R1-2500565
discussion
Discussions on AI/ML for beam management
From LG Electronics
Summary
LG Electronics presents 26 proposals and 5 observations regarding AI/ML enhancements for NR beam management, focusing on data collection, inference mechanisms, performance monitoring, and consistency between training and inference. The document argues for overhead reduction in UE-sided models by configuring only Set B resources and proposes specific reporting formats for temporal beam prediction, including differential RSRP and confidence metrics. It also addresses the need for clear definitions of beam similarity and introduces the concept of AI/ML Processing Units (APUs) for managing UE-side inference complexity.
Position
LG Electronics proposes extending the Rel-18 NES sub-configuration mechanism for flexible Set B beam reporting and supports configuring only Set B resources for UE-sided inference to minimize overhead. They require the definition of 'similar properties' for beam consistency to be based on maintaining the same downlink spatial domain transmission filters. LG opposes supporting Aperiodic CSI-RS for Set B in BM-Case 2 due to excessive DL resource overhead and instead supports only Periodic CSI-RS for Set A. They propose introducing AI/ML Processing Units (APUs) distinct from CPU for managing inference complexity and suggest using differential RSRP reporting with confidence/probability metrics for temporal beam prediction. Furthermore, they argue for event-triggered performance monitoring to reduce signaling overhead compared to periodic reporting.
Key proposals
- Proposal #1 (Data Collection): Support UE reporting enhancements for NW-sided AI/ML data collection, specifically past/present best N beams per time stamp and tendency/variance of best N beams.
- Proposal #3 (Data Collection): Extend the Rel-18 NES sub-configuration mechanism for Set B beam measurement and reporting, allowing different Set A/B associations per sub-configuration.
- Proposal #4 (Inference - NW-sided): For L1 signaling of more than 4 beam-related information, support legacy CRI/SSBRI for beam information along with corresponding L1-RSRPs.
- Proposal #5 (Inference - NW-sided): To indicate a beam in Set A not in Set B, support indicating multiple neighboring beams from Set B to help the UE find its Rx beam.
- Proposal #6 (Inference - UE-sided): Support configuring only the resource set for Set B in CSI-ResourceConfig for inference result reporting (Alt 1) to reduce signaling overhead.
- Proposal #7 (Inference - UE-sided): Provide assistance information on the relation between Set A and Set B beams using linear combining coefficients or 2D/3D coordinates to preserve proprietary information.
- Proposal #11 (Inference - UE-sided): Support differential RSRP reporting among multiple beams over multiple time instances for BM-Case 2, with an indicator for the time instance containing the largest predicted RSRP.
- Proposal #18 (Inference - UE-sided): Introduce AI/ML Processing Units (APUs) for CSI reports, managed separately from CPU, to handle UE-side AI/ML inference operations.
- Proposal #19 (Performance Monitoring): For UE-assisted performance monitoring (Option 2), support at least the full Set A for the dedicated resource set for monitoring.
- Proposal #20 (Performance Monitoring): Specify a time limit or buffering window for inference results when linked to a monitoring report to prevent indefinite UE buffering.
- Proposal #24 (Consistency): Define 'similar properties' of a DL Tx beam set as maintaining the same downlink spatial domain transmission filters for each beam in different transmission instances.
- Proposal #26 (Applicable Functionality): UE reports value ranges for Functionality Group components in Step 2 and actual supported values in Step 4 based on NW configuration.