R1-2500465
discussion
On specification for AI/ML-based beam management
From OPPO
Summary
OPPO presents a comprehensive set of proposals for AI/ML-based beam management in NR Rel-19, covering both NW-side and UE-side models for BM-Case 1 and BM-Case 2. The document contains approximately 35 distinct proposals and observations, focusing on specification impacts for inference, data collection, performance monitoring, and consistency mechanisms.
Position
OPPO proposes clarifying beam information as CRI/SSBRI indices for NW-side models and supports implicit temporal reporting for BM-Case2 to reduce overhead. They argue that UE-side additional conditions on Rx beamforming are unnecessary for NW-side models, as consistency can be achieved via mixed datasets during training. For UE-side models, OPPO supports reporting probability information (Opt 3) and adopting Option 3b (Top-K/M) for beam prediction accuracy metrics, while also supporting Type 2 performance monitoring where the UE makes LCM decisions. They propose configuring {associated ID, Set B, Set A} triples within the CSI framework to ensure consistency across training and inference. Additionally, OPPO argues for separate CPU counting for AI/ML-based CSI reporting versus legacy CSI reporting, citing hardware separation between sequential and parallel computing units.
Key proposals
- Proposal (Sec NW-side Inference): For NW-side model, clarify beam information as beam reporting index (CRI/SSBRI) in CSI reporting.
- Proposal (Sec NW-side Inference): For BM-Case2 with NW-side model, support UE reporting multiple measurement instances of Set B in a single beam reporting instance.
- Proposal (Sec NW-side Inference): For BM-Case2 with NW-side model, support indicating multiple TCI states for multiple future instances via one-shot beam indication.
- Proposal (Sec NW-side Data Collection): For NW-side model training, UE reports L1-RSRPs of fixed Set B as model inputs and Top-K L1-RSRP(s) and Top-1 Tx beam index as labels.
- Proposal (Sec NW-side Data Collection): For BM-Case2 with NW-side model, temporal domain information for training data should be reported implicitly without explicit timestamps.
- Proposal (Sec NW-side Consistency): For BM-Case1 and BM-Case2 with NW-side model, it is NOT necessary to specify UE-side additional condition on UE Rx beamforming.
- Proposal (Sec UE-side Inference): For UE-side model inference, support reporting (Opt 3) beam information on predicted Top K beam(s) and probability information of predicted Top K beam(s).
- Proposal (Sec UE-side Inference): For BM-Case2 with UE-side model, enhance differential RSRP reporting where UE packs predicted beams/RSRPs from 1st to N-th time instances in a single report.
- Proposal (Sec UE-side Inference): For BM-Case2 with UE-side model, support indicating multiple TCI states for up to N future time instances with one-shot beam indication.
- Proposal (Sec UE-side Configuration): For BM-Case2 with UE-side model, support adopting Option 2 (CSI reference resource) as the reference time of the earliest time instance for predicted results.
- Proposal (Sec UE-side Monitoring): For beam prediction accuracy metric, support Option 3b (Top-K/M): at least one of the Top-K predicted beams is among Top M beam(s) with largest M measured L1-RSRP(s).
- Proposal (Sec UE-side Monitoring): For Type 1 Option 2 performance monitoring, support the probability of model output (Alt.4) as one performance metric.
- Proposal (Sec UE-side Monitoring): For performance monitoring with UE-side model, support Type 2 (indication/request/report from UE to gNB) performance monitoring.
- Proposal (Sec UE-side Consistency): Support configuring multiple sets of {associated ID, Set B, Set A} within the CSI framework to ensure flexibility on NW-side additional conditions.
- Proposal (Sec Other Aspects): For BM-Case1 and BM-Case2 with UE-side model, the overall CPU should be separately counted between legacy CSI reporting and AI/ML-based CSI reporting.