R1-2410817
discussion
Summary #1 of CSI prediction
From LG Electronics
Summary
This 3GPP RAN1 document from LG Electronics summarizes CSI prediction evaluation results and consistency issues between training and inference for UE-sided AI/ML models. The document contains numerous observations from multiple companies regarding generalization performance across different network conditions, with the majority concluding that tilt angle and TXRU mapping have negligible impact on CSI prediction performance.
Position
LG Electronics, as the moderator, takes a consensus-building approach advocating for concluding that tilt angle and TXRU mapping have negligible impact on CSI prediction generalization performance based on majority company results. They push for drawing high-level conclusions rather than requiring additional simulations, and support moving forward with the study completion while identifying that specification enhancements may not be needed for these specific network-side additional conditions.
Key proposals
- Observation (Sec 8.1): For CSI prediction using UE-sided model, for consistency between training and inference regarding tilt angle at NW-side, 6 sources observe 0~-2.6% performance degradation in generalization case 2
- Observation (Sec 8.1): For CSI prediction using UE-sided model, for consistency between training and inference regarding TXRU mapping at NW-side, 7 sources observe negligible performance degradation while 2 sources observe non-negligible performance degradation
- Agreement (Sec 9.1): For consistency between training and inference, study to identify which potential NW-side additional conditions may impact on UE assumption for CSI prediction using UE-sided model
- Agreement (Sec 9.1): For generalization evaluation to identify potential NW-side additional conditions, consider various tilt angle (e.g., 102 degrees, 110 degrees) and various TXRU mapping configurations
- Observation (Sec 9.2): For generalization Case 2, generalized performance may be achieved for some certain combinations of UE speed but not for others, with degradation ranging from -72.37% to -2.7%
- Observation (Sec 9.2): For generalization Case 3, generalized performance of the AI/ML model can be achieved in general (0%~-3.8% loss) if training dataset includes multiple UE speeds
- Observation (Sec 9.2): For AI/ML based CSI prediction over non-AI/ML based CSI prediction, 0%~7.8% gain in mean UPT and 3.8%~20.7% gain in 5% UE UPT is observed
- Agreement (Sec 9.2): For computational complexity of both AI/ML and non-AI/ML based CSI prediction, report the number of FLOPs assuming whole bandwidth and one prediction sample
- Agreement (Sec 9.2): From RAN1 perspective, study of CSI prediction has been completed and performance improvement is observed with increased complexity
- Agreement (Sec 9.3): For the boundary between Type 3 and Type 1 performance monitoring, the difference is whether UE reports performance metric or performance monitoring output to NW
- Agreement (Sec 9.4): For AI/ML based CSI prediction, adopt UE speed 30km/h and 60km/h as baseline, with observation window 5/5ms and 10/5ms
- Agreement (Sec 9.4): For CSI report, adopt N4 value 1 and 4 as baseline for evaluation purpose
- Agreement (Sec 9.4): For evaluation of UE-sided model based CSI prediction, UE distribution of 80% indoor, 20% outdoor can be optionally simulated