R1-2410899
discussion
Summary #3 of CSI prediction
From LG Electronics
Summary
This 3GPP RAN1 document (R1-2410899) from LG Electronics summarizes evaluation results for AI/ML based CSI prediction, focusing on consistency between training and inference regarding network-side conditions like antenna tilt angles and TXRU mapping. The document contains numerous observations and conclusions across multiple sections evaluating generalization performance.
Position
LG Electronics, as the document moderator, presents a comprehensive summary showing mixed industry consensus on CSI prediction consistency issues. They advocate for concluding that antenna tilt angles have negligible impact and don't require additional signaling, while acknowledging that TXRU mapping shows more varied results with no clear industry consensus, suggesting the need for continued study rather than immediate standardization.
Key proposals
- Conclusion (Sec 4.4): For CSI prediction using UE-sided model, for consistency between training and inference, tilt angle at NW-side is not considered as NW-side additional condition
- Conclusion (Sec 6.1.1): For CSI prediction using UE-sided model, for consistency between training and inference, tilt angle at NW-side has negligible (0%~-4%) degradation for generalization case 2
- Conclusion (Sec 6.1.2): For CSI prediction using UE-sided model, RAN1 has no consensus on whether tilt angle or TXRU mapping results in non-negligible performance degradation on model generalization
- Conclusion (Sec 6.2.1): For CSI prediction using UE-sided model, for consistency between training and inference, tilt angle at NW-side is not considered as NW-side additional conditions
- Conclusion (Sec 6.2.2): For CSI prediction using UE-sided model, RAN1 has no consensus on specifying consistency of training/inference due to NW-side additional conditions
- Conclusion (Sec 7.1): For CSI prediction using UE-sided model, for consistency between training and inference, tilt angle at NW-side has negligible performance degradation on model generalization
- Conclusion (Sec 7): For CSI prediction using UE-sided model, RAN1 has no consensus on specifying consistency of training/inference due to NW-side additional conditions
- Observation (Sec 4.2): For generalization case 2, 6 sources observe 0~-2.6% performance degradation in terms of SGCS, 1 source observes -26%~-28% performance degradation
- Observation (Sec 4.5): 7 sources observe negligible performance degradation for TXRU mapping generalization case 2/3, while 2 sources observe non-negligible performance degradation
- Observation (Sec 6.1.1): In terms of UPT, 2 sources observe 0% ~-2% performance degradation in terms of mean UPT and 0% ~-4% performance degradation in terms of 5% UE UPT
- Observation (Sec 6.1.2): For N4=1, 4 sources observe -0.1~-1.7% performance degradation, 1 source observes -0.6%~-7% performance degradation, 2 sources observe -13%~-41% performance degradation
- Observation (Sec 6.2.1): 2 sources observe 0% ~-2% performance degradation in terms of mean UPT and 0% ~-4% performance degradation in terms of 5% UE UPT for generalization case 2
- Observation (Sec 7.1): 1 source observes 18dB~29dB performance degradation in terms of NMSE for TXRU mapping generalization case 2