Page 27 - Proceedings of the 2018 ITU Kaleidoscope
P. 27
Machine learning for a 5G future
Figure 3. Machine-learning techniques and paradigms in 5G [11]
In this sense, a detailed description of the different ML In this context, ITU-T SG-13 has developed several
algorithms and their application in the 5G environment is Recommendations related to 5G QoS frameworks and ML
presented in [11] (Figure 3). The authors of this paper note techniques [4, 12] (Figure 4). In addition, ITU-T SG-12, the
the importance of selecting the appropriate learning type expert group responsible for the development of
when using ML techniques since the goal of ML is to international standards for QoS and QoE, has been active in
predict the output of an input turning observational data the editing and updating of related Recommendations to
into a model that can be used for this prediction. Therefore, work around new mobile scenarios [6, 7]. The 5G-PPP
depending on the nature of the observational data, different partnership, initiative between the European Commission
types of learning can be distinguished: and European ICT industry, is also working actively to
define architectures, technologies and standards for NGW
In supervised learning, each input value of the systems.
observational data is given with the corresponding
output to form a training set. This set is used to In spite of the significant advances in the definition of the
learn a predictive function. Supervised learning is 5G ecosystem and the identification of new QoE models
useful for classification and for regression and ML techniques for this environment, there is still a
problems.
In unsupervised learning, only the input values are need for global QoS management models to be deployed in
real-world NGW scenarios. In the following sections, an
included in the observational data. The most approach to solving this shortage is presented.
common application of unsupervised learning is
cluster analysis to find similarities between the IMT-2020 QoS Monitoring and Data Collection
input values and extract hidden patterns to group Static and dynamic QoS information data
them into clusters. Machine learning based IMT-2020 assurance
Reinforcement learning is based on dynamic functional entity
iterative learning and decision-making processes. Machine learning based IMT-2020 QoS
The learner is not told which actions to take but assurance modeling
instead must determine those that yield the output Machine learning based IMT-2020 QoS
closest to the target by successive trials. assurance training
2.2 Standardization: Key to 5G IMT-2020 QoS anomaly detector
IMT-2020 QoS anomaly predictor
Standardization bodies also envision the importance of
defining new standards on QoS and ML for NGW IMT-2020 QoS reporting of QoS anomaly
ecosystems. In January 2018, a workshop on "Machine detection and prediction results IMT-2020
Learning for 5G and beyond" was held in Geneva IMT-2020 Control Plane Management Plane
(Switzerland) in the context of the first meeting of the IMT-2020 QoS Control Plane Function QoS Management
IMT-2020
Anomaly
recently-launched ITU “Focus Group on Machine Learning QoS Data QoS Control Detection QoS Capabilities Plane Function
Modeling
Exposure
Collection
for Future Networks including 5G” [5]. Three different &Training &Prediction Interface IMT-2020 QoS
Policy
working groups were established: IMT-2020 User Plane IMT-2020
SLA
WG1: Use cases, services & requirements IMT-2020 QoS User Plane Function QoS QoS IMT-2020 Plane
QoE
QoS Function
QoS
WG2: Data formats & ML technologies Mapping Awareness Enforcement Monitoring Management
WG3: ML-aware network architecture
Figure 4. ITU model of ML-based QoS assurance [4, 12]
– 11 –