R1-2409581
discussion
Discussion for supporting AI/ML based beam management
From Samsung
Summary
Samsung presents 30 proposals and 1 observation regarding AI/ML-based beam management for NR, covering both NW-side and UE-side models. The document addresses data collection for training and inference, spatial and temporal enhancements for beam reporting, consistency mechanisms via DL Tx IDs and associated IDs, performance monitoring metrics like BAI, and CPU/timeline considerations for UE-side inference.
Position
Samsung proposes specific data collection contents for NW-side training, including L1-RSRPs for Set A and Set B and timestamps, conveyed via high-layer signaling. For UE-side inference, they support configurability between Alt 1 and Alt 3 for CSI-ReportConfig and introduce DL Tx IDs to ensure consistent spatial domain transmission filters between Set A and Set B. They propose introducing a Beam Accuracy Indicator (BAI) for Type 1 Option 2 performance monitoring, calculated over X CSI reports, and prefer dedicated CSI report configurations for monitoring resources. Regarding CPU handling, Samsung proposes separate CPU counting for AI/ML-based CSI reports and scaling the legacy Z timeline based on UE capability.
Key proposals
- Proposal 1 (Sec 2.1.1): For NW-side training data collection, support L1-RSRP(s) for all beams of Set A and Set B, Top-K Beam ID(s) for Set A, related timestamps, and information to facilitate model training.
- Proposal 2 (Sec 2.1.2): For NW-side training data collection, support the enhancement to use high layer signaling (e.g., MAC-CE, RRC) to convey data collection content.
- Proposal 3 (Sec 2.2.1): For NW-side inference with >4 beams, support differential L1-RSRP reporting with larger quantization steps/smaller ranges and consider two-part CSI to reduce overhead.
- Proposal 5 (Sec 2.2.2): For NW-side inference, support CSI-ReportConfig with measurements for multiple past time instances in one reporting instance.
- Proposal 6 (Sec 2.3): Support single beam indication for multiple future time instances using the unified TCI framework.
- Proposal 7 (Sec 3.1.1): For UE-side inference (BM-Case 1), support configurability between Alt 1 (one CSI-ResourceConfigId for Set B, size of Set A configured by CSI-ReportConfig) and Alt 3 (two CSI-ResourceConfigIds for Set A and Set B separately).
- Proposal 8 (Sec 3.1.1): Introduce DL Tx IDs for identifying downlink spatial domain transmission filters, where each beam in Set A and Set B is associated with a DL Tx ID, and UE assumes beams with the same DL Tx ID share the same filter.
- Proposal 11 (Sec 3.1.1): For UE-side inference report content, support Option 3: Beam information on predicted Top K beam(s) and probability information of predicted Top K beam(s).
- Proposal 14 (Sec 3.1.1): For BM-Case 2, determine N future time instances where each consists of P consecutive slots, with reference time based on the uplink slot for the report, and offsets D1 and D2 for time instance positioning.
- Proposal 18 (Sec 3.1.2): For consistency of NW-side additional conditions, UE assumes consistent physical beam characteristics (e.g., same DL spatial domain transmission filters) and consistent mapping for DL Tx beams of Set A and Set B at least at a cell level.
- Proposal 21 (Sec 3.2.1): For Type 1 Option 2 monitoring, introduce a new quantity 'beam accuracy indicator' (BAI) in the CSI report, defined by the number of CSI reports with 'true' comparison results from X reports.
- Proposal 22 (Sec 3.2.1): For Type 1 Option 2 monitoring, support Option 2 where dedicated resource sets and report configurations for monitoring are configured in a dedicated CSI report configuration.
- Proposal 28 (Sec 3.3): Consider separately counting CPU between legacy non-AI/ML based CSI reports and AI/ML based CSI reports.
- Proposal 30 (Sec 3.3): Support scaling of the legacy Z timeline (Z3/Z3') for beam management with a scaling factor depending on UE capability.