IEEE Computational Intelligence Magazine - May 2018 - 56

Processing complexity is shifted to the creation
of the prediction model. By using the patterns
found in the historical data, we can make intelligent
decisions faster. This allows to perform multi-objective
optimization in real-time.
1) machine learning for data analytics,
and 2) multi-objective optimization.
Each is depicted in Figures 3 and 4,
respectively, and described in the subsequent subsections.

with the links between each cell and the
SON coordinator and they must pass
through the interface-N (Itf-N). We
propose to solve the challenge of coordination among SON functions by analyzing their interactions based on the
measurement history of the network. To
do this, we design the self-coordination
framework based on two main functions:

Collecting
Data

a. Machine Learning for Data analytics

The main objective of this function is to
analyze large amounts of data sets to

Processing
Data

Partitioning
Data

Training Data

Testing Data

Bagged-SVM

Model
Validation

Model

Figure 3 Machine learning for data analytics.

s1, . . ., s psize
No

Reach
Stopping
Criteria

Collect n ′ Measurements
(New Data)
Yes
Model

Prediction Function
Pareto Front
Metric Predicted

Genetic Operations

Figure 4 Multi-objective optimization.

IEEE ComputatIonal IntEllIgEnCE magazInE | may 2018

uncover hidden patterns, correlations
and other useful information to make
better decisions. This is done by considering the UE measurement reports,
which contain power and quality measurements from the serving and neighbor base stations:
❏ Reference Signal Received Power
(RSRP): Average power received
from LTE reference signals, used for
channel estimation and handover/
cell selection.
❏ Reference Signal Received Quality
(RSRQ): Ratio between the total
power received from reference signals
and the total power received in the
full bandwidth. The RSRQ measurement provides additional information about interference levels to
make a reliable handover/cell reselection decision.
Once the data have been collected,
we perform a data preparation process
to extract and analyze the radio measurements. To do that, we propose to
make use of ML techniques, which have
been demonstrated to be very effective
for making predictions based on observations. Among the most important
ML tools for data analysis, we focus on
regression analysis. It is important to
note that evaluating the network performance for a given configuration requires a dynamic system-level simulation,
which requires high computational time.
If several configurations need to be
assessed during the optimization, the
computational cost is indeed prohibitive.
So, regression analysis shifts the processing complexity to the model creation
stage and allows fast performance evaluations when optimization is running.
Regression analysis is an ML technique, which allows us to predict the
performance metric of each SON function. Regression takes an input vector
(x) and an output value (y) to develop a
predictive model, returning the predicted
output yt . We represent the input space
by an n-dimensional input vector
x = (x (1), f, x (n)) T ! R n, where each
dimension is an input variable. The size
of the input space is [x # n]. The number of rows is the number of UEs at any
place and anywhere, and the number of

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2018