R1-2508740
discussion
Views on AI/ML in 6GR air interface
From Huawei
Summary
This Huawei/HiSilicon contribution to 3GPP RAN1#123 provides 11 proposals and 4 observations on AI/ML in the 6G air interface, covering updates on use cases, a functional framework, and future study directions. The document advocates for continued study of AI/ML without premature down-selection, proposes distribution of use cases to specific agendas, and provides evaluation results for use cases including RAN digital twin, token traffic, SRS overhead reduction, CSI prediction, and DMRS overhead reduction.
Position
Huawei opposes performing use case down-selection at the current phase of 6GR SI, arguing that common understandings on applicable scenario, achievable performance, and complexity have not been achieved, and that supporting companies for 5G AI/ML extension use cases does not mean their values are well justified. Huawei proposes studying four specific use cases at the early phase: sensing-based RAN digital twin with NW-side or distributed model, improved RAN solutions for token traffic, frequency/spatial domain CSI prediction with AI/ML, and low overhead DMRS with AI/ML receiver. For CSI prediction, Huawei questions the applicability of UE-side models for sparse CSI-RS and proposes studying NW-side models with a new CSI report type based on long-term multi-path power/angle/delay information. For UL DMRS overhead reduction, Huawei identifies PAPR increase as a critical issue requiring study of AI/ML PAPR reduction solutions, showing PAPR increases from 5.1dB to 6.3-7.4dB for various DMRS overhead reduction schemes.
Key proposals
- Proposal 1 (Sec 2.1.1): Update the RAN1#122bis observation for site specific learning for AI/ML and RAN Digital Twin as Table 2.
- Observation 1 (Sec 2.2.1): Enabling RAN awareness of the differentiated importance on Token level can benefit the efficient transmission of the token traffic.
- Proposal 2 (Sec 2.3): Update the RAN1#122bis observation for low overhead SRS with AI/ML as Table 4.
- Proposal 3 (Sec 3.5): Study the AI/ML framework aspects across different use cases, including data collection, training methods, model location, RAN awareness and assistance on AI/ML services, and UE processing and memory on AI/ML functionalities.
- Proposal 4 (Sec 4.1): For the study of AI/ML use cases, consider principles including bringing new revenues for operators, wide applicability, significant performance improvement with acceptable UE complexity, UL coverage and user experience criteria, and impact on fundamental 6GR design or interaction with other WGs.
- Proposal 5 (Sec 4.2.1): For sensing based RAN digital twin based on NW-side model or distributed model, study KPIs of sensing accuracy (RMSE, IoU, edge detection probability) and overhead of feedback payload size, with benchmarks of NW side mono-static sensing or UE-side bi-static sensing without NW side AI/ML.
- Proposal 6 (Sec 4.2.2): For improved RAN solutions for token traffic, study evaluation methodology at RAN1 by extracting and parameterizing characteristics/requirements of Token success rate, Token delay budget, Token arrival rate, and Token size from specific application layer models, then evaluating SLS network performance.
- Proposal 7 (Sec 4.2.3): For frequency/spatial domain CSI prediction with AI/ML, study justification on applicability and performance analysis between UE-side model and NW-side model, new CSI report type for CSI prediction under ultra-sparse CSI-RS.
- Proposal 8 (Sec 4.2.4): For low overhead DMRS with AI/ML receiver, evaluate UL benefit of improved user experience and coverage including PAPR reduction, and for DL further study the tradeoff between performance and complexity.
- Proposal 9 (Sec 4.3): No need to perform use case down selection at this phase of 6GR SI.
- Proposal 10 (Sec 4.3): Distribute the categorized use cases to specific agendas including Control/data scheduling/HARQ, MIMO operation, Waveform, Channel coding/modulation, Initial access, and Sensing.
- Proposal 11 (Sec 4.3): RAN1 AI/ML agenda is still preserved after RAN1#123 meeting to guarantee consistency among different use cases, handle cross WG issues, and discuss remaining issues such as cross agenda impact.