R1-2409627
discussion
Discussion on AIML for CSI compression
From Spreadtrum
Spreadtrum's prior position on
9.1.4.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Strongly advocates for AI-based CSI Spatial-Temporal-Frequency compression showing superior performance gains, while opposing option 5a for inter-vendor collaboration due to increased complexity.
Summary
Spreadtrum presents evaluation results for AI-based CSI Spatial-Temporal-Frequency (S-T-F) compression, demonstrating superior SGCS and UPT performance over Rel-16 and Rel-18 baselines. The document contains 8 proposals and 7 observations addressing inter-vendor training collaboration directions, CQI determination, historical CSI misalignment handling, and performance monitoring options.
Position
Spreadtrum supports extending CSI compression to Spatial-Temporal-Frequency (S-T-F) domains, presenting technical evidence that S-T-F compression yields higher SGCS and UPT gains than S-F compression or Rel-16/18 codebooks. They propose using SGCS as the primary performance metric for inter-vendor training collaboration in Direction A and suggest mitigating data mismatch issues by having UEs report UE-side data to the network. They argue that Direction B suffers from unaddressable overhead concerns due to timeliness requirements for on-device inference. For CQI determination, they support Option 1b, calculating CQI based on target CSI with realistic channel measurement and potential adjustment to account for compression errors. They recommend using 3GPP statistical channel models for reference model training in Direction C to avoid field data collection complexities. Finally, they propose NW-triggered signaling for reporting historical CSI when UCI drops and support specific monitoring options based on ground-truth CSI and NW-indicated recovery CSI.
Key proposals
- Proposal 1 (Sec 3.2 Direction A): For additional information in Direction A (Options 3a-1 and 4-1), the performance target including performance metric should at least consider SGCS, with application to encoder/decoder pairs depending on UE training method.
- Proposal 2 (Sec 3.2 Direction A): To alleviate overhead concerns in Direction A, it can be considered to transfer a small amount of data each time rather than bulk transfer.
- Proposal 3 (Sec 3.2 Direction A): If NW-side data does not contain UE-side data, performance may deteriorate; UE can report UE-side data to alleviate performance degradation caused by data mismatch.
- Proposal 4 (Sec 3.2 Direction B): Direction B has an overhead concern that cannot be addressed due to the requirement for immediate inference after parameter transfer.
- Proposal 5 (Sec 3.2 Direction C): For Direction A, 3GPP’s statistical channel model should be used for reference model(s) training instead of field data.
- Proposal 6 (Sec 3.3 CQI determination): Support Option 1b for CQI determination in inference, where CQI is calculated based on target CSI with realistic channel measurement and potential adjustment.
- Proposal 7 (Sec 3.3 Handling of misalignment): Consider reporting historical CSI information via NW-triggered signaling when UCI is missing or dropping to handle misalignment.
- Proposal 8 (Sec 3.3 Monitoring): Support NW-side monitoring based on ground-truth CSI reported by UE and UE-side monitoring based on recovery CSI indicated by NW.
- Observation 1 (Sec 2.2): Both AI-based CSI S-T-F compression case 2 and AI-based CSI S-F compression achieve better SGCS performance than Rel-16 eType II codebook.
- Observation 2 (Sec 2.2): AI-based CSI S-T-F compression case 2 achieves better UPT performance than Rel-16 eType II codebook and AI-based CSI S-F compression.
- Observation 3 (Sec 2.3): AI-based CSI S-T-F compression case 3 achieves better SGCS performance than Rel-18 doppler eType II codebook.
- Observation 5 (Sec 3.2 Direction B): Whether to use common or multiple encoders depends on UE capabilities; common encoders for similar capabilities, multiple for different capabilities.
- Observation 7 (Sec 3.2 Direction B): For option 3b-1, there is no proprietary information concern of disclosing encoder parameters from NW side to UE side.