R1-2409987
discussion
AI/ML for CSI Prediction
From Nokia
Summary
Nokia presents simulation results demonstrating robustness of their AI/ML CSI predictor against training/inference mismatches and proposes leveraging Release 18 non-AI/ML CSI prediction framework for AI/ML-based CSI prediction. The document contains 5 observations and 1 proposal across sections on generalization performance and normative phase preparation.
Position
Nokia advocates FOR reusing existing Release 18 non-AI/ML CSI prediction framework and UE-sided beam management agreements to minimize standardization effort for AI/ML CSI prediction. They demonstrate that their AI/ML architecture design (operating per tap/port without spatial domain correlations) provides robust generalization performance across different antenna configurations and downtilt angles, positioning against the need for complex additional network-side conditions or associated IDs for consistency.
Key proposals
- Observation 1 (Sec 2.1): For a mismatch in downtilt configuration of training and inference, there was no performance degradation observed for the AI/ML CSI predictor
- Observation 2 (Sec 2.1): The SGCS generalization results for the AI/ML CSI predictor trained on a dataset with antenna-port layout (8,8,2,1,1,2,8) and doing inference on a dataset with antenna-port layout (8,8,2,1,1,2,4) has shown no significant performance degradation
- Observation 3 (Sec 2.1): Different AI/ML architecture designs may lead to different generalization results. Our AI/ML design does not rely on the correlations in the spatial domain. Hence, its performance is not degrading
- Observation 4 (Sec 2.2): It has been agreed to reuse agreements from the UE sided BM sub use case as far as possible. An analysis of the agreements so far indicates that reusing the specified framework for non AI/ML channel prediction from Release 18 covers already most of the BM agreements
- Observation 5 (Sec 2.2): The need for associated IDs as agreed for the BM use case is an ongoing discussion and the performance monitoring details will be part of the WI phase
- Proposal 1 (Sec 2.2): Use Rel-18 non-AI/ML CSI prediction as a starting point for configuration of AI/ML-based CSI prediction. Reuse applicable agreements from UE sided beam management when possible