This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
1

Diagnosis based on genetic fuzzy algorithms for LTE Self-Healing

E. J. Khatib1 , R. Barco1 , A. Gomez-Andrades1 , I. Serrano2
1

Universidad de Malaga, Communications Engineering Dept., Malaga, Spain

2 Ericsson, PBO RA Continuous Analysis

Self-Organizing Networks (SON) mechanisms reduce Operational Expenditure (OPEX) in cellular networks, whilst enhancing
the offered quality of service. Within SON, self-healing aims to autonomously solve problems in the radio access network and to
minimize their impact on the user. Self-healing comprises automatic fault detection, root cause analysis, fault compensation and
recovery. This paper presents a root cause analysis system based on fuzzy logic. A genetic algorithm is proposed for learning the
rule base. The proposed method is adapted to the way of reasoning of troubleshooting experts, which ease knowledge acquisition
and system output interpretation. Results show that the obtained results are comparable or even better than those obtained when
the troubleshooting experts define the rules, with the clear benefit of not requiring the experts to define the system. In addition, the
system is robust, since fine tuning of its parameters is not mandatory.
Index TermsSelf-Organizing Networks, self-healing, troubleshooting, root cause analysis, genetic algorithms, supervised learning,
fuzzy systems

I. I NTRODUCTION
In the last decade, cellular mobile networks have grown
rapidly in size and complexity. Operation and maintenance
of these networks is a difficult challenge because of the evergrowing number of users and a demand for reliability. For this
reason, the NGMN Alliance [1] came up with the definition
of Self-Organizing Networks (SON) as a set of principles and
concepts to add automation to mobile networks so that they
require less maintenance than traditional networks while improving service quality. Subsequently, 3GPP proposed a set of
use cases and specific SON functionalities for next generation
mobile communication networks, that is Long Term Evolution
(LTE) [2], and identified SON as one of the key features
of LTE-Advanced (LTE-A) [3]. Three main aspects conform
the SON definition: self-configuration, self-optimization and
self-healing. This study contributes to the research effort in
self-healing [4] [5]. Self-healing encompasses the automatic
functionality aimed to solve problems (traditionally called
troubleshooting) as they emerge in the network. It is a substitute of manual troubleshooting, which currently is the only
way network problems are being confronted. Automatic troubleshooting is a great competitive advantage since offloading
the Operation and Maintenance (O&M) experts from the task
of solving common problems, especially related to the access
networks, greatly reduces Operational Expenditure (OPEX).
The reduced response time of automatic troubleshooting also
increases the quality of service.
Self-healing consists of fault detection, root cause analysis
(diagnosis), compensation and fault recovery. Recently, some
research effort has started to be carried out on these topics. However, so far there are no commercial products that
implement a totally automated troubleshooting tool, although
there is an emerging research effort into the topic. In [6] fault
detection is based on monitoring the throughput. A sector is
considered problematic if it falls under a certain threshold that
is calculated according to the recorded average throughput.

There is no diagnosis stage. The neighboring sectors are

then modified to compensate the problematic sector. A neural
network based approach is described in [7] to detect problems
in 3G networks. The neural networks are trained to recognize
the normal behavior.
In [8] a framework for diagnosis based on bayesian networks is proposed. A Knowledge Acquisition Tool for defining
the parameters of the bayesian network is defined in [9].
The UniverSelf project [10] proposes to combine bayesian
networks with case-based reasoning [11] [12].
The COMMUNE project [13] is targeted towards the study
of uncertainty in mobile networks. It highlights uncertainty as
the main cause that self-x (self-healing amongst others) functions are underdeveloped. In [14] it is determined that bayesian
networks are a complex approximation for troubleshooting,
given the difficulty to obtain expert knowledge as the required
conditional probabilities. Thus, COMMUNE is inclined to
case based reasoning, such as the algorithm described in [15],
where a PI (Performance Indicator) vector is compared against
a database of known cases, and the most similar is selected
as the diagnosed cause. It has an implicit supervised learning
capability, as experts add cases to the database. Some research
effort has also been put into compensation. [16] describes the
problem of cell outage detection and compensation. [17] and
[18] describe a method of compensation based on modifying
the antenna tilt.
Among self-healing functions diagnosis is especially critical, since operators spend long periods of time finding out the
root causes of problems. Based on the previous references,
it can be concluded that although some approaches have
been proposed for diagnosis, they have not been accepted
for implementation in real networks. The main cause of this
failure is that these systems require a high collaboration degree
from the troubleshooting experts, since the knowledge transfer
forces them not only to dedicate time, but also the effort of
learning the topics related to the selected technique. Even if
the system is built from solved cases or from a knowledge

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
2

acquisition tool, operators feel reluctance to use the automatic

diagnosis system because they do not understand its way of
reasoning, which is far from the workflow normally used by
troubleshooting experts.
In this paper, diagnosis system based on fuzzy logic [19]
is used. Fuzzy logic models the process of human thinking,
translating the system input values to sets (fuzzy sets) that are
easy to understand and to link through heuristic rules (fuzzy
rules). Therefore, since the expert knowledge is represented by
rules, converting the flowcharts that troubleshooting experts
normally use into the rules required by the fuzzy system
and vice versa is straightforward. This is a large advantage
compared to other solutions, both for defining the system and
for the operators acceptance in real networks. Fuzzy logic
has been applied to troubleshooting in other fields such as
industrial processes and machinery, among others. In mobile
networks, it has been used for optimization [20] [21], but not
for self-healing.
Traditionally, the preferred method for creating the fuzzy
rules has been by knowledge acquisition. But as the systems
controlled by fuzzy logic become more complex, this task is
harder and the help of automated methods is required. Since
learning algorithms use heavy CPU and memory resources,
the advent of faster and cheaper computers in the last decade
is helping this trend. This paper proposes to use a genetic
learning algorithm to obtain troubleshooting rules from solved
troubleshooting cases. These rules will then be used on a fuzzy
logic based troubleshooting system. The learning task will be
done by observing the work of the troubleshooting experts
through solved problems. This method requires less intervention from the experts compared to the methods previously
proposed in the literature, that is either interviews with experts
or requiring the experts to define the parameters related to the
given artificial intelligence technique (fuzzy logic in this case).
This will increase the expert collaboration degree, since the
effort required from them is much smaller.
This paper is organized as follows. In Section II the basic
concepts used in this paper are introduced. In Section III, the
proposed genetic algorithm and the method for diagnosis in
mobile networks are described. In Section IV, the proposed
system is evaluated. Finally, the conclusions are discussed in
Section V.
II. P ROBLEM FORMULATION
Troubleshooting in mobile networks is a procedure that has
traditionally been manual. It consists of three main tasks:
1) Detection: to identify the sectors with degraded performance.
2) Diagnosis: to determine the cause of the problem.
3) Recovery: to take actions to solve the problem (repair
HW, configuration corrections, etc).
While the fault is identified and solved, since this may
take some time, compensation may be triggered, that is
neighboring cells may provide service to the users under the
problematic cell.
Manual troubleshooting is usually a process of application
of certain heuristic rules for determining a diagnosis given

