R1-2500160 discussion

Discussion on AIML for CSI prediction

From Spreadtrum
Status: not treated
WI: NR_AIML_air
Agenda: 9.1.3
Release: Rel-19
Source: 3gpp.org ↗
Spreadtrum's prior position on 9.1.3 at RAN1#118bis · AI-synthesized, paraphrased
verify sources →
Advocates for reusing existing AI beam management conclusions and associated ID mechanisms for CSI prediction to minimize workload. Opposes performance monitoring-based approaches, arguing they would cause significant performance loss due to trial-and-error processes.

Summary

Spreadtrum presents eight proposals for 3GPP RAN1 regarding AI/ML-based CSI prediction in NR, focusing on ensuring consistency between training and inference, defining data collection procedures, and establishing performance monitoring mechanisms. The document argues for reusing existing Rel-18 CSI-RS configurations and prioritizing UE-side data collection to minimize specification impact and signaling overhead.

Position

Spreadtrum proposes using an associated ID, configured within the CSI framework, to ensure consistency between training and inference for UE-sided CSI prediction models, arguing that performance monitoring-based approaches cause unacceptable performance loss due to trial-and-error processes. They prefer UE-side data collection over network-side collection to avoid significant reporting overhead and model transfer complexities, suggesting that NW configuration or UE requests should trigger this collection. For inference, they propose reusing Rel-18 MIMO CSI-RS configurations to reduce specification impact. Regarding monitoring, they support Type 1 and Type 3 monitoring using intermediate KPIs like SGCS, while explicitly deprioritizing Type 2 monitoring due to the high overhead of reporting ground-truth CSI to the gNB. They further suggest that gNB should indicate the association between prediction and ground-truth CSI-RS resources to facilitate accurate metric calculation.

Key proposals

Your notes

Private to your account