R1-2409928
discussion
Additional study on AI/ML-based CSI compression
From CATT
Summary
This CATT document provides extensive analysis and evaluation results for AI/ML-based CSI compression inter-vendor training collaboration, containing 27 proposals and 25 observations covering various approaches including Direction A (parameter/dataset sharing with UE offline engineering), Direction B (NW encoder parameter sharing), temporal domain aspects, and comprehensive specification requirements.
Position
CATT strongly advocates FOR Direction B (Option 3b) over Direction A, supporting direct NW encoder parameter sharing to UE without offline engineering based on their evaluation showing better performance and less complexity. They push FOR standardized OTA signaling with RRC as starting point, spatial-frequency domain model input, transformer backbone for Case 0, and CSI-RS measurement based data collection. They are AGAINST localized models due to implementation complexity, UCI loss specification support for Case 2, and UE-side performance monitoring based on reference/proxy models.
Key proposals
- Proposal 2 (Sec 2.2): Regarding inter-vendor training collaboration of AI/ML-based CSI compression using two-sided model, for parameter/dataset/model exchange between UE-side and NW-side(Option 3/4), prioritize the solutions with standardized over-the-air signaling, and RRC signaling can be considered as a starting point
- Proposal 3 (Sec 2.3): Prioritize model input on spatial-frequency domain
- Proposal 7 (Sec 2.4): For inter-vendor collaboration direction A and B, prioritize direction B (Option 3b)
- Proposal 8 (Sec 2.5): For Case 2, specification support on dealing with non-ideal UCI feedback (i.e., UCI loss) is not needed
- Proposal 9 (Sec 2.8.1): In CSI compression using two-sided model use case, for NW-side data collection for model training, focus on CSI-RS measurement based data collection
- Proposal 17 (Sec 2.8.2.1): In CSI compression using two-sided model use case, performance monitoring at UE-side based on reference model or proxy model can be deprioritized
- Proposal 19 (Sec 2.8.2.1): In CSI compression using two-sided model use case, support UE-side monitoring based on precoded RS (e.g., CSI-RS, DMRS) transmitted from NW based on the output of the CSI reconstruction model
- Proposal 22 (Sec 2.8.2.2): In CSI compression using two-sided model use case, if NW-side performance monitoring based on intermediate KPI is adopted, NW-side monitoring for identifying whether performance degradation is caused by UE-side issues can be considered, based on target CSI reported by the UE
- Proposal 24 (Sec 2.8.3): In CSI compression using two-sided model use case, quantization alignment between UE-side and NW-side in a 3GPP non-transparent manner is supported
- Proposal 25 (Sec 2.8.4): In CSI compression using two-sided model use case, legacy CSI reporting principles are reused as much as possible
- Proposal 26 (Sec 2.8.4): In CSI compression using two-sided model use case, if CQI in CSI report is configured, for CQI determination in CSI report, one of the sub options of Option 1 is adopted - CQI is NOT calculated based on the output of CSI reconstruction part from the realistic channel estimation
- Proposal 4 (Sec 2.3): SF-domain AI/ML-based CSI compression (temporal domain Case 0) is supported for AI/ML-based CSI compression
- Proposal 5 (Sec 2.3): For temporal domain Case 0, adopt transformer as the backbone structure, other model structures are not precluded if RAN4 verifies their feasibility
- Proposal 10 (Sec 2.8.1): In CSI compression using two-sided model use case, both L1 signaling based reporting and RRC signaling based reporting are supported for ground-truth CSI for NW-side data collection for model training
- Proposal 20 (Sec 2.8.2.1): In CSI compression using two-sided model use case, for temporal domain aspects Case 3 and Case 4 with separate prediction and compression adopted, support monitoring the performance of the model for prediction and the performance of the model for compression separately