R1-2409583 discussion

Views on AI/ML based CSI prediction

From Samsung
Status: not treated
WI: NR_AIML_air
Agenda: 9.1.3
Release: Rel-19
Source: 3gpp.org ↗

Summary

Samsung presents observations and proposals for AI/ML-based CSI prediction in NR, highlighting the performance degradation when models are trained on mismatched TRP antenna settings and the need for network assistance to ensure training/inference consistency. The document contains 2 key observations regarding site-specific model benefits and TXRU mapping impacts, and 5 proposals covering network-side condition indication, data collection frameworks, and specific configurations for Type 1, 2, and 3 model monitoring.

Position

Samsung argues that site-specific AI/ML models significantly outperform generic models, particularly at large prediction horizons, and demonstrates that data distribution mismatch due to different TRP antenna settings (e.g., [2,8,2] vs [4,4,2] TXRU mapping) causes up to 77% performance loss. They propose considering TRP-related aspects for network-side additional condition indication to ensure consistency between training and inference, noting that TCI frameworks are insufficient due to flexible mapping changes over time. Samsung proposes studying specific data collection procedures, including distinguishing resources for CSI reporting versus internal data collection, and addressing the unfair advantage periodic measurements give to linear baselines like Kalman filters. For model monitoring, they propose detailed configurations for Type 1 (UE-side metric calculation with baseline/thresholds), Type 3 (CMR/MMR separation and window configuration), and Type 2 (network-side monitoring using enhanced Type II CSI for ground truth across multiple time instances) to handle complexity and performance verification.

Key proposals

Your notes

Private to your account