IEEE Computational Intelligence Magazine - May 2018 - 58

Ensemble learning produces more accurate predictions
than a single regression model would. The bagged
support vector machines method is used to create
our predictor.
(GAs), which are heuristic algorithms
that can adapt their objective functions
to the multi-objective problem. GAs are
able to improve partial solutions since
they create new individuals by performing selection, crossover and mutation,
but at the same time they keep the best
individuals (i.e., the elitism). In this
regard, NSGA-II allows less computational complexity than other evolutionary algorithms, and it also prevents the
loss of good solutions once they are
found since it preserves the elitism.
The concept of GA was introduced
by Holland in [23]. GAs are inspired by
the evolutionary theory explaining the
natural selection. In GAs, a solution
vector is called a chromosome, which is
represented by a series of genes made of
discrete units. Each unit controls one or
more of the chromosome features. A
chromosome corresponds to a unique
solution s in the solution space S. GAs
manage a set of chromosomes or population, which is usually initialized by
random valid solutions. As the search
evolves, the population eventually converges to a single solution. To do that,
the two most important operators are
crossover and mutation. Crossover combines two parent chromosomes to form
two new solutions called offspring. By
iteratively applying the crossover operator, the best offsprings appear more frequently in the population, leading to
convergence toward a good solution.
On the other hand, mutation introduces
genetic diversity in the population by
means of random alterations in the offspring. The probability of mutation is
very small and depends on the length of
the chromosome. Therefore, the new
chromosome produced by mutation
will not be very different from the original one.
Given this, the procedure is adapted
to our particular problem as follows:

NSGA-II starts with a random initial
population s 1, f, s psize of size p size .
Each feasible solution corresponds to a
vector of size M, where each value
denotes the network parameter to be
optimized, which is associated with
each cell. In order to evaluate the fitness of each chromosome, in each iteration t, we collect nl measurements at
some arbitrary points in the scenario.
These measurements are obtained as a
consequence of having configured the
parameters of the scenario according to
each chromosome s. These nl measurements (new data) and the built
model already described in Section
III-A, are the inputs to the prediction
function, which gives us the UE performance. As a result, for each s ! S,
we obtain a predicted performance
metric. The algorithm applies crossover
and mutation to create an offspring
population. In generation t, an offspring population of size p size is created
from the parent population, and nondominated fronts F1, F2, f, FK are
identified in the combined population.
The next population is filled starting
from solutions in F1, then F2, and so
on. A high-level overview of the process is depicted in Figure 4.
In summary, we first exploit the huge
amount of data already available in the
network to predict future network performance. We build a prediction model
based on historical UE measurements
and we apply regression analysis techniques to predict network performance.
Second, we use the built model as an
input of the multi-objective evolutionary algorithm to solve potential conflicts
by finding a set of solutions that satisfies
the objectives without being dominated
by any other solution. In this way, the
gain in time is substantial since the built
model allows fast performance evaluation when optimization is running.

IEEE ComputatIonal IntEllIgEnCE magazInE | may 2018

IV. MLB and MRO Function Conflict

In this section, we analyze in detail the
MLB-MRO SON conflict. First, we
explain the handover triggering procedure. Second, we explain the reason for
the conflict, and finally, the considered
performance indicators.
a. Handover Triggering Procedure

Among the different handover triggering procedures in LTE, we focus on
Event A3, which is the main criterion to
manage intra-LTE mobility. Event A3 is
defined as the situation in which the UE
perceives that a neighbor cell's RSRP is
better than the serving cell's RSRP by a
certain margin [24]. In order to reduce
the ping-pong effect, measurements
used to assess the event are averaged,
hysteresis margins are introduced and
the conditions must be met during the
so-called TTT. Hence, the event entering condition is defined as
RSRPnc 2 RSRPsc + offset
+ Hysteresis - CIO
where RSRPnc and RSRPsc are the averaged reference signal power strengths
of the neighbor and the serving cell,
respectively, while offset and hysteresis
parameters cause the serving and neighbor cells to be more and less attractive,
respectively [24]. The combination of
both defines a net hysteresis margin
that delays notification of the event to
guarantee that the neighbor cell is now
the dominant one. These values are
used as a basis and can be modified to
adapt the condition to a particular UE
mobility status, for example, being
reduced for a high-speed UE. Finally,
the cell individual offset, or CIO, is a
cell-specific parameter set by the serving cell for each of its neighboring
cells. It is used for load management,
since the higher its value, the more
attractive the corresponding neighbor
will be. The whole process can be appreciated in Figure 5a, which shows
the trace of average RSRP measurements from the serving cell and a
neighbor before and after a handover.
On the other hand, Figure 5b illustrates
the difference between the hysteresis

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