the observation of some symptoms. In mobile networks, the

symptoms are observed as abnormal values of Performance
Indicators (PIs).
Experts use a top-down approach for the detection stage;
they monitor some high level Key PIs (KPIs, such as the
number of dropped and blocked calls, failed handovers, etc)
aggregated on the whole network (or on a group of sectors).
When a degradation is detected in these high level KPIs, a
list of the worst offenders is obtained identifying the sectors
that are causing degradation. Once those sectors are identified,
the diagnosis stage further screens additional PIs, active measurements and configuration audits. Once a possible cause has
been determined out of the observations and the experience
of the expert, a correction is applied, and depending on its
outcome, the problem may be considered solved.
In this paper, fuzzy logic is used for automating the diagnosis stage of a troubleshooting system. A set of representative
PIs is selected and fed to a fuzzy logic controller (FLC)
[22] that contains the appropriate rules for determining the
problems. Fuzzy sets and fuzzy rules are modeled to imitate
the process of thinking of the troubleshooting experts. This is
achieved in two possible ways:
Knowledge acquisition: the rules are derived directly
from the experts by interviewing them [9].
Learning algorithms: the rules are derived by an algorithm
from a training set. The training set consists of a list of
specific cases that are either labeled (supervised learning)
[23] or not labeled (unsupervised learning) [24]. In this
study, only supervised learning is considered.
The aim of the learning algorithm proposed in Section III-C
is to obtain the rules from a set of training pairs (vector of
PIs and a label identifying the problem) fed by experts. This
changes the focus of the experts in the process of knowledge
acquisition. Instead of dealing with the parameters of the fuzzy
system (which requires training them in the details of fuzzy
logic controller design), they are only required to choose the
most representative troubleshooting cases.
III. FAULT MANAGEMENT BASED ON GENETIC
ALGORITHMS

A. Fuzzy logic controllers

Fuzzy logic is a branch of artificial intelligence modeling
human thinking [19]. To do this, it transforms numerical values
of variables into descriptive values (such as high or low).
Several fuzzy sets (S1 , S2 , ...) are defined over the domain of
a crisp variable. A crisp variable is a common numerical nonfuzzy variable, and its domain is the universe of discourse (U ).
A fuzzy set comprises the values of U that have a common
characteristic as perceived by a human (for instance, high or
low values for a PI). Each fuzzy set Si has an associated
membership function Si [0, 1] that defines the degree of
truth of each crisp value belonging to that fuzzy set. In Fig.
1 an example of how membership functions work is depicted.
A crisp variable x has two fuzzy sets defined on its domain:
S1 and S2 . For a given crisp value x1 of x, the membership
degree for each set is given by S1 (x1 ) and S2 (x1 ), the
membership functions of sets S1 and S2 , respectively. A fuzzy

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
3

Fig. 1. Fuzzification of a crisp variable

variable is formed by the linguistic labels identifying a fuzzy

set and their membership functions. The process of converting
a crisp value to a fuzzy value is called fuzzification. In Fig. 1,
the fuzzified value of x1 is S1 in S1 (x1 ) degree and S2 in
S2 (x1 ) degree.
Fuzzy logic controllers (FLC) are algorithms that use the
principles of fuzzy logic to assign values to output variables
based on some input values. FLCs are often used in control
systems where the input variables reflect the state of the
system and the output variables are control actions. In this
paper, a FLC is used for automatic diagnosis in mobile
communications networks, the inputs being PIs, alarms and
configuration parameters, and the outputs the root cause and
possible solutions. A FLC applies three consecutive processes:
fuzzification, a fuzzy reasoning that assigns fuzzy values to
output variables based on fuzzy values of input variables and
defuzzification, that transforms the fuzzy value of the output
variable into a crisp value.
Fuzzy reasoning in a FLC is done through If ... Then ...
rules. These fuzzy rules are composed of two main parts: the
antecedent (the if part) and the consequent (the then part). The
degree of truth of the consequent is obtained by calculating
the degree of truth of the antecedent.
The antecedent contains assertions about input variables
belonging to fuzzy sets (for example x1 is S1 ). The degree
of truth of these assertions is the degree of membership of the
variables (S1 (x1 )). Several assertions can be done in the same
antecedent, joined by AND or OR operators. Usually in these
cases, the degree of truth of the antecedent is the minimum
of all the individual assertions (with AND operators) or the
maximum (with OR operators).
The consequent contains an assertion about an output
variable. The degree of truth of the antecedent modifies the
membership function of the fuzzy set of the value assigned
in the assertion, either by truncating it or by obtaining the
product. The full process of assigning a degree of truth to a
variable in the consequent based on the degree of truth of the
antecedent is depicted in Fig. 2. The antecedent of the rule has
two assertions: x is S1 and y is S3 . The minimum degree
of truth of both assertions (in this case, S3 (y1 )) is assigned
to the consequent. The degree of truth of the consequent
truncates the membership function SO1 (z) of the fuzzy set
SO1 assigned by the assertion z is SO1 .
The crisp value of an output variable z can be inferred
through the aggregation of the outputs of individual rules.
(T )
(T )
A truncated membership function SO1 (z), SO2 (z), ... is
obtained for each rule on the domain of the output variable

Fig. 2. Linguistic reasoning

Fig. 3. Defuzzification process

according to the results of the linguistic reasoning. For each

point of the domain of the output variable, the maximum
degree of truth among the output membership functions is
taken, that is a combined function O (z) is defined as
(T )
O (z) = max(SOi (z)). A crisp value is then obtained from
this function in the defuzzification process. The crisp value can
be taken according to a specific policy:
SOM (Smallest Of Maximum): The smallest point with
the maximum degree of truth.
Centroid: The average of all points with the maximum
degree of truth.
LOM (Largest Of Maximum): The largest point with the
maximum degree of truth.
Fig. 3 depicts this process. The output of two individual
(T )
(T )
rules (functions SO1 (z) and SO2 (z)) are aggregated, creating a combined membership function. The crisp value for
variable z is one of the points where this new function is maximum. z1 , z2 and z3 are the values for z if the defuzzification
method is SOM, centroid and LOM, respectively.
B. Genetic algorithms
This paper proposes a genetic algorithm to automatically
learn the rules of the FLC used for diagnosis. Genetic algorithms [25] [26] imitate the process of natural selection, that is,
they seek the solution of problems by a trial and error method,
repeated generation after generation. All genetic algorithms
have three common elements:
A population of individuals, each being a possible
solution to the problem. Each individual is a collection

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
4

Fig. 4. Format of training data

(genome) of traits (gens) that determine the behavior

of the solution and are subject to variation when the
individual reproduces.
Operators that define the birth of new individuals.
Specifically, there are two big groups of operators:
Crossing: when the traits of two individuals are
combined to produce a new improved individual.
Mutation: when a new individual is created by
copying another individual and introducing minor
(usually random) changes.
A fitness function that assigns a score to each solution
based on its quality.

