R1-2508643
discussion
AI/ML use cases and framework for 6GR Air Interface
From AT&T
Summary
AT&T presents 8 observations and 30 proposals across 5 sections addressing the 6G Radio AI/ML framework and use cases, totaling 38 actionable items. The document argues the 5G LCM framework is insufficient for native AI/ML integration in 6GR, proposes a unified LCM framework, identifies specific use cases for study (CSI-RS overhead reduction, DMRS overhead reduction, DPoD, beam management), and prioritizes network-side benefits and operator relevance.
Position
AT&T argues that the 5G LCM framework for AI/ML does not allow for native AI/ML integration in 6GR and is not scalable, proposing instead a unified LCM framework including data and model management, model transfer, and model training. They require that 6GR AI/ML use case evaluation use final system performance metrics (throughput, BLER) rather than intermediate metrics like SGCS, and that generalization performance be prioritized under wide-ranging realistic conditions. They support studying frequency and/or spatial domain CSI prediction with sparse/low overhead CSI-RS, sparse orthogonal DMRS overhead reduction, and DPoD/Non-linearity compensation at the network side. AT&T deprioritizes SRS overhead reduction (insisting uplink coverage remains the principal objective for SRS enhancements) and AI/ML beam prediction for initial access (citing fundamental discrepancies with connected-mode beam prediction and risk of complicating 6GR design with separate procedures for AI/ML-capable and incapable UEs). They request a dedicated AI/ML agenda item post-RAN1#123 to prevent the LCM and data collection framework from being fragmented across disparate use case agendas.
Key proposals
- Proposal 1 (Sec 2): Study requirements/design of a unified LCM framework in 6GR, including data and model management, handling NW-side/UE-side additional conditions, framework for AI/ML processing unit and memory, and considering 5G NR LCM framework as starting point.
- Proposal 2 (Sec 3.1): Consider use cases that provide high impact practical relevance to operators for 6GR design.
- Proposal 7 (Sec 3.2): When evaluating AI/ML-based use cases on CSI-RS overhead reduction, DMRS design with AI receiver, interference prediction and cross frequency CSI prediction, consider site-specific aspects and realistic assumptions on interference modelling, traffic modelling and RF impairment modeling.
- Proposal 8 (Sec 3.2.1): Study frequency and/or spatial domain CSI prediction with sparse/low overhead CSI-RS with AI/ML, with NR Rel-20 MIMO framework for CSI-RS overhead reduction considered as a baseline.
- Proposal 9 (Sec 3.2.2): Study AI/ML-based DMRS overhead reduction, with sparse orthogonal DMRS in frequency and/or time domain as a starting point.
- Proposal 10 (Sec 3.2.4): Study AI/ML-based DPoD/Non-linearity compensation at the network side.
- Proposal 11 (Sec 3.2.5): The AI/ML-based SRS overhead reduction study for 6GR is deprioritized.
- Proposal 12 (Sec 3.2.6): Further study AI/ML-based interference prediction and management in the beam domain.
- Proposal 13 (Sec 3.2.6): AI/ML beam prediction for initial access sub-use case is deprioritized in the 6GR study.
- Proposal 14 (Sec 4): RAN1 Chair guidance is needed on handling the distribution of work across AI/ML-based framework and non-AI/ML-based framework of a given feature.
- Proposal 15 (Sec 4): A dedicated AI/ML agenda item is needed to continue discussing 6GR AI/ML LCM framework and data and model management, generalizable and applicable to all the use cases.
- Proposal 3 (Sec 3.1): Use system performance metrics (e.g. throughput, overhead) or link performance metrics (e.g. BLER) for performance evaluation/conclusions on AI/ML use cases, when applicable.
- Proposal 4 (Sec 3.1): Consider complexity and performance tradeoffs for evaluating AI/ML use cases, including model/computational complexity.
- Proposal 5 (Sec 3.1): AI/ML use cases should provide clear benefits over non-AI/ML use cases in terms of performance, overhead reduction, power savings.
- Proposal 6 (Sec 3.1): Generalization performance should be prioritized under a wide range of conditions, including realistic channels/deployment scenarios, when applicable.