IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO.

XX, XXX 2019 1

A QoE-driven traffic steering algorithm for LTE

networks
Marı́a Luisa Marı́-Altozano, Salvador Luna-Ramı́rez, Matı́as Toril, Carolina Gijón
Email: {mlma, sluna, mtoril, cgm}@ic.uma.es
Department of Communication Engineering, University of Málaga, 29071, Málaga, Spain.

Abstract—Due to the diversity of mobile services and raising selecting paths in software-defined architectures in future 5G
user expectations, mobile network management has changed its systems [6].
focus from Quality of Service (QoS) to Quality of Experience The above paradigm change will make network mana-
(QoE). As a consequence, classical network optimization proce-
dures must be updated accordingly. One of these procedures gement more complicated. The need for an efficient net-
is traffic sharing, whose aim is to redistribute traffic among work management has caused intense research on automation
adjacent cells so as to provide an adequate QoE to subscribers. In techniques, referred to as Self Organizing Networks (SON).
this work, a novel QoE-driven traffic sharing algorithm based on SON procedures are classified into three use cases: self-
mobility load balancing is proposed for LTE networks offering planning, self-healing and self-optimization [7]. In particular,
services of very different nature. Unlike previous approaches,
where the aim was to balance some QoS indicator, the aim here self-optimization includes those techniques designed to cope
is to equalize the QoE provided by all cells in the network. with network changes so that optimal network performance is
For this purpose, the handover margins between adjacent cells always ensured during the operational stage. Traffic steering
are tuned on a per-adjacency or per-service basis based on (a.k.a. traffic sharing or load balancing) is one of the key
QoE measurements collected in the network management system. use cases of self-optimization [8]. The aim of traffic sharing
Method assessment is based on a dynamic system-level simulator
implementing a realistic LTE scenario. Results show that the is to alleviate congestion problems due to the uneven traffic
proposed QoE-driven traffic sharing algorithm alleviates QoE demand by redistributing users among neighbor cells. This
problems by equalizing user QoE throughout the scenario. is achieved by changing cell service areas with new base
Index Terms—Long Term Evolution (LTE), self organizing station parameter settings, such as, e.g., transmit power [9],
network (SON), self-tuning, fuzzy, quality of experience. cell reselection offset [10], antenna tilt angle [11] or HandO-
ver (HO) margin [12], the latter being the preferred option
(referred to as mobility load balancing, MLB). Likewise, load
I. I NTRODUCTION
balancing algorithms can be classified into static or dynamic
Over the last few years, there has been an exponential approaches [13]. Static approaches can make use of analytical
growth in the demand of mobile services. At the same time, approaches to ensure optimal performance proactively in the
the success of smartphones and tablets has changed traffic long term [14] [15]. In contrast, dynamic approaches rely on
patterns in mobile networks due to the introduction of new simple reactive schemes, prone to instabilities. First dynamic
services [1]. These changes will continue in the coming years MLB algorithms designed for Long Term Evolution (LTE)
with the deployment of 5G systems, which will introduce new were based on a simple proportional controller driven by
mobile use cases [2]. the load imbalance between adjacent cells [16] [17]. As
In parallel, technological advances have raised users’ expec- shown in [15], these algorithms may lead to severe network
tations, forcing operators to change the way they manage their performance degradation due to the tight frequency reuse in
networks. Traditionally, network management has been based LTE. More sophisticated algorithms use fuzzy logic controllers
on objective performance indicators (Quality of Service, QoS) with reinforcement learning [18] or combine MLB with remote
measuring user or network performance (e.g., accessibility, electrical tilting [11] or power re-planning [19].
retainability, integrity. . . ). Recently, operators have shifted All the above-mentioned MLB algorithms are driven by
their focus from network performance to end user opinion simple indicators, such as average cell load or call blocking
(a.k.a. Quality of Experience, QoE). In this context, QoE is ratio. Thus, QoE is not taken into account. In current mobile
defined as the overall satisfaction of a service as subjectively networks, QoS control is carried out by packet scheduling
perceived by the user [3]. Customer experience management (PS) algorithms, dynamically assigning radio resources to
(CEM) will be even more important in 5G, as services with user data requests based on QoS constraints [20] [21]. More
very different requirements will coexist (e.g., high-definition sophisticated schedulers exploit multiuser diversity gain to
television, virtual/augmented reality, autonomous vehicle, sen- achieve optimal system performance and ensure user fair-
sor networks . . . ) [4]. Thus, maximizing the QoE should be the ness [22]. Several QoE-aware schedulers have been proposed
main criterion for assigning radio resources [5] or dynamically in the literature to optimize the overall QoE, while ensuring a
minimum QoE for all users. Such advanced schedulers are
Copyright (c) 2015 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be designed for specific services (e.g., web [23], progressive
obtained from the IEEE by sending a request to [email protected]. video streaming [24] or adaptive video streaming [25], [26],
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 2

[27]). However, the aim of most schedulers is to ensure a

minimum QoS/QoE for the worst users or equalize the average
QoS/QoE per service within a cell, rather than equalizing
the average QoE of services across cells in the network.
Thus, QoE balance between users or services in the spatial
domain is not guaranteed. A QoE imbalanced network implies
an unfair allocation of radio resources: while fully satisfied
users in underutilized cells waste resources without increasing
their QoE, there are completely unsatisfied users in congested
cells who lack these resources. In the end, such an unfairly
distributed users’ satisfaction may turn into higher churn rates
and revenue loss. Moreover, implementing the aforementioned
advanced QoE-aware schedulers would require upgrading net- Fig. 1: Traffic sharing by changing handover margins [15].
work equipment, which is not desired by network operators
that have already made an important investment to upgrade
to the latest radio access technology. Alternatively, a standard where Prx (j) is the pilot signal level received from neighbor
scheduler can be tuned to improve the system QoE. In [28], cell j, Prx (i) is the pilot signal level received from the
a self-tuning algorithm for a classical multi-service packet serving cell i, and HOM is the HO margin, defined on a
scheduler is proposed to balance QoE across services by re- per-adjacency basis (i.e., one value for each pair of cells and
prioritizing users in a LTE cell. Similarly, in [29], a different direction of the adjacency). In most cases, HO margins are set
self-tuning algorithm for the same scheduler is designed based complementarily in both directions of the adjacency to prevent
on optimality criteria to ensure the best overall system QoE ping-pong effect, so that
driven by network performance statistics. However, none of
these self-tuning schemes manages to equalize the QoE across
HOM (i, j) + HOM (j, i) = H , (2)
cells in the system. Even if MLB may potentially reduce QoE
differences between cells, to the best of authors’ knowledge,
no MLB algorithm in the literature explicitly takes QoE into where H represents the hysteresis value.
account. Figure 1 illustrates how modifying HO margins can be
In this work, a novel QoE-aware MLB algorithm is proposed used as a traffic sharing technique [15]. An increase of ∆
for LTE systems. Unlike previous approaches, the proposed dB in HOM (i, j) enlarges the serving area of cell i while
algorithm aims to minimize QoE differences across cells decreasing that of cell j. Thus, the carried traffic (and load) in
and services by adjusting handover parameters on a per- cell i increases, whereas the carried traffic (and load) of cell j
adjacency basis. Parameter tuning is performed by a fuzzy decreases. Conversely, a decrease in HOM (i, j) reduces the
logic controller driven by QoE estimates obtained from key serving area of cell i, while increasing that of cell j.
service performance indicators. The algorithm is validated in Classical Load Balance (LB) schemes modify HO margins
a dynamic system-level simulator implementing a realistic to equalize the load between neighbor cells in the hope that the
macrocellular LTE scenario. The main contributions of this overall call blocking ratio (or some other global QoS indicator)
work are: a) uncovering the limitations of traditional traffic is improved [16][18]. Such a positive effect is often achieved at
sharing schemes from a QoE perspective, b) a novel self-tuning the expense of deteriorating network spectral efficiency, since
algorithm for balancing QoE by modifying handover margins, users are reassigned to cells that do not prove the largest signal
and c) the validation of the algorithm via simulations in a level [15].
realistic macrocellular LTE scenario. The rest of the work is An evenly loaded network does not necessarily imply a
organized as follows. Section II discusses the limitations of QoE-balanced network. User satisfaction is highly dependent
classical load balancing schemes in terms of QoE. Section III on the required service, causing that the QoE of users of
describes the system model used in this work. Section IV different services can differ significantly even if they receive
describes the proposed QoE balancing algorithm. Section V the same amount of resources. More importantly, balancing
presents algorithm assessment. Finally, Section VI summarizes the load between adjacent cells does not necessarily reduce
the main conclusions. QoE differences among users in the cells, since some services
are more sensitive to load increments than others (as will be
shown later). As a consequence, equalizing the load between
II. P ROBLEM FORMULATION neighbor cells does not necessarily reduce QoE differences if
the service mix (i.e., ratio of connections for each service) is
In mobile networks, the HO process ensures a seamless not exactly the same in both cells.
connection between neighbor cells when the user moves.
The previous considerations suggest that a classical QoS-
Specifically, a HO is triggered when the following condition
based traffic sharing scheme does reach an evenly loaded
is fulfilled
network, but might lead to a more unevenly balanced QoE
distribution among cells in the system. This is the main
Prx (j) − Prx (i) ≥ HOM (i, j) , (1) hypothesis that will be tested in this paper.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 3

