R1-2407690
discussion
Specification Support for AI/ML for Beam Management
From Kyocera
Summary
This Kyocera document proposes comprehensive specification support for AI/ML-enabled beam management in 5G NR Rel-19, covering both network-sided and UE-sided models for spatial and temporal beam prediction. The document contains 24 proposals and 2 observations addressing configuration, reporting, consistency, and performance monitoring aspects.
Position
Kyocera advocates for a comprehensive AI/ML beam management framework that leverages existing CSI infrastructure while introducing minimal new specifications. They strongly support virtual Set A configuration for UE-side models, flexible Set B definition through new IEs, and practical approaches like 4-bit probability quantization. They push against overly complex reporting mechanisms and favor implementation-based solutions for performance monitoring, particularly opposing probability-only performance metrics (Alt 4) that lack ground truth validation.
Key proposals
- Proposal 1 (Sec 2.1): For UE-side AI/ML models, Set A is virtually configured as reference for mapping model output to CRIs, while Set B is explicitly configured for UE measurements
- Proposal 2 (Sec 2.2): Association between Sets A and B requires new IE within CSI-ResourceConfig or AI/ML specific configuration to inform UE about beam membership
- Proposal 4 (Sec 3): Support CSI framework for training data collection with new IEs like 'CSI-AIML-ResourceSet' and 'AIML-info-RSRP' for flexible Set B definition
- Proposal 5 (Sec 4.1): For NW-sided inference reporting, support L1-RSRPs and beam info of top M beams, with maximum M supporting 256 beams (Set A size)
- Proposal 6 (Sec 4.2): For UE-side inference results, support options 1-3 and further study benefits of option 3 considering quantization effects and overhead
- Proposal 7 (Sec 4.2): Define confidence information as range within which predicted RSRP falls certain percentage of time for same input sample
- Proposal 11 (Sec 4.2): Support one decimal precision probabilities using 4 bits per beam for probability information reporting
- Proposal 13 (Sec 5.1): For UE-side model consistency, same associated ID should maintain consistent beam properties, set sizes, and resource indexing between training and inference
- Proposal 14 (Sec 5.1): Introduce associated ID as new IE within RRC messaging of nzp-CSI-RS-ResourceSet and csi-SSB-ResourceSet
- Proposal 17 (Sec 6.1): For Type 1 option 1 performance monitoring, support reporting model outputs (beam info, predicted RSRP, probability) and performance measurements (CRI/SSBRI, L1-RSRP)
- Proposal 19 (Sec 6.1): Deprioritize Alt 4 (probability-based) for Type 1 option 2 performance monitoring in BM Cases 1 and 2
- Proposal 20 (Sec 6.1): Performance monitoring set size should be ≤ Set A size and network-defined
- Proposal 22 (Sec 6.1): Event-based reporting not applicable to option 1 but relevant to option 2 where UE calculates performance metrics
- Proposal 23 (Sec 6.1): Type 2 performance monitoring has no specification impact and should be left to UE vendor implementation
- Proposal 24 (Sec 6.2): NW-side AI/ML performance monitoring requires no specification impact and can be vendor implementation
Revision chain
1 versions in this meeting · oldest first
-
R1-2407690 ← you are here discussion revised