R1-2500201
discussion
Discussion on AI/ML-based beam management
From CATT
CATT's prior position on
9.1.1
at
RAN1#119
· AI-synthesized, paraphrased
verify sources →
Advocates for flexible AI/ML beam management solutions that reuse existing CSI framework mechanisms while introducing minimal specification impact. Supports both UE-sided and network-sided models with comprehensive performance monitoring. Opposes overly complex new signaling mechanisms, favoring practical approaches like reusing legacy TCI delay requirements and CPU mechanisms rather than creating entirely new frameworks.
Summary
This document from CATT presents 49 proposals regarding AI/ML-based beam management for NR, covering consistency issues, UE-sided and NW-sided model inference, performance monitoring, and CSI processing enhancements. It addresses specific configuration mechanisms for Set A and Set B resources, reporting formats for predicted beams, and methods for validating model performance across different cells and time instances.
Position
CATT concludes that defining similar properties for DL Tx beams associated with an ID is unnecessary for UE-sided models, proposing instead that the associated ID be configured at the CSI-Report level. They argue that aligning Rx beam information between the network and UE is beneficial for NW-sided models to maintain prediction accuracy. For BM-Case2, CATT opposes extending Rel-17 TCI state activation for multiple future time instances, citing complexity and limited overhead benefits. They propose introducing an enhanced CPU pool for AI/ML processing, distinct from legacy CPUs, and support event-based performance monitoring linked to inference reports. Additionally, they suggest using predefined mappings for monitoring resources when full Set A measurement is not feasible.
Key proposals
- Proposal 1 (Sec 2.1): Conclude that for UE-sided models, there is no need to define the similar properties of a DL Tx beam associated with an ID.
- Proposal 5 (Sec 2.2): For NW-sided models, align the Rx information of measurements between the network and UE to ensure consistency.
- Proposal 7 (Sec 3.1.1): Support configuring one CSI-ResourceConfigId for Set B, with Set A determined via an associated ID for UE-sided inference.
- Proposal 10 (Sec 3.1.2): For BM-Case2, consider options for aperiodic CSI-RS configuration where multiple transmissions of a resource set are triggered by one instance.
- Proposal 15 (Sec 3.1.3): Allow gNB to configure whether reported RSRP for predicted beams is predicted L1-RSRP or measured L1-RSRP.
- Proposal 19 (Sec 3.1.4): Oppose extending Rel-17 TCI state activation methods to indicate TCI states for multiple future time instances in BM-Case2.
- Proposal 21 (Sec 3.2.2): Define successful inference instances for performance monitoring based on Top-K beam prediction accuracy within an X dB margin.
- Proposal 26 (Sec 3.2.2): Support predefined mapping between full Set A resources and monitoring resources for performance metric calculation.
- Proposal 33 (Sec 3.3): Introduce a new type of processing unit (enhanced CPU) for AI/ML-based CSI processing, counted separately from the legacy CPU pool.
- Proposal 38 (Sec 4.1): Support reporting L1-RSRPs and beam information for up to M beams within an X dB gap to the largest measured value for NW-sided models.
- Proposal 43 (Sec 4.2): Support differential L1-RSRP reporting with legacy quantization steps for data collection via higher layer signaling.
- Proposal 46 (Sec 4.3): Allow the network to dynamically indicate active resources or QCL relationships for Top-K beam measurements (P2 procedure).