R1-2500566
discussion
Discussions on CSI prediction
From LG Electronics
LG Electronics's prior position on
9.1.3
at
RAN1#119
· AI-synthesized, paraphrased
verify sources →
Concludes that gNB antenna tilt angle variations have negligible impact on CSI prediction performance, arguing no additional specification support is needed for tilt angle consistency. Pushes for prioritizing Type 1 and Type 3 performance monitoring over Type 2 due to lower reporting overhead and higher accuracy, explicitly opposing Type 2 monitoring due to payload and quantization issues. As document moderator, facilitates consensus that tilt angle impacts are negligible, while acknowledging a lack of industry consensus on TXRU mapping effects, suggesting continued study rather than immediate standardization for the latter.
Summary
LG Electronics presents a contribution on AI/ML-based CSI prediction, focusing on data collection frameworks, inference reporting mechanisms, performance monitoring types, and consistency between training and inference. The document contains 9 proposals and 2 observations, arguing for the reuse of legacy Rel-18 frameworks to minimize specification overhead while addressing specific UE implementation concerns regarding buffering and CPU processing.
Position
LG Electronics proposes reusing the Rel-18 'typeIIDoppler-r18' codebook for AI/ML-based CSI prediction inference reporting to avoid the effort of defining new feedback mechanisms. They propose reusing the existing CSI Processing Unit (CPU) mechanism as a starting point for handling the additional computational load of inference algorithms. Regarding performance monitoring, LG Electronics prefers deprioritizing Type 2 monitoring due to large payload overhead, but proposes reusing legacy codebooks for ground truth CSI if Type 2 is supported. They highlight a critical UE implementation issue, proposing that a specific time limit or buffering window be specified for inference results linked to monitoring reports to prevent indefinite buffering. Finally, based on simulation results showing marginal generalization loss, they propose that no additional specification support is required for consistency regarding gNB antenna tilt angles.
Key proposals
- Proposal 1 (Data collection): Proposes using the legacy CSI framework and BM agenda framework as a starting point for data collection discussions.
- Proposal 2 (Data collection): Proposes that RAN1 discuss data type/format for training, while leaving dataset delivery/transfer mechanisms to RAN2.
- Proposal 3 (Inference and related CSI reporting): Proposes reusing the Rel-18 codebook ('typeIIDoppler-r18') for inference reporting in AI/ML-based CSI prediction.
- Proposal 4 (FFS on potential refinement of codebook parameter): Proposes reusing the existing CSI Processing Unit (CPU) mechanism as a starting point for AI/ML-based CSI prediction.
- Proposal 5 (Performance monitoring): Proposes deprioritizing Type 2 performance monitoring due to high reporting overhead.
- Proposal 6 (Performance monitoring): Proposes that if Type 2 performance monitoring is supported, the legacy codebook should be reused to carry ground truth CSI.
- Proposal 7 (Performance monitoring): Proposes discussing how to calculate performance monitoring metrics/outputs when N4>1 for Type 1 and Type 3 monitoring.
- Proposal 8 (Performance monitoring): Proposes specifying a time limit or buffering window for inference results when linked to a monitoring report to prevent excessive UE buffering.
- Proposal 9 (Consistency between training and inference): Proposes concluding that specification support is not needed to ensure consistency regarding gNB antenna tilt angles for UE-sided models.
- Observation 1 (Consistency between training and inference): Observes that generalization performance is achieved for various gNB antenna tilt angles in Generalization Case 2.
- Observation 2 (Consistency between training and inference): Observes marginal loss in generalization performance for various TXRU mappings in Generalization Case 2.