Artificial Intelligence / Machine Learning Model Inference

Description

AI/ML Model Inference is the operational phase of the AI/ML lifecycle in which a previously trained model is applied to new input data to generate predictions, classifications, or control decisions. In the context of 3GPP networks, inference can be hosted at various locations: within the gNB (e.g., for scheduling, beam management, or mobility optimization), at the core network (e.g., for network slicing management or anomaly detection), or at an external AI/ML server. The 3GPP architecture defined in TS 38.401 and studied in TR 38.843 specifies how inference functions interact with network functions through standardized interfaces, enabling interoperability between vendors.

The inference process begins with the collection of input data, which may include radio measurements, UE context information, traffic statistics, or network performance metrics. These inputs are pre-processed and fed into the trained model, which then computes an output. The output could be a prediction (e.g., future load), a classification (e.g., failure type), or a recommended action (e.g., handover decision). The inference result is then consumed by the requesting network function or application. Latency, accuracy, and resource consumption are key performance considerations, especially for real-time RAN use cases where inference must complete within milliseconds.

3GPP defines several architectural options for AI/ML inference. In the RAN, inference can be executed inside the gNB-CU or gNB-DU, or split across them. The study in TR 38.843 identifies functional splits and data collection mechanisms, including the possibility of using the E2 interface for near-real-time RIC-based inference in O-RAN contexts. For core network AI/ML, inference may be hosted in a dedicated network function (e.g., NWDAF) or in an application function. Standardized service-based interfaces allow AI/ML capabilities to be exposed as services.

A critical aspect of inference is the management of model lifecycle. 3GPP specifies procedures for model deployment, activation, deactivation, and monitoring. The network may perform inference using a model trained internally or provided by an external party. Model performance monitoring is essential to detect degradation and trigger retraining. The inference function may also provide feedback to the training function, closing the loop for continuous improvement.

In Rel-19, AI/ML Model Inference is a foundational concept for the broader AI/ML framework in 5G-Advanced. It enables use cases such as network energy saving, load balancing, mobility optimization, and beam management. By standardizing inference procedures, 3GPP ensures that AI/ML can be integrated into multi-vendor networks, paving the way for autonomous network operation and the evolution toward 6G.

Purpose & Motivation

AI/ML Model Inference exists to bring the benefits of machine learning into live network operations. Traditional rule-based algorithms struggle to adapt to the dynamic and complex conditions of 5G networks, where traffic patterns, radio environments, and service requirements change rapidly. Inference allows networks to make data-driven decisions in real time, improving efficiency and user experience.

The motivation for standardizing inference in 3GPP is to enable interoperability and multi-vendor deployments. Without common procedures, AI/ML would remain proprietary and siloed. By defining how inference functions interact with network functions, 3GPP ensures that an AI/ML model trained by one vendor can be executed on another vendor's infrastructure, fostering a competitive ecosystem.

Furthermore, inference addresses the need for low-latency, localized decision-making. Centralized AI/ML may introduce transport delays, so 3GPP supports distributed inference at the edge, such as in the RAN. This is particularly important for use cases like beam management and scheduling, where decisions must be made within a few milliseconds. Overall, AI/ML Model Inference is a key enabler for the intelligent, automated networks of the future.