R1-2409752
discussion
Discussion on AI/ML for CSI Compression
Summary
Tejas Networks Limited presents 11 proposals and 5 observations regarding AI/ML-based CSI compression for NR Release 19, focusing on evaluation methodologies, inter-vendor collaboration, and monitoring frameworks. The document proposes specific assumptions for CSI-Net model architecture, defines scenarios for handling UCI loss, and recommends prioritizing specific inter-vendor collaboration sub-options (3a/5a-1 and 4-1) based on performance and proprietary protection. It also addresses data collection strategies, suggesting precoding matrix reporting over channel matrix, and outlines monitoring approaches using legacy codebooks.
Position
Tejas Networks proposes specific architectural assumptions for CSI-Net, including a convolutional encoder with a 33 kernel size and RefineNet-based decoder, to standardize evaluation metrics for temporal domain CSI compression. They require the use of Model IDs to identify models robust to UE/NW data distribution mismatches, specifically noting robustness to antenna tilt angles. Regarding inter-vendor collaboration, they prioritize sub-options 3a/5a-1 and 4-1 for down-selection, citing their balance of performance, proprietary protection, and manageable overhead. For data collection, they prefer UE reporting of the precoding matrix over the channel matrix as ground truth to reduce complexity. Finally, they propose leveraging legacy eT2 codebooks for NW-side monitoring and UE-side reconstruction outputs for UE-side monitoring to ensure accuracy.
Key proposals
- Proposal 1 (Sec 2.1): Defines specific assumptions for CSI-Net encoder/decoder architecture for temporal domain CSI compression evaluation, including convolutional layers with 33 kernel size and RefineNet units for reconstruction.
- Proposal 2 (Sec 2.1): Proposes using Model ID to represent models that are robust to data distribution mismatch between UE and Network sides.
- Proposal 3 (Sec 2.2): Proposes considering a 10% missing rate for individual CSI report occasions to model UCI loss in Case 2.
- Proposal 4 (Sec 2.2): Outlines evaluation scenarios for UCI loss: Scenario A (No Loss) as baseline, Scenario B (NW mitigation only) to assess NW effectiveness, and Scenario C (Coordinated mitigation) as the preferred approach if overhead is manageable.
- Proposal 5 (Sec 2.3): Clarifies that the upper bound for temporal domain aspects Case 3/4 should be calculated based on ideal CSI prediction without CSI compression.
- Proposal 6 (Sec 2.3): Proposes studying the impact on LCM aspects (data collection, training, monitoring, model control) of separate vs. joint prediction and compression.
- Proposal 7 (Sec 2.4): Proposes that for N different local regions with N localized models, average performance should be considered over the N regions.
- Proposal 8 (Sec 2.5): Prioritizes sub-options 3a/5a-1 and 4-1 for further study in CSI compression, balancing performance, proprietary protection, and overhead.
- Proposal 9 (Sec 2.6): Proposes prioritizing UE reporting of the precoding matrix rather than the channel matrix as ground truth for NW-side data collection.
- Proposal 10 (Sec 2.7): Proposes using target CSI reported by UE via legacy eT2 codebook or eT2-like high-resolution codebook for NW-side monitoring accuracy.
- Proposal 11 (Sec 2.7): Proposes using Case 2-1 (based on UE-side CSI reconstruction model output) for UE-side monitoring accuracy.