R1-2410734
discussion
FL summary #1 for AI/ML in beam management
From Samsung
Summary
This Samsung-moderated FL summary document for RAN1#119 contains over 250 proposals and observations across 9 main sections covering AI/ML beam management, including RAN2 LS handling, performance monitoring, configuration aspects, and inference reporting. The document addresses both UE-sided and NW-sided AI/ML models for spatial (BM-Case1) and temporal (BM-Case2) beam prediction use cases.
Position
Samsung advocates for a flexible approach supporting multiple options for UE-sided model applicability reporting, favoring dedicated monitoring configurations separate from inference configurations, and supporting enhanced quantization steps for overhead reduction. They push for practical solutions that maintain consistency between training and inference while enabling both spatial and temporal beam prediction capabilities.
Key proposals
- Proposal 1.1B (Sec 1.1): Potential working framework combining Option 1 and Option 2 for applicability reporting - configure CSI-ReportConfig and/or inference related parameters in Step 3, with applicability reported via RRCReconfigurationComplete or UAI in Step 4
- Proposal 2.1-1a (Sec 2.1): Introduce beam accuracy indicator (BAI) for beam prediction accuracy in CSI report for monitoring, with BAI = Np/N where Np is number of correct predictions out of N total instances
- Proposal 2.2-1a (Sec 2.2): Support dedicated resource set(s) for monitoring configured in separate CSI report configuration, linked to inference report configuration via CSI-ReportConfigID
- Proposal 3.1a (Sec 3.1): For BM-Case2 reference time, support Option 3 based on latest transmission occasion of CSI-RS/SSB resource in Set B for measurement
- Proposal 4.1b (Sec 4.1): For UE-sided model BM-Case1 inference results, ranking information of predicted Top-K beams conveyed by order of beam information based on model output
- Proposal 5.2 (Sec 5.2): For NW-sided model training data collection, support collecting L1-RSRPs and beam-IDs as baseline content
- Proposal 5.4 (Sec 5.4): Support Y dB quantization steps (Y=3 and/or 4) for differential L1-RSRP reporting, larger than legacy quantization
- Proposal 7.1 (Sec 7.1): Study extending Rel-17 TCI state activation methods to indicate N joint TCI states for N future time instances in BM-Case2
- Proposal 8.1 (Sec 8.1): Use existing CPU mechanism as starting point for AI/ML-based CSI processing, with FFS on sharing between legacy and AI/ML reporting
- Proposal A (Sec 9): For BM-Case2 UE-side model, resources for Set B across different time instances are fixed
- Proposal B (Sec 9): For UE-sided model inference with two resource sets, configure separate CSI-ResourceConfigIds for Set A and Set B
- Proposal C (Sec 9): Support three types of training data collection content for NW-sided model - L1-RSRPs only, L1-RSRPs plus beam information, or L1-RSRPs plus beam information and RSRPs
- Agreement (Sec 2.1): Support Alt 1 for UE-assisted performance monitoring - Top 1 or Top K beam prediction accuracy comparing prediction results with measurements from monitoring resource set