A typical genetic algorithm has three processes:

Reproduction: new individuals are created either by

crossing two chosen rules, by mutation, or a combination
of both. Usually, a high mutation rate produces a high
exploration ratio, since random changes are introduced.
In some algorithms, the parent rules are eliminated once
the reproduction is done.
Evaluation: the fitness of each individual is obtained
using the fitness function. This value will determine the
probability of the individual to survive and reproduce in
future generations.
Selection: This process decides which rules survive to
pass on to the next generation. The rules with a higher
fitness have higher chances of surviving.

C. Genetic rule learning for diagnosis in cellular networks

In this section the genetic algorithm proposed for learning
troubleshooting fuzzy rules from previously diagnosed examples (hereafter called cases) is described.
The training data is a set of cases following the format
described in Fig. 4. A case is a vector consisting of crisp
values for several PIs, and a class label corresponding to the
diagnosed fault cause.
The statistical behaviour of each PI may be influenced by
diverse aspects, including (but not limited to) the location
of the originating sector, the distribution of the users or the
configuration. Also, the requirements for each sector may vary,
so what is considered a normal value for a certain sector, may
be an abnormal value for another sector. Since the data vectors
are fuzzified before being used for diagnosis or learning, this
variability is managed in the fuzzification process. The fuzzy
sets may vary from sector to sector, adapting to what is the
meaning of the PI values for each sector. These fuzzy sets
may also be parameterized to adapt to variables such as the
time of the day to reflect the time-dependent behaviour of
the PIs. The output of the fuzzification process is therefore a
normalized version of the input vectors, that may then be used
by the FLC or the learning process regardless of the original
numerical value.

Fig. 5. Flowchart of the algorithm

In cellular networks, there are thousands of counters and

indicators which can be used as symptoms for diagnosis.
However, typically in the process of troubleshooting experts
concentrate only on a limited subset of PIs to identify the
root causes. In addition, the analyzed PIs may be different
depending on the fault cause under study. Due to this fact,
a fuzzy system designed for diagnosis in cellular networks
should allow incomplete rules [27]. Incomplete fuzzy rules
are rules that do not require a value for every input variable.
The FLC can detect cases with more than a fault cause simultaneously only if incomplete rules are included. Therefore
troubleshooting rules must only contain information about the
PIs that are affected by a certain fault cause.
In case a new problem that was not present in the training
set is fed to the FLC, the diagnosis will be wrong, since there
is no information in the rules. It may either be a different
diagnosis, or diagnosed as not being a problem. To prevent
this situation, it is necessary to execute the learning algorithm
whenever a new class of problems is found, either by manual
inspection of the problems or by observing an unusual increase
in diagnosis errors of the FLC.
The proposed algorithm, which includes the elements and
processes described in Section III-B, is shown in Fig. 5. The
features of the algorithm have been devised as follows:

Individuals: each individual is a rule. The genome of an

individual has two components:
Antecedent (the if ... part): a vector,
{a1 , a2 , ..., aL }, of length L equal to the number
of inputs of the diagnosis system. Each entry ai in
this vector represents the fuzzy value that the PI ki
should take for the rule to be activated. If ki should
be low, then ai = 1; if it should be high, then
ai = 2, and if the rule does not require any particular
value for ki , then ai = 0. The last value of ai allows
the existence of incomplete rules. Incomplete rules
are calculated at the end of the algorithm, and are
used in the process of reproduction, but there are
no individuals representing incomplete rules during
the execution of the algorithm.
Consequent: (the then ... part): an integer, c,
representing the problem diagnosed by the rule for
the fuzzy values of the antecedent.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
5

Fig. 7. Example of fusion of two rules.

The base (B) represents the statistic relevance of the rule

and the success rate (SR ) represents the accuracy of the
rule, that is the percentage of cases fulfilling the rule. The
base is defined as:
1
(3)
B =1
1 + Nc

Fig. 6. Example of crossing two rules to obtain a child.

Each individual has an associated score W that represents

its fitness. The initial population is generated with random
valid values in the antecedent vector and the consequent.
Reproduction: the rules are chosen randomly to reproduce. Crossing and mutation are done separately. For each
individual, the probability of being chosen for reproduction by crossing is given by Pcross = W R, where
W is the score of the rule, and R is the reproduction
ratio. Only the best rules are chosen for crossing, so the
best traits are passed on. For mutation, the probability is
Pmut = 1Pcross , so the worst rules (that have not been
killed) are given a chance to improve.
Crossing (Fig. 6) happens between parent rules that have
the same consequent. A random cutting point cp [0, L
1] is chosen. Each parent rule antecedent is divided into
two fragments [a0 ...acp ] (first fragment) and [acp +1 ...aL ]
(second fragment). A score is given to each of the
four fragments (two per parent). For this purpose, each
fragment is extended with zeros into a valid antecedent of
its own, and a partial rule is created with this antecedent.
The partial rules are then evaluated with the training
cases to see which of them are activated more often by
calculating the average degree of activation (DoA). The
DoA of a rule for an input vector is the degree of truth
of the antecedent. An attractiveness (a) value is assigned
to each fragment:
a = DoAf ragment Wparent

(1)

where Wparent is the score of the parent rule. The child

is then created with an antecedent that is the combination
of the best (most attractive) first fragment with the best
second fragment. The parents and the child are kept in
the population.
The mutation process creates several copies of an individual, each with one gene randomly altered. Either
one of the components of the antecedent vector or the
consequent are given a random value between the valid
values it can take.
Evaluation: the score W of the rules is calculated as the
product of two separate terms:
W = B SR

(2)

Where is a parameter that adjusts the sensitivity of the

base to uncommon cases (the higher is , the easier it is
to obtain a high base even with a low number of covered
cases), c is the number of covered cases and N is the total
number of cases in the training set. A case is considered
covered when the DoA of the antecedent for that case is
greater than a predefined threshold Minimum Degree of
Activation (M DA).
Selection: the algorithm finds the rules that have a score
equal to 0 and eliminates them.
Postprocessing: we propose to add an extra stage to
the standard genetic scheme, where the final population
is processed to create a rule set containing incomplete
rules. Rules are grouped according to their consequents,
and they are fused into incomplete rules. Two rules
are fused if they are both correct. A rule is considered
to be correct if its score is higher than the Minimum
Inclusion Degree (M ID). In that case, the contradictions
between them are resolved by ignoring the non-common
parts of the antecedent. Rules whose score is lower
than M ID are removed. An example of rule fusion is
shown in Fig. 7. Since Rule 1 and Rule 2 have different
restrictions over PI1, and if both rules are correct, then
it will be considered that the value of PI1 is not relevant
for detecting the cause Problem 1. PI2 has the same
restrictions in both rules, so the observation of a HIGH
value in this PI leads to the detection of Problem 1.

