R1-2410720
discussion
Summary#2 of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This 3GPP RAN1 technical document (Tdoc R1-2410720) presents a draft summary of AI/ML-based CSI compression studies, containing over 300 proposals from various companies covering temporal domain aspects, inter-vendor training collaboration, monitoring, and inference aspects. The document focuses on addressing UCI loss mitigation, separate vs. joint prediction/compression for Case 3, and standardized model structure specification.
Position
Qualcomm, as the document moderator, advocates for prioritizing Direction A Option 4-1 (dataset exchange) for inter-vendor collaboration while supporting Direction C for minimum performance assurance. They push for dataset sharing over parameter sharing due to lower specification impact, emphasize the necessity of NW-side target CSI sharing to address UE/NW data distribution mismatch, and strongly support signaling-based UCI loss mitigation. They oppose Direction B due to feasibility concerns with common/specific encoders and deprioritize Option 3a-1 without target CSI sharing.
Key proposals
- Proposal 21a (UCI loss): Conclude that under 10% UCI loss, Case 2 shows small to large performance drop, and signaling support to mitigate UCI loss can improve performance through NW-signaling to reset historical CSI information and NW-triggered CSI retransmission
- Proposal 22a (SPC vs JPC): For temporal domain aspects Case 3, take separate prediction and compression (SPC) as baseline and study its LCM aspects, with options for training data collection based on predicted vs. measured CSI
- Proposal 41a (LS on exchange method): Send LS to RAN2 for study of feasibility of parameter/dataset exchange using over-the-air and other signaling approaches for Direction A
- Proposal 45b (Direction C performance): Conclude that fully specified models in Direction C may have limited performance compared to Direction A, but can be improved if either side trains using field data
- Proposal 47b (model structure): Adopt Transformer as backbone structure for spatial-frequency domain input in Case 0, with specific adaptation options for Cases 2 and 3
- Proposal 49a (way forward): Support Direction A Option 4-1 at minimum, further study Option 3a-1 and Direction C, with feasibility contingent on scalable model structure specification
- Proposal 51b (NW-side data collection): Confirm necessity of ground-truth reporting using Rel-16/18 eType2 codebook with potential enhancements for NW-side data collection
- Proposal 52b (UE-side data collection): Confirm necessity of UE-side data collection with NW configuration and RS configuration for target CSI measurement
- Proposal 61a (monitoring study): Confirm necessity of NW-side monitoring based on target CSI reported via eType2 codebook and at least one additional monitoring option for UEs with limited capability
- Proposal 62a (monitoring evaluation): Evaluate monitoring decision performance using false alarm and missed detection probability as KPIs for accuracy assessment
- Proposal 1 (Huawei, Sec Inter-vendor): Study performance target including end-to-end, encoder-only, and decoder-only performance metrics with testing dataset from NW
- Proposal 1 (Samsung, Sec Temporal): Consider past CSI information generated by UE-part/network-part or SD/FD basis vectors for temporal domain Case 2
- Proposal 8 (Apple, Sec Monitoring): Further study RLF/BFD like mechanism for UE initiated report in CSI compression monitoring
- Proposal 20 (QC, Sec Inference): Conclude that CQI calculation option 2a using UE-side CSI reconstruction model can be employed considering higher processing requirements
- Proposal 24 (QC, Sec Complexity): Study complexity reduction through angle-delay domain representation considering SGCS performance, model complexity, and overhead for SD/FD basis signaling