R1-2500068
discussion
Discussion on specification support for AI CSI prediction
From ZTE
ZTE's prior position on
9.1.3
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Proposes a systematic approach that prioritizes identifying additional conditions before developing detailed consistency solutions. Supports reusing proven AI beam prediction mechanisms but opposes introducing new associated IDs for scenarios/carrier frequency.
Summary
This document from ZTE analyzes the specification support for AI-based CSI prediction in Rel-19, concluding that down tilt angle and TXRU mapping do not require additional network-side conditions due to model generalization capabilities. It presents seven proposals covering the reuse of Rel-18 codebooks, specific CSI-RS resource separations, performance monitoring options, and separate CPU counting criteria for AI tasks.
Position
ZTE argues that down tilt angle and TXRU mapping should not be treated as network-side additional conditions for AI CSI prediction, citing simulation results showing that mixed datasets guarantee performance generalization with minimal SGCS impact. They propose reusing the Rel-18 MIMO CSI prediction codebook design and CQI calculation mechanism to minimize specification changes. For resource configuration, they suggest considering m=4 or m=5 slots for the separation between consecutive aperiodic CSI-RS resources to align with baseline assumptions. Regarding performance monitoring, ZTE supports Type 2 and Type 3 monitoring and proposes supporting both periodic reports after the monitoring window and event-triggered reports. Finally, they propose considering separate CSI CPU counting for AI prediction tasks from legacy CSI processing criteria to account for distinct computational loads on dedicated units.
Key proposals
- Proposal 1 (Sec 2.1): RAN1 concludes that neither down tilt angle nor TXRU mapping is considered as additional condition for AI CSI prediction because performance can be guaranteed by model generalization based on mixed dataset.
- Proposal 2 (Sec 3.1): RAN1 reuses the Rel-18 MIMO CSI prediction codebook design and CQI calculation mechanism for Rel-19 AI-based CSI prediction.
- Proposal 3 (Sec 3.1): For AI/ML CSI prediction, consider m=5 and/or m=4 as candidates for the separation between two consecutive aperiodic CSI-RS resources.
- Proposal 4 (Sec 3.2): Rel-19 AI-based CSI prediction supports at least opt 1 (report after monitoring window) and opt 2 (event-triggered report) for report configuration for performance monitoring.
- Proposal 5 (Sec 3.2): Further study whether the existing aperiodic CSI-RS triggering configuration and mechanism is sufficient for the performance monitoring of the AI-based CSI prediction.
- Proposal 6 (Sec 3.2): RAN1 supports type 2/3 performance monitoring for AI CSI prediction.
- Proposal 7 (Sec 3.4): Regarding AI CSI prediction, consider separate CSI CPU counting from the legacy CSI processing criteria.
- Observation 1 (Sec 2.1): AI CSI prediction can generalize well across different down tilt angles, where using a dataset generated from a different angle only incurs 0.28% SGCS impact.
- Observation 2 (Sec 2.1): Regarding lower UE speed (30km/h), model trained in one TXRU mapping configuration may incur performance loss in a scenario with different TXRU mapping, but mixed dataset can well address the potential performance loss.
- Observation 3 (Sec 2.1): Regarding higher UE speed (60km/h), model trained in one TXRU mapping configuration may incur performance loss in a scenario with different TXRU mapping, but mixed dataset can well address the potential performance loss.
- Observation 4 (Sec 3.3): Compared with performance monitoring, there is no specific specification impact for CSI-RS configuration and CSI report regarding data collection for model training for AI CSI prediction.