R1-2601758
LS in
LS on RAN4 down-selected AI/ML use cases
From Samsung
Summary
This document is a Liaison Statement (LS) from RAN WG4 to RAN, informing RAN of RAN4's down-selection and prioritization of ten AI/ML use cases for the Rel-20 6G study item. RAN4 requests RAN to consider this prioritization, which sequences work on AI-based non-linearity compensation, SRS power imbalance compensation, and five AI-RRM sub-cases based on meeting bandwidth.
Position
Samsung (as source RAN WG4, contact person He (Jackson) Wang) presents RAN4's agreed prioritization for Rel-20 6G AI/ML use cases. RAN4 prioritizes studying AI-based DPoD at NW first, explicitly deferring the study of AI-based DPD at UE until the DPoD study provides a performance KPI benchmark for gain, complexity, and power efficiency comparison. RAN4 agrees to start AI-RRM studies with Sub-Case 1 (FR2-1 L3 spatial domain beam-level prediction for Tx, intra-cell) and Sub-Case 2 (FR1 L3 frequency domain cell-level prediction, inter-cell, non-collocated), while sequencing Sub-Case 3 (Spatial domain Rx prediction), Sub-Case 4 (AI/ML-based prioritization for MOs), and Sub-Case 5 (Non-collocated frequency domain prediction in beam level) for later consideration contingent on completion of the first two. RAN4 identifies potential specification impacts on RAN4 requirements including EVM, MPR, and BS/UE demodulation performance for the non-linearity compensation use cases and on prediction accuracy/delay or measurement delay requirements for the AI-RRM use cases.
Key proposals
- Proposal 1 (Sec 2): RAN4 respectfully asks RAN to take the above conclusion into consideration.
- Conclusion (Sec 1): RAN4 has agreed to introduce AI-NC Sub-Case 1 (AI-based DPoD at NW) as a 6G RAN4-led AI use case to study first, with Sub-Case 2 (AI-based DPD at UE) to be studied subsequently only after the DPoD study completes, using its performance KPIs as a benchmark.
- Conclusion (Sec 1): RAN4 has agreed to introduce the Use Case-2 (AI-based SRS power imbalance compensation) as a 6G RAN4-led AI use case.
- Conclusion (Sec 1): For AI-RRM, RAN4 has agreed to begin the study on AI-RRM Sub-Case 1 (FR2-1 spatial beam prediction) and Sub-Case 2 (FR1 frequency cell prediction) first.
- Conclusion (Sec 1): Upon completion of AI-RRM Sub-Case 1 and 2, RAN4 will start Sub-Case 3 (Spatial domain Rx prediction) and then decide whether to study Sub-Case 4 (AI/ML-based MO prioritization) and Sub-Case 5 (Non-collocated frequency beam prediction).
- Note (Sec 1): It may require cross-WG study on channel model for AI-RRM use cases involving non-collocated prediction.
- Conclusion (Sec 4.1): For AI-NC Sub-Case 1 (AI-DPoD at NW), RAN4 foresees potential RAN4 spec impact on EVM, MPR, and BS demodulation performance requirements.
- Conclusion (Sec 4.2): For AI-NC Sub-Case 2 (AI-DPD at UE), RAN4 foresees potential RAN4 spec impact on UE RF requirements, e.g., MPR, and notes other requirements are not precluded.
- Observation (Sec 4.3): For AI-based SRS power imbalance compensation, companies provided evaluation results based on a model complexity of 73 M FLOPs, though other options are not precluded.
- Observation (Sec 4.1): For AI-DPoD at NW, some companies provided evaluation results based on Lightweight NN (~3M FLOPs) up to ~100M FLOPs, with other options not precluded.