R1-2409985
discussion
AI/ML for Beam Management
From Nokia
Summary
Nokia presents a comprehensive technical document with 22 proposals and 7 observations addressing AI/ML for beam management in 5G NR, covering both UE-sided and NW-sided models for spatial (BM-Case1) and temporal (BM-Case2) beam prediction. The document systematically addresses inference operations, performance monitoring, beam indication, and data collection aspects.
Position
Nokia strongly advocates FOR reusing existing CSI reporting frameworks as the foundation for AI/ML beam management while proposing systematic enhancements for beam prediction capabilities. They push FOR practical parameter values (K=1,2,4 for beams, N=1,2,4 for time instances), flexible resource set configurations, and comprehensive performance monitoring options. Nokia is notably pushing AGAINST overly complex new signaling mechanisms, favoring incremental enhancements to proven legacy frameworks over entirely new protocols. They specifically support Option 1 for inference applicability and advocate for separate but coordinated CPU limits for AI/ML processing.
Key proposals
- Proposal 1 (Sec 2.1.1.1): For BM-Case1 beam prediction reporting, support K=1,2,4 configurable values for Top-K beams with CRIs corresponding to Set A, RSRP reporting similar to legacy L1-RSRP, and probability information for beam prediction confidence
- Proposal 2 (Sec 2.1.1.2): For BM-Case2, support N=1,2,4 future time instances with duration values of 40ms, 80ms, 160ms, 320ms, 640ms and reference time as the slot carrying the inference report
- Proposal 3 (Sec 2.2): Configure two separate resource sets for Set A and Set B in CSI report configuration for inference results reporting
- Proposal 5 (Sec 2.3): Extend Rel-17 TCI state activation/indication signalling to activate/indicate N TCI states corresponding to future time N instances for BM-Case2
- Proposal 8 (Sec 2.4.2): Support dedicated resource set for monitoring with dedicated CSI report configuration that links to inference report configuration for UE-assisted performance monitoring
- Proposal 12 (Sec 2.6): Configure associated ID separately for Set A and Set B within CSI framework, with assignment left to NW implementation
- Proposal 13 (Sec 2.7.1): Consider separate CPU limit for AI/ML-enabled CSI reports with both legacy and AI/ML-specific limits considered for CPU occupancy
- Proposal 15 (Sec 3.1.2): Enhance L1-RSRP quantization by increasing differential L1-RSRPs to X dB quantization step (X=3 or 4) to reduce UCI reporting overhead
- Proposal 16 (Sec 3.1.2): For M>4 beams, support configurable M values of 4,6,8,12,16,32 with options for M CRIs/SSBRIs or resource subset indication
- Proposal 18 (Sec 3.2): Extend Rel-17 TCI state activation/indication signalling for NW-sided BM-Case2 to activate/indicate N TCI states for future time instances
- Proposal 19 (Sec 3.3): Support two variants for NW-sided performance monitoring - either different CSI report or same CSI configuration for monitoring and inference
- Proposal 20 (Sec 3.4): Support NW-sided data collection via CSI reporting with options for L1-RSRP reporting on one/two measurement sets or combined L1-RSRP and Top-K beam information
- Proposal 21 (Sec 4.1): Support Option 1 for applicability of inference for UE-side model among three discussed options
- Proposal 22 (Sec 4.1): Any CSI-ReportConfig reported as applicable via RRCReconfigurationComplete can be used for inference after reporting, with UAI reporting not impacting CSI-ReportConfig use