Tech-invite  3GPPspace  IETFspace
21  22  23  24  25  26  27  28  29  31  32  33  34  35  36  37  38  4‑5x
Top   in Index   Prev   Next

TR 38.843
Study on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface

3GPP‑Page  
V18.0.0 (Wzip)  2023/12  187 p.
Rapporteur:
Dr. Montojo, Juan
Qualcomm Germany

full Table of Contents for  TR 38.843  Word version:  18.0.0

Here   Top
1 Scope  2 References  3 Definitions of terms, symbols and abbreviations  4 General AI/ML framework  5 Use cases  6 Evaluations  7 Potential specification impact assessment  8 Conclusions  $ Change history 

 

1  Scope  p. 7

The application of AI/ML to wireless communications has been thus far limited to implementation-based approaches, both, at the network and the UE sides. A study on enhancement for data collection for NR and ENDC (FS_NR_ENDC_data_collect) has examined the functional framework for RAN intelligence enabled by further enhancement of data collection through use cases, examples etc. and identify the potential standardization impacts on current NG-RAN nodes and interfaces. In SA WG2 AI/ML related study, a network functionality NWDAF (Network Data Analytics Function) was introduced in Rel-15 and has been enhanced in Rel-16 and Rel-17.
This study explores the benefits of augmenting the air-interface with features enabling improved support of AI/ML. The 3GPP framework for AI/ML is studied for air-interface corresponding to each target use case regarding aspects such as performance, complexity, and potential specification impact.
Through studying a few carefully selected use cases, assessing their performance in comparison with traditional methods and the associated potential specification impacts that enable their solutions, this study lays the foundation for future air-interface use cases leveraging AI/ML techniques.
Sufficient use cases are targeted to enable the identification of a common AI/ML framework, including functional requirements of AI/ML architecture, which could be used in subsequent projects. The study also serves identifying areas where AI/ML could improve the performance of air-interface functions.
The study serves identifying what is required for an adequate AI/ML model characterization and description establishing pertinent notation for discussions and subsequent evaluations. Various levels of collaboration between the gNB and UE are identified and considered.
Evaluations to exercise the attainable gains of AI/ML based techniques for the use cases under consideration are carried out with the corresponding identification of KPIs with the goal to have a better understanding of the attainable gains and associated complexity requirements.
Finally, specification impact are assessed in order to improve the overall understanding of what would be required to enable AI/ML techniques for the air-interface.
The central objective of this project is to study the 3GPP framework for AI/ML for air-interface corresponding to each target use case regarding aspects such as performance, complexity, and potential specification impact.
The use cases to focus include:
  • CSI feedback enhancement
    • Spatial-frequency domain CSI compression using two-sided AI model
    • Time domain CSI prediction using UE sided model
  • Beam management
    • Spatial-domain Downlink beam prediction for Set A of beams based on measurement results of Set B of beams
    • Temporal Downlink beam prediction for Set A of beams based on the historic measurement results of Set B of beams
  • Positioning accuracy enhancements
    • Direct AI/ML positioning
    • AI/ML assisted positioning
This study also introduces AI/ML model terminology and description to identify common and specific characteristics for framework investigations, namely to:
  • Characterize the defining stages of AI/ML related algorithms and associated complexity:
    • Model generation, e.g., model training (including input/output, pre-/post-process, online/offline as applicable), model validation, model testing, as applicable
    • Inference operation, e.g., input/output, pre-/post-process, as applicable
  • Identify various levels of collaboration between UE and gNB pertinent to the selected use cases, e.g.,
    • No collaboration: implementation-based only AI/ML algorithms without information exchange [for comparison purposes]
    • Various levels of UE/gNB collaboration targeting at separate or joint ML operation.
  • Characterize lifecycle management of AI/ML model: e.g., model training, model deployment, model inference, model monitoring, model updating
  • Dataset(s) for training, validation, testing, and inference
  • Identify common notation and terminology for AI/ML related functions, procedures and interfaces
For the use cases under consideration:
  1. Performance benefits of AI/ML based algorithms for the agreed use cases are evaluated:
    • Methodology based on statistical models (from TR 38.901 and TR 38.857 [positioning]), for link and system level simulations.
      • Extensions of 3GPP evaluation methodology for better suitability to AI/ML based techniques should be considered as needed.
      • Whether field data are optionally needed to further assess the performance and robustness in real-world environments should be discussed as part of the study.
      • Need for common assumptions in dataset construction for training, validation and test for the selected use cases.
      • Consider adequate model training strategy, collaboration levels and associated implications
      • Consider agreed-upon base AI model(s) for calibration
      • AI model description and training methodology used for evaluation should be reported for information and cross-checking purposes
    • KPIs: Determine the common KPIs and corresponding requirements for the AI/ML operations. Determine the use-case specific KPIs and benchmarks of the selected use-cases.
      • Performance, inference latency and computational complexity of AI/ML based algorithms should be compared to that of a state-of-the-art baseline
      • Overhead, power consumption (including computational), memory storage, and hardware requirements (including for given processing delays) associated with enabling respective AI/ML scheme, as well as generalization capability should be considered.
  2. Potential specification impact, specifically for the agreed use cases and for a common framework, is assessed:
    • PHY layer aspects, e.g., (RAN1)
      • Considering aspects related to, e.g., the potential specification of the AI Model lifecycle management, and dataset construction for training, validation and test for the selected use cases
      • Use case and collaboration level specific specification impact, such as new signalling, means for training and validation data assistance, assistance information, measurement, and feedback
    • Protocol aspects, e.g., (RAN2) - RAN2 only starts the work after there is sufficient progress on the use case study in RAN1
      • Considering aspects related to, e.g., capability indication, configuration and control procedures (training/inference), and management of data and AI/ML model, per RAN1 input
      • Collaboration level specific specification impact per use case
    • Interoperability and testability aspects, e.g., (RAN4) - RAN4 only starts the work after there is sufficient progress on use case study in RAN1 and RAN2
      • Requirements and testing frameworks to validate AI/ML based performance enhancements and ensuring that UE and gNB with AI/ML meet or exceed the existing minimum requirements if applicable
      • Considering the need and implications for AI/ML processing capabilities definition
Up

