IEEE Computational Intelligence Magazine - May 2018 - 53

functions, the self-coordination concept
was introduced in Release 10 [3].
In this paper, we discuss the selfcoordination problem. In particular, we
focus on the output parameter SON
conflict, i.e., when two or more SON
functions aim at adjusting the same output parameter with opposite values. So, a
SON coordinator controlling the actions
of the SON functions during operation
is considered a necessity [4]. In order to
evaluate the performance of the proposed scheme, we focus on the handover
management issue. We consider a wellknown SON conflict between mobility
load balancing (MLB) and mobility
robustness optimization (MRO). MLB
and MRO are two of the most important self-optimization functions that deal
with mobility management. Both mechanisms modify the behavior of handover,
which is a procedure that allows connections to be transferred between base stations in a seamless manner. However,
each one pursues a different objective:
❏ MLB aims at balancing traffic load
among cells so that cells with excess
traffic (congested cells) can transfer
some of their users to less loaded
neighboring cells and vice versa. This
can be done by changing the parameters that govern UE cell selection,
such as handover thresholds, hysteresis
margins and times to trigger a handover event. The primary goal is to
achieve a higher system capacity, and
this is done by distributing UE traffic
across the available radio resources in
the system.
❏ On the other hand, MRO is designed to improve mobility robustness:
* Minimization of call drops due to
radio link failures: Depending on
how the handover parameters are
adjusted, too-early handovers may
happen for some users. This means
that the communication fails due
to high propagation losses with
the new cell. Similarly, too-late
handovers may imply that communication with the serving base
station is lost before a new connection is established.
* Minimization of unnecessary handovers: Too-short time-of-stays in the

to the use of a variety of techniques
from data mining, statistics and ML, it is
possible to analyze historical data to
make predictions about unknown future
events. This is known as predictive analytics and for the current work, it turns
the network management from reactive
to predictive. In this context, big data
analytics are currently receiving big
attention due to their capability to provide insightful information from the
analysis of high volumes of data that are
readily available for operators.
Based on that, the motivation behind
this paper is to provide a tool that allows
mobile network operators to become
proactive by anticipating behaviors and
making decisions accordingly. We focus
on building a tool for efficient self-coordination that is based on the network
performance prediction that is made by
doing a proper data analysis of UE measurements. The tool works in two steps
as graphically depicted in Figure 1:
1) In the first step, the proposed scheme
learns from past experience to obtain
a network performance predictor for
each SON function individually. In
particular, ML is used for data analytics. Available measurements are globally utilized in a learning process that
yields an estimator by means of
regression models.
2) Then, performance predictions are
the inputs of a multi-objective optimization process, which searches for a
set of non-dominated solutions or
Pareto front.That is to say, the solutions

new cell or ping-pong (quick handover back to the previous cell) should
not happen. For example, a vehicular user moving in a city where
macrocells coexist with a layer of
small cells should utilize a larger
timer to trigger handovers toward
picocells. This should avoid handovers from macro- to picocells, where
vehicular users with high speed
would have very short time-ofstays. Note that this would cause
data throughput degradation due to
the high volume of signaling transmission required to repeatedly update the serving cell.
Based on the information above, it is
clear that MLB and MRO are two independent functions with independent
objectives. However, if both are applied
in parallel and without coordination, the
actions requested by each SON function
may be different. They may cause opposite changes in handover parameters, and
so, they may enter into a conflict. Under
these circumstances, the system would
enter into a cycle of constant re-configuration, thereby causing performance
degradation due to excessive signaling
that requires radio resources.
In this context, machine learning
(ML) is proposed as a candidate tool that
allows the network to learn from experience and solve conflicts in an effective
manner. The main feature is that the
network is able to run different SON
functions in parallel and improve the
system performance. In particular, thanks

Self-Optimization
Self-Coordination Framework
Data

Machine Learning for
Data Analytics
Predictive Model
Multi-Objective Optimization

Self-Organizing Network
Self-Configuration

Decisions

Pareto Front

Self-Healing

Figure 1 Self-coordination framework.

may 2018 | IEEE ComputatIonal IntEllIgEnCE magazInE

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