R1-2409502
discussion
Discussion on AI/ML for CSI compression
From CMCC
Summary
CMCC presents evaluation results for AI/ML-based temporal domain CSI compression (Cases 2 and 3), demonstrating significant SGCS gains over Rel-16 and Rel-18 benchmarks. The document proposes prioritizing inter-vendor collaboration Directions A and B, specifies requirements for dataset/parameter exchange, and recommends NW-side monitoring and enhanced codebook structures for ground-truth data collection.
Position
CMCC prioritizes inter-vendor collaboration Directions A and B for further study, arguing they offer better performance and lower specification effort than Direction C. They propose that for Direction A, the NW-side must share dataset information and performance targets to enable UE-side offline engineering, while acknowledging that over-the-air parameter exchange introduces extra overhead. For Direction B, they assert there is no proprietary information disclosure concern and that common encoders are feasible for UEs with similar conditions within a cell. Regarding performance monitoring, CMCC prioritizes NW-side monitoring based on ground-truth CSI reports and UE-side monitoring based on recovery CSI indication. Finally, they propose using Rel-16 eType-II and Rel-18 Doppler codebooks as starting points for ground-truth CSI data collection, with enhanced parameter values to ensure high resolution.
Key proposals
- Proposal 1 (Sec 3.1.1): For Direction A, NW-side should share dataset information and performance targets to enable UE-side encoder training, validation, and testing.
- Proposal 3 (Sec 3.1.3): For Direction A, dataset/parameter exchange over the air is necessary to resolve inter-vendor collaboration complexity, acknowledging the resulting extra overhead.
- Proposal 4 (Sec 3.2.1): For Direction B, parameter exchange over the air is necessary to resolve inter-vendor collaboration complexity, which will bring extra overhead.
- Proposal 5 (Sec 3.2.3): For Direction B, UEs in one cell with similar NW-side conditions and model structures can use a common encoder.
- Proposal 9 (Sec 3.2.5): For Direction B, there is no concern for NW’s or UE’s proprietary information disclosure since only model parameters are shared.
- Proposal 10 (Sec 3.3.1): For Direction C, 3GPP’s statistical channel model should be a starting point for reference model training, with field data added to better match real-world environments.
- Proposal 13 (Sec 3.4): Direction A and Direction B should be prioritized for further study over Direction C due to better performance and lower specification effort.
- Proposal 14 (Sec 4.1): For performance monitoring, prioritize NW-side monitoring based on ground-truth CSI reports and UE-side monitoring based on recovery CSI indication.
- Proposal 15 (Sec 4.2): For NW-side data collection, Rel-16 eType-II codebook and Rel-18 Doppler codebook can be used as a starting point for ground-truth CSI format.
- Proposal 17 (Sec 4.2): The exact supported values of codebook parameters should be studied to ensure high-resolution data reports for ground-truth CSI collection.