R1-2409986
discussion
AI/ML for Positioning Accuracy Enhancement
From Nokia
Summary
Nokia's comprehensive technical document on AI/ML for positioning accuracy enhancement presents 51 detailed proposals and 11 observations covering performance monitoring, training-inference consistency, data collection, and inference operations for various AI/ML positioning use cases.
Position
Nokia advocates FOR comprehensive standardization of both label-free and label-based monitoring approaches with strong LMF control over functionality decisions, supporting both path-based and sample-based representations while pushing AGAINST CIR support for inference input due to overhead concerns. They strongly promote PRU-assisted monitoring solutions and Associated ID mechanisms for training-inference consistency while opposing autonomous UE functionality reconfiguration.
Key proposals
- Proposal 1 (Sec 2): RAN1 to standardize both monitoring approaches, label-free based monitoring and label-based monitoring for first priority cases.
- Proposal 9 (Sec 2): For label-based monitoring of Case 1, for UE to derive monitoring metric, support Option A-3: LMF provides the UE necessary data for monitoring over LPP which contains measurements collected from PRU(s) and associated ground truth including quality indicators.
- Proposal 18 (Sec 3): RAN 1 to consider the following approaches to ensure consistency between training and inference regarding NW-side additional conditions: (i) consistency based on Assistance data, (ii) based on Associated ID, and (iii) consistency assisted by monitoring.
- Proposal 25 (Sec 4): The quality of content of the dataset used for training, RAN1 considers the sample's density (#samples/m2) and the distribution similarity between the target dataset and the uniform distribution.
- Proposal 28 (Sec 4): For training data collection of AI/ML based positioning, quality indicator of label is represented by a value between a predefined range (e.g. between 0 and 1 with 1 indicating the highest quality).
- Proposal 32 (Sec 4): In data collection for ground truth generation by the target UE for Case 1, LMF may indicate UE positioning method(s) with necessary criteria for an estimation to be used as ground truth.
- Proposal 34 (Sec 5): RAN1 to consider both: path-based and sample-based representation for AI/ML positioning cases.
- Proposal 39 (Sec 5): CIR is not supported for inference input when the model running at the LMF-side, gNB-side or UE-side cases.
- Proposal 42 (Sec 5): During model inference for Case 3a, AIML model confidence is taken into account for the quality of timing information assessment.
- Proposal 46 (Sec 6): RAN 1 to discuss the potential approaches on reusing or enhancing the current LPP to enable the functionality framework for AI/ML positioning.
- Proposal 50 (Sec 6): For all use cases (Cases 1, 3a, 3b) LMF is the only entity to determine functionality decision based on monitoring outcome.
- Proposal 51 (Sec 7): The AI/ML multi-RTT positioning method may be built on top of legacy multi-RTT positioning by including at least new assistance data sent by the LMF to the UE/gNB.