R1-2500512
discussion
Discussion on AI/ML for beam management
From Ofinno
Summary
Ofinno presents 15 proposals and 1 observation regarding AI/ML beam management for NR, focusing on UE-sided model data collection, inference reporting enhancements, and performance monitoring. The document argues for UE-initiated control over data collection to reduce overhead, proposes enhancing inference reports with prediction quality metrics, and suggests extending the unified TCI framework to support multi-time-instance beam indications.
Position
Ofinno proposes supporting UE requests to initiate and terminate data collection for UE-sided models to prevent unnecessary transmission of Set A reference signals when training is complete or not required. They propose enhancing inference result reports for BM-Case 1 by including prediction quality information, such as probability or confidence, potentially filtered by an RRC-configured threshold. For BM-Case 2, they propose allowing the UE to select from multiple configured duration values and specify a UE-initiated reporting mechanism to override previous predictions if channel conditions change significantly. Regarding beam indication, Ofinno proposes extending the unified TCI framework to allow a single beam indication for N future time instances and suggests synthesizing QCL properties for unmeasured beams using spatially correlated measured beams from Set B. For performance monitoring, they propose using a subset of Set A beams and supporting aperiodic CSI reporting for beam prediction accuracy, while also considering monitoring for multiple models associated with one functionality. Finally, they propose defining specific priority rules for AI/ML reports and enhancing CPU occupancy definitions to account for long observation windows.
Key proposals
- Proposal 1 (Overall procedure): Support UE requests to initiate/terminate data collection for a UE-sided model to avoid unnecessary transmission of Set A reference signals.
- Proposal 2 (Inference result report BM-Case 1): Enhance reporting options to allow informing prediction quality, such as probability or confidence information.
- Proposal 3 (Inference result report BM-Case 1): Specify RRC configuration of a threshold for prediction quality to filter reported beams.
- Proposal 5 (Inference result report BM-Case 2): Support multiple duration values for BM-Case 2 inference reports, allowing the UE to select and inform the value.
- Proposal 6 (Inference result report BM-Case 2): Specify a UE-initiated inference results report mechanism to override previously reported results in case of channel changes.
- Proposal 8 (Beam indication): Further discuss beam sweeping to address QCL assumptions for predicted beams, considering flexible indication based on gNB configuration.
- Proposal 9 (Beam indication): Consider an approach where the gNB indicates spatially correlated Tx beams from Set B for each Tx beam of Set A to synthesize QCL properties.
- Proposal 10 (Beam indication): Support extending the unified TCI framework to allow single beam indication for N future time instances.
- Proposal 11 (Performance monitoring): Support using a subset of Set A beams for performance monitoring, considering Alt 3 (Top K’ among subset) and Alt 2 (predefined mapping).
- Proposal 12 (Performance monitoring): Support at least aperiodic CSI reporting of beam prediction accuracy for performance monitoring.
- Proposal 13 (Performance monitoring): Consider monitoring on multiple models associated with one AI/ML functionality to support model switching.
- Proposal 14 (Priority rules): Define priority rules between AI/ML-based reports and non-AI/ML-based reports, considering use cases, LCM purposes, and report quantity types.
- Proposal 15 (CSI processing): Enhance CPU and CPU occupancy definitions considering long observation windows for performance monitoring and inference.