D. Parameters
The algorithm has 7 parameters:
Uncommon case sensitivity () (0, +): adjusts the
sensitivity of the algorithm to uncommon cases. A too
low value may disregard cases that are important although
they are uncommon. A too high value may give relevance
to cases that are wrongly diagnosed.
Minimum inclusion degree (M ID) (0, 1): minimum
score a rule must have to be part of the final rule set.
Minimum degree of activation (M DA) (0, 1): indicates
the minimum DoA of a rule on a case to consider that
it covers it.
Initial rules [1, +): Number of random initial rules.
If none of the initial rules survives the first iteration of the
algorithm, an extinction occurs and the algorithm stops.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
6

Number of loops [1, +): Number of iterations of

the genetic algorithm. The number of loops must be high
enough for the algorithm to converge to a correct solution.
Reproduction ratio (R) (0, 1]: the proportion of cases
that will reproduce (supposing all the cases have W = 1).
This parameter is a population growth control.
Mutation multiply factor (M M F ) [1, +): Number
of children that a rule produces when mutation takes
place. A small value will produce a low exploration rate,
whereas a too high value will increase the execution time.
IV. E VALUATION

A. Case study
In cellular networks, the O&M can provide historical
records containing the value of PIs and configuration parameters. However, after troubleshooting experts perform a
root cause analysis of problematic cells, the diagnosed cause
is normally not saved together with the value of the PIs.
Therefore, there are no existing databases of classified fault
cases in which to test new diagnosis systems, especially in
LTE, which is starting to be deployed in current networks.
Due to these reasons, and in order to be as close as possible
to the behavior of real networks (consequently, discarding
using a simulator), a network emulator was built based on the
knowledge of troubleshooting experts. The network emulator
is simply a case generator, i.e. it produces fault causes and
their associated PIs (labeled cases). The advantage of using
the network emulator is that it can provide as many cases
as needed, still being close to the real network behavior.
In order to define the LTE network emulator, the following
methodology was followed. Firstly, troubleshooting experts
defined the most common categories of fault causes and
related PIs for LTE. Secondly, the frequency of occurrence
of each fault cause was defined (prior probabilities). Finally,
the probability density function (PDF) of each PI conditioned
to each fault cause was modeled either as a normal or a beta
probability density function, which has been demonstrated to
be an adequate distribution in cellular networks [8]. The parameters of the distributions were defined by the experts. The
network emulator works by generating new cases according to
the defined prior and conditional probabilities. The network
emulator has been used to generate both training cases and
testing cases.
The defined fault categories, each covering a broader group
of causes, are the following:

SW problem: There is a problem with the software in

the site. Performance is degraded, but radio conditions
are normal.
Coverage: The sector has problems reaching some regions in its covering area.
Quality: The uplink or downlink signal in the sector
suffers interference.
Mobility: Problem with handovers to/from neighboring
sectors.

In addition, a normal category is included, for the cases

where there are no problems in the network.

TABLE I
P RIOR PROBABILITIES
Fault category
SW Problem
Coverage
Quality
Mobility

Proportion (%)
13
25
34
28

TABLE II
PI MODELING . PDF OF EACH PI CONDITIONED TO AN EXISTING PROBLEM

PI

Type

Acc.

beta

Ret.

beta

HOSR

beta

RSRP

norm

RSRQ

norm

avg

avg

Nor
2
0.1
17
0.5
4
0.02
-70
3
-6.5
1.1

Parameters/Cause
SW
Cov
Qual
12
1.391
450.3
3
0.028
23. 7
11.756
10
11
1.306
1.5
1.9
4.62
3
5
0.024
0.02
0.04
-75
-107
-72
6
5
7
-6
-10
-13
2
5
2

Mob
2
0.5
9
2
42.5
7.5
-80
10
-11
3

For this study, the probability of occurrence of each fault

case is given in Table I. These probabilities are conditioned
to the existence of a problem.
The most important PIs used to identify the previous fault
causes are the following:
Accessibility: it reflects the ability to establish a connection in the network. It is the inverse of the blocking rate.
Its values range between 0 and 100%, and it is considered
normal above 99% and low below 98%. These values are
a commercial requirement.
Retainability: it reflects the ability to end a call correctly.
It is the inverse of the dropped call rate. Its values
range between 0 and 100%, and it is considered normal
above 99% and low below 98%. These values are also a
commercial requirement.
Handover Success Rate (HOSR): percentage of initiated
handovers that end successfully. Its values range between
0 and 100%. It is considered normal above 98.5% and
low below 95%. These values are also a commercial
requirement.
95 percentile of RSRP: value of RSRP under which 95
percent of the samples fall. Experts consider it low if it
is below -100 dBm and high if it is above -80 dBm.
95 percentile of RSRQ: value of RSRQ under which 95
percent of the samples fall. Experts consider it low if it
is below -23 dB and high if it is above -12 dBm.
The probability density functions of the PIs conditioned to
the causes, together with their parameters are shown in Table
II. As an example, Fig. 8 shows the PDF of the 95 percentile
of RSRP conditioned to each fault cause.
B. Experimental design
With the emulator described in Section IV-A, 2000 training
cases (containing 600 problems and 1400 normal cases) and

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
7

TABLE III
PARAMETER VALUES
Test

Fig. 8. PDF of the 95 percentile of RSRP PI.