TABLE I: Traffic model parameters. B. QoE models

Service Main features QoE is often measured using the Mean Opinion Score
VoIP Coding rate 16 kbps (MOS) scale, ranging from 1 (bad) to 5 (excellent). In absence
Session time: exponential distribution (avg. 60 s). of surveys, QoE can be estimated from QoS measurements.
Call dropped after 1 s without resources. λV oIP ' 0.
VIDEO H.264/MPEG-4 AVC
For this purpose, QoS measurements gathered on a session
VBR (Variable bit rate) basis are mapped into QoE figures by utility functions [35]. A
720p resolution, utility function describes the relationship between objective
25 frames per second.
Video duration: uniform distribution between 0 and
QoS performance indicators and subjective QoE for each
540 s. Frame size according to real traces (avg. 9.2 service. Utility functions provide an estimate of the user QoE,
MB). although they miss contextual factors. (e.g., location, time of
Connection dropped when stalling lasts for twice the
video duration. λV IDEO = 4 · 10−3 .
day, · · · ). Thus, network operators that do not take explicit
FTP File size: log-normal distribution (avg. 20 MB) [30]. QoE measurements can estimate user QoE by processing
λF T P = 2.5 · 10−3 . passive measurements of key performance indicators.
WEB Web page size: log-normal distribution (avg. 20 MB).
No. pages per session: log-normal (avg. 4).
For VoIP service, user QoE can be estimated as [36]:
Waiting time: exponential distribution (avg. 107 s) [30].
λW EB = 3.7 · 10−3 .
QoE (V oIP ) = 1+0.035R+R(R−60)(100−R)7·10−6 , (3)

III. S YSTEM MODEL where QoE (V oIP ) is the MOS value for a VoIP connection,
and R is a parameter representing the connection quality, with
This section outlines the traffic and QoE models of the values from 0 (minimum) to 93 (maximum), that only depend
mobile services covered in this work. on the delay experienced by VoIP packets (mouth-to-ear
delay). Note that max(QoE (V oIP ) ) = 4.4054 (when R = 93),
i.e., MOS never reaches the value of 5, showing that, even with
A. Traffic models
the best possible network performance, some individuals may
Table I shows the main characteristics of the four ser- not score their experience as excellent. Likewise, QoE is set
vices considered in this work: voice over Internet Protocol to the minimum (i.e., QoE (V oIP ) = 1) when the connection
(VoIP), progressive video streaming (VIDEO), file download is dropped.
service via File Transfer Protocol (FTP) and web browsing VIDEO utility function is defined as [28]:
(WEB) [30]. VoIP is a guaranteed bit rate (GBR) service with
low data rates. VoIP is modeled to generate 20 bytes of voice
every 10 ms, with a bit rate of 16 kbps. In contrast, VIDEO, QoE (V IDEO) = 4.23 − 0.0672Lti − 0.742Lf r − 0.106Ltr ,
FTP and WEB services are non-guaranteed bit rate (non- (4)
GBR) services. The video service model (inspired in [31]) where QoE (V IDEO) is the MOS estimated for the video
corresponds to buffered live video streaming with fixed quality connection, Lti denotes the initial buffering time (in seconds),
(720 p) and variable bit rate. For this purpose, a simple model Lf r is the average stalling frequency (s−1 ) (i.e., number of
of the player’s buffer at the client side is implemented. In times per second that the video player is paused due to an
live video streaming, content generation and playback request empty client buffer), and Ltr is the average stalling duration
occur at the same time (unlike in video on demand, where the (in seconds). The maximum QoE value for a video connection
whole content is available at the start of the session); thus, is upper limited to 4.23. As in VoIP, QoE (V ıdeo) = 1 if
the video server starts sending frames to the client as they connection is dropped.
are generated, which are stored in client buffer until reaching The utility function for FTP service is [37]
a minimum video content (3 seconds, in this work). This is
modeled as a fixed video playback start delay (i.e., initial
buffering time, LTI, of 3 seconds). Later, if the buffer runs QoE (F T P ) = max(1, min(5, 6.5 · T H − 0.54)) , (5)
out, the video stops (i.e., stalling event) and the player waits where T H denotes the average user throughput in Mbps.
until the buffer is re-filled again. Video duration follows a
Finally, the utility function for WEB service is [37]
uniform distribution between 0 and 540 s. Obviously, videos
of less than 3 s do not experience stalling. Frame sizes are 578
taken from a real H.264 video trace [32]. A video session drop QoE (W EB) = 5 − , (6)
1+ ( T H+541.1
45.98 )2
model is also simulated, where the connection is terminated
if session time is more than twice the video content duration. where T H is the average user throughput in kbps. Note that,
The other two data services FTP and WEB are best-effort max(QoE (W EB) ) = 5. No dropping of web connections is
services. FTP is a file download service and WEB consists considered, so that low MOS values for web are reached when
of downloading several web pages with different sizes with T H is zero (i.e., QoE (W EB) = 1 when T H ' 0 kbps).
reading time between them. Note that the above-described QoE models do not depend
Traffic appears as data bursts; therefore, new connections on traffic model parameters (e.g., video sequence duration or
follow a Poisson distribution for all services [33] [34]. file/web page size).
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 4

IV. E XPERIENCE BALANCING ALGORITHM

In this section, a new QoE balancing algorithm with two
variants is presented.

