Network-side Artificial Intelligence / Machine Learning Model

Description

A Network-side (AI/ML) model is a machine learning model that operates entirely within the network infrastructure, performing inference (the process of using a trained model to make predictions or decisions) at network nodes such as the gNB, core network functions, or management systems. This contrasts with UE-side AI/ML models, where inference is executed on the user device. The network-side approach centralizes intelligence, allowing for comprehensive data collection from multiple sources, including UEs, base stations, and core network elements, to train and run models that optimize various aspects of network performance.

In 3GPP, the concept of AI/ML for the radio access network (RAN) was studied in Release 18 and specified in Release 19 under the broader AI/ML framework for NG-RAN. The network-side model is one of the key deployment scenarios. The architecture typically involves an AI/ML entity that can reside in the gNB (e.g., in the RAN Intelligent Controller or as part of the gNB's internal functions) or in the core network (e.g., in a Network Data Analytics Function, NWDAF). The model uses input data such as channel state information, traffic patterns, UE measurements, and network configuration to produce outputs like beamforming weights, handover decisions, resource allocation, or energy-saving parameters. The inference is performed at the network, meaning the computational load is on the network side, which is advantageous because network nodes typically have more processing power and energy resources compared to UEs.

The network-side model can be trained offline using historical data or online using real-time data. Training may occur in a centralized location (e.g., a cloud server) and then the model is deployed to network nodes for inference. Alternatively, online learning can be performed at the edge. The 3GPP specifications (TS 38.300 and TS 38.843) outline the functional framework, including data collection, model management, and inference execution. The model can be used for a variety of use cases: network energy saving, load balancing, mobility optimization, beam management, and quality of experience (QoE) enhancement. For example, a network-side model can predict traffic congestion and proactively adjust handover parameters or antenna tilts to balance load across cells.

Key components of a network-side AI/ML model include the model itself (e.g., a neural network or decision tree), the inference engine that executes the model, and the data pipeline that feeds input features. The model may be hosted in a centralized unit (CU) or distributed unit (DU) of the gNB, or in a core network function. The 3GPP framework supports model lifecycle management, including model training, testing, deployment, monitoring, and updating. Interoperability is ensured through standardized interfaces for data collection and model transfer. The network-side model can also collaborate with UE-side models, where the UE performs some inference and reports results to the network, but the network-side model remains the primary decision-maker.

The introduction of network-side AI/ML models in Release 19 reflects the industry's push towards autonomous networks and the use of AI/ML to handle the increasing complexity of 5G and beyond. By performing inference entirely at the network, operators can leverage powerful computing resources and global data views, leading to more accurate and timely optimizations. This approach also simplifies UE design and reduces UE power consumption, as UEs are not burdened with heavy AI/ML computations. As networks evolve towards 6G, network-side AI/ML models are expected to become even more integral, enabling self-configuring, self-optimizing, and self-healing networks.

Purpose & Motivation

The network-side AI/ML model exists to bring the benefits of artificial intelligence and machine learning to network operations without imposing computational burdens on user devices. Traditional network optimization relies on rule-based algorithms and manual configuration, which are increasingly inadequate for the dynamic and complex nature of 5G networks with diverse services, massive connectivity, and varying channel conditions. AI/ML models can learn from data and adapt to changing conditions, but running these models on UEs would consume battery and processing power, and would limit the scope of data available. By performing inference entirely at the network, operators can aggregate data from many sources, train sophisticated models, and make centralized or distributed decisions quickly. This approach addresses the limitations of previous approaches that were static and reactive. The Release 19 framework for network-side AI/ML models provides a standardized way to integrate AI/ML into NG-RAN, ensuring interoperability and enabling multi-vendor deployments. It solves the problem of how to manage and execute AI/ML models in a standardized manner, facilitating innovation and efficiency in network management.