R1-2409852
discussion
Discussion on specification support for AIML based positioning accuracy enhancement
From NEC
Summary
This NEC contribution provides a comprehensive discussion on AI/ML for NR positioning accuracy enhancement, covering model input/output, training data collection, lifecycle management, and consistency between training and inference. The document contains 42 detailed proposals and 6 observations addressing various technical aspects of the standardization work.
Position
NEC strongly advocates FOR sample-based measurements over path-based measurements, arguing that sample-based approach provides better standardization control and avoids vendor-specific path detection algorithms. They push FOR supporting both UE-side and network-side AI/ML model deployment with comprehensive quality indicators and consistency mechanisms. NEC is AGAINST limiting phase information usage and advocates FOR flexible implementation-dependent decisions on phase information per case. They strongly support semi-supervised learning and mixed dataset approaches to improve model robustness.
Key proposals
- Proposal 1 (Sec 2.1): For sample-based measurement definition, value range of Nt' should be ≤9 or 9< Nt' ≤256
- Proposal 6 (Sec 2.1): Support compromised scheme with both sample-based and path-based measurement, prioritizing sample-based if down-selection needed
- Proposal 7 (Sec 2.2): For Case 1, confirm measurement generated by PRU/Non-PRU UE and label generated by PRU/Non-PRU UE/LMF
- Proposal 9 (Sec 2.3): Study necessary signaling exchange between UE/gNB and LMF to pair Part A and Part B as valid training data
- Proposal 10 (Sec 2.4): Endorse integration of mixed datasets from diverse scenarios to train robust generalized models
- Proposal 13 (Sec 2.5): For Case 3b, support phase information reporting from gNB to LMF using Rel-18 phase info as starting point
- Proposal 15 (Sec 3.1): For Case 3a, no need to report LOS/NLOS indicator if model output is timing information
- Proposal 19 (Sec 4.1): Define quality indicator for data samples based on associated measurement and ground truth label quality
- Proposal 22 (Sec 4.2): Use 'UTC time + SFN + slot index' as baseline for timestamp design
- Proposal 25 (Sec 5.1): Support semi-supervised learning for AI/ML based positioning
- Proposal 26 (Sec 5.2): Support both Option A and B for Case 1 monitoring - both LMF and UE can perform metric calculation
- Proposal 30 (Sec 6.2): Support associated ID as implicit way to ensure training/inference consistency with further study on scope
- Proposal 33 (Sec 7): Specify gNB-initiated RS configuration request for model training/monitoring at gNB side
- Proposal 34 (Sec 7): Specify distinct RS configuration patterns for different LCM stages with different time domain periods