Artificial Intelligence / Machine Learning Model Training

Description

AI/ML Model Training is the phase in the AI/ML lifecycle where a model is developed by learning patterns from historical or simulated data. In 3GPP networks, training can be performed within the network (e.g., at a gNB, a core network function, or a dedicated AI/ML training server) or externally by a third party. The 3GPP architecture in TS 38.401 and the study in TR 38.843 define how training functions interact with network functions, including data collection, model training, and model transfer to inference hosts.

The training process involves several steps: data collection, data preprocessing, feature engineering, model selection, training algorithm execution, and validation. Data may be collected from various sources, including UE measurements, gNB performance counters, core network analytics, and OAM systems. The quality and quantity of data are critical to model performance. 3GPP specifies mechanisms for data collection, such as the E2 interface for RAN data or the service-based interfaces for core network data.

Training can be centralized or distributed. Centralized training uses a single training node with access to a large dataset, while distributed training may involve multiple nodes collaborating, e.g., via federated learning. 3GPP TR 38.843 studies both approaches, highlighting trade-offs in terms of privacy, latency, and resource efficiency. Federated learning, in particular, allows models to be trained across multiple gNBs without sharing raw data, addressing privacy concerns.

Once training is complete, the model is validated and then deployed to inference hosts. 3GPP defines procedures for model transfer, activation, and version management. The training function may also monitor model performance in the inference phase and trigger retraining when degradation is detected. This closed-loop operation is essential for maintaining model accuracy over time.

In Rel-19, AI/ML Model Training is a cornerstone of the AI/ML framework, enabling use cases such as network energy saving, load balancing, and mobility optimization. By standardizing training procedures, 3GPP ensures that AI/ML models can be developed and shared across multi-vendor networks, facilitating innovation and interoperability.

Purpose & Motivation

AI/ML Model Training exists to create models that can make intelligent decisions in live networks. Without training, AI/ML models cannot learn the complex relationships between network inputs and desired outputs. Traditional network optimization relies on manually crafted rules, which are time-consuming to develop and often fail to generalize. Training automates this process, enabling networks to adapt to changing conditions.

The motivation for standardizing training in 3GPP is to enable multi-vendor AI/ML interoperability and to ensure that training can leverage data from across the network. By defining common procedures and interfaces, 3GPP allows network operators to train models using data from multiple vendors' equipment and deploy them on any compliant infrastructure. This reduces vendor lock-in and fosters a healthy ecosystem.

Additionally, training addresses the need for continuous improvement. As network conditions evolve, models must be retrained to maintain performance. 3GPP specifies mechanisms for monitoring and retraining, ensuring that AI/ML remains effective over the network's lifetime. Overall, AI/ML Model Training is a key enabler for autonomous, self-optimizing networks.