Page 95 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

service is very important for 5G because the number of overview of AI and machine learning related technological
configurable parameters in 5G would be fairly large, 2000 discussion being carried out in ITU, ETSI, ISO/IEC, and
or more [3]. Bandwidth control for eMBB and latency TM Forum.
control for URLLC services can be performed with
reinforcement learning and deep learning. Reinforcement 5.1 ITU FG-ML5G
learning, especially multi-armed bandit, can be used to The ITU Focus Group on Machine Learning for Future
model a resource arbitration problem in which a fixed
limited amount of resources have to be proportionally Networks including 5G (FG-ML5G) [13] has been
established in November 2017 to study network
allocated to competitive services whose properties are only architectures, protocols, interfaces, use cases, algorithm and
partially known at the time of resource allocation [9].
data formats for the adoption of machine learning methods
in 5G and future networks. It is an open platform for
Monitoring, fault detection and security involve functions experts from ITU members and non-members to quickly
for sensing the system, collecting huge amounts of syslog progress studies on machine learning methods for networks.
and performance data, classification and clustering of data In its lifetime of one year (which can be extended if
for contextualization, forecasting usual/unusual user necessary), FG-ML5G is mandated to hold meetings
behavior and network resource utilization trend. For these several times to review the contributions received from
purposes, unsupervised learning techniques such as spectral participants and develop deliverables. It has formed three
clustering, K-mean clustering, and principal component working groups to progress work simultaneously on (1) use
analysis, supervised learning techniques such as support cases, services and requirements; (2) data format and
vector machine and logistic regression, as well as deep techniques; and (3) machine learning aware network
learning and reinforcement learning are applicable. architecture. FG-ML5G has not produced any publicly
available document yet. Its deliverables will be handled
For the detection of network anomaly, principal and over to ITU-T Study Group 13 for further study and
independent component analysis techniques can be used development of formal standards (called ITU-T
because they can easily identify statistically unusual Recommendations) on the basis of these deliverables.
behavior from the system operation data. Similarly, traffic
analysis by various unsupervised learning techniques can 5.2 ETSI ISG ENI
help in the detection of intrusion and spoofing attacks. ETSI has created the Industrial Specification Groups (ISG)
Logistic regression of supervised learning can be used to “Experiential Network Intelligence” (ENI) in February
predict unusual behavior of devices or users based on their 2017 with the purpose of defining a cognitive network
traffic characteristics. Deep learning would help in management architecture based on the “observe-orient-
detecting unprecedented security issues that may come up decide-act” control model [14]. It uses AI techniques and
with new types of services. context-aware policies for dynamically adjusting network
services in response to changing user demands, network
Although in Table 1 and above discussion we have conditions and business goals. It envisions making the
mentioned only the classical methods of machine learning, system capable of learning from its own operations and
they can be modified for improving accuracy, reducing instruction given by human operators (thus called
complexity or their trade-off, when applying to network experiential). It would be instrumental in the automation of
slicing functions. Many variations of their extensions are network configuration and monitoring processes, thus
available in literature. However, their applicability in reducing the operational cost, human errors, and time to
network slicing is yet to be studied. market the service. ISG ENI aims at studying various AI
methods for enabling model-driven architecture for
5. AI AND MACHINE LEARNING IN SDOs AND adaptive and intelligent service operations. The architecture
FORUMS accommodates different types of policies and selects the
best-fitted one to adaptively drive the network system
In several standards development organizations (SDO) and according to the changes in user behavior, service
industrial forums, AI and machine learning techniques are requirements, network conditions, and business goals. ISG
being investigated for enabling systems to make autonomic ENI is open for participants from all ETSI member as well
decision by processing large amounts of data, learning from as non-member organizations that sign ISG Agreements.
own operations, and adapting to changing environment. ISG ENI has already released five specifications (as of June
Although both AI and machine learning are often used 2018) on use case, requirements, context-aware policy
interchangeably, they are subtly different. AI comprises management, terminology and proof-of-concept (PoC)
multi-disciplinary techniques such as machine learning, framework. These deliverables can be accessed freely from
optimization theory, game theory, control theory, and meta- its website [14].
heuristic analytics [7]. Thus, machine learning can be
considered as a form of or subfield of AI that enables Closely related another ISG is Zero-touch network and
machines to learn by themselves by providing them access Service Management (ISG ZSM), which focuses on
to large amounts of data. In this section, we provide an

– 79 –