A. Experience balancing algorithm

In this section, a new self-tuning algorithm to equalize
the QoE across cells in a LTE network is presented. The
proposed algorithm, referred to as Experience Balancing (EB),
is inspired in the MLB algorithm described in [19] (hereafter
denoted as LB). Similar to LB, EB is implemented by Fuzzy
Logic Controllers (FLC) that decide whether to increase or Fig. 2: Structure of fuzzy logic controller [19].
decrease HOM on a per-adjacency basis. Compared to classi-
cal proportional-integrative-derivative (PID) controllers, fuzzy
logic controllers are simpler and easier to understand, since negative, zero. . . ) [38]. Such a mapping is carried out by input
they are described in natural language, taking advantage of membership functions, µx , as shown in Figure 3a (x denotes
operator experience. Unlike LB, EB aims to equalize average the specific linguistic variable). For simplicity, triangular input
user QoE, instead of cell load, among cells. For this purpose, membership functions are selected. Experience shows that, in
users are sent from a cell with a lower QoE to an adjacent mobile networks, the flexibility provided by more complex
cell experiencing higher QoE. functions in the controller is not translated into a finer control
Two variants of EB are defined. In a first variant, referred due to the stochastic noise of input measurements. Also note
to as EB-C (for cell), the indicator to be balanced is the cell that function overlapping causes that a crisp input value
average QoE, defined as is simultaneously assigned to one or more adjectives with
P different degrees.
QoE(i, s)
s In the inference stage, a set of “IF-THEN” rules define the
QoE(i) = , (7)
Ns (i) mapping of the input to the output in linguistic terms. A rule
takes the form ’if x is A, then y is B’, where A and B are
where QoE(i, s) is the average QoE for users demanding adjectives associated to input and output, respectively (i.e.,
service s in cell i, defined as very negative, negative, zero, positive and very positive). The
P
QoE (s) (u) first part of the rule (x is A) is the antecedent, while the second
∀u∈i ,S(u)=s part of the rule (y is B) is the consequent. Unlike traditional
QoE(i, s) = , (8) expert systems, several rules can be fired at the same time
Nu (i, s)
in a fuzzy inference engine. The firing strength of each rule
where QoE (s) (u) is the quality of experience for user u depends on the degree in which its antecedents are satisfied
demanding service s, computed as in (3)-(6), Ns is the number (referred to as the truth value of the rule). This feature of
of services in cell i and Nu (i, s) is the number of users in cell fuzzy controllers ensures smooth control actions.
i demanding service s. In (7), it is implicitly assumed that all Figure 3c summarizes the set of rules that describe the
services are equally important for the operator. This is aligned tuning process in EB. Roughly, HOM is decreased (i.e.,
to the goal of ensuring that all users have the same service 4HOM (i, j) is N or VN) when the average QoE in the target
experience, regardless of the particular selected service. cell is better than in the source cell (i.e., QoEdif f (i, j) is P
Finally, the global average QoE is defined as or VP).
Nc
P Finally, in the defuzzification stage, the output value is
QoE(i) obtained from the aggregation of rules. In this work, the
i=1
QoE = , (9) centre-of-gravity method [39] is applied to compute the final
NC
output value as a weighted average. Weights are calculated
where NC is the number of cells in the network. from the truth value of each rule, computed from the degree
As stated before, EB-C aims to balance the cell average of fulfillment of their antecedents. For simplicity, a Takagi-
QoE (7). Thus, a difference indicator is defined for EB-C as Sugeno approach [38] is used, where the output membership
functions are constants, as shown in Figure 3b, leading to
QoEdif f (i, j) = QoE(j) − QoE(i) . (10) a more compact representation easier to adjust and reducing
Such an indicator is used as an input for EB-C FLC. Figure 2 computational load. More complex output membership functi-
shows the structure of the FLC. The output variable is the ons tend to give similar results.
increment/decrement of the HOM between neighbor cells i FLCs are periodically executed after a fixed period, referred
and j, ∆HOM (i, j). As shown in the figure, FLC consists of to as reporting output period (ROP). The value of HOM (i, j)
three stages: fuzzification, inference and defuzzification. for the next iteration, referred to as optimization loop, is
In the fuzzification stage, the value of the input, QoEdif f (i, j), calculated from performance measurements in the previous
is broadly termed with linguistic variables (e.g., very negative, iteration as
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 5

some services may have larger QoE than others within a cell
(i.e., QoE(i, s1 )6=QoE(i, s2 )) or different to the same service
in other cells (i.e., QoE(i, s1 ) 6= QoE(j, s1 )). To solve this, a
second variant of the algorithm, referred to as EB-CS, aims to
balancing the average service QoE of a cell against the average
user QoE of its adjacent cells. EB-CS takes advantage of the
fact that some LTE vendors now give operators the flexibility
to set different HOM values for services within a cell. Thus,
HOM can be tuned on a per-adjacency and per-service basis
(a) Input membership functions. driven by a new difference indicator

QoEdif f (i, j, s) = QoE(j) − QoE(i, s) . (12)

Similarly to EB-C, QoEdif f (i, j, s) in EB-CS is defined per

adjacency, but, differently to EB-C, it is segregated by service.
Thus, in EB-CS, one FLC is needed per service and adja-
cency, each proposing a service-specific margin modification,
HOM (i, j, s). EB-CS can be understood as four different QoE
balancing mechanisms per adjacency (i.e., one per service),
which might push users in different directions, e.g., WEB users
are handed over from cell i to j, but VIDEO users are handed
(b) Output membership functions.
over from j to i. Membership functions and inference rules
are identical to those in EB-C, shown in Figure 3, with the
QoEdif f (i, j) ∆HOM (i, j)
only change of the input parameter, QoEdif f (i, j, s) instead
VP VN of QoEdif f (i, j).
P N Another significant difference of EB-CS compared to EB-
Z Z C is the loss of symmetry. The difference indicator in (12)
N P is calculated by subtracting QoE(i, s) from QoE(j), so that
VN VP QoEdif f (i, j, s) 6= QoEdif f (j, i, s). Thus, unlike EB-C, two
(c) Rules. FLCs must be executed in EB-CS for the same adjacency, one
per direction. This separate FLC executions in two adjacent
Fig. 3: Fuzzy logic controller for EB-C algorithm. cells i and j might lead to changes of different magnitude for
both directions of the same adjacency, ∆HOM (i, j, s) and
∆HOM (j, i, s), causing that (2) is not fulfilled. To enforce
a constant hysteresis level, an average ∆HOM value is
calculated in EB-CS for both directions of the adjacency as
HOM (n+1) (i, j) =
=min(max(round(HOM (n) (i, j)+
(n) (n)
∆HOM (n) (i, j)), −7), 13) , (11) ∆HOM (i, j, s) = −∆HOM (j, i, s) =
(n)
∆HOM (i, j, s) − ∆HOM (n) (j, i, s)
where superscripts n and n + 1 denotes the iteration number, = . (13)
and all parameters are in dB. Note that HOM changes are 2
limited to the range −7 to 13 dB. The lower limit is the It can be argued that the proposed schemes are based on
minimum signal-to-interference ratio (SINR) needed for the QoS rather than on QoE, since QoE metrics used to drive the
scheduler to assign any radio resource to a connection. The tuning process are derived from QoS measurements. Note that
upper limit is calculated with (2) to ensure a hysteresis level of the utility functions mapping QoS into QoE are not linear, and
H = 6 dB. From (2), it is also deduced that any modification a large QoS increment does not necessarily lead to a large QoE
in HOM (i, j) automatically implies an opposite change in increase. Such a non-linearity is critical when evaluating the
HOM (j, i). As a consequence, a single FLC is needed for overall system performance gain (in terms of user satisfaction)
both directions of the adjacency. Thus, the required number of reassigning users to a different cell. This issue is avoided
of FLC equals to the number of adjacencies in the system. by explicitly computing QoE.
Due to their similarities, EB-C and LB would work the same
(i.e., propose the same HOM changes) if the service mix was
V. P ERFORMANCE ANALYSIS
the same for all cells. However, as will be shown later, service
mix differs from cell to cell in live networks, causing that EB- The proposed algorithm is tested in a dynamic system-level
C and LB take different tuning actions. LTE simulator [40]. For clarity, the simulation set-up is pre-
Note that, even if EB-C ensures that the cell average QoE sented first and results are shown later. Then, implementation
is balanced across the network (i.e., QoE(i) ' QoE(j) ∀ i, j), issues are discussed.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 6

