R1-2410048
discussion
Discussion on specification support on AI/ML for beam management
From Fujitsu
Summary
Fujitsu presents a comprehensive technical document with 35 proposals for AI/ML-based beam management in 5G NR Release 19, covering RAN2 liaison responses, training data collection, UE-side and network-side model inference, and performance monitoring across both spatial (BM Case-1) and temporal (BM Case-2) beam prediction scenarios.
Position
Fujitsu advocates FOR Option-3 in applicable functionality reporting (UE-driven parameter reporting rather than network-driven configuration), supports aperiodic CSI-RS for temporal beam prediction, prefers separate configurations for monitoring vs inference, and pushes for high-resolution quantization in training data. They argue AGAINST Option-1/Option-2 for functionality reporting due to feasibility concerns and advocate for L1-RSRP difference as the preferred performance metric for network-side models over beam prediction accuracy.
Key proposals
- Proposal 1 (Sec 2): Support Option-3 for applicable functionality reporting where UE reports inference parameters via UAI after receiving associated IDs from network
- Proposal 8 (Sec 3.1): Training data collection should include reference signal ID and beam quality (L1-RSRP) as minimum information
- Proposal 12 (Sec 3.2): Configure reference signals same as Set B for model input data and Set A for ground truth data, with single resource set if Set B is subset of Set A
- Proposal 15 (Sec 3.3): Apply same UE Rx beam for measurements on both Set B (model input) and Set A (ground truth) reference signals during training data collection
- Proposal 17 (Sec 4.1.1): Prefer Option 2 for BM Case-1 UE-side model reporting - beam information and RSRP of predicted Top K beams, with beam info including CRI/SSBRI and CC ID
- Proposal 19 (Sec 4.1.2): Support aperiodic CSI-RS for BM Case-2 UE-side model inference to provide network flexibility
- Proposal 25 (Sec 4.2): Prefer Option-2 for UE-assisted performance monitoring using dedicated resource sets and report configurations separate from inference
- Proposal 29 (Sec 5.1.1): For BM Case-1 NW-side model, beam information in Set B measurement reporting should include CRI/SSBRI
- Proposal 33 (Sec 5.2): Prefer L1-RSRP difference between measured and predicted values as performance metric for NW-side model monitoring
- Proposal 11 (Sec 3.1): Consider high-resolution quantization and non-differential RSRP for training data to improve model quality
- Proposal 21 (Sec 4.1.2): Allow UE to report preferred measurement and prediction patterns including number of instances and intervals for BM Case-2
- Proposal 24 (Sec 4.1.2): Support beam indication enhancement where TCI states for multiple time instances can be indicated via single DCI
- Proposal 28 (Sec 4.2): Use L1-RSRP difference between predicted and measured values as additional performance metric beyond beam prediction accuracy
- Proposal 34 (Sec 5.2): Consider high-resolution quantization and non-differential RSRP for ground truth data in NW-side performance monitoring
- Proposal 35 (Sec 6): Further discuss application of associated ID across different cells for training-inference consistency