R1-2410537
discussion
AI/ML - Specification support for CSI Prediction
From MediaTek
Summary
MediaTek's document evaluates AI/ML model generalization for CSI prediction across different network configurations, showing that models can generalize well across tilt angles and TXRU mappings with less than 2% performance loss. The document contains 8 proposals and 4 observations focused on ensuring consistency between training and inference phases.
Position
MediaTek advocates FOR demonstrating that AI/ML models for CSI prediction can generalize well across different network configurations (tilt angles and TXRU mappings), supporting the use of associated ID as a baseline approach for training-inference consistency while proposing performance monitoring as an alternative. They are positioned AGAINST overly complex solutions and push FOR leveraging existing beam management frameworks to reduce specification overhead.
Key proposals
- Proposal 1 (Sec 3): Study which kinds of NW-side additional conditions will impact the generalization of AI/ML-based CSI prediction models
- Proposal 2 (Sec 3): For consistency between training and inference for CSI prediction using UE-sided model, slightly propose the use of the associated ID as the baseline
- Proposal 3 (Sec 3): Leverage and reuse the existing discussion and agreements on the beam management agenda for the associated ID framework and procedures as much as possible to reduce the discussion required for CSI prediction
- Proposal 4 (Sec 3): Consider performance monitoring-based methods as an alternative approach, allowing the UE to dynamically select and switch between multiple models based on real-time performance monitoring
- Proposal 5 (Conclusion): Study which kinds of NW-side additional conditions will impact the generalization of AI/ML-based CSI prediction models
- Proposal 6 (Conclusion): For consistency between training and inference for CSI prediction using UE-sided model, slightly propose the use of the associated ID as the baseline
- Proposal 7 (Conclusion): Leverage and reuse the existing discussion and agreements on the beam management agenda for the associated ID framework and procedures
- Proposal 8 (Conclusion): Consider performance monitoring-based methods as an alternative approach for UE to dynamically select between models