4.4 two stages. In a first stage, the aim is to evaluate the impact
WEB
VIDEO of network load on the QoE of each service. With this aim,
VoIP offered traffic is swept by increasing the number of users in
4 one of the three cells (e.g., WEB users in C1) while traffic
C1
intensity in the others is kept constant (e.g.,VIDEO and VoIP
C3 in C2 and C3, respectively). For each offered traffic value,
3.6
cell load is measured as the average PRB utilization, U (i),
y [km]

for the whole simulated period (i.e., 1 hour of network time).

3.2 Likewise, offered traffic in cells with constant traffic (VIDEO
C2
in C2 and VoIP in C3 in the example) is fixed to a sufficiently
large value (U (C2)'U (C3)' 80 %) to generate a background
2.8 interference level. Users are uniformly distributed within every
cell. The same test is repeated by sweeping traffic intensity in
2.4
the other two cells. Note that, in this first stage, due to the
1.5 2 2.5 3 3.5 specific spatial user distribution, QoE(i) coincides with the
x [km] average QoE for the service demanded in the cell under study
Fig. 4: Naı̈ve scenario. (e.g., QoE(C1) = QoE(C1, W EB)).
In a second stage, the aim of the experiment is to illustrate
TABLE II: Simulation parameters. the limitations of classical load balancing in terms of QoE.
With this aim, HOM (i, j) is swept in 1 dB steps in a cell (e.g.,
Time resolution 10 TTI (10 ms) C3). Margins from that cell to the other two cells are simul-
Propagation Pathloss Okumura-Hata, slow fading
model (log-normal σ = 8 dB, dcorr = 20 m), fast taneously swept (e.g., HOM (C3, C1) = HOM (C3, C2)).
fading (ETU model) Hysteresis is maintained by synchronizing changes in both
Base station Tri-sectorized antennas, MIMO 2x2, directions of the adjacencies to ensure (2).
model BW = 5 MHz (25 PRB), fcarrier = 2 GHz,
EIRPmax = 68 dBm. 2) Second experiment (algorithm assessment): The aim
Scheduler Classical exponential/proportional fair [22] of the second experiment is: to check the behavior of the
Link adaptation CQI-based proposed experience balancing schemes and compare their
Cell geometry cell radius = 0.5 km,
Inter-site distance = 1.5 km performance against other traffic steering algorithms.
During the analysis, five self-tuning approaches are compa-
red. The first two are the proposed schemes that aim to equa-
A. Assessment methodology lize user experience across cells, EB-C, or cells and services,
EB-CS. A third scheme is the legacy MLB algorithm [19], LB,
Two main experiments are performed. In a first experiment, whose aim is to equalize the average PRB utilization between
the aim is to check how balancing load and balancing QoE neighbor cells. For a fair comparison, a fourth scheme, referred
behave differently in a naı̈ve scenario. Then, in a second to as throughput-based balancing, TB [41], is also included,
experiment, the proposed experience balancing schemes are to show the benefit of explicitly considering QoE instead
compared with other self-tuning approaches in a realistic of QoS (user throughput). TB aims to balancing the mean
scenario to assess their performance gain. In both experiments, user throughput, T (i), across cells by tuning handover margin
only the downlink is simulated to reduce the computational settings on a per-adjacency basis. For this purpose, a fuzzy
load. controller is implemented to steer users from cells with lower
1) First experiment (proof of concept): This experiment is user throughput to cells with larger user throughput. The input
a proof of concept whose aim is to reveal that load balance to the controller is the difference of mean user throughput
between two neighbor cells does not imply QoE balance. between adjacent cells, and the output is the handover margin
Figure 4 shows the naı̈ve scenario used in the first experiment, of the adjacency. Finally, a fifth scheme referred to as QoE-
consisting of a regular tri-sectorized scenario. Table II shows based reprioritization, QR [28], is included to show the benefit
relevant simulation parameters. of redistributing users between cells (as in EB-C and EB-
Only for this proof of concept, traffic demand is confined to CS) instead of reprioritizing services inside a cell by packet
three cells (denoted as C1, C2 and C3) and it is forced that all scheduling. QR aims to balance the QoE of users within a
users in a cell demand the same service (WEB, VIDEO and cell by reprioritizing services in a classical scheduler. For this
VoIP for C1, C2 and C3, respectively), as shown in Figure 4. purpose, a set of four proportional controllers (1 per service)
User locations are represented by a different symbol depending are implemented per cell to tune service priority on a long-term
on the requested service. Likewise, it is assumed that users are basis so that users of services with worse QoE are prioritized.
static (only in this naı̈ve scenario) to ensure that users do not The input to each controller is the average QoE difference of
change cell for mobility reasons. Thus, it is easier to segregate a service against other services, and the output is the service
users per service, which facilitates the analysis. Note that, once priority parameter of that service in the scheduler.
users are handed over to a different cell as a result of HOM A live LTE network scenario is implemented in the si-
tuning, different services can be found in the same cell. mulator for this second experiment. Figure 5 shows the si-
In the above-described scenario, the first experiment has mulated scenario, consisting of 108-macrocells (36 sites with
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 7

