R1-2409780
discussion
Specification support for AI-enabled beam management
From NVIDIA
Summary
NVIDIA presents a comprehensive framework for AI/ML-enabled beam management in 5G-Advanced, focusing on spatial (BM-Case 1) and temporal (BM-Case 2) downlink beam prediction. The document contains 11 proposals and 1 observation, advocating for specification support for beam set associations, model lifecycle management, and UE capability signaling to facilitate both offline and continuous online training.
Position
NVIDIA proposes specification support for associating Set A of beams with Set B of beams for both spatial (BM-Case 1) and temporal (BM-Case 2) DL beam prediction, establishing the foundational mapping for AI/ML inputs. They require specification support for using L1-RSRP measurements and historical optimal beam indices from Set B as AI/ML model inputs, alongside UE reporting of inference results for Set A. NVIDIA emphasizes the necessity of assistance signaling for full model lifecycle management, including configuration, activation/deactivation, recovery/termination, and selection. They further propose specification support for model performance monitoring and update/tuning procedures to address dynamic environment changes. Finally, they call for UE capability signaling for AI/ML beam prediction and the definition of conditions for Feature/FG availability and additional conditions in model description information during identification.
Key proposals
- Proposal 1 (Sec 2.1): For BM-Case 1, introduce specification support for associating Set A of beams with Set B of beams.
- Proposal 2 (Sec 2.1): For BM-Case 1, at least introduce specification support for using L1-RSRP measurement based on Set B of beams as AI/ML model input.
- Proposal 3 (Sec 2.1): For BM-Case 1, introduce specification support for UE to report inference results on L1-RSRP and beam information in the resource set for Set A.
- Proposal 4 (Sec 2.2): For BM-Case 2, introduce specification support for associating Set A of beams with Set B of beams.
- Proposal 5 (Sec 2.2): For BM-Case 2, at least introduce specification support for using historical optimal beam index and/or L1-RSPR measurement based on Set B of beams as AI/ML model input.
- Proposal 6 (Sec 2.2): For BM-Case 1, introduce specification support for UE to report inference results on L1-RSRP and beam information in the resource set for Set A for a number of future time instances.
- Proposal 7 (Sec 3): For AI/ML based beam prediction in spatial/time domain, introduce specification support for assistance signalling and procedure for model configuration, model activation/deactivation, model recovery/termination, and model selection.
- Proposal 8 (Sec 3): For AI/ML based beam prediction in spatial/time domain, introduce specification support for assistance signalling and procedure for model performance monitoring and model update/tuning.
- Proposal 9 (Sec 3): For AI/ML based beam prediction in spatial/time domain, introduce specification support for UE capability signaling for AI/ML based beam prediction including model training, model inference and model monitoring.
- Proposal 10 (Sec 3): For AI/ML based beam prediction in spatial/time domain, introduce specification support for conditions of a Feature/FG available for functionality.
- Proposal 11 (Sec 3): For AI/ML based beam prediction in spatial/time domain, introduce specification support for additional conditions to include them in model description information during model identification.