R1-2508689
discussion
Discussion on AI/ML in 6GR interface
From Xiaomi
Summary
This document from Xiaomi contains 24 Observations and 17 Proposals addressing AI/ML use cases, framework, and evaluation for the 6G radio interface. It prioritizes use cases for study, recommends joint AI/non-AI discussions in dedicated sessions, and proposes specific extensions to the AI/ML framework including multi-cell associated IDs, AI engine power states, and data collection for DMRS design.
Position
Xiaomi proposes a structured three-category approach to AI/ML use case selection (5GA-supported, extensions, and new use cases) with distinct handling per type. They propose moving specific use cases—AI-based Inter-cell/M-TRP beam prediction, Cross-frequency beam prediction, CSI prediction in frequency/spatial domain, DMRS design, and constellation generation—to their corresponding non-AI session (MIMO or Modulation) for joint AI/non-AI discussion. Regarding the AI/ML framework, Xiaomi proposes extending associated ID applicability from single-cell to multi-cell domains to support generalized UE models, and requires defining standardized AI engine power states (Deep Sleep, Light Sleep, Active Execution) with transition mechanisms. They propose data collection framework extensions enabling acquisition of dedicated bits/symbols/sequence for AI/ML-based data processing use cases like DMRS design. For evaluation methodology, Xiaomi proposes adopting TR 38.840 as baseline while removing calibration requirements using agreed-upon base AI models, and introduces energy efficiency as a new common KPI defined through operational power consumption modeling.
Key proposals
- Proposal 1 (Sec Introduction): Candidate use cases for selection can be categorized as: 5GA-supported use cases, extensions of 5GA use cases, and new use cases. Distinct approaches should be applied to handle each type.
- Proposal 3 (Sec Support of use cases in 5GA): Support 5G-A AI-based beam management and CSI prediction as 6G day 1 features, and only carry out potential specification impact study during the later phase of the 6G SI.
- Proposal 4 (Sec Extension on AI/ML for beam management): Move AI-based Inter-cell/M-TRP beam prediction to the MIMO session for joint discussion, studying common evaluation assumptions, performance comparison with non-AI solutions, generalization, and spec impact.
- Proposal 6 (Sec AI based CSI prediction): Move AI-based CSI prediction in frequency and/or spatial domain to the MIMO session for joint discussion with non-AI CSI prediction, covering evaluation assumptions, performance comparison, generalization, and spec impact.
- Proposal 7 (Sec AI assisted DMRS design): Move AI-based DMRS design to Agenda 11.8 (MIMO operation) to enable joint study with non-AI DMRS design, further evaluating sparse/superimposed DMRS with different AI receiver types for both NW and UE sides.
- Proposal 8 (Sec AI assisted constellation generation): Move the AI-based constellation design to the modulation session to discuss AI-based and non-AI-based design jointly.
- Proposal 9 (Sec AI-based CSI compression enhancement): Enable AI/ML-based CSI compression in 6GR by minimizing or eliminating inter-vendor collaboration in model training and leveraging physical layer procedures similar to traditional CSI compression.
- Proposal 10 (Sec AI/ML Framework): Guide framework extension studies by controlling UE complexity/cost, maintaining excellent user experience (energy efficiency, privacy), and extending data collection for new sample types (e.g., transmission data bits/symbols).
- Proposal 11 (Sec AI/ML Framework): Establish a dedicated agenda item for AI/ML framework discussions in future meetings.
- Proposal 12 (Sec AI/ML Framework): Unique associated ID among multiple cells should be supported to ensure more efficient network condition check.
- Proposal 13 (Sec AI/ML Framework): Extend the AI/ML framework by defining standardized power states of the AI/ML engine and transition mechanisms among different power states.
- Proposal 14 (Sec AI/ML Framework): Consider data collection extension for data processing related use cases, e.g., defining dedicated bits/symbols/sequence for data collection.
- Proposal 15 (Sec AI/ML Evaluation Methodology): Adopt TR 38.840 Section 6 methodology as baseline, incorporating field data for training/testing and removing the requirement for calibration using agreed-upon baseline AI model(s).
- Proposal 16 (Sec AI/ML Evaluation Methodology): Consider common KPIs from TR 38.843 Section 6.1 as starting point, and further consider energy efficiency of AI operation as one additional common KPI.
- Proposal 17 (Sec AI Energy consumption evaluation): Define an energy efficiency KPI based on power states including active execution energy and light sleep energy.