TABLE III: Parameters in real scenario. where Uimb (i) is the average PRB utilization imbalance of cell
Bandwidth 10 MHz (50 PRB)
i, computed by comparing its average PRB utilization against
Base station model EIRPmax = 67 dBm that of its neighbors, A(i) is the set of neighbor cells of cell
fcarrier =1850 MHz i, Nadj (i) is the number of neighbor cells of cell i and Nc is
Traffic model Spatial traffic distribution and
service mix based on live statistics
the number of cells in the scenario. An overall intra-cell QoE
collected on a per-cell basis imbalance indicator is defined as
Mobility model Random direction, constant speed,
3km/h
HOM (0) (i, j) 3 dB ∀ (i, j) 1 X
SP I (0) (i, s) 7 ∀ (i, s) QoEimb,f = |QoEimb,f (i)| =
NC i
1 1 XX
= 4QoE(i, sk ) , (16)
3 tri-sectorized antennas per site). Table III shows the main NC Ns (i) i
k
simulation parameters, taken from the live network. In this
where
experiment, users move at 3 km/h in a straight path randomly
selected. Likewise, the default HOM and SPI settings are 3 P
dB and 7, respectively. QoE(i, s)
s6=sk
The five self-tuning algorithms (LB, TB, QR, EB-C and 4QoE(i, sk ) = QoE(i, sk ) − , (17)
EB-CS) are tested along 15 optimization loops. It is checked Ns (i) − 1
a posteriori that the system reaches stability after 15 iterations and QoEimb,f (i) is the QoE imbalance among services in cell
in the five algorithms. The duration of every optimization loop i, calculated as the mean value of the difference between the
(i.e., the ROP) is 1 hour, long enough to ensure reliable perfor- QoE of a service and the mean QoE for the rest of the services,
mance statistics. At the end of each loop, the indicators used as in (17).
as drivers (U (i), T (i), QoE(i, s), QoE(i) and QoE(i, s)) are Similarly, an overall throughput imbalance indicator is de-
collected and algorithms are triggered. After each optimization fined as
loop, the system updates HOM or SPI values and a new
optimization loop begins. For a fair comparison, it is ensured
that all optimization loops for the five algorithms are executed 1 X
Timb = |Timb (i)| =
under identical conditions by pre-generating a realization of NC i
all random variables. Thus, performance differences between T (j)
P
loops are only due to the different HOM/SPI settings, and not 1 X j∈A(i)
to the stochastic nature of simulation. Network performance = T (i) − , (18)
NC i Nadj (i)
with the default HOM/SPI settings is considered as a baseline.
The aim of the proposed algorithms (EB-C and EB-CS) is to
where Timb (i) is the throughput imbalance indicator of cell i,
reduce differences between users of different cells and servi-
computed by comparing its average throughput against that of
ces. This is achieved by improving the worst users/services
its neighbors, T (j) is defined as
at the expense of deteriorating the best users/services. For
consistency, the main figure of merit is the 5th percentile of P
T (i, s)
the QoE distribution across cells and services in the network, T (i) = s , (19)
(5%−th)
QoE (i, s). Ns (i)
A secondary figure of merit is the overall QoE, computed as where
the average of all services and cells in the scenario, P
T (u)
u∈(i,s)
T (i, s) = . (20)
P
1 X 1 X QoE(i, s)
s
Nu (i, s)
QoE = QoE(i) = . (14)
NC i NC i
Ns (i) and T (u) is the connection throughput of user u. Likewise, an
overall QoE imbalance indicator across cells in the scenario
Five additional key performance indicators are defined to
is defined as
check performance differences across cells and services in the
network. An overall cell load imbalance indicator is defined
as 1 X
QoEimb,c = |QoEimb,c (i)| =
NC i
1 X P
QoE(j)
Uimb = |Uimb (i)| =
NC i 1 X j∈A(i)
= QoE(i) − , (21)
P
U (j) NC i Nadj (i)
1 X j∈A(i)
= U (i) − , (15)
NC i Nadj (i) where QoEimb,c (i) is the average cell imbalance indicator of
cell i, computed by comparing its average QoE against that
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 8

Fig. 5: Real scenario.

TABLE IV: Baseline network performance. 5

Indicator Avg. Max Min
4.5
Nu (i, V oIP )/Nu (i) [%] 1.8e-4 2.7e-3 0
Nu (i, V IDEO)/Nu (i) [%] 37.21 54.96 5.52
Nu (i, F T P )/Nu (i) [%] 27.09 72.14 1.83 4
Nu (i, W EB)/Nu (i) [%] 35.79 43.22 22.33
U (i) [%] 75 100 11.9
3.5
QoE(i) 3.02 4.42 1.89

3
C3 (VoIP)
of its neighbors. Finally, an overall QoE imbalance indicator
2.5 C2 (VIDEO)
across services in the scenario is defined as
C1 (WEB)
2
1 X
QoEimb,s = |QoEimb,s (i)| = 1.5
NC i
P 0 20 40 60 80 100
QoE(j)
1 1 X X j∈A(i)
= QoE(i, s) − , (22) Fig. 6: QoE dependence on cell load.
NC Ns (i) i s Nadj (i)

where QoEimb,s (i) is the average cell and service imbalance B. Results
indicator of cell i, computed by comparing its average QoE
per service against the average QoE of its neighbors. 1) First experiment (proof of concept): Figure 6 shows the
For clarity, Table IV shows some relevant network perfor- sensitivity of cell average QoE, QoE(i), to cell load, U (i)
mance indicators with the default HOM/SPI settings. Both (i.e., PRB utilization). Each curve represents one of the three
spatial user distribution and service mix, giving the probability cells (services) in the naı̈ve scenario. As expected, similar load
of a user initiating a connection of a service in a cell, are conditions do not lead to the same QoE values in the three
taken from real statistics. Only the call arrival rate is artificially services. Specifically, QoE(C3) > QoE(C1) > QoE(C2)
modified (i.e., increased) to generate a highly loaded scenario. when U (i) > 68 %. Thus, it is inferred that, for the scheduling
From the table, it is deduced that VIDEO is the most popular algorithm in the simulator, VoIP has better experience than
service in the area, followed by WEB. Likewise, with the WEB or VIDEO for high cell load. It can also be observed
default HOM and SPI settings, cell load may differ in up that the QoE of VoIP keeps almost constant and high up to
to 88.1 % and cell-average QoE may differ in up to 2.53 a very large cell load (i.e., QoE(C3) ' 4.4 ∀ U (C3) > 97
MOS points, justifying the need for the tuning process. It %). The same holds for VIDEO and WEB services, but with
should be pointed out that, in the considered live scenario, lower load thresholds (U (i) ≈ 58 % and 56 %, respectively).
VoIP traffic is extremely low and scattered in a few cells in Large markers in Figure 6, represent the working point
the network. As this might cause unreliable QoE statistics, selected for the next stage of the experiment, whose aim is to
EB-CS is not allowed to change HOM settings for this service show the benefit of QoE balancing. Such settings correspond to
(i.e., HOM (i, j, V oIP ) = 3 in EB-CS) as well as QR is not a situation where the three cells have a similar cell load close
allowed to change SPI settings (i.e., SP I(i, V oIP ) = 7 in to 90.8%, but completely different QoE values (QoE(C3) =
QR). 4.4, QoE(C2) = 3.21 and QoE(C1) = 2.71). This situation
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 9

5 100 0.5

4.5
95 0.4
4

3.5 90 0.3

3
85 0.2
2.5

2 C1 LB
C2 80 0.1 TB
C3 QR
1.5
EB-C
EB-CS
1 75 0
3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
loop index
Fig. 7: Sensitivity of cell load and QoE to HOM changes. Fig. 8: Evolution of QoE imbalance.

