R1-2500057
discussion
AI/ML for CSI prediction
From Ericsson
Ericsson's prior position on
9.1.3
at
RAN1#119
· AI-synthesized, paraphrased
verify sources →
Argues the inapplicability of the inconsistency issue identified for UE-sided spatial beam prediction to CSI prediction, concluding no specification enhancements are needed for UE speed, deployment scenario, carrier frequency, antenna tilt, or TXRU mapping due to negligible performance degradation. Proposes studying specification impacts on CSI-RS configuration to indicate association between observation and prediction window resources for data collection. Opposes Type 2 based performance monitoring, citing large reporting overhead and significant network complexity.
Summary
Ericsson presents proposals for the normative phase of UE-sided CSI prediction in Rel-19, focusing on functionality-based LCM, data collection, and performance monitoring. The document contains 14 proposals and 18 observations, arguing that no specification enhancements are needed to ensure consistency between training and inference due to negligible performance degradation from NW-side conditions like antenna tilt and TXRU mapping.
Position
Ericsson proposes reusing the functionality-based LCM framework from UE-sided beam management use cases for CSI prediction, specifically leveraging legacy CSI-RS configurations with enhancements for training data association. They require the use of 'typeII-Doppler-r18' codebook format for both predicted CSI and ground-truth labels to ensure testable performance metrics, and they oppose supporting Type 2 performance monitoring due to high overhead and NW complexity. Ericsson argues inapplicability of consistency specification enhancements for training/inference, presenting technical evidence that NW antenna tilt and TXRU mapping variations cause negligible degradation (0-4% for tilt, -1~4% for TXRU). They propose supporting combined periodic and aperiodic CSI-RS configurations to address performance issues with long 20 ms periodicities, and they restrict the intermediate KPI to SGCS only, excluding NMSE.
Key proposals
- Proposal 1 (Sec 2.2.2): Support indication of the association between CSI-RS resources used for measurements in an observation window (model input) and CSI-RS resource(s) used for measurements in a prediction window (ground-truth label) for training data collection.
- Proposal 2 (Sec 2.2.2): Support configuration of a pair of CSI-RS resource sets for training, where the first set has K resources (K>1) for model input and the second has N4 resources (N4>=1) for ground-truth labels, reusing Rel-18 'typeII-Doppler-r18' candidate values.
- Proposal 3 (Sec 2.2.2): Support Periodic (P), Semi-persistent (SP), and Aperiodic (AP) CSI-RS for the two CSI-RS resource sets used in training data collection.
- Proposal 4 (Sec 2.2.3): Support configuration of a pair of CSI-RS resource sets for monitoring data collection, allowing the second set (prediction window) to have a smaller size (n <= N4) to save overhead compared to training.
- Proposal 5 (Sec 2.3.1): Type 2 based performance monitoring (where NW calculates metrics) is not supported for UE-sided CSI prediction.
- Proposal 6 (Sec 2.3.2): Support using 'typeII-Doppler-r18' as the ground-truth format for calculating the intermediate KPI in performance monitoring.
- Proposal 7 (Sec 2.3.2): Only support SGCS (Squared Generalized Cosine Similarity) as the intermediate KPI, excluding NMSE.
- Proposal 9 (Sec 2.3.2): Define the intermediate KPI per monitoring data sample as the SGCS per MIMO layer per prediction instance between predicted CSI and ground-truth label.
- Proposal 10 (Sec 2.3.2): Support defining performance metrics as either the intermediate KPI per sample or statistics of the intermediate KPI over multiple samples in a monitoring window.
- Proposal 12 (Sec 2.3.2): Support UE reporting monitoring output as an intermediate KPI range indicator or an intermediate KPI statistics range indicator based on configured thresholds.
- Proposal 13 (Sec 2.4): Support combined periodic CSI-RS and aperiodic (or semi-persistent) CSI-RS configuration for channel measurement to improve prediction performance with long periodicities (e.g., 20 ms).
- Proposal 14 (Sec 3.3.2): Conclude that there is no need for specification enhancement to ensure consistency between model training and model inference for the CSI prediction use case.