R1-2500091
discussion
Discussion on AI/ML for CSI prediction
From Huawei
Huawei's prior position on
9.1.3
at
RAN1#119
· AI-synthesized, paraphrased
verify sources →
Presents a technical case against introducing network-side indications, such as associated IDs, for CSI prediction consistency, citing feasibility issues and proprietary disclosure risks regarding network planning information. Argues that generalized AI/ML models achieve satisfactory performance using mixed datasets (Generalization Case 3), rendering explicit consistency indications unnecessary. Proposes ensuring consistency via UE-side performance monitoring in an implementation manner, allowing the UE to select the best-matching model without network signaling overhead.
Summary
This Huawei contribution analyzes the specification impacts of AI/ML-based CSI prediction in NR, presenting 4 observations and 16 proposals across training consistency, inference configuration, performance monitoring, and Life Cycle Management (LCM). The document argues that mixed datasets achieve sufficient generalization, eliminating the need for explicit network indications for training/inference consistency, and proposes reusing Rel-18 MIMO mechanisms and Beam Management LCM procedures for CSI prediction.
Position
Huawei argues that generalized performance in CSI prediction can be achieved with mixed datasets (Generalization Case 3), thereby proposing that no explicit NW indications or associated IDs are needed for training/inference consistency. They propose reusing Rel-18 MIMO mechanisms for CSI report configuration and suggest investigating new priority rules where monitoring reports take precedence over inference reports. For performance monitoring, Huawei prefers SGCS as the primary metric and advocates for L1 signaling for monitoring results due to latency constraints, while proposing separate CSI report configurations for inference and monitoring processes. Regarding LCM, they propose separating CPU counting between legacy and AI/ML CSI reporting, introducing activation delays into processing timelines based on functionality state, and implementing memory occupancy alignment mechanisms. Finally, they support reusing the Beam Management functionality alignment signaling procedure (Steps 1-5) for CSI prediction, including reporting UE requirements for CPU, memory, and timeline in Step 4.
Key proposals
- Proposal 1 (Sec 2.1): No need to introduce indications from NW (e.g., explicit indication or associated ID) for ensuring consistency between training/inference for CSI prediction, as mixed datasets achieve generalized performance.
- Proposal 2 (Sec 3.1): For the configuration of CSI measurement and report for AI/ML based CSI prediction, the mechanism of Rel-18 MIMO may be reused.
- Proposal 3 (Sec 3.2): RAN1 to investigate the priority rule for AI/ML-based CSI reporting types, considering that monitoring results may be more critical than inference reports.
- Proposal 4 (Sec 4.1.1): For monitoring Type 1, consider SGCS with higher priority for the metric type, and define events triggered when prediction accuracy is lower than a threshold satisfying a timer/counter.
- Proposal 5 (Sec 4.1.2): For monitoring Type 2, consider the legacy codebook/PMI as the ground-truth CSI to minimize reporting overhead.
- Proposal 6 (Sec 4.1.3): For monitoring Type 3, consider SGCS with higher priority and allow reporting per sample, for a set of samples, or statistical values.
- Proposal 7 (Sec 4.2): Study the type of predicted CSI and ground-truth CSI for calculating intermediate KPI metrics, comparing channel matrix/raw eigenvectors against legacy codebook/PMI.
- Proposal 8 (Sec 4.3): For the report of monitoring results, consider L1 signaling with higher priority due to stringent latency requirements.
- Proposal 9 (Sec 4.4): For monitoring Type 1/3, adopt two separate CSI reports configured to handle the inference process and monitoring process respectively, linking them via associated CSI-ReportConfigId.
- Proposal 10 (Sec 5.1.1): Focus on the solution where overall CPU is separately counted between legacy CSI reporting and AI/ML-based CSI reporting, but shared among CSI report related AI/ML functionalities.
- Proposal 11 (Sec 5.1.2): The CSI processing timeline for A-CSI reports should be impacted by the active/inactive state of the functionality, adding activation delay if inactive.
- Proposal 12 (Sec 5.1.2): Determine active/inactive state via NW-configured activation windows or based on SP-CSI reports subject to the same functionality.
- Proposal 13 (Sec 5.1.3): Discuss a memory occupancy alignment mechanism to align the availability of memory storage for updating AI/ML-based CSI reports.
- Proposal 14 (Sec 5.2): Reuse the general signaling procedure (Steps 1-5) and Direction A/B agreed for beam management for CSI prediction functionality alignment.
- Proposal 16 (Sec 5.2): UE should report performance and requirement related information (target performance, timeline, CPU, memory) along with the applicable functionality in Step 4.