reflects an evenly balanced scenario in terms of cell load, but HOM offset is computed for EB-C and EB-CS as the average
an unevenly balanced scenario in terms of QoE. It is thus deviation of HOM values from the initial default settings, i.e.,
expected that a load balancing algorithm would not modify X
HOM values, even if large QoE differences exist among cells. δHOM (i, j, s)
In contrast, a QoE-driven balancing algorithm would change (i,j,s)
δHOM = =
HOMs to equalize QoE among cells. Nadjs
P
In the second stage of the experiment, HOMs are tuned to |HOM (i, j, s) − 3|
steer traffic from C1 and C2 (the two cells with the worst (i,j,s)
= , (23)
QoE) to C3 (the cell with the best QoE). Such an effect is Nadjs
achieved by enlarging the cell service area of C3, produced where Nadjs is the total number of adjacencies in the network.
by increasing HOMs in the outgoing adjacencies of C3. Figure 10 illustrates the average HOM deviation evolution
Figure 7 shows cell load (solid lines) and QoE (dashed across iterations in all the schemes. At the 15th loop, δHOM
lines) values for the three cells, when HOM (C3, C1) and reaches 7 dB, 5.6 dB and 0 dB and for LB, TB and QR,
HOM (C3, C2) are simultaneously swept in 1 dB steps respectively, showing that LB and TB produce a significant
from 3 (the default setting) to 11 dB. In the figure, it is displacement of HOMs in many adjacencies, while, as ex-
observed that load imbalance increases as QoE imbalance pected, QR does not change HOM. EB-C and EB-CS also
decreases. Specifically, load imbalance increases from Uimb = produce a deviation of HOMs in many adjacencies. In EB-C,
0.3 % to 12.7 % while QoEimb,c decreases from 0.96 % δHOM reaches 4.6 dB at the 15th optimization loop, which is
to 0.36 %. Thus, a load balancing algorithm would end up less than the deviation needed by LB and TB to reach load or
in HOM (C3, C1)=HOM (C3, C2)= 3 dB, while a QoE throughput balance. This proves again that an evenly balanced
balancing algorithm would set a completely different balance load or user throughput across the network do not necessarily
point with HOM (C3, C1/C2)= 11 dB. This is clear evidence imply an evenly balanced QoE. In EB-CS, δHOM ranges
that load balancing and QoE balancing might drive the system from 0 to 6 dB depending on the service. VIDEO service
to very different states in the presence of different service requires a larger HOM deviation, indicating that, with the
mixes in cells. current service mix, it is needed (on average) to hand over
2) Second experiment (algorithm assessment): Figure 8 a larger amount of VIDEO users than FTP or WEB users to
shows the impact of LB, TB, QR and EB algorithms on QoE reach QoE balance among services of neighbor cells.
imbalance among cells along the 15 optimization loops. As For comparison purposes, Table V summarizes the main
illustrated, LB does not change QoEimb,c significantly, QR performance indicators at the beginning (column Initial) and
increases QoEimb,c and TB achieves a slight reduction of the end of the tuning process (15th optimization loop) for
QoEimb,c . In contrast, both EB-C and EB-CS more than half the different schemes. As expected, LB achieves the best load
the initial imbalance. balance, Uimb , TB achieves the best throughput balance among
To spot the difference between EB-C and EB-CS, Figure 9 services within a cell, Timb (0.28 Mbps) and QR achieves
shows the evolution of the imbalance between cells and servi- the smallest QoE imbalance among services within a cell,
ces, QoEimb,s , across iterations in both schemes. As expected, QoEimb,f (0.32). EB-C gets a better QoE balance than LB,
EB-CS better equalizes QoE among cells and services due to TB or QR, QoEimb,c (0.18). However, EB-CS achieves the
its service-based design. To clarify this capability, an average smallest QoE imbalance across cells (QoEimb,c = 0.14) and
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 10

0.7 1
EB-C (VIDEO)
EB-C (FTP)
0.6 EB-C (WEB) 0.8 Initial
EB-CS (VIDEO)
EB-CS (FTP) EB-C
EB-CS (WEB)
0.5 0.6 EB-CS

CDF
0.4 0.4

0.3 0.2

0.2 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 1.5 2 2.5 3 3.5 4 4.5 5
loop index
Fig. 9: Evolution of QoE imbalance per service. Fig. 11: QoE distribution for services across cells function.

8 TABLE V: Main performance indicators.

7 Indicator Initial LB TB QR EB-C EB-CS

Uimb [%] 15.2 6.3 15.2 14.9 17 16.6
6 Timb [M bps] 1.01 4.57 0.28 0.52 4.09 0.79
QoEimb,f 0.51 0.50 0.49 0.32 0.46 0.33
5 QoEimb,c 0.4 0.45 0.30 0.35 0.18 0.14
QoEimb,s 0.51 0.56 0.45 0.42 0.35 0.26
QoE 3.02 2.94 2.96 2.99 2.94 2.93
4 QoE(i, s) 5th tile 1.89 1.90 2.01 2.07 2.10 2.36
LB
3 TB
QR
EB-C
2 modifies HOM only on a per-adjacency basis. Thus, the cell
EB-CS (VoIP)
EB-CS (VIDEO) service area is shared by all services. Figure 12b depicts
1 EB-CS (FTP) the service area of the same cell with EB-CS at the 15th
EB-CS (WEB) optimization loop. EB-CS algorithm modifies HOM on a per-
0 adjacency and per-service basis. Therefore, the cell service
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 area is different depending on the service (VIDEO, FTP or
loop index WEB). VoIP case is not shown because VoIP traffic in the
Fig. 10: Evolution of average HOM offset. network under analysis is negligible. Note that the cell service
area produced by EB-C, shown in Figure 12a, is completely
different from those of EB-CS, shown in Figure 12b. This
services (QoEimb,s = 0.26). This is the result of adjusting flexibility of changing cell service areas on a per-adjacency
HOM on a per-adjacency and per-service basis. and per-service basis is the reason for the superiority of EB-
Finally, Figure 11 shows the cumulative distribution CS when equalizing the QoE among cells and services.
function of QoE(i, s) achieved by EB and EB-CS. It is obser- Another experiment has been carried out to check the
ved that both balancing approaches deteriorate the QoE of the ability of the proposed iterative algorithm (EB-CS) to adapt
best cells/services to improve that of the worst cells/services. to changes in the number of users and the traffic mix. For
Focusing on the worst cells and services (lower left), it is this purpose, once the system is stable and has reached QoE
observed that EB-CS achieves the best improvement for those balance, the number of users is modified by increasing the
cells and services experiencing the lowest QoE values. This Poisson arrival rate of each service by 4%. This change
is also shown in Table V, where EB-CS has the highest modifies the balance point, so imbalance is expected to appear
value for the QoE indicator representing the worst users (i.e., again and EB-CS starts to modify margins searching for
(5%−tile)
QoE (i, s) = 2.36). the new balance point. After this second balance stage, the
The stronger QoE balancing effect with EB-CS is achieved traffic mix is modified now by changing the percentage of
by modifying cell service areas on a per-adjacency and per- FTP, VIDEO and WEB. WEB traffic is decreased by 24%,
service basis. Figure 12a illustrates how the service area of a whereas FTP and VIDEO traffic is increased by 12%. Table
particular cell is modified by EB-C at the end of the tuning VI summarizes the results for these traffic changes, showing
process (15th optimization loop). Note that EB-C algorithm the value of the QoE imbalance indicator, QoEimb,c , and the
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 11

