R1-2500276
discussion
Discussion on AI/ML for CSI prediction
From CMCC
CMCC's prior position on
9.1.3
at
RAN1#119
· AI-synthesized, paraphrased
verify sources →
Proposes studying two options for ensuring consistency: one based on associated ID and another based on performance monitoring, suggesting a combined solution where the ID preliminarily guarantees consistency without exposing proprietary information. Requires that if associated ID is supported, the UE assumes consistency of network-side additional conditions with the same ID at least within a cell, configured within the CSI framework. Proposes using intermediate KPIs for Type 3 monitoring and reusing data collection mechanisms from AI/ML beam management while adapting Rel-18 CSI parameters.
Summary
CMCC discusses specification impacts for AI/ML-based CSI prediction in Rel-19, focusing on training/inference consistency, performance monitoring, data collection, and inference parameters. The document presents five proposals, arguing against specifying NW-side condition consistency unless significant degradation is proven, favoring intermediate KPIs for monitoring, and suggesting reuse of Rel-18 mechanisms with modifications.
Position
CMCC prefers not to specify consistency of training/inference regarding NW-side additional conditions (such as tilt angle and TXRU mapping) unless sufficient simulation results demonstrate non-negligible degradation on model generalization performance, citing current observations of negligible impact. For performance monitoring, CMCC proposes using intermediate KPIs (e.g., SGCS, NMSE) as the starting point for Type 3 monitoring, arguing that eventual KPIs are inappropriate due to dependency on external factors like scheduling. For Type 1 monitoring, CMCC presents two alternatives for performance output: calculation/comparison outcomes based on thresholds or recommended LCM decisions reported to the NW. CMCC supports reusing the data collection mechanism from AI/ML-based beam management for CSI prediction to minimize specification overhead. Finally, CMCC proposes reusing Rel-18 CSI parameters (e.g., measurement and reporting windows) with necessary modifications to adapt to AI/ML-enabled CSI prediction, rather than defining entirely new parameters.
Key proposals
- Proposal 1 (Sec 2.1): Consistency of training/inference due to NW-side additional conditions is not specified, unless sufficient simulation results can show non-negligible degradation on model generalization performance regarding NW-side additional conditions.
- Proposal 2 (Sec 2.2): Regarding the definition of performance metric in Type 3 performance monitoring, intermediate KPI can be considered as starting point.
- Proposal 3 (Sec 2.2): For the definition of performance metric in Type 1 performance monitoring, there might be two alternatives: Alt1: Calculation or comparison outcome based on the performance metric; Alt2: Recommended LCM decision to NW.
- Proposal 4 (Sec 2.3): For data collection of AI/ML based CSI prediction, data collection mechanism of AI/ML based beam management can be reused as much as possible, if necessary.
- Proposal 5 (Sec 2.4): For CSI prediction, the CSI related parameters in Rel-18 can be reused with modification to adapt AI/ML-enabled CSI prediction, e.g. the parameters about measurement window, reporting window.
- Observation (Sec 2.1): Simulation results indicate that tilt angle and TXRU mapping at NW-side have negligible degradation on model generalization performance for most sources, supporting the preference not to specify consistency.
- Observation (Sec 2.2): Eventual KPIs like throughput are deemed inappropriate for AI functionality monitoring as they depend on external factors like scheduling and mobility, whereas intermediate KPIs reflect model performance directly.
- Observation (Sec 2.2): Type 2 monitoring is noted as a candidate for further discussion because it reduces UE calculation overhead by having the NW calculate metrics from reported predicted and ground-truth CSI.
- Observation (Sec 2.4): Legacy feedback mechanism using codebook type 'typeII-Doppler-r18' is identified as a starting point, but parameters like measurement windows may need revisiting to maximize AI/ML gains.
- Observation (Sec 2.1): Specific simulation data shows mixed results for TXRU mapping generalization, with some vendors observing up to -41% SGCS degradation, though CMCC argues the overall trend supports non-specification of consistency.