R1-2500066
discussion
Discussion on AI/ML-based beam management
From ZTE
ZTE's prior position on
9.1.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Strongly advocates for functionality-based lifecycle management over model-ID-based approaches and pushes for bitmap-based beam reporting methods to significantly reduce signaling overhead while leveraging existing UE capability frameworks.
Summary
ZTE proposes functionality-based LCM without model ID signaling for AI/ML beam management, emphasizing overhead reduction through bitmap reporting and threshold-based beam selection. The document outlines specific mechanisms for NW-side data collection, UE-side inference reporting, performance monitoring metrics, and the optional use of associated IDs for NW-side additional conditions.
Position
ZTE proposes utilizing functionality-based LCM without model ID based signaling for AI/ML beam management, arguing that model transfer challenges diminish the need for model-ID-based approaches. They support bitmap-based beam information reporting and threshold-based beam selection to significantly reduce overhead, citing up to 63.5% reduction in typical settings. For NW-side data collection, ZTE proposes supporting L1 signaling and flexible resource configuration (one or two RS sets) depending on Set A/Set B construction. Regarding performance monitoring, they prefer beam prediction accuracy and RSRP prediction accuracy as primary metrics and support L1 signaling for NW-side monitoring. Finally, ZTE argues that the configuration of associated ID for NW-side additional conditions should be optional, with UE assuming similar properties of DL Tx beam sets for the same ID, while also proposing separate CPU counting for AI/ML processing.
Key proposals
- Proposal (General Views): Functionality-based LCM without specifying any model ID based signaling should be utilized for AI/ML based beam management.
- Proposal (Data Collection - Issue 1): For NW-sided model training via higher layer signaling, three types of content (all L1-RSRPs, beam ID of Top-K, or beam ID and RSRP of Top-K) can be considered to serve different training strategies.
- Proposal (Data Collection - Issue 2): One or two RS resource sets can be configured for NW-side training data collection, dependent on whether Set B is a subset of Set A or distinct.
- Proposal (Data Collection - Issue 3 Method 1): Support threshold-based beam reporting where up to M beams within X dB gap to the largest L1-RSRP are reported, reusing legacy two-part CSI encoding.
- Proposal (Data Collection - Issue 3 Method 2): Specify bitmap-based method for reporting beam information to reduce overhead by up to 63.5% compared to legacy CRI/SSBRI methods.
- Proposal (Data Collection - Issue 3 Method 3): Support differential L1-RSRP reporting with larger quantization step sizes (e.g., 4 dB) for differential values to reduce bit width.
- Proposal (Data Collection - Issue 4): Support L1 signaling for NW-side data collection irrespective of purpose (training, inference, monitoring).
- Proposal (Model Inference - Issue 1 BM-Case2): Support enhancements to report measurements of multiple past time instances in one reporting instance, including timestamp, reference beam, and common beam information indications.
- Proposal (Model Inference - Issue 2): For BM-Case 2 UE-side model, support Option 3 for reference time (based on latest transmission occasion of CSI-RS/SSB in Set B) due to configuration simplicity.
- Proposal (Model Inference - Issue 3): If Set B is a subset of Set A, configure only Set A resources and indicate Set B as a subset via assistance information like a bitmap.
- Proposal (Model Inference - Issue 4): For UE-sided model inference results, report measured RSRP if available for the beam, as it is more reliable than predicted RSRP.
- Proposal (Performance Monitoring - Issue 1): Support beam prediction accuracy (Alt 1) and RSRP prediction accuracy (Alt 3) as primary performance metrics for UE-sided AI/ML model monitoring.
- Proposal (Performance Monitoring - Issue 2): Support L1 measurement (CSI reporting) for collecting data to enable Type 1 NW-side performance monitoring due to latency and payload constraints.
- Proposal (NW-side Additional Conditions): The configuration of associated ID is optional and can be configured by the NW if necessary; UE may assume similar properties of DL Tx beam set/list for the same ID.
- Proposal (CSI Processing Criteria): The overall CPU can be separately counted between legacy CSI reporting and AI/ML-based CSI reporting to accommodate hardware flexibility.