R1-2410466
discussion
Specification support for AI-ML-based beam management
From Qualcomm
Summary
This Qualcomm document presents 19 proposals and 3 observations for AI/ML-based beam management in 5G NR, focusing on ensuring consistency between training and inference phases for UE-side models, performance monitoring mechanisms, and signaling configurations within the CSI framework.
Position
Qualcomm strongly advocates FOR associated ID-based consistency mechanisms over performance monitoring approaches, pushing for joint Set A/Set B configuration within existing CSI framework to minimize specification impact. They argue AGAINST performance monitoring-based consistency (Option 2) due to UE complexity and advocate FOR PLMN-wide associated ID uniqueness to enable cross-cell model generalization. Their position emphasizes practical implementation considerations and leveraging existing 3GPP frameworks rather than creating new signaling mechanisms.
Key proposals
- Proposal 1 (Sec 2.1.1): Ensure consistency between training and inference for Set A/Set B beams through order/indexing consistency (resource index consistency) and beam shape consistency (pointing direction and beamwidth differences under predefined tolerances)
- Proposal 3 (Sec 2.2): Deprioritize performance monitoring-based consistency (Opt2) due to extra UE burden, hit/fail methodology, and significant delays in determining applicable functionality
- Proposal 5 (Sec 2.3): Configure associated ID for Set A and Set B pairs jointly rather than separately within CSI framework to reduce signaling overhead
- Proposal 7 (Sec 2.4): NW-side additional conditions with the same associated ID should be consistent within a PLMN to enable model generalization across multiple cells
- Proposal 8 (Sec 3.1): Support Alt 3 - configure two separate CSI-ResourceConfigIds for Set A and Set B to leverage existing CSI framework with minimal changes
- Proposal 10 (Sec 3.1): Support Option 4 for inference results reporting - include beam information, RSRP of predicted Top-K beams, and confidence information of RSRP predictions
- Proposal 11 (Sec 4.1): Specify signaling details for reference signals used in performance monitoring that span entire Set A or subsets of beams from Set A
- Proposal 13 (Sec 4.2.1): Define Top-K beam prediction accuracy metric with L1-RSRP margin as ratio N/M where N is successful predictions within margin and M is total monitorable instances
- Proposal 15 (Sec 4.2.2): Support L1-RSRP difference metric between measured Top-1 predicted beam and best measured beam from performance monitoring set, per monitoring instance
- Proposal 17 (Sec 4.3): Study three specific events for event-triggered performance monitoring reports: Top-K beams within monitoring set, Top-1 measured beam not among Top-K predicted, and L1-RSRP not within X dB margin
- Proposal 18 (Sec 5.1): Associated ID must be indicated from NW to UE during both training and inference - determining applicability without associated ID is not feasible
- Proposal 19 (Sec 5.2): Support Option 1 for applicable functionality reporting with clarifications that one or more associated IDs representing current NW-side conditions are shared, and associated ID must be PLMN-unique