Machine Learning Model Activation

Description

ML model activation is a management procedure defined in 3GPP Release 19 (21.905) that enables a specific machine learning (ML) model for a particular ML-enabled feature. In 3GPP networks, ML models are trained and deployed to enhance various features, such as beam management, positioning, mobility optimization, and network slicing. A model is typically trained offline or online, validated, and then stored. Activation is the step that associates a model with a feature and makes it the active model for that feature, meaning that when the feature requires inference, this model will be used. This is distinct from ML inference activation, which enables the inference capability itself; model activation selects which model is used.

The activation process involves several steps: selecting a model from a repository, verifying its compatibility with the target feature and network entity, loading the model into the inference function, and configuring parameters. It may also include validation checks to ensure the model meets performance requirements. Once activated, the model is used for inference until it is deactivated or replaced. Activation can be triggered by management functions (e.g., OAM), by the network entity itself (e.g., based on performance monitoring), or by policy. The process is standardized to ensure interoperability across vendors and to support multi-vendor deployments.

In 3GPP architecture, ML model activation is relevant to any network function that uses ML, including gNB, NWDAF, and core network functions. The model lifecycle includes training, testing, deployment, activation, deactivation, and retirement. Activation is a key state in this lifecycle. It allows operators to manage multiple models for the same feature, switching between them as needed (e.g., for different scenarios or as models improve). The activation state is per model per feature; a model can be active for one feature and inactive for another.

Key technical aspects include the model identifier, the feature identifier, and the activation state. The procedure may involve signaling between the ML model management function and the inference function. It also considers resource implications: activating a model may require loading it into memory and allocating compute resources. Deactivation frees these resources. The separation of model activation from inference activation allows flexible management: an inference function can be active but no model active (resulting in no inference), or a model can be active but inference deactivated (model ready but not executing).

Overall, ML model activation is a fundamental concept for managing ML models in 3GPP networks, enabling dynamic selection and use of models to optimize network performance.

Purpose & Motivation

ML model activation exists to provide a standardized way to select and enable a specific ML model for a given feature in a 3GPP network. Before this concept, model deployment was often manual and vendor-specific, leading to interoperability issues and lack of dynamic control. The activation process solves this by defining a clear state transition that network operators can manage, ensuring that the right model is used for the right feature at the right time. It also supports model lifecycle management, allowing models to be updated or replaced without disrupting the feature.

The motivation for model activation is to enable flexibility and adaptability in ML-enabled features. Different scenarios may require different models (e.g., a model trained for high-mobility vs. low-mobility users). Activation allows switching between models based on conditions. It also facilitates A/B testing and gradual rollouts, where a new model can be activated for a subset of entities. Furthermore, activation is essential for multi-vendor environments, where models from different vendors may be used; standardized activation ensures they can be integrated seamlessly.

Another key motivation is resource efficiency. Not all models need to be active at all times; activating only the necessary models saves memory and compute. This is particularly important in edge and RAN nodes with limited resources. The separation from inference activation provides additional control: operators can prepare models (activate them) but keep inference off until needed. Overall, ML model activation is a cornerstone of 3GPP's AI/ML management framework, enabling dynamic, efficient, and interoperable use of ML models.