.

Key proposals

Your notes

Private to your account