R1-2500406
discussion
Specification support for CSI Prediction
Summary
Tejas Networks discusses AI/ML-based CSI prediction for Rel-19, focusing on consistency between training and inference, data collection mechanisms, and performance monitoring. The document presents 16 proposals and 3 observations addressing issues such as interference variations, TxRU mapping impacts, and the need for specific CSI-RS configurations.
Position
Tejas Networks proposes using an associated ID to align training and inference conditions, arguing that variations in interference and network-side conditions like TxRU mappings degrade AI model performance. They support AI/ML model identification via Model ID in LCM mode, requiring the Network to identify models and the UE to indicate support. For data collection, they propose reusing Rel-18 Type II Doppler codebook CSI-RS configurations and separating resources for inference from those for training/monitoring. They prioritize Type 1 performance monitoring, where the UE calculates metrics and reports output only if below a Network-assigned threshold, utilizing intermediate KPIs like SGCS or NMSE.
Key proposals
- Proposal 1 (Consistency): Use an associated ID-based solution to ensure consistency between training and inference phases, extending methods from AI-based beam management.
- Proposal 2 (Consistency): Acknowledge that tilt angle and/or TXRU mapping variations will cause performance degradation in CSI prediction.
- Proposal 3 (Model Inference): Consider the existing CSI framework for resources and reports of AI/ML-based CSI prediction.
- Proposal 4 (LCM mode): Support AI/ML model identification where models are identified by Model ID by the Network and UE indicates supported models.
- Proposal 5 (Data collection): Consider options for initiating CSI-RS transmission for prediction, including NW-triggered or UE-initiated transmission.
- Proposal 6 (Data collection): Consider that CSI-RS transmission for prediction can be periodic or aperiodic.
- Proposal 7 (Data collection): Consider that the UE should report the minimum number of CSI-RS instances required for CSI prediction to the Network.
- Proposal 8 (Data collection): Reuse the CSI-RS configuration from Rel-18 Type II Doppler codebook for model inference.
- Proposal 9 (Data collection): Consider defining a maximum time gap allowed between the prediction and observation window to ensure better prediction accuracy.
- Proposal 10 (Data collection): Consider the existing CSI framework for data collection, separating CSI resources for inference from those for training/monitoring.
- Proposal 11 (Performance monitoring): Consider specification impacts for performance monitoring, including CSI-RS configuration, threshold criteria, and fallback operations.
- Proposal 12 (Performance monitoring): Consider that the Network assigns the threshold criterion for the UE.
- Proposal 13 (Performance monitoring): Consider that the UE only reports performance monitoring output when it falls below the configured threshold.
- Proposal 15 (Performance monitoring): Consider using intermediate KPIs (e.g., SGCS, NMSE) or statistical values within the window for Type 1 and Type 3 performance monitoring.
- Proposal 16 (Performance monitoring): Prioritize Type 1 performance monitoring.