R1-2500159
discussion
Discussion on AIML for beam management
From Spreadtrum
Spreadtrum's prior position on
9.1.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Advocates for UE-initiated data collection for UE-side models, supports up to 16 beams reporting per instance, and favors reusing existing CSI frameworks to minimize specification impact while opposing larger quantization steps to maintain accuracy.
Summary
Spreadtrum presents 15 proposals and 6 observations regarding AI/ML for NR Beam Management, focusing on data collection, inference reporting, and performance monitoring for both UE-side and Network-side models. The document argues against configuring only Set B for UE-side inference, supports reusing existing TCI frameworks for BM-Case 2, and rejects probability-based metrics for performance monitoring due to reliability concerns.
Position
Spreadtrum opposes configuring only Set B for UE-side inference (Proposal 1) to prevent ambiguity in input-output correspondence between training and inference phases. They prefer reusing existing standard mechanisms, such as CRI/SSBRI for beam information (Proposal 2) and TCI indication frameworks for BM-Case 2 (Proposal 5), to minimize specification impact. For performance monitoring, they present a technical case against using probability information (Alt. 4) as a metric (Proposal 13), arguing it is unreliable and limited to classification models. They require the associated ID to be configured in CSI-ReportConfig (Proposal 11) and propose using it to ensure consistency between training and inference (Proposal 15). Additionally, they argue that implicit time reporting (Proposal 8) and adaptive beam selection (Proposal 10) are sufficient for BM-Case 2 without introducing new reference time definitions or fixed reporting structures.
Key proposals
- Proposal 1 (Sec 2.1.1): For UE-side model inference results reporting, do not support the configuration where only the resource set for Set B is configured, to ensure consistency between training and inference.
- Proposal 2 (Sec 2.2.1): For NW-side model reported beam information, only support CRI/SSBRI of a measurement resource set, avoiding the complexity of introducing bitmap schemes.
- Proposal 3 (Sec 2.2.2): For NW-side model data collection for training, support both L1 and higher-layer signaling to accommodate different latency and data volume requirements.
- Proposal 4 (Sec 2.2.2): For NW-side model data collection via higher-layer signaling, support all three types of content (Type 1, 2, and 3) for each instance to cover regression and classification models.
- Proposal 5 (Sec 3.2): For BM-Case 2, reuse the TCI indication framework by gNB, allowing beams from multiple time instances to be indicated via multiple beam indications respectively.
- Proposal 6 (Sec 3.2): For BM-Case 2 with NW-side model, reporting multiple past time instances in one reporting instance is not needed, as it adds overhead without improving AI performance.
- Proposal 7 (Sec 3.2.1): For BM-Case 1 and 2 with NW-side model, larger quantization steps should not be considered for model inference to maintain accuracy and reduce standardization impact.
- Proposal 8 (Sec 3.3.1): For BM-Case 2 with UE-side model, support implicit reporting of time information, where the order of reported beams corresponds to the predicted time instances.
- Proposal 9 (Sec 3.3.1): For BM-Case 2, support Option 1 (based on uplink slot for the report) for the reference time of the earliest time instance for predicted results.
- Proposal 10 (Sec 3.3.2): For BM-Case 2 with UE-side model, allow the UE to adaptively select the number of reported beams for each N instance based on channel conditions.
- Proposal 11 (Sec 3.3.3): Configure the associated ID in CSI-ReportConfig to efficiently manage beam properties for a pair of Set A and Set B.
- Proposal 12 (Sec 3.3.4): For UE-side model inference results, ranking information for Top K beams is conveyed by the order of beam information, with the method of obtaining ranking left to UE implementation.
- Proposal 13 (Sec 4.1): Do not support Alt. 4 (probability information) as a performance metric for UE-assisted monitoring, citing questionable reliability and limitation to classification models.
- Proposal 14 (Sec 4.2): Deprioritize the combination of aperiodic inference reports and periodic/semi-persistent monitoring reports for UE-side model monitoring.
- Proposal 15 (Sec 4.3): For UE-side model inference, consider Option 1 (using associated ID) to ensure consistency between training and inference regarding NW-side additional conditions.