5000 testing cases (1500 problems and 3500 normal) have

been generated. Both sets contain the proportions for each
problem defined in Table I. The algorithm proposed in this
paper is first trained with the training cases, and afterwards, it
is tested with the testing cases. This process is repeated 100
times and the average values are calculated.
The tests evaluate the influence of the most important
parameters in the proposed algorithm (Table III). Initially,
a random population of 5 rules is generated to seed the
algorithm. The genetic process is repeated for 50 generations;
that is, 50 cycles of reproduction, evaluation, selection and
death.
For the evaluation, the average on the 100 runs of the
following measurements are assessed:
Diagnosis Error Rate: proportion of incorrect diagnosis
over the total number of problems. This measurement
counts neither the problems that are not diagnosed (i.e.:
are detected as being normal) nor the false positives. It
is given by:
Nde
(4)
Ed =
Np

Where Nde is the number of problems diagnosed as a

different problem and Np the total number of problems
(1500 in the case of these tests). This measurement
indicates the accuracy of the classifier.
Undetected Rate: proportion of problems that are not
diagnosed at all. It is given by:
Eu =

Nun
Np

(5)

Where Nun is the number of problematic cases that

are classified as normal. This measurement indicates the
reliability of the fuzzy controller. It indicates its ability
to actually detect a problem.
False Positive Rate: proportion of normal cases that are
diagnosed as a problem. Given by:
Nf p
(6)
Nn
Where Nf p is the number of normal cases diagnosed as
having a problem and Nn is the total number of normal
cases (3500 in the case of these tests). A high False
Positive Rate indicates that the fuzzy logic controller is
not very usable; since there is a high chance of false
alarms. Note that even a relatively low Ef p can be
translated into a high absolute number of false positives
in the output, since the number of normal cases is usually
much higher than the number of problems.
Ef p =

Default
value
24

MID

0.2

MDA

0.4

0.5

reproduction ratio
MMF

Training cases

Only
problematic

Expert
elicited
rules
Live network example

Variable

Scalability test

Tested values
6, 12, 18, 24, 30, 36,
42, 48, 54, 60
0.1, 0.2, 0.3, 0.4, 0.5,
0.6, 0.7, 0.8, 0.9
0, 0.1, 0.2, 0.3, 0.4, 0.5,
0.6, 0.7, 0.8, 0.9, 1
0, 0.1, 0.2, 0.3, 0.4, 0.5,
0.6, 0.7, 0.8, 0.9, 1
1, 2, 3, 4, 5, 6, 7, 8, 9,
10
Only
problematic,
problematic + normal
Learned rules, expert
rules
Default and Enhanced
(=27,M ID=0, number of initial rules = 20,
M M F =20)
Default and Enhanced
(=42,M DA=0.1,
loops = 100)

Note that Ep = Ed + Eu constitutes the total error rate over

the input problems (that is, the probability that a problematic
case in the input of the FLC produces a wrong diagnosis in
the output). The overall error (that is, the probability that a
specific diagnosis is wrong) is given by:
E = Pn E f p + Pp E p

(7)

Where Pn and Pp are respectively the proportion of normal

and problematic cases in the validation set. Nevertheless, the
results given in Section IV-C are given as the three separate
terms described earlier, because each one reflects an important
characteristic. Throughout the study of the results, it will
be considered that there is a previous detection stage that
separates normal cases from problematic cases, thus, the main
objective will be to minimize Ep . There is a trade-off between
the Undetected Rate and the False Positive Rate. If there is
no previous filter to detect problems (i.e. the detection also
relies on the fuzzy controller, along with the functionality of
diagnosis), it is very important to minimize the Undetected
Rate. On the other hand, this will normally increase the
False Positive Rate. Considering that in a normal scenario,
the number of cases showing a normal state outnumber those
showing problems, the absolute number of false positives
may be higher than the number of correct detections. The
probability that a given positive diagnosis is a false positive
(the complementary of the Positive Predictive Value) is given
by:
Pf p = 1 P P V =

Pn E f p
Pn Ef p + Pp (1 Eu )

(8)

C. Results
Test 1:

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
8

0.45

4
Diagnosis Error Rate
False Positive Rate
Undetected Rate

0.4

3.5

0.35

0.3

2.5

0.25
2
0.2
1.5

0.15

0.1

0.5

0.05
0

10

20

30

40

50

60

Diagnosis Error Rate

False Positive Rate
Undetected Rate

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1

0.2

0.3

0.4

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Fig. 11. Number of output rules for variable M ID.

0.8

0.2

MID

Fig. 9. Error rates for variable .

0.9

0
0.1

0.5

0.6

0.7

0.8

0.9

MID

Fig. 10. Error rates for variable M ID.

This experiment finds the influence of . This variable

regulates the sensitivity of the algorithm to rare cases. A small
value of gives a low score to rules that cover uncommon
cases. A higher value lets the score of a rule grow rapidly as
its base increases.
The Diagnosis Error Rate, Undetected Rate and False Positive Rate are depicted in Fig. 9.
The Diagnosis Error Rate is not very sensitive to the value
of . The False Positive Rate increases with , whereas the
Undetected Rate decreases with . A value of between 20
and 30 provides a good compromise in the values of both
the Undetected Rate and the False Positive Rate. However, if
minimizing the Undetected Rate is considered more important
than minimizing the false positive rate, then a high value of
should be selected.
Test 2: MID
The M ID parameter determines the minimum score that a
rule should have to be fused with other rules to produce more
general rules. The results are shown in Fig. 10.
The Diagnosis Error Rate and False Positive Rate decrease
as M ID increases. For M ID = 0.9, both error rates are
zero. Nevertheless, the Undetected Rate is 1, which means
that absolutely no problem is diagnosed; there is no actual
diagnosis system, as shown in Fig. 11, where the average

number of rules is depicted. As M ID increases, the number

of rules decreases. A high M ID is more restrictive, thus
including less rules in the final rule set. For M ID = 0.9
there is no rule with the required score to be included in
the final rule set, so nothing is diagnosed. There is a direct
relation between the number of rules and the Undetected Rate.
Likewise, as the number of rules decreases, the False Positive
Rate also decreases, as less rules produce a smaller chance
of a wrong diagnosis. Given the values Pn = 0.7, Pp = 0.3,
and taking the values of Ef p and Eu for the default value
of M ID = 0.2 (0.3179 and 0.0075, respectively), according
to Eq. (8), Pf p = 0.4277. That is, 42.77 percent of the
times, when the system indicates positive, it is a false alarm.
This is for a network where 30 percent of the sectors have
problems. For a better network, that number would increase.
This stresses the need of a detection phase that filters normal
cases, avoiding their analysis by the diagnosis fuzzy logic
controller. Supposing that an ideal detection phase is used,
Pn = 0 and Pp = 1 by definition. Applying Eq. (8), Pf p = 0.
Anyway, a detection phase might still have a small error rate
that lets some false positives pass, so it still makes sense to try
to keep a low False Positive Rate, even though this criterion
is not the priority. Since the Diagnosis Error Rate is relatively
insensitive to the variation of M ID, the main criterion will
be to minimize the Undetected Rate. As a conclusion, to keep
a low Undetected Rate, M ID should be low (around 0.2).
Test 3: MDA
This experiment evaluates the influence of the M DA
parameter over the error rate. This parameter regulates the
minimum degree of truth of an antecedent for a case to
consider it covered. This modifies the base of the rules, and
consequently their scores. Fig. 12 depicts the results.
Again, the tradeoff between undetected rate and False Positive Rate is observed. Since a higher M DA is more restrictive,
the effect in the number of rules is similar to the case of M ID,
as Fig. 13 shows.
The best value of M DA depends on the priorities when
tuning the algorithm. To obtain a minimum Undetected Rate,
the value for M DA should be lower or equal to 0.3. Nevertheless, for values 0.1 and 0.2, both the Diagnosis Rate and False

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
9

0.45

0.8
Diagnosis Error Rate
False Positive Rate
Undetected Rate

0.4
0.35

Diagnosis Error Rate

False Positive Rate
Undetected Rate

0.7
0.6

0.3

0.5

0.25
0.4
0.2
0.3

0.15

0.2

0.1

0.1

0.05
0
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0
0

0.2

MDA

0.4

0.6

0.8

0.8

Reproduction Ratio

Fig. 12. Error rates for parameter M DA.

