R1-2409582
discussion
Discussion for supporting AI/ML based positioning accuracy enhancement
From Samsung
Summary
Samsung presents a comprehensive discussion on AI/ML-based positioning accuracy enhancement, outlining 29 observations across triggering, model selection, data collection, inference, monitoring, and consistency checks. The document emphasizes the need for processed channel measurements rather than raw data, defines specific roles for data generation entities, and proposes mechanisms for model performance monitoring and training-inference consistency.
Position
Samsung argues that full-size raw channel measurements are unsuitable for data collection due to prohibitive overhead and storage costs, proposing instead that truncated or feature-extracted measurements be used. They support explicit signaling mechanisms where UEs can notify the network of their willingness or ability to act as data providers, including notifications when specific quality conditions like SNR are not met. Samsung proposes that consistency checks between training and inference phases must occur before model selection to ensure alignment. They support the inclusion of timestamps and quality indicators alongside model outputs and define distinct monitoring metrics based on model output, input, or other measurements. Furthermore, they specify that post-monitoring behaviors, such as finetuning or fallback to legacy methods, must be clearly defined in the specification.
Key proposals
- Observation 1 (Sec 2.2): RAN1 considers trigger conditions for enabling AI/ML based positioning, including model applicable conditions or environment/channel applicable conditions.
- Observation 2 (Sec 2.3): RAN1 supports the indication of the determined model from one entity to another entity.
- Observation 4 (Sec 3.1.1): For paired timing and power information, two types are considered: (b-1) separate values with joint reporting, and (b-2) joint time and power value reporting.
- Observation 6 (Sec 3.1.2.2): Full-size (or raw) channel measurement is deemed unsuitable for collected data samples due to overhead and storage constraints.
- Observation 7 (Sec 3.1.2.2): Both truncated channel measurement and feature extracted channel measurement are considered as candidates for channel measurement in collected data samples.
- Observation 8 (Sec 3.1.2.2): Path-based measurement could be considered with potential overhead reduction methods.
- Observation 11 (Sec 3.2): RAN1 supports the UE notifying the network whether it is a data provider or not.
- Observation 12 (Sec 3.2): RAN1 supports the UE notifying the network that it is inapplicable to provide data samples when certain conditions (e.g., SNR/RSRP requirements) are not satisfied.
- Observation 13 (Sec 3.3.1): RAN1 supports collected data being signalled from a PRU (other UE) to a UE deployed with an AI model.
- Observation 15 (Sec 3.4): Training requires relatively large data set sizes, while finetuning and some monitoring require medium sizes, and inference requires small sizes.
- Observation 18 (Sec 4): Timestamp and quality information are supported to be attached to the model output.
- Observation 21 (Sec 5.1): RAN1 considers model output based, model input based, and other (measurement) based monitoring metrics.
- Observation 27 (Sec 5.4): Post monitoring behavior should be specified, including actions like finetuning, terminating the model, or falling back to legacy methods.
- Observation 28 (Sec 6.1): Consistency checks between training and inference should be done when or before the entity decides which model to use.