User Equipment-side Artificial Intelligence/Machine Learning Model

Description

A UE-side (AI/ML) model is an artificial intelligence or machine learning model that is executed entirely on the User Equipment (UE), with all inference computations performed locally. This contrasts with network-side AI/ML models, where inference occurs at the base station or core network. The UE-side model leverages the UE's own computational resources to make predictions, classifications, or decisions that can optimize various aspects of the UE's operation, such as beam management, mobility, channel state information (CSI) feedback, and power control.

In the context of 3GPP, the introduction of AI/ML capabilities has been a major theme in Releases 17 and 18, with Release 19 further specifying the framework for UE-side models. The specifications 38.300 (NR overall description) and 38.843 (AI/ML for NR air interface) provide the foundational definitions and requirements. The UE-side model is trained either offline or online, and its parameters are either pre-configured, downloaded from the network, or learned on the device. The model's inputs are typically derived from UE measurements and sensor data, and its outputs are used to adjust UE behavior.

The architecture for UE-side AI/ML models involves several key components: the model itself (e.g., a neural network), the inference engine, and the interface to the UE's radio protocols. The model may be specified by the network (e.g., a standardized model) or may be proprietary to the UE vendor. In the latter case, the network may not be aware of the model's internal details, but it can still benefit from the improved performance. The 3GPP framework supports both cases, with mechanisms for the network to assist or control the UE-side model if needed.

One of the main challenges of UE-side AI/ML is the limited computational power and battery life of UEs. Therefore, the models must be lightweight and energy-efficient. Techniques such as model quantization, pruning, and knowledge distillation are often employed. Additionally, the UE must manage the trade-off between inference accuracy and resource consumption. The 3GPP specifications define performance requirements and testing methodologies to ensure that UE-side models meet the necessary standards.

UE-side AI/ML models can be applied to a variety of use cases. For example, in beam management, the UE can use a model to predict the best beam direction based on past measurements, reducing beam sweeping overhead. In mobility, the UE can predict handover events and prepare in advance. In CSI feedback, the model can compress channel information more efficiently. These applications can lead to significant improvements in spectral efficiency, latency, and reliability.

The standardization of UE-side AI/ML models in Release 19 is a significant step towards intelligent, adaptive UEs that can optimize their own performance in real-time. It complements network-side AI/ML and enables a distributed intelligence architecture where both the network and the UE contribute to overall system optimization.

Purpose & Motivation

The UE-side AI/ML model was introduced to harness the increasing computational capabilities of modern UEs to improve radio performance and user experience. Traditional UE operations rely on predefined algorithms and rules that may not adapt well to dynamic environments. AI/ML models can learn from data and make more accurate predictions, leading to better decisions in areas such as beam selection, mobility, and link adaptation.

By performing inference locally, the UE can react quickly to changing conditions without the latency of sending data to the network for processing. This is particularly important for time-critical applications and for reducing signaling overhead. Moreover, UE-side models can be personalized to the specific usage patterns and environment of the user, offering a level of optimization that network-side models cannot easily achieve.

The motivation for standardizing UE-side AI/ML in 3GPP stems from the need to ensure interoperability and performance across different UE vendors and network deployments. Without standards, each vendor might implement proprietary solutions that could lead to inconsistent behavior. The 3GPP framework provides a common understanding of how UE-side models interact with the network, including signaling for model management, performance monitoring, and fallback mechanisms.

Release 19 builds on the study phase in Release 18, where the use cases and requirements were identified. The normative work in Release 19 specifies the procedures and signaling for UE-side AI/ML, enabling commercial deployment. This is part of the broader trend towards AI-native air interfaces in 5G-Advanced and 6G.