R1-2409782
discussion
Specification support for AI-enabled CSI prediction
From NVIDIA
Summary
NVIDIA presents a technical contribution on specification support for AI-enabled CSI prediction, highlighting four key proposals and two observations regarding inference location, training/inference consistency, and post-deployment monitoring. The document argues for evaluating gNB-side inference alongside UE-side models and emphasizes the need for digital twin-based training data to ensure model generalization in real-world environments.
Position
NVIDIA proposes that inference for one-sided AI/ML CSI prediction models be evaluated at both the gNB and UE sides to assess comparative gains. They argue that inconsistency between training and inference arises when using stochastic channel models, and therefore propose specifying solutions to ensure consistency, specifically by leveraging network digital twins with ray tracing to generate realistic training data. Furthermore, NVIDIA proposes studying post-deployment performance monitoring mechanisms, including three types of fallback operations (Type 1, 2, and 3), to detect performance degradation and non-compliance in the field. They suggest that legacy CSI-RS configuration and 'typeII-Doppler-r18' feedback mechanisms can serve as starting points for inference specifications.
Key proposals
- Proposal 1 (Sec 2): The inference of one-sided AI/ML model for CSI prediction can be performed at either gNB or UE, with companies encouraged to study gNB-side prediction to understand potential gains versus UE-side.
- Proposal 2 (Sec 3): Conclude that there is a need for consistency of training/inference in AI/ML-based CSI prediction.
- Proposal 3 (Sec 3): Specify solutions to ensure consistency of training/inference for AI/ML-based CSI prediction.
- Proposal 4 (Sec 4): RAN1 to study post-deployment performance monitoring mechanisms to detect performance degradation and non-compliance to guarantee satisfactory performance of AI/ML-based CSI prediction in the field.