R1-2500600
discussion
Discussion on specification support for CSI prediction
From NEC
Summary
NEC presents 17 proposals and 1 observation regarding the normative specification of CSI prediction with UE-sided models, focusing on consistency between training and inference, data collection mechanisms, model inference procedures, and performance monitoring types. The document argues for reusing associated IDs for consistency, defining specific performance metrics for UE-side monitoring, and establishing reporting procedures for model switching and data collection requests.
Position
NEC proposes reusing the associated ID from BM-Case 1/2 to ensure consistency of NW-side additional conditions for CSI prediction, while simultaneously arguing that performance monitoring serves as a fundamental LCM procedure to guarantee inference performance regardless of consistency mechanisms. They require support for UE-initiated data collection requests and the provision of preferred observation and prediction window periods to the NW. NEC proposes configuring distinct observation and prediction windows for P/SP CSI reports to reduce signaling overhead and suggests autonomous updates to CSI reporting periodicity based on significant points of variation. Regarding performance monitoring, NEC focuses specification efforts on Type 1 and Type 3 monitoring, proposing intermediate KPIs like SGCS and NMSE, and defines specific fallback criteria based on threshold violations for current and candidate models. They further require the UE to report changes in model input/output parameters, such as window lengths, to the NW during model switching or updates.
Key proposals
- Proposal 1 (Sec 2.1): Reuse the associated ID introduced in BM-Case 1/2 to ensure consistency of NW-side additional conditions across training and inference for CSI prediction with UE-sided model, if such consistency is deemed necessary.
- Proposal 2 (Sec 2.1): Use performance monitoring to ensure consistency of NW-side additional conditions across training and inference, irrespective of whether explicit consistency mechanisms like associated IDs are used.
- Proposal 3 (Sec 2.2.1): Support the UE to report a request for data collection to the Network (NW) to initiate CSI-RS transmission for model training or update.
- Proposal 4 (Sec 2.2.1): Support the UE to provide preferred or expected CSI measurement periods in the observation window and CSI prediction periods in the prediction window to the NW.
- Proposal 5 (Sec 2.2.1): Support enhanced CSI-RS configurations for data collection to adapt to changing channel patterns and reduce overhead.
- Proposal 6 (Sec 2.2.2): For P/SP CSI reports based on CSI prediction, support the configuration of observation windows and prediction windows to align NW and UE behaviors without excessive signaling overhead.
- Proposal 7 (Sec 2.2.2): Allow CSI reporting periodicity to be updated autonomously upon reaching a significant point of variation determined by time, location, or distance.
- Proposal 8 (Sec 2.2.2): Reuse two-part CSI reporting to report predicted CSI(s) for one or multiple future time instances.
- Proposal 9 (Sec 2.2.3): Focus the definition of performance metrics on Type 1 and Type 3 performance monitoring, where calculation is performed at the UE side.
- Proposal 10 (Sec 2.2.3): Define performance metrics for Type 1 and 3 monitoring as intermediate KPIs (e.g., SGCS, NMSE) or input/output data distribution-related metrics (e.g., similarity, divergence, distance).
- Proposal 11 (Sec 2.2.3): Define Type 1 performance monitoring output to facilitate NW fallback decisions by indicating states where current or candidate AI/ML models do not meet NW-provided thresholds.
- Proposal 14 (Sec 2.2.3): Identify spec impacts for performance monitoring, including threshold definition for Type 1, quantization for Type 2, and association between predicted and ground-truth CSI reporting for Type 3.
- Proposal 16 (Sec 2.2.3): Study how to refine performance monitoring procedures when the target timing of predicted CSI is not aligned with the timing of available ground-truth CSI.
- Proposal 17 (Sec 2.2.4): Require the UE to report changes in AI/ML model input/output parameters (e.g., observation/prediction window lengths) to the NW during model switching, selection, or update.