R1-2410042
discussion
On AI/ML-based CSI prediction
From InterDigital
Summary
InterDigital presents evaluation results on AI/ML-based CSI prediction for NR air interface, demonstrating that UE-sided models can generalize well across different network conditions without requiring complex associated ID mechanisms. The document contains 1 formal proposal and 3 observations focused on simplifying CSI prediction implementation.
Position
InterDigital advocates FOR simplifying CSI prediction by eliminating the need for associated ID mechanisms, arguing that UE-sided models can generalize well across different network conditions without complex consistency checking. They are pushing AGAINST the adoption of associated ID-based approaches for CSI prediction, positioning this as unnecessarily complex when model performance monitoring and out-of-distribution detection can achieve the same goals more efficiently.
Key proposals
- Observation 1 (Sec 2.2): Case 2 generalization performance of the UE-sided CSI prediction over various antenna down tilts shows negligible SGCS change relative to Case 1
- Observation 2 (Sec 2.3): For the simulated scenario (one local region and a generalized region generated separately from the local) the localized AI/ML CSI prediction model provides a minor increase of the SGCS gain over benchmark #1, as compared to the generalized model
- Observation 3 (Sec 3.1): At least for UE-sided CSI prediction models, the associated ID based approach for determining the consistency of training/inference (for NW-side additional conditions) may be of high complexity and is not essential
- Proposal 1 (Sec 3.1): For CSI prediction, the associated ID is not needed to ensure training/inference consistency regarding NW-side additional conditions