7.75 C. Implementation issues

Initial The EB-C algorithm is executed on a per-adjacency basis,
7.5 and, therefore, its worst-case time complexity is O(Nadjs ).
EB-C
In contrast, The EB-C algorithm is executed on a per-
7.25 adjacency and per-service basis, so that its time complexity
is O(Nadjs ∗ Ns ). Both algorithms have been implemented
y [km]

7 with the Fuzzy Logic Toolbox in Matlab. For the considered

scenario, consisting of 108 cells, 11664 adjacencies and 3
6.75 services, the average execution time of 1 iteration of EB-C
and EB-CS is 2.8 and 6.5 seconds in a personal computer
6.5 with a 3.6-GHz octa-core processor and 16 GB of RAM.

6.25 VI. C ONCLUSIONS

7.25 7.5 7.75 8 8.25 8.5 8.75 In this paper, a self-tuning algorithm for adjusting handover
x [km] margins in a LTE network has been proposed. The aim of the
(a) Initial and EB-C. algorithm is to balance the QoE between cells and services.
7.75 The proposed iterative algorithm changes handover margins
between adjacent cells to push users from cells with a lower
Initial
EB-CS (VIDEO) QoE to neighbor cells with a higher QoE. Two variants have
7.5
EB-CS (FTP) been presented, depending on whether margins are tuned on a
EB-CS (HTTP) per-adjacency or per-adjacency and per-service basis. Method
7.25
assessment has been carried out in a dynamic system-level
LTE simulator implementing a realistic macrocellular scenario.
y [km]

7 Results have shown that the average QoE of cells becomes less
imbalanced after parameter tuning. Specifically, the average
6.75 QoE difference among cells is reduced by 0.22 and 0.26 MOS
points with EB-C and EB-CS, respectively, with an average
6.5 HOM change between 0 and 6 dB depending on the service.
The proposed algorithm is conceived as a centralized solu-
6.25
tion for the network management system, since QoE statistics
7.25 7.5 7.75 8 8.25 8.5 8.75 needed by the algorithm are currently obtained by packet
x [km] inspection techniques in selected core network interfaces [42].
(b) EB-CS. The underlying iterative algorithm is devised to be executed
Fig. 12: Cell service area per service. after each reporting output period (e.g., 1 hour). Such a time
window ensures reliable QoE measurements for long video
streaming sessions. If faster changes are needed, the proposed
algorithm could be executed with a shorter periodicity (e.g.,
TABLE VI: System performance to network changes. minutes), provided that reliable QoE estimates are available.
Stage Initial network Network load Change in traffic It is envisaged that such information will be delivered to SON
conditions increase, ∆λ mix, ∆λ(s) frameworks as part of big data generated by future 5G mobile
Loop index 1 15 16 26 27 45 communication systems.
QoEimb,s 0.51 0.26 0.43 0.26 0.48 0.29
QoEimb,c 0.4 0.14 0.29 0.15 0.32 0.15 ACKNOWLEDGMENTS
This work has been partially funded by the Spanish Mini-
stry of Economy and Competitiveness (TEC2015-69982-R),
the Spanish Ministry of Science, Innovation and Universi-
value of the QoE imbalance per service, QoEimb,s , at the ties (RTI2018-099148-B-I00) and the Horizon 2020 project
beginning/end of each stage. It is observed that any change ONE5G (ICT-760809), receiving funds from the European
in network conditions produce a temporary QoE imbalance Union.
among cells and services, (QoEimb,s = 0.43 and QoEimb,c =
0.29 when traffic is generally increased, and QoEimb,s = 0.48 R EFERENCES
and QoEimb,c = 0.32 when traffic mix is changed). These
[1] Nokia Siemens Networks, ”Understanding Smartphone Behavior in the
imbalances are successfully corrected by EB-CS after a few Network,” White paper, 2011.
iterations, leading network performance to a similar balance [2] Ericsson AB, ”Ericsson Mobility Report”, Nov. 2017.
point than that of the first stage of the experiment in the [3] S. Barakovic and L. Skorin-Kapov, “Survey and Challenges of QoE
Management Issues in Wireless Networks,” vol. 2013, Mar. 2013.
15th iteration. Thus, it is shown that EB-CS can cope with [4] NGMN, ”Next Generation Mobile Networks Recommendation on SON
fluctuations of traffic demand during a day. and O&M requirements”, 2008.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 12

