R1-2500390
discussion
Specification Support for AI/ML for Beam Management
From Kyocera
Kyocera's prior position on
9.1.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Advocates for a comprehensive AI/ML beam management framework leveraging existing CSI infrastructure with minimal new specifications, strongly supporting virtual Set A configuration for UE-side models and flexible Set B definition through new information elements.
Summary
Kyocera submits 25 proposals and 4 observations regarding AI/ML beam management for NR Rel-19, focusing on the configuration of Sets A and B, inference reporting formats, and performance monitoring mechanisms. The document addresses consistency requirements via associated IDs, defines specific metrics for UE-side model monitoring (Type 1 Option 2), and proposes overhead reduction techniques for NW-side models.
Position
Kyocera proposes distinct configuration strategies for UE-side and NW-side AI/ML models, specifically requiring Set B to be explicitly configured for UE measurements while allowing Set A to be virtually configured for reference mapping. They argue against reporting full beam indices for NW-side models when M equals the resource set size, proposing instead to report only the index of the strongest beam to reduce overhead. Regarding consistency, Kyocera requires that beam properties, set sizes, and resource indexing remain consistent between training and inference phases for the same associated ID, and prioritizes defining these UE assumptions before extending associated ID applicability across cells to prevent revealing proprietary network information. For performance monitoring, they support Type 1 Option 2 metrics based on ground truth measurements (Alt 1-3) and explicitly deprioritize Alt 4 (probability-only metrics) due to the lack of ground truth validation. They further propose specific Beam Accuracy Indicator (BAI) definitions and handling mechanisms for monitoring sets that are subsets of Set A.
Key proposals
- Proposal 1 (Sec 2.1): For UE-side models, support configuring Set B explicitly for measurements while Set A is optionally virtually configured as a reference for mapping classification outputs to CRIs.
- Proposal 2 (Sec 2.2): For NW-side models, if M equals the size of the measurement resource set, report all L1-RSRPs and only one beam index (CRI/SSBRI) for the largest value, omitting indices for other beams.
- Proposal 3 (Sec 3.1): For UE-side inference reports, support options 1-3 and further study option 4 (including probability/confidence info) considering quantization effects and overhead.
- Proposal 4 (Sec 3.2): Define confidence information for option 4 as a range within which the predicted RSRP is expected to fall a certain percentage of time if the same input sample is reused.
- Proposal 6 (Sec 3.4): Support one decimal precision for probability information reporting, requiring only 4 bits per beam, with rounding left to UE implementation.
- Proposal 7 (Sec 3.5): Support two methods for ranking Top-K beams: CRI-based vector reporting or Bitmap reporting (K times), with FFS on feasibility/overhead comparison.
- Proposal 8 (Sec 4.1): Ensure consistency for UE-side models by assuming beam properties (shape, width), set sizes, and resource indexing remain consistent between training and inference for the same associated ID.
- Proposal 9 (Sec 4.2): Configure the associated ID within the CSI framework (nzp-CSI-RS-ResourceSet and csi-SSB-ResourceSet) for each Set A and Set B.
- Proposal 10 (Sec 4.3): Prioritize finalizing UE assumptions for the same associated ID before addressing its applicability across different cells to avoid revealing proprietary network information.
- Proposal 12 (Sec 5.1): For NW-side models, study ensuring consistency by having the UE report receive beam index or assuming the best receive beam via QCL configuration.
- Proposal 13 (Sec 6.1.1): For Type 1 Option 1 monitoring, report model outputs (beam info, predicted RSRP, probability, confidence) alongside performance measurements (CRI, L1-RSRP).
- Proposal 15 (Sec 6.1.2): For Type 1 Option 2 monitoring, support Alt 1-3 (beam accuracy, L1-RSRP diff, RSRP diff) and deprioritize Alt 4 (probability only) as it lacks ground truth.
- Proposal 16 (Sec 6.1.2): Define Beam Accuracy Indicator (BAI) options for sample accuracy: Top-1/1, 1/Top-K, and Top-K/K, reported either per sample or as an average over a window T.
- Proposal 17 (Sec 6.1.2): For monitoring subsets of Set A, study alternatives including calculating metrics only if Top-K beams are included, using down-sampling mapping, or counting uncovered beams as inaccurate.
- Proposal 19 (Sec 6.1.2): Support specific time domain combinations for inference and monitoring reports, such as periodic inference with periodic/semi-persistent/aperiodic monitoring.