R1-2500533
discussion
On AI/ML-based CSI prediction
From InterDigital
Summary
InterDigital presents evaluation results on AI/ML-based CSI prediction, demonstrating that UE-sided models generalize well across antenna down-tilt variations and that localized models offer only minor gains over generalized ones. The document contains 4 proposals and 5 observations, arguing against the necessity of Associated IDs for training/inference consistency and recommending Type 3 performance monitoring combined with out-of-distribution metrics.
Position
InterDigital argues that the Associated ID mechanism is unnecessary for UE-sided CSI prediction, citing high complexity and negligible generalization degradation across NW-side conditions like antenna down-tilt. They present technical evidence that Case 2 generalization shows negligible SGCS change relative to Case 1, and that localized models provide only minor SGCS gains over generalized models. Consequently, they propose dropping the Associated ID requirement (Proposal 1) and instead rely on model performance monitoring. They explicitly oppose Type 2 monitoring due to overhead and support Type 3 monitoring (Proposal 3), further proposing that out-of-distribution metrics be used alongside intermediate KPIs (Proposal 4) to reduce unnecessary model switching overhead.
Key proposals
- Proposal 1 (Sec 3.1): For CSI prediction, the associated ID is not needed to ensure training/inference consistency regarding NW-side additional conditions.
- Proposal 2 (Sec 3.2): Type 2 model performance monitoring is not further considered due to excessive feedback overhead.
- Proposal 3 (Sec 3.2): Type 3 is supported for UE-side model performance monitoring, allowing the UE to report performance metrics like NMSE or SGCS to the NW.
- Proposal 4 (Sec 3.2): For UE-side model monitoring, out-of-distribution metrics should be used in conjunction with intermediate KPIs to improve monitoring accuracy.