R1-2500467
discussion
On specification for AI/ML-based CSI prediction
From OPPO
Summary
OPPO presents 13 proposals and 2 observations regarding AI/ML-based CSI prediction, arguing that UE-sided models face no consistency issues and prioritizing UE-side data collection over NW-side model transfer. The document proposes reusing Rel-18 CSI frameworks for data collection, specifying raw channel reporting with float32 format, and defining performance monitoring metrics based on average NMSE for Type 1 and Type 3 monitoring.
Position
OPPO argues that there is no consistency issue for UE-sided CSI prediction, rendering the associated ID framework unnecessary. They propose prioritizing UE-side data collection because it avoids the complexity of model transfer and high uplink overhead associated with raw channel reporting to the network. For NW-side collection, they support reusing Rel-18 MIMO CSI frameworks and MDT-based reporting, while preferring raw channel reporting with float32 format over CSI eigenvectors. Regarding monitoring, OPPO opposes Type 2 monitoring due to overhead and proposes using average NMSE for Type 3, with configurable averaging over K monitoring instances. They also raise an FFS on whether monitoring should be UE-level or cell-level.
Key proposals
- Proposal 1 (Consistency Discussion): There is no consistency issue for UE-sided CSI prediction, as performance loss from NW-side conditions is slight.
- Proposal 2 (Data Collection - NW-side): Support for NW-side data collection and model transfer to UE depends on the outcome of agenda item 9.1.4.2.
- Proposal 3 (Data Collection - NW-side): Support MDT-based data reporting for NW-side data collection and model training.
- Proposal 4 (Data Collection - NW-side): Data reporting content should include both input (CSI measurements in observation window) and label (CSI measurements in prediction window).
- Proposal 5 (Data Collection - NW-side): The observation window size, prediction window size, and time offset from last measurement to first prediction should be configured in CSI measurement resource configurations.
- Proposal 6 (Data Collection - NW-side): Support reusing current CSI framework and CSI-RS configurations from Rel-18 MIMO; no enhancements supported in Rel-19 WI phase.
- Proposal 7 (Data Collection - NW-side): Study and specify raw channel reporting instead of CSI eigenvector reporting, using float32 as the basic data format.
- Proposal 8 (Data Collection - UE-side): Support studying and specifying UE-side data collection for UE-side CSI prediction with higher priority than NW-side approaches.
- Proposal 9 (Data Collection - UE-side): Support reusing current CSI framework and CSI-RS configurations from Rel-18 MIMO for UE-side data collection.
- Proposal 10 (Model Monitoring): Prioritize Type 1 and Type 3 CSI prediction performance monitoring; Type 2 is not supported due to high overhead.
- Proposal 11 (Model Monitoring): For Type 3 monitoring, use average NMSE between predicted raw channels and ground-truth as the performance metric.
- Proposal 12 (Model Monitoring): Monitoring output/metric can be reported based on one time or K times of monitoring, with K>1 configurable by the network.
- Proposal 13 (Model Monitoring): Assess whether model monitoring should be handled at the UE level or the cell level.