Page 6 - Highlight_Issue_2_FLIP_BOOK

TECHNICAL HIGHLIGHTS

ARTIFICIAL INTELLIGENCE
AND MACHINE LEARNING IN

NG-RAN: NEW STUDY IN RAN3

By the 3GPP Working Group
RAN3 leadership

(Gino Masini, Sasha Sirotkin, Yin Gao)

5G brings more stringent requirements for Key Performance AI can be broadly defined as getting computers to perform tasks
Indicators (KPIs) like latency, reliability, user experience, and regarded as uniquely human. ML is one category of AI techniques:
others; jointly optimizing those KPIs is becoming more challenging a large and somewhat loosely defined area of computer algorithms
due to the increased complexity of foreseen deployments. able to automatically improve their performance without explicit
programming. AI algorithms were first conceived circa 1950,
Operators and vendors are now turning their attention to but only in recent years ML has become very popular partly due
Artificial Intelligence and Machine Learning (AI/ML) to address to massive advancements in computational power and to the
this challenge. For this reason, following RAN plenary approval, possibility to store vast amounts of data. ML techniques have made
3GPP RAN3 has recently started a new Release-17 study on the tremendous progress in fields such as computer vision, natural
applications of AI/ML to RAN.
language processing, and others.

ML algorithms can be divided into the following types:
• Supervised learning: given a training labeled data and desired • Reinforcement learning (RL): unlike the other types, which
output, the algorithms produce a function which can be used include a training phase (typically performed offline) and
to predict the output. In other words, supervised learning an inference phase (typically performed in “real time”), this
algorithms infer a generalized rule that maps inputs to outputs. approach is based on “real-time” interaction between an agent
Most Deep Learning approaches are also based on supervised and the environment. The agent performs a certain action
learning. changing the state of the system, which leads to a “reward”
or a “penalty”.
• Unsupervised learning: given some training data without
pre-existing labels, the algorithms can search for patterns to
uncover useful information.

Perhaps the most obvious candidate for AI/ML in RAN is Self- The study has just begun, and at the time of writing we can
Organizing Networks (SON) functionality, currently part of LTE and only provide initial considerations. According to the mandate
NR specifications (it was initially introduced in Rel-8 for LTE). With received from RAN, our study focuses on the functionality and the
SON, the network self-adjusts and fine-tunes a range of parameters corresponding types of inputs and outputs (massive data collected
according to the different radio and traffic conditions, alleviating from RAN, core network, and terminals), and on potential impacts
the burden of manual optimization for the operator. While the on existing nodes and interfaces; the detailed AI/ML algorithms
algorithms behind SON functions are not standardized in 3GPP, are out of RAN3 scope. Within the RAN architecture defined in
SON implementations are typically rule-based. One of the main RAN3, this study prioritizes NG-RAN, including EN-DC. In terms
differences between SON and an AI-based approach is the switch of use cases, the group has agreed to start with energy saving, load
from a reactive paradigm to a proactive one. balancing, and mobility optimization. Although the importance of
avoiding a duplication of SON was recognized, additional use cases
may be discussed as the study progresses, according to companies’
contributions. The aim is to define a framework for AI/ML within
the current NG-RAN architecture, and the AI/ML workflow being
discussed should not prevent “thinking beyond”, if a use case
requires so.

Stay tuned for further updates as the study progresses in RAN3, or
consider joining us in our journey into the “uncharted” territory of
AI/ML in NG-RAN.
https://www.3gpp.org/specifications-groups

|
06 3GPP Highlights newsletter