R1-2500529
discussion
Discussion on AI/ML for beam management
From InterDigital
Summary
InterDigital presents 27 proposals and 14 observations regarding AI/ML for beam management in NR, focusing on data collection configuration, beam prediction reporting overhead reduction, and performance monitoring. The document argues for shared CPU resources between legacy and AI/ML CSI processing, dynamic switching of inference location, and specific mechanisms to handle QCL parameters for unmeasured predicted beams.
Position
InterDigital proposes clarifying data collection configuration signaling before finalizing associated ID working assumptions, supporting dynamically activated sub-configurations. They require that Set A and Set B be configured via two separate CSI-ResourceConfig Ids and support Set A resources with no physical transmission parameters. They argue for a shared CPU counter between legacy and AI/ML CSI reporting to avoid inefficient resource allocation. InterDigital supports overhead reduction mechanisms, specifically reporting beams within an X dB gap of the best beam, and proposes enhanced reporting of up to 64 RSRP values for network-side models. They propose TCI-state based activation for performance monitoring and specific metrics like 'Top-1 in Top-K' or 'margin-based' accuracy. Finally, they support a procedure to dynamically switch the location of AI/ML inference between UE and network based on workload and performance.
Key proposals
- Proposal 1 (Associated ID): Clarify the definition and detailed signaling of data collection related configuration before confirming the working assumption on the support of associated IDs in CSI framework.
- Proposal 2 (Associated ID): Data collection related configuration supports sub-data collection configurations, which can be dynamically activated, considering different scenarios/configurations associated with one associated ID.
- Proposal 3 (Applicability of inference): For configuration of each set of inference related parameters for applicability report, down select from Option 1 (simplified parameters) or Option 2 (reuse existing IEs).
- Proposal 5 (Configuration of Set A and Set B): No configuration method is supported other than configuration of two CSI-ResourceConfig Ids for Set A and Set B separately.
- Proposal 8 (QCL measurements): Support configuration of RS resources to estimate QCL-parameters for a RS resource without transmission in Set A.
- Proposal 9 (Pattern based beam indication): Support a beam indication mechanism with a beam pattern and corresponding TCI states required for the indicated beam pattern.
- Proposal 10 (Reporting beam measurements): Support a shared CPU counter between legacy and AIML-based CSI reporting.
- Proposal 12 (Reporting beam measurements): Support reporting of L1-RSRPs and corresponding beam information of up to M beams within X dB gap to the largest measured value of L1-RSRP.
- Proposal 15 (Reporting beam measurements): Reporting prediction results of multiple future time instances in one report should be supported, with periodicity aligned with CSI-Reporting periodicity.
- Proposal 19 (Reporting beam measurements): For network sided model, support enhanced UE reporting to report up to 64 RSRP values for whole Set A over multiple time instances without CRIs/SSBRIs.
- Proposal 21 (LCM and Consistency): Support reporting of UE selected Set B based on a rule (e.g., subset of best measured beams).
- Proposal 23 (Performance monitoring): For UE-assisted performance monitoring, down select from Option 2 (1/Top-K), Option 3 (Top-K/M), or Option 4 (best of Top-K/1 with margin) for performance metrics.
- Proposal 27 (Location of AIML Inference): Support a procedure to dynamically switch AIML inference location (e.g., based on NW workload and/or UE’s prediction performance).