Fig. 14. Error rates for the reproduction ratio.

2.5

3.5
2

3
2.5

1.5

2
1

1.5
1

0.5
0.5
0
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

MDA

Fig. 13. Average number of rules for M DA parameter.

Positive Rate are at their maximum. For 0.3, the Diagnosis

Error Rate decreases to almost a third of its maximum value.
Therefore, 0.3 is a good choice that has a low Undetected
Rate and a relatively low Diagnosis Error Rate. To obtain a
low False Positive Rate a good choice would be a value of
0.5 for M DA, which has a slightly worse Undetected Rate.
Since the addition of a detection stage would eliminate (or
at least reduce) the number of normal cases in the input
of the FLC, the False Positive Rate would no longer be an
important factor. The Undetected Rate would be transformed
into the proportion of cases that are known to be problematic,
but cannot be diagnosed. Therefore, the minimization of the
Undetected Rate should be prioritized over the minimization
of the False Positive Rate.
Test 4: Reproduction ratio
The reproduction ratio adjusts the probability of reproduction for each rule. A high value of this parameter increases the
number of combinations tested in the algorithm, at the cost of
an increased execution time. In Fig. 14 the measured errors are
represented. The execution time with different reproduction
ratios relative to the execution time taken with the default
configuration is shown in Fig. 15.
The results show that for a low reproduction ratio (lower
than 0.5), the Undetected Rate decreases as the reproduction

0
0

0.2

0.4

0.6

Reproduction Ratio

Fig. 15. Relative execution times for different values of the reproduction
ratio.

ratio increases. At the same time, the False Positive Rate

increases. For values higher than 0.5, all the rates are stagnant.
On the other hand, the relative execution time increases as
the reproduction ratio grows. Therefore, a value of 0.5 for
the reproduction rate is the optimal, since it offers results as
good as those obtained for higher ratios, without increasing
the execution time.
Test 5: MMF
The M M F adjusts the number of rules that are created
each time that a rule reproduces by mutation. A high value
of M M F provides an increased exploration capacity, but also
increases the execution time as it adds rules to the system.
Fig. 16 depicts the error rates as a function of the M M F and
Fig. 17 shows the relative execution time as compared to the
configuration with a default value for M M F (M M F = 5).
As the number of mutated offspring grows, the algorithm
produces better results. The Diagnosis Error Rate is more or
less stagnant all along the executions. The Undetected Rate
shows a large decrease as the number of mutated offspring
grows, and the False Positive Rate grows at the same time.
The increased exploration provided by this method greatly
influences the results.
For values of M M F higher than 6, the Undetected Rate

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
10

0.45
Diagnosis Error Rate
False Positive Rate
Undetected Rate

0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
1

10

Fig. 18. Results comparing the presence and lack of normal cases in the
training set.

MMF

Fig. 16. Error rates for M M F .

3.5
3
2.5
2
1.5
1
0.5
0
1

10

MMF

Fig. 17. Relative execution times for different configurations of M M F .

reaches its minimum. Therefore, for values lower than 6,

the higher M M F is the better the solution, at the cost of
computing time. For values higher than 6, the quality of the
solution does not improve, despite the increase of execution
time.
Test 6: Training using normal cases
In all the previous tests, the training was done using only
cases that had problems. In this experiment, the training will
be done with the default values of all parameters (Table III),
but using normal cases in the training phase together with the
problematic cases. Also, to isolate the effects of , the default
N +N
value is used when not using normal cases, and pNp n
when using normal and problematic cases in the training.
When using this modified value of in Eq. (3), the absolute
number of cases c that a rule must cover in order to obtain
a certain base B is the same. Therefore, the variations in the
results are only dependent on the fact of using normal cases
in the training. The results are depicted in Fig. 18.
The Diagnosis Error Rate and Undetected Rate suffer a
slight degradation, whereas the False Positive Rate is clearly
improved. This is due to the the fact that the algorithm evolves
the rules with knowledge about normal cases. Therefore, it
will be less likely that rules that classify normal cases as
problematic are considered valid. Thus, the rules are more

evolved towards avoiding false positives. The improvement

of the False Positive Rate comes at the cost of an increased
execution time, due to the increase in the number of individual
applications of rules over cases when the evaluation in the
genetic loop is done (appr. 3.5 times higher execution time
when including normal cases). When no normal cases are used
in training, each individual rule is evaluated on 30 percent
of 2000 cases (600), and when normal cases are used, each
rule is evaluated over all the cases. The decision of using
normal cases in training or not will depend mostly on whether
there will or will not be a preliminary detection phase. In case
there is, since false positives are not a problem, the training
should be done without normal cases, because of the reduced
execution time and slightly better error rates. On the other
hand, if no detection phase is used, reducing the False Positive
Rate will result in a more usable diagnosis algorithm.
Test 7: Comparison between expert elicited and learned
rules
Some experts were requested to manually define the rules
relating the selected causes and symptoms (Table IV). In this
experiment, the three error measurements were compared for
the expert elicited rules versus the learned rules using the
genetic algorithm (default parameters). Results are shown in
Fig. 19.
TABLE IV
E XPERT ELICITED RULES
Acc.
LOW
LOW

Ret.
LOW
LOW
LOW
LOW

HOSR
HIGH
HIGH
LOW

RSRP
HIGH
LOW
HIGH

RSRQ
HIGH
HIGH
LOW

Cause
SW Problem
Coverage
Quality
Mobility

It can be seen that there are significant reductions in the

Undetected Rate when using learned rules. On the other hand,
False Positives are much greater when using learned rules,
while Diagnosis Error Rate is similar. It can be concluded that
the parameters used in these tests favor the reduction of the
Undetected Rate. This might be an advantage depending on
the requirements of the final system. In any case, the learning
algorithm serves its purpose of obtaining a rule set without
the direct intervention of experts, which is an advantage, as
exposed in Section I.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
11

TABLE VI
A DDITIONAL PI MODELS . PDF OF EACH PI CONDITIONED TO AN
EXISTING PROBLEM

PI

Fig. 19. Results comparing expert elicited and learned rules.

TABLE V

Type

SINR

norm

Dist.

norm

Thp.

norm

CPU

negative
binomial

RSSI

norm

avg

avg

avg

n
p
avg

Nor
15
0.8
0.9
0.2
11.4
3.1
0.052
0.01
-116.7
0.5

Parameters/Cause
SW
Cov
Qual
13.6
15
7
1.4
1
1.7
0.9
1.1
1
0.3
0.3
0.2
7.5
7.9
6
4.25
6.1
5.25
2.07
0.052
0.024
6.3e-3
9.6e-4
4.7e-4
-112
-114
-109
10
10
11

Mob
14.2
1
0.9
0.3
9.7
4.25
5.5e-3
2.3e-3
-119
12

R ESULTS OF THE LIVE NETWORK TESTS

Test
Default parameters
Enhanced parameters

Diagnosis
Error Rate
0%
0%

Undetected
Rate
62.41 %
19.19 %

False
Positive Rate
10.08 %
21.67 %

Test 8: Live network example

A small test with 72 live network cases was carried out
following the same steps as with the simulated cases. Although
there were not enough real examples for an in-depth analysis
such as in the previous tests, a small amount of available cases
may show the behavior of the presented learning algorithm
in a real network. The cases are organized in 4 categories:
missing neighbor, interferenece, high CPU usage and normal
behavior. There are 18 vectors representing problems in each
category. The PIs used for diagnosing these problems are
accessibility, retainability, HOSR, average RSSI and number
of CPU overload alarms. A cross validation technique is used
where each category (except normal) is divided in two random
partitions. One of these partitions is aggregated with partitions
of other categories to create a training set, and the remaining
partitions, along with the normal cases are used as a validation
set. The algorithm is trained with the training set and then the
error measurements are taken with the testing set, following
the methodology used with the simulated cases. This process is
repeated 100 times and the averages of the errors are extracted.
The results of the test are shown in Table V.
Using the default values for the parameters, the obtained
results are very poor. This is mainly due to the small number of
available cases. This has a complex effect on the evolution of
the population. Since the number of rules is low, a rule with a
low success rate has less probabilities of surviving. Therefore,
bad rules die earlier. This reduces the initial population rapidly,
and, therefore, the gene pool, that is, the reserve of not-sogood rules that may produce very successful rules via a small
mutation. In this population bottleneck, only the best rules
survive, and they may not cover all the cases. Since the best
rules have a lower probability of mutating, there is a lower
chance that new rules covering the gaps in the training sets
appear, and therefore, the Undetected Rate increases. To fix
this, it is important that the algorithm has the opportunity of
exploring many possibilities. Therefore the immediate actions

that can be taken are increasing the M M F to increase the

exploration. Also, in this situation of a small gene pool, it
will help to increase the chances of exploring rules that would
otherwise be ignored. Therefore, the number of initial rules is
also increased. A new set of enhanced parameters is created
where the number of initial rules is 20 and the M M F is 10.
The results are clearly improved, although still the Undetected
Rate is higher than in the tests due to the very small number
of training cases.
Test 9: Scalability test
In all these tests, we have used 5 PIs to diagnose the
problems. In real networks, the number of PIs is usually much
larger, and the interactions much more complex. Therefore,
it would be useful to know if the proposed algorithm can
scale up and manage an increased number of PIs. To test
this, an extended dataset has been created, containing the same
problems as the default dataset, and adding 5 PIs to the original
set of modeled PIs:
95 percentile of SINR: value of SINR under which 95
percent of the samples fall. It is considered low when the
value is below 10 dB and high when it is above 12 dB.
95 percentile of the distance: radius of the circumference
that encompasses 95% of the users of the sector. Distances below 1 Km are low, whereas distances above 1.1
Km are high.
Throughput: Data rate that the sector successfully transmits to the users (downlink). Throughput values below 7
Gbps are low and values above 7.5 Gbps are high.
Number of CPU overload alarms: Number of alarms per
hour indicating that the CPU of the sector is overloaded.
Below 20, the values are considered low and above 40,
high.
Average RSSI: Total signal power received in the operational bandwidth. Below -110 dBm, it is considered low,
and above -108 dBm, high.
The model parameters of these PIs are summarized in Table
VI. The model parameters and the requirements are all based
on expert input.
The configuration of the training and testing sets are kept
on the same values as described in Section IV-B, and a test
with the default parameters is executed. The results are shown
in Table VII.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
12

TABLE VII
R ESULTS OF THE SCALABILITY TESTS
Test
Default parameters
Enhanced parameters

Diagnosis
Error Rate
0.6 %
1.2 %

Undetected
Rate
50.13 %
1.4 %

False
Positive Rate
3.26 %
4.2 %

of the parameters. The selection criterion should be more

dependent of the Undetected Rate, since the impact of the
parameters on this measurement might be high. Nevertheless,
the method is still robust, since the allowed parameter ranges
for low Undetected Rate are broad.
ACKNOWLEDGMENT

A degradation in the performance of the obtained rule set

is clearly observed. This is due to more complex individuals
(rules) involved in the evolutionary process. Since the Undetected Rate has grown, it can be concluded that not enough
rules are being created to cover all cases. Therefore, a new
set of improved parameters is tested. The number of loops
is increased from 50 to 100, to ensure that the evolutionary
process has enough time to process the increased complexity
of the individuals. The parameter is increased from 24 to 42,
so that the sensitivity to unusual cases is increased. Finally, the
M DA parameter is reduced to 0.1, so that lower scored rules
survive to have a chance of improving. With these parameters,
the performance is largely improved (Table VII). In fact, these
results improve the Diagnosis Error Rate over the default
number of PIs, since there is more information for the FLC
to process and discriminate among the diagnoses. Also, there
is a reduction in the False Positive Rate due to the increase
in the number of variables, that reduces the probability that a
normal case has the KPI values that are typical of a specific
problem. Nevertheless, this gain comes at a cost of execution
time and a slight increase in the Undetected Rate. Although
the implementation of the algorithm is not speed-optimized,
the new data set takes more than 36 times more time than the
default data set.
V. C ONCLUSIONS
A novel approach for obtaining the rule base of fuzzy
controllers for mobile network self-diagnosis has been presented, using a learning method instead of direct knowledge
acquisition. The proposed method is a supervised, genetic
learning algorithm, whose inputs are vectors of PIs, alarms
and configuration parameters labeled with the fault cause.
The method has been evaluated using a reduced model
comprising some fault causes and PIs.
The experiments cover the main configuration parameters of
the algorithm and the presence or absence of normal cases in
the training set. The conclusions are that the parameters must
allow diversity in the rule base; that is, they must not filter rare
cases or cases that loosely follow a rule. If the parameters are
too restrictive, the output rule base is unreliable, in the sense
that it will often be unable to diagnose (or even detect, if
the fuzzy controller is assigned this task) a proportion of the
problems. On the other hand, the increase in reliability comes
at a cost; the fuzzy rule set will detect some normal cases
as problems. This problem is solved if there is a detection
stage that filters out normal cases prior to its analysis by the
diagnosis stage. Results show that the proposed method is
quite robust to the selection of the parameters, since there are
only slight variations in the diagnosis error for different values

This work has been partially funded by Optimi-Ericsson,

Junta de Andaluca (Agencia IDEA, Consejera de Ciencia,
Innovacion y Empresa, ref. 59288; and Proyecto de Investigacion de Excelencia P12-TIC-2905) and ERDF.
R EFERENCES
[1] NGMN, Use Cases Related to Self-Organising Network, Overall
Description, Next Generation Mobile Networks (NGMN) Alliance,
http://www.ngmn.org, April 2007.
[2] 3GPP, Evolved Universal Terrestrial Radio Access (E-UTRA) and
Evolved Universal Terrestrial Radio Access Network (E-UTRAN); Overall description; Stage 2, ts 36.300 ed., Next Generation Mobile Networks
(NGMN) Alliance, December 2012.
[3] , Telecommunication management; Self-Organizing Networks
(SON); Concepts and requirements, ts 32.500 ed., Next Generation
Mobile Networks (NGMN) Alliance, September 2012.
[4] R. Barco, P. Lazaro, and P. Munoz., A unified framework for selfhealing in wireless networks, IEEE Communications Magazine, vol. 50,
pp. 134142, December 2012.
[5] 3GPP, Telecommunication management; Self-Organizing Networks
(SON); Self-healing concepts and requirements (Rel 11), ts 32.541 ed.,
Next Generation Mobile Networks (NGMN) Alliance, September 2012.
[6] M. Asghar, S. Hamalainen, and T. Ristaniemi, Self-healing framework
for LTE networks, in IEEE 17th International Workshop on Computer
Aided Modeling and Design of Communication Links and Networks
(CAMAD), September 2012.
[7] G. A. Barreto, J. C. M. Mota, L. G. M. Souza, R. A. Frota,
and L. Aguayo., Condition monitoring of 3G cellular networks
through competitive neural models, IEEE Trans. Neural Networks, vol.
Vol.16(5), pp. 10641075, 2005.
[8] R. Barco, P. Lazaro, L. Dez, and V. Wille., Continuous versus
discrete model in autodiagnosis systems for wireless networks, IEEE
Transactions on Mobile Computing, vol. Vol.7 (6), pp. 673681, June
2008.
[9] R. Barco, P. Lazaro, V. Wille, L. Dez, and S. Patel, Knowledge
acquisition for diagnosis model in wireless networks, Expert systems
with Applications., vol. Vol.36, Issue 3, Part 1, pp. 47454752, April
2009.
[10] Univerself, Univerself project, http://www.univerself-project.eu/, 2012.
[11] L. Bennacer, L. Ciavaglia, A. Chibani, Y. Amirat, and A. Mellouk,
Optimization of fault diagnosis based on the combination of bayesian
networks and case-based reasoning, in IEEE Network Operations and
Management Symposium (NOMS), April 2012.
[12] C. Hounkonnou and E. Fabre., Empowering self-diagnosis with selfmodeling, in 8th international conference Network and service management (CNSM), and 2012 workshop on systems virtualization management (SVM), October 2012.
[13] COMMUNE, Commune (COgnitive network ManageMent under UNcErtainty), http://projects.celtic-initiative.org/commune/, 2012.
[14] C. Project, Specification of knowledge-based reasoning algorithms,
2012, deliverable 4.1.
[15] P. Szilagyi and S. Novaczki, An automatic detection and diagnosis
framework for mobile communication systems, IEEE Transactions on
Network and Service Management, vol. 9, no.2, pp. 184197, June 2012.
[16] M. Amirijoo, L. Jorguseski, T. Kurner, R. Litjens, M. Neuland, L. C.
Schmelz, and U. Turke, Cell outage management in LTE networks.
in 6th International Symposium on Wireless Communication Systems
(ISWCS), 2009.
[17] H. Eckhardt, S. Klein, , and M. Gruber., Vertical antenna tilt optimization for lte base stations. in IEEE 73rd Vehicular Technology
Conference (VTC Spring), 2011.
[18] R. Razavi, Self-optimisation of antenna beam tilting in LTE networks..
in IEEE 75th Vehicular Technology Conference (VTC Spring), 2012.
[19] L. A. Zadeh, Fuzzy sets, Information and control, 1965.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2015.2414296, IEEE Transactions on Vehicular Technology
13

[20] P. Munoz, R. Barco, and I. de la Bandera, On the potential of

handover parameter optimization for self-organizing networks, IEEE
Transactions on Vehicular Technology, vol. 62, no.5, pp. 18951905,
June 2013.
[21] , Optimization of load balancing using fuzzy Q-learning for next
generation wireless networks, Expert Systems With Applications, vol.
40, no. 4, pp. 984994, March 2013.
[22] C. Lee, Fuzzy logic in control systems: fuzzy logic controller, IEEE
Transactions on I. Systems, Man and Cybernetics, pp. 404418, 1990.
[23] S. Kotsiantis, Supervised machine learning: A review of classification
techniques, Informatica, vol. 31, pp. 249268, 2007.
[24] N. Grira, M. Crucianu, and N. Boujemaa, Unsupervised and semisupervised clustering: a brief survey, A review of machine learning
techniques for processing multimedia content, Report of the MUSCLE
European Network of Excellence (FP6), 2004.
[25] O. Cordon, F. Herrera, and P. Villar, Generating the knowledge base
of a fuzzy rule-based system by the genetic learning of the data base,
IEEE Transactions on Fuzzy Systems, vol. 9, no. 4, pp. 667674, August
2001.
[26] A. Gonzalez and R. Perez, SLAVE: A genetic learning system based
on an iterative approach, Fuzzy Systems, IEEE Transactions on, vol. 7,
no. 2, pp. 176191, April 1999.
[27] S. Guillaume, Designing fuzzy inference systems from data: an
interpretability-oriented review, Fuzzy Systems, IEEE Transactions on,
vol. 9, no. 3, pp. 426443, June 2001.

0018-9545 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.