[5] V. F. Monteiro, D. A. Sousa, T. F. Maciel, F. R. M. Lima, E. B. Rodri- [27] Z. Yan, J. Xue, and C. W. Chen, “Prius: Hybrid Edge Cloud and Client
gues, and F. R. P. Cavalcanti, “Radio resource allocation framework for Adaptation for HTTP Adaptive Streaming in Cellular Networks,” IEEE
quality of experience optimization in wireless networks,” IEEE Network, Transactions on Circuits and Systems for Video Technology, vol. 27,
vol. 29, no. 6, pp. 33–39, Nov. 2015. no. 1, pp. 209–222, Jan. 2017.
[6] H. Nam, K. H. Kim, J. Y. Kim, and H. Schulzrinne, “Towards QoE-aware [28] P. Oliver-Balsalobre, M. Toril, S. Luna-Ramı́rez, and J. M. Ruiz Avilés,
video streaming using SDN,” in 2014 IEEE Global Communications “Self-tuning of scheduling parameters for balancing the quality of
Conference, Dec. 2014, pp. 1317–1322. experience among services in LTE,” EURASIP Journal on Wireless
[7] S. Hämäläinen, H. Sanneck, and C. Sartori, “LTE Self-Organizing Communications and Networking, vol. 2016, no. 1, p. 7, Jan. 2016.
Networks (SON): Network Management Automation for Operational [29] P. Oliver-Balsalobre, M. Toril, S. Luna-Ramı́rez, and R. G. Garaluz,
Efficiency Hardcover,” 2012. “Self-Tuning of Service Priority Parameters for Optimizing Quality
[8] O. G. Aliu, A. Imran, M. A. Imran, and B. G. Evans, “A Survey of of Experience in LTE,” IEEE Transactions on Vehicular Technology,
Self Organisation in Future Cellular Networks,” IEEE Communications vol. 67, no. 4, pp. 3534–3544, Apr. 2018.
Surveys and Tutorials, vol. 15, no. 1, pp. 336–361, 2013. [30] Scenarios, requirements and KPIs for 5G mobile and wireless system
[9] J. Kojima and K. Mizoe, “Radio mobile communication system wherein ICT-317669 METIS project, D6.1 v1-2, 2013.
probability of loss of calls is reduced without a surplus of base station [31] 3GPP TSG RAN WG1, 3GPP TR 25.892 v6.0.0; Feasibility Study
equipment, U.S. Patent 4435840,” vol. 54, no. 5, pp. 1875–1886, Sep. for Orthogonal Frequency Division Multiplexing(OFDM) for UTRAN
1984. Enhancement(Rel-6), pp. 62-63, 2004.
[10] N. Papaoulakis, D. Nikitopoulos, and S. Kyriazakosin, “Practical radio [32] P. Seeling and M. Reisslein, “Video Transport Evaluation With H.264
resource management techniques for increased mobile network perfor- Video Traces,” IEEE Communications Surveys Tutorials, vol. 14, no. 4,
mance,” 12th IST Mobile and Wireless Communications Summit, Jun. pp. 1142–1165, Apr. 2012.
2003. [33] Y. Fang, I. Chlamtac, and Yi-Bing Lin, “Channel occupancy times
[11] A. J. Fehske, H. Klessig, J. Voigt, and G. P. Fettweis, “Concurrent and handoff rate for mobile computing and PCS networks,” IEEE
Load-Aware Adjustment of User Association and Antenna Tilts in Transactions on Computers, vol. 47, no. 6, pp. 679–692, Jun. 1998.
Self-Organizing Radio Networks,” IEEE Transactions on Vehicular [34] R. R. Tyagi, F. Aurzada, K. Lee, and M. Reisslein, “Connection
Technology, vol. 62, no. 5, pp. 1974–1988, Jun. 2013. Establishment in LTE-A Networks: Justification of Poisson Process
[12] V. Wille, S. Pedraza, M. Toril, R. Ferrer, and J. Escobar, “Trial Results Modeling,” IEEE Systems Journal, vol. 11, no. 4, pp. 2383–2394, Dec.
from Adaptive Hand-Over Boundary Modification,” IEE Electronics 2017.
Letters, vol. 39, pp. 405–407, Feb. 2003. [35] P. Reichl, B. Tuffin, and R. Schatz, “Logarithmic laws in service quality
[13] T. L. Casavant and J. G. Kuhl, “A taxonomy of scheduling in general- perception: where microeconomics meets psychophysics and quality of
purpose distributed computing systems,” IEEE Transactions on Software experience,” Telecommunication Systems, vol. 52, no. 2, pp. 587–600,
Engineering, vol. 14, no. 2, pp. 141–154, Feb. 1988. Feb. 2013.
[14] M. Toril and V. Wille, “Optimisation of Handover Parameters for Traffic [36] ITU-T G.114 Recommendation, ”One-Way Transmission Time,”, 2003.
Sharing in GERAN,” Wireless Personal Communications, vol. 74, no. 3, [37] J. Navarro-Ortiz, J. M. Lopez-Soler, and G. Stea, “Quality of experience
pp. 315–336, Feb. 2008. based resource sharing in IEEE 802.11e HCCA,” in 2010 European
[15] J. M. Ruiz-Avilés, M. Toril, S. Luna-Ramı́rez, V. Buenestado, and M. A. Wireless Conference (EW), Apr. 2010, pp. 454–461.
Regueira, “Analysis of Limitations of Mobility Load Balancing in a Live [38] T. Ross, Fuzzy logic with engineering applications. McGraw-Hill, 1995.
LTE System,” IEEE Wireless Communications Letters, vol. 4, no. 4, pp. [39] C. C. Lee, “Fuzzy logic in control systems: fuzzy logic controller.” IEEE
417–420, Aug. 2015. Transactions on Systems, Man, and Cybernetics, vol. 20, no. 2, pp. 404–
[16] R. Kwan, R. Arnott, R. Paterson, R. Trivisonno, and M. Kubota, 418, Mar. 1990.
“On Mobility Load Balancing for LTE Systems,” in 2010 IEEE 72nd [40] P. Muñoz, I. de la Bandera Cascales, F. Ruiz, S. Luna-Ramı́rez, R. Barco,
Vehicular Technology Conference - Fall, Sep. 2010, pp. 1–5. M. Toril, P. Lázaro, and J. Rodrı́guez, “Computationally-Efficient Design
[17] A. Lobinger, S. Stefanski, T. Jansen, and I. Balan, “Load Balancing in of a Dynamic System-Level LTE Simulator,” International Journal of
Downlink LTE Self-Optimizing Networks,” in 2010 IEEE 71st Vehicular Electronics and Telecommunications, vol. 57, pp. 347–358, Sep. 2011.
Technology Conference, May. 2010, pp. 1–5. [41] L. Gimenez, I. Z. Kovács, J. Wigard, and K. I. Pedersen, “Throughput-
[18] P. Muñoz, R. Barco, and I. de la Bandera, “Optimization of load Based Traffic Steering in LTE-Advanced HetNet Deployments,” Sep.
balancing using fuzzy Q-learning for next generation wireless networks,” 2015.
Expert Systems with Applications, vol. 40, no. 4, pp. 984 – 994, 2013. [42] A. Bär, P. Casas, A. D’Alconzo, P. Fiadino, L. Golab, M. Mellia, and
[19] J. M. Ruiz-Avilés, S. Luna-Ramı́rez, M. Toril, and F. Ruiz, “Traffic steer- E. Schikuta, “Dbstream: a Holistic Approach to Large-scale Network
ing by self-tuning controllers in enterprise LTE femtocells,” EURASIP Traffic Monitoring and Analysis,” Computer Networks, vol. 107, part 1,
Journal on Wireless Communications and Networking, vol. 2012, no. 1, pp. 5–19, 2016.
p. 337, Nov. 2012.
[20] K. I. Pedersen, T. E. Kolding, F. Frederiksen, I. Z. Kovacs, D. La-
selva, and P. E. Mogensen, “An overview of downlink radio resource
management for UTRAN long-term evolution,” IEEE Communications
Magazine, vol. 47, no. 7, pp. 86–93, Jul. 2009.
[21] F. R. M. Lima, T. F. Maciel, W. C. Freitas, and F. R. P. Cavalcanti,
“Resource Assignment for Rate Maximization With QoS Guarantees
in Multiservice Wireless Systems,” IEEE Transactions on Vehicular
Technology, vol. 61, no. 3, pp. 1318–1332, Mar. 2012.
[22] J. Rhee, J. M. Holtzman, and D.-K. Kim, “Scheduling of Real/Non-
real Time Services: Adaptive EXP/PF Algorithm,” in The 57th IEEE Marı́a Luisa Marı́-Altozano received her M.S.
Semiannual Vehicular Technology Conference, VTC 2003-Spring, vol. 1, degree in Telecommunication Engineering from the
Apr. 2003, pp. 462–466 vol.1. University of Málaga, Spain, in 2012. From 2013
[23] P. Ameigeiras, J. J. Ramos-Munoz, J. Navarro-Ortiz, P. Mogensen, and to 2016, she was with Ericsson in a collaborative
J. M. Lopez-Soler, “QoE oriented cross-layer design of a resource project with the University of Málaga. Since 2017,
allocation algorithm in beyond 3G systems,” Computer Communications, she has been working toward the Ph.D with the
vol. 33, no. 5, pp. 571 – 582, 2010. Communication Engineering Department, Univer-
[24] F. Wamser, D. Staehle, J. Prokopec, A. Maeder, and P. Tran-Gia, sity of Málaga. Her interests are focused on self-
“Utilizing Buffered Youtube Playtime for QