R1-2500637
discussion
Specification support for CSI prediction
From Lenovo
Summary
Lenovo proposes reusing the AI/ML framework for Beam Management (BM) for CSI prediction, emphasizing an urgent decision on network-side additional conditions and defining performance monitoring types. The document contains 18 proposals and 12 observations covering scope, feedback formats, and training/inference consistency.
Position
Lenovo proposes reusing the AI/ML framework for Beam Management (BM) for CSI prediction, including RAN2 procedures and functionality definitions. They require an urgent decision on the necessity of network-side additional conditions to facilitate specification support. Lenovo argues against model training at the network side with transfer to the UE, citing significant challenges in model transfer overhead. They prefer deprioritizing model identification-based techniques for training/inference consistency, instead focusing on additional condition associated ID techniques and monitoring-based techniques (Type 1 and Type 3). They propose using the legacy Type-II CSI feedback format as a baseline while studying enhancements for time-domain representation and evaluating CPU calculation impacts.
Key proposals
- Proposal 1 (Sec 2.1): Reuse the AI/ML framework for BM whenever applicable as a first step, including RAN2 procedures and functionality definitions.
- Proposal 2 (Sec 2.1): An urgent decision on whether network-side additional conditions are required for AI/ML for CSI prediction is needed.
- Proposal 3 (Sec 2.2): Evaluate potential configurations of the observation window and the prediction window for UE-based CSI prediction.
- Proposal 4 (Sec 2.3): The legacy Type-II CSI feedback format is used as the baseline for design of CSI feedback for AI-based CSI prediction.
- Proposal 5 (Sec 2.3): Study potential enhancements of the CSI feedback format for predicted CSI, at least with respect to time-domain representation.
- Proposal 6 (Sec 2.3): Evaluate the CPU calculation for AI/ML-based CSI report(s) for UE-based CSI prediction.
- Proposal 7 (Sec 2.4): Evaluate the specification impact corresponding to AI/ML model monitoring, considering decisions like no model change, parameter update, switching, or fallback.
- Proposal 8 (Sec 2.4): Study and evaluate the pros and cons of Implicit vs. Explicit performance monitoring alternatives.
- Proposal 9 (Sec 2.5): Adopt specific definitions for ‘performance metric’ (implicit for Type 1, explicit for Type 3) and ‘monitoring output’.
- Proposal 10 (Sec 3): Verify whether UE-sided AI/ML-based CSI prediction performance degradation resembles that of two-sided AI/ML-based CSI compression regarding TXRU virtualization.
- Proposal 11 (Sec 4.1): Maintaining consistency via model training at the NW side followed by model transfer to UE side is not considered for the study.
- Proposal 12 (Sec 4.2): Further study maintaining consistency via indication of NW-side additional conditions, focusing on scenarios with few parameters or localized models.
- Proposal 13 (Sec 4.3): Deprioritize model identification-based techniques for training/inference consistency for UE-sided CSI prediction.
- Proposal 14 (Sec 4.4): Further study Type 1 and Type 3 performance monitoring techniques for CSI prediction for the purpose of training/inference consistency.
- Proposal 15 (Sec 4.4): For training/inference consistency issues, if specified, scope is limited to additional condition associated ID techniques and monitoring-based techniques.