2  References  p. 9

3  Definitions of terms, symbols and abbreviations  p. 9

3.1  Terms  p. 9

3.2  Symbols  p. 11

3.3  Abbreviations  p. 11

4  General AI/ML framework  p. 12

4.1  Description of AI/ML stages  p. 12

4.2  Life cycle management  p. 12

4.3  Collaboration levels  p. 15

4.4  Functional framework details  p. 16

5  Use cases  p. 16

5.1  CSI feedback enhancement  p. 16

5.2  Beam management  p. 19

5.3  Positioning accuracy enhancements  p. 21

6  Evaluations  p. 21

6.1  Common evaluation methodology and KPIs  p. 22

6.2  CSI feedback enhancement  p. 23

6.2.1  Evaluation assumptions, methodology and KPIs  p. 23

6.2.2  Performance results  p. 33

6.2.2.1  1-on-1 joint training for CSI compression  p. 35

6.2.2.2  Generalization evaluations for CSI compression  p. 45

6.2.2.3  Scalability evaluations for CSI compression  p. 48

6.2.2.4  Multi-vendor joint training for CSI compression  p. 50

6.2.2.5  Separate training for CSI compression  p. 51

6.2.2.6  Basic performance for CSI prediction  p. 55

6.2.2.7  Generalization evaluations for CSI prediction  p. 58

6.2.2.8  Summary of Performance Results for CSI feedback enhancement  p. 59

6.3  Beam management  p. 59

6.3.1  Evaluation assumptions, methodology and KPIs  p. 59

6.3.2  Performance results  p. 69

6.3.2.1  Basic performance for BM-Case1  p. 69

6.3.2.1.1  Performance when Set B is a subset of Set A for DL Tx beam prediction  p. 70
6.3.2.1.2  Performance when Set B is different than Set A for DL Tx beam prediction  p. 72
6.3.2.1.3  Performance when Set B is a subset of Set A for DL Tx-Rx beam pair prediction  p. 73
6.3.2.1.4  Performance when Set B is different to Set A for DL Tx-Rx beam pair prediction  p. 76

6.3.2.2  Basic performance for BM-Case2  p. 76

6.3.2.2.1  Performance when Set A = Set B  p. 76
6.3.2.2.2  Performance when Set B is a subset of Set A  p. 85

6.3.2.3  Performance under different assumptions/scenarios for BM-Case1 and/or BM-Case2  p. 93

6.3.2.4  Generalization Performance for BM-Case1 and BM-Case2  p. 101

6.3.2.5  Summary of Performance Results for Beam Management  p. 106

6.4  Positioning accuracy enhancements  p. 108

6.4.1  Evaluation assumptions, methodology and KPIs  p. 108

6.4.2  Performance results  p. 115

6.4.2.1  Training Data Collection  p. 116

6.4.2.2  Generalization Aspects  p. 116

6.4.2.3  Fine-tuning  p. 120

6.4.2.4  Model-input Size Reduction  p. 127

6.4.2.5  Non-ideal label(s)  p. 130

6.4.2.6  Summary of Performance Results for Positioning accuracy enhancements  p. 131

7  Potential specification impact assessment  p. 131

7.1  General observations  p. 131

7.2  Physical layer aspects  p. 131

7.2.1  Common framework  p. 131

7.2.2  CSI feedback enhancement  p. 132

7.2.3  Beam management  p. 135

7.2.4  Positioning accuracy enhancements  p. 139

7.3  Protocol aspects  p. 145

7.3.1  Common framework  p. 145

7.3.2  CSI feedback enhancement  p. 145

7.3.3  Beam management  p. 145

7.3.4  Positioning accuracy enhancements  p. 145

7.4  Interoperability and testability aspects  p. 145

7.4.1  Common framework  p. 145

7.4.2  CSI feedback enhancement  p. 145

7.4.3  Beam management  p. 145

7.4.4  Positioning accuracy enhancements  p. 145

8  Conclusions  p. 145

$  Change history  p. 146

Up   Top