R1-2410722
discussion
Summary#4 of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This 3GPP RAN1 technical document (Tdoc R1-2410722) from Qualcomm presents a comprehensive draft summary on AI/ML for NR air interface CSI compression, containing approximately 120+ proposals across four major sections covering temporal domain aspects, localized models, inter-vendor training collaboration, and monitoring approaches. The document consolidates company proposals and technical discussions from RAN1 meetings #116-119 to advance the standardization of two-sided AI/ML models for CSI compression in Release 19.
Position
Qualcomm, as the moderator, advocates FOR prioritizing Case 2 and Case 3 temporal domain aspects with separate prediction and compression as baseline, supporting Direction A Option 4-1 (dataset exchange) over Option 3a-1 (parameter exchange) due to lower specification impact, and emphasizing the necessity of NW-side target CSI sharing to address data distribution mismatch. They push AGAINST deprioritizing Direction B and advocate for comprehensive monitoring solutions including both NW-side and UE-side approaches to ensure robust performance in field deployment.
Key proposals
- Proposal 21b (Sec 1): 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 at both UE and NW, and NW-triggered CSI retransmission
- Proposal 22b (Sec 1): For temporal domain aspects Case 3, take separate prediction and compression (SPC) as baseline with Option 1 using predicted CSI and Option 2 using measured CSI for training data collection and monitoring targets
- Proposal 44a (Sec 2): Deprioritize direction B in Rel-19 study item considering there is no consensus on feasibility of developing UE-common or UE-specific encoder
- Proposal 41a (Sec 4): For direction A, send LS to RAN2 for study of feasibility of parameter/dataset exchange using over-the-air and other signaling approaches
- Proposal 42b (Sec 4): Conclude that performance impact due to NW/UE data distribution mismatch can be addressed by timely NW-side data collection and UE-side encoder training using NW-side target CSI or dataset shared by NW-side
- Proposal 43b (Sec 4): Regarding target CSI sharing for option 3a-1, NW-side target CSI sharing is needed for UE-side encoder performance validation/testing, with FFS on need for training
- Proposal 45b (Sec 4): 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, yet still worse than achievable performance of Direction A
- Proposal 49a (Sec 4): For Direction A and C, at least support Direction A Option 4-1, with feasibility of Option 3a-1 and Direction C contingent on scalable model structure specification
- Proposal 47b (Sec 4): For studying standardized model structure, adopt Transformer as backbone for Case 0 spatial-frequency domain input, with adaptation options for Case 2 including Conv-LSTM, LSTM, and latent adaptation
- Proposal 51b (Sec 5): Confirm necessity of ground-truth reporting for NW-side data collection using codebook-based Rel-16 eType2 or Rel-18 eType2, with FFS on enhancement needs
- Proposal 52b (Sec 5): Confirm necessity of UE-side data collection from UE to UE-side server for training, considering NW configuration or UE request with RS configuration for target CSI measurement
- Proposal 61a (Sec 6): Confirm necessity of NW-side monitoring based on target CSI reported via legacy eT2 codebook or eT2-like high-resolution codebook, contingent on UE capability
- Proposal 62a (Sec 6): Evaluate monitoring decision performance using false alarm probability and missed detection probability as KPIs, with KPIActual obtained based on N samples large enough to average out channel variations
- Proposal 64b (Sec 6): For NW-first training in inter-vendor collaboration options 3, 4, and 5, conclude that root cause identification can be achieved by ground-truth CSI reporting
- Proposal 81a (Sec 7): Study performance-complexity trade-off by comparing different AI/ML models and precoder representations in spatial-frequency vs angle-delay domains, considering performance, FLOP complexity, and latency/power consumption