R1-2409879
discussion
Discussion on AI/ML model based CSI prediction
From Xiaomi
Summary
Xiaomi proposes methods to ensure consistency between AI/ML model training and inference for UE-side CSI prediction, presenting simulation results showing that models trained with consistent TXRU mapping configurations achieve significantly better performance. The document contains 2 proposals and 1 observation focused on network-side condition indication and leveraging beam prediction methods.
Position
Xiaomi advocates FOR explicit network indication of additional conditions (like TXRU mapping) to UEs to maintain AI/ML model performance consistency, and FOR leveraging beam prediction consistency methods as a foundation for CSI prediction. They push FOR high-priority study of network-side condition information provision and performance monitoring options, demonstrating through simulation that mixed training datasets significantly degrade performance compared to consistent configurations.
Key proposals
- Proposal 1 (Sec Discussion): Suggest NW's additional conditions indicated to UE to ensure optimal AI/ML model performance for CSI prediction
- Proposal 2 (Sec Discussion): For UE-side AI/ML model-based CSI prediction, the methods to ensure consistency of training and inference, i.e., associated ID or performance monitoring, for UE-side AI/ML-based beam prediction could be studied as a starting point
- Observation 1 (Sec Discussion): The trained model according to training dataset generated by using one TXRU mapping configuration could achieve obviously better performance than generation model trained by using two different TXRU mapping configurations