Skip to main content
Technology

AI/ML Evaluation Framework

At a glance
  • What it is: AI and machine learning in 5G, from network-side analytics in the NWDAF to 3GPP studies of AI/ML in media processing and delivery.
  • Why it matters: It shows how network data and AI models can support media services and their quality.
  • In this section: This page covers the NWDAF, data collection, the media studies in two technical reports, and related studies feeding 6G.

How it works​

NWDAF works from network-side data and produces analytics for network and application functions to consume. SA4's studies look at applying AI/ML directly to media processing and delivery. The sections below cover NWDAF's own structure and SA4's study work (TS 23.288 V20.2.0 clause 4.2.0; TS 23.288 V20.2.0 clause 5.1; TR 26.927 V19.0.0 clause 4.2; TR 26.847 V19.0.0 clause 1).

Key specifications: 3GPP TS 23.288 (NWDAF architecture, SA2), TR 26.847 (Evaluation of AI and ML in 5G media services) (TR 26.847 V19.0.0 front page (title)).

NWDAF: network-side analytics (SA2)​

The Network Data Analytics Function (NWDAF) is the 5G core function through which network functions and other data sources expose measurements, and through which consumers request analytics or predictions derived from that data (TS 23.288 V20.2.0 clause 4.1; TS 23.288 V20.2.0 clause 4.2.0; TS 23.288 V20.2.0 clause 5.1). TS 23.288 splits the NWDAF's internal role in two:

  • AnLF (Analytics Logical Function): produces the analytics or prediction output that a consumer (for example the PCF, for QoS policy, or an Application Function (AF)) subscribes to or requests (TS 23.288 V20.2.0 clause 5.1; TS 23.288 V20.2.0 clause 6.4.1; TS 23.288 V20.2.0 clause 6.4.4; TS 23.288 V20.2.0 clause 6.3.1).
  • MTLF (Model Training Logical Function): trains the ML models that an AnLF uses for inference, and can expose those models to other NWDAF instances (TS 23.288 V20.2.0 clause 5.1; TS 23.288 V20.2.0 clause 4.2.0; TS 23.288 V20.2.0 clause 6.2A.1; TS 23.288 V20.2.0 clause 7.5).

Release differences in TS 23.288: V17.12.0 has the ML model provisioning procedure (clause 6.2A, Nnwdaf_MLModelProvision in clause 7.5) and the AnLF and MTLF split (TR 21.917 clause 9.4.3). ML model training (clause 6.2F, Nnwdaf_MLModelTraining in clause 7.10), accuracy monitoring of analytics and ML models (clauses 5C, 6.2D and 6.2E, Nnwdaf_MLModelMonitor in clause 7.9), federated learning among NWDAFs (clauses 5.3 and 6.2C) and ML model storage in the ADRF (clause 10.3) are in V18.13.0 and not in V17.12.0. Vertical federated learning (clauses 5.4 and 6.2H) and the monitoring of AI/ML positioning models (clause 6.2E.4) are in V19.8.0 and not in V18.13.0 (TR 21.919 clause 8.1.3). In V20.2.0, clause 5C adds a loss value and its compute method to the model accuracy information, which V19.8.0 does not have; the change history lists CR 1542 "Model Evaluation Metrics Enhancement" in V20.0.0.

Splitting training from inference means that an NWDAF containing an AnLF can use trained ML models provisioned by another NWDAF containing an MTLF. An AF can be a consumer of analytics such as slice load level and QoS sustainability; an AF that consumes Service Experience analytics may adjust service parameters, for a video service for example bit rate, frame rate or codec format, to better match the network conditions. TS 23.288 does not name the 5GMS AF; that it can act as such a consumer is 5G-MAG's reading. See Standards: Media Streaming and Advanced Media Delivery for where the 5GMS AF sits in that path (TS 23.288 V20.2.0 clause 4.2.0; TS 23.288 V20.2.0 clause 6.2A.1; TS 23.288 V20.2.0 clause 6.3.1; TS 23.288 V20.2.0 clause 6.4.4; TS 23.288 V20.2.0 clause 6.9.1).

Data collection for NWDAF​

TS 26.531, clause 4.1: “Clause 6.2.8 of TS 23.288 [4] envisages a set of high-level procedures by which data is collected by a Network Data Analytics Function (NWDAF) from UE Application(s) via an intermediary Application Function that provides an Event Exposure Service to event consumers.” The project for it is User Data Collection and Reporting (TS 26.531 V19.0.0 clause 4.1; TS 23.288 V20.2.0 clause 6.2.8). In TS 23.288 the clause on data collection from the UE Application (6.2.8) and the clause on user consent for analytics (6.2.9) are already in V17.12.0, from the Release 17 work item eNA_Ph2 (TR 21.917 clause 9.4.3; TS 23.288 V17.12.0 clauses 6.2.8 and 6.2.9).

The SA4 AI/ML media studies: TR 26.927 and TR 26.847​

SA4's study of applying AI/ML techniques directly to media processing and delivery, as distinct from the network-analytics use of AI/ML in NWDAF above, is captured in two companion technical reports, both published at version 19.0.0 in June 2025 (Release 19): (TR 26.927 V19.0.0 clause 4.2; TR 26.847 V19.0.0 clause 1; TS 23.288 V20.2.0 clause 4.2.0; TR 26.927 V19.0.0 change history; TR 26.847 V19.0.0 change history and clause 1; TR 26.927 V19.0.0 front page; TR 26.847 V19.0.0 front page)

  • TR 26.927 (Study on Artificial Intelligence and Machine Learning in 5G media services) covers the functional side: media-based AI/ML use cases (object recognition in image and video, video quality enhancement in streaming, crowd-sourced media capture, natural language processing on speech) and the media service architecture for AI/ML, including split-inference configurations and model delivery (TR 26.927 V19.0.0 front page (title); TR 26.847 V19.0.0 clause 1; TR 26.927 V19.0.0 clause 4.2.2; TR 26.927 V19.0.0 clauses 4.2.2, 4.2.3, 4.2.4, 4.2.5; TR 26.927 V19.0.0 clause 4.2.5; clause 5; TR 26.927 V19.0.0 clause 5.1; clause 5.2.2; clause 5.2.3).
  • TR 26.847 (Evaluation of Artificial Intelligence and Machine Learning in 5G media services) is the evaluation companion: testbed architectures and anchors for split inferencing and model-data transmission, evaluation metrics, and scenarios such as compressed AI/ML model transfer for automatic speech recognition, split inferencing for object detection and labelling, bit-incremental model transmission and deployment, and real-time translation over IMS (TR 26.847 V19.0.0 clauses 5.2 to 5.5). The AI/ML Evaluation Framework reference tooling is aligned with this report (TR 26.847 V19.0.0 front page (title); TR 26.847 V19.0.0 clauses 4.1, 4.2, 4.3; TR 26.847 V19.0.0 clause 4.4; clause 5.2).

TR 22.874 (SA1, Study on traffic characteristics and performance requirements for AI/ML model transfer in 5GS) is a companion study, on the network-transport side of AI/ML rather than the media-processing side: it looks at how model-transfer traffic itself behaves on the 5G system. Both this study and TR 26.847 feed 5G-MAG's early input to 6G, where "Artificial Intelligence and Communication" is one of the six usage scenarios of IMT-2030, a scenario that "would support distributed computing and AI applications" (Recommendation ITU-R M.2160-0 clause 3); see Towards 6G Media for that wider context (TR 22.874 V18.2.0 front page (title); clause 1; TR 22.874 V18.2.0 clause 1; clause 4). The Stage 1 requirements for AI/ML model transfer are in TS 22.261, clause 6.40 (requirements) and clause 7.10 (KPIs), from Release 18 (work item AIML_MT, TR 21.918 clause 11.1; TS 22.261 V18.19.0); TS 22.261 V19.15.0 adds requirements and KPIs for direct device connection (clauses 6.40.2.2 and 7.10.2), recorded as AIML-Ph2 change requests CR 0713 and CR 0714 in V19.4.0.

Related: Towards 6G Media