Page 20 - Proceedings of the 2018 ITU Kaleidoscope

Other searches for patterns will be “Unsupervised”. In these cases, we have no information about the
exits that the systems could bring, an example of this are the models of congestion because the call
is an aleatory event with no previous information about the channel status.

How do we use machine learning in 5G planning and deployment?
5G is a change of concept in mobile services; this new concept will assume new frequencies to
provide different kinds of services routing communications based on the speed needed for the service
looking to obtain the maximum performance combining coverage, propagation and the penetration
bandwidth compatible with the kind of service to be used. This service will combine different
techniques such as frequency allocation, carrier aggregation or MIMO (multiple in multiple out) as
an example.
The knowledge obtained from 4G analysis performs the basis for the 5G model.

To design the new model the definition of segmented models is highly significant, where we will test
the obtained predictive functions, testing the needs of coverage and based on service areas and
demand of services.

Once the type of demand of service can be managed by changing the variable values, we will
introduce new frequencies and services looking for the reinforcement of services in the mapped areas
where the model shows there is a lack of services.

The definition of this new model gives us the opportunity of introducing two new learning algorithm
types, such as reinforcement learning. This introduces feedback to correct and learn the ways to
optimize selected variables and to multitask the learning of algorithms using homologous theory
where the system learns over other solutions obtained in the past based on similar conditions.
Once this kind of model has been developed, we could develop Bayesian networks to predict the most
probable behavior of the designed network.

Network deployment using this tool will allow the measuring of the values predicted for the system
and the values obtained will be used to provide feedback to the systems providing new entrances to
validate the model and reinforce the learning capabilities of the system based in machine learning.
It is very important to define the scenario of where we are working, starting with the definition of the
components of the system identifying the variables with incidence in the work with the functions that
predict the behavior of these variables.
All these variables with an undefined standard set of values or logical status will be played using
Montecarlo models, Markov chains or other statistical sources, based in the universe of the variable
and its statistical patterns.

The use of described techniques will provide three stages of learning: “Grow”, learning from the
environment; “Restructuring”, learning from corrections and obtaining new knowledge form this
action; and adjustment generalizing concepts and adjusting from the values obtained from the
experience sensing of the real world.
5G will need a great quantity of sites with different frequencies and services. This situation will create
a complex and multivariable scenario.
Machine learning gives us the tools to define the patterns in this multivariable scenario, showing even
those patterns for which we do not know of their existence.
This planification, testing adjustment and network deployment, looking for the better balance
between power, height of antennas, location and density of the network crossed with the aleatory
demand of services in the universe of individual users and types of devices could not be solved
without described tools.

– xvi –