R1-2409449
discussion
AI/ML for CSI prediction
From Ericsson
Summary
Ericsson presents evaluation results for UE-sided AI/ML CSI prediction, concluding that no specification enhancements are needed to ensure consistency between training and inference regarding UE speed, deployment scenario, carrier frequency, NW antenna tilt, or TXRU mapping, as generalization performance degradation is negligible. The document proposes studying specification impacts on CSI-RS configuration to support data collection for training and monitoring, specifically requiring indications of association between observation and prediction window resources, and supports combining periodic and aperiodic CSI-RS for inference to improve performance with long periodicities.
Position
Ericsson argues inapplicability of the inconsistency issue identified for UE-sided spatial beam prediction to the UE-sided CSI prediction use case, which relies on temporal domain correlation rather than spatial beam sets. They conclude that no specification enhancement is required to ensure consistency between training and inference regarding UE speed, deployment scenario, carrier frequency, UE distribution, NW antenna tilt, or NW TXRU mapping, citing negligible performance degradation (0-4% for tilt, -1~4% for TXRU mapping) in generalization evaluations. They propose studying specification impacts on CSI-RS configuration to indicate the association between CSI-RS resources in the observation window and ground-truth labels in the prediction window for both training and monitoring data collection. They support specification impacts for inference to enable combined periodic and aperiodic (or semi-persistent) CSI-RS configurations to improve prediction performance with long periodicities like 20 ms. They oppose Type 2 based performance monitoring, citing large reporting overhead and significant NW complexity.
Key proposals
- Proposal 1 (Sec 2): Conclude that there is no need of specification enhancement to ensure consistency between model training and model inference for CSI prediction use case regarding UE speed, deployment scenario, carrier frequency, and UE distribution.
- Proposal 2 (Sec 3): Conclude that there is no need of specification enhancement to ensure consistency between model training and model inference for localized models for CSI prediction use case.
- Proposal 3 (Sec 4.1): Conclude that there is no need of specification enhancement to ensure consistency between model training and model inference for CSI prediction use case regarding NW antenna tilt value.
- Proposal 4 (Sec 4.2): Conclude that there is no need of specification enhancement to ensure consistency between model training and model inference for CSI prediction use case regarding NW TXRU mapping.
- Proposal 5 (Sec 5.1.1): For the CSI prediction use case, at least for training data collection, study potential specification impacts on CSI-RS configuration, including indication of the association between CSI-RS resources used for measurements in an observation window and CSI-RS resource(s) used for ground-truth labels in a prediction window.
- Proposal 6 (Sec 5.1.2): For the CSI prediction use case, at least for monitoring data collection, study potential specification impacts on CSI-RS configuration, including indication of the association between CSI-RS resources used for measurements in an observation window and CSI-RS resource(s) used for ground-truth labels in a prediction window.
- Proposal 7 (Sec 5.1.3): At least for inference, for UE-sided model based CSI prediction, support potential specification impacts including combined periodic CSI-RS and aperiodic (or semi-persistent) CSI-RS configuration for channel measurement to improve prediction performance over only periodic CSI-RS configuration with long periodicity (e.g., 20 ms).
- Proposal 8 (Sec 5.2.1): For performance monitoring for CSI prediction use case with UE side model, Type 2 based monitoring is not supported.