Page 29 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

As stated in the framework in Recommendation ITU-T from the context, system and human influence factors,
E.802, for any QoS management model to be successful, it together with the subjective information about the user’s
is crucial that the identification of the QoS criteria is requirements and expectations gathered from surveys (i.e.
relevant to the users (based on their the training set in Figure 8). Once the KPIs have been
requirements/expectations). In our previous work, we have determined, the related KPPs will be specified for NP
considered the four models proposed in Recommendation measurements based on the CoS and the ITU-T framework
ITU-T E.802 for the identification of the QoS criteria. In guidelines. At this stage, once the KPPs have been
the various NGW scenarios, it is essential to determine the identified, the control of the “intrinsic QoS” may start
different QoE dimensions and influence factors that may (Control Plane in Figure 8). It is recommended to enable
aid in identifying the user´s requirements. In the 5G unsupervised ML techniques from the NP-gathered data to
wireless scenario, the network´s response will be dependent infer adequate radio/channel selection and detect faults and
on the users’ behavior and many other contextual and non- anomalies in the network behavior. This constitutes the first
contextual agents not controlled by the providers, which intervention point where corrective actions may be
may have a considerable effect on the user´s final implemented to enhance the QoS (Anomaly
satisfaction. For this reason, the first step proposed in the Detection/Fault* in Figure 8).
methodology is to understand the users’ behavior in each of
the scenarios to identify their requirements/expectations For the study of the QoE (User Plane in Figure 8), the
and determine the relevant QoS criteria. Therefore, since proposed method considers using also both the objective
this first step is critical for the QoS management model to and subjective gathered data. Supervised machine-learning
succeed, a combination of both contextual and non- algorithms are suggested (regression model) for the
contextual information is to be gathered through big data correlation of the NP/QoE [10] using the survey’s results
analysis. Unsupervised ML techniques (clustering) are and the NP collected data from probes, the correlation
proposed for inferring the different scenarios/profiles and function between the QoE and NP can be deduced (see
for finding the user’s context influence factors (the context NP/QoE correlation model in Figure 8). The context and
extraction in Figure 8). In addition, inductive supervised non-context influence factors will be also crucial when
learning is suggested to infer the rules to identify the QoS analyzing the user’s QoE and for this reason, they feed
criteria and KQI relevant for the users. In this way, the again the training set to learn the rules that will provide the
complex procedure of surveying can be avoided except for predicted QoE. The ML will avoid repeating the surveying
the initial training period for the ML models to capture complex process except for the time to capture the
influence factors and user’s requirements. Thus, the survey necessary observation data that will feed the training set
results will provide the particular cases of observation to process. This training set will be used to deduce the rules
draw the general rules predicted from a training set drawn that will control the NP/QoE correlation model.

Policy & Business
Monitoring
KBO* Entity
* (Management Plane)

QoE/SAT KRI* Requirements/Exp. Contextual
Data Collection Data Collection
(Surveys) (Surveys) &
Expectation
Monitoring
Predicted QoE KQI QoS Criteria Entity
(User Plane)
Rules NP/QoE correlation Model Rules
Criteria Criteria
NP
NP Data Collection
Monitoring
KPI Entity
Control & Training Context extraction
(Control Plane)
- Profiles/Scenarios - Human Influence Factors
- Radio/Channel Learning - Context Influence Factors Training Machine
- Anomaly Detection/Fault* - System Influence Factors Set Learning
Entity
Unsupervised Learning Supervised Learning
Figure 8. Proposed methodology for machine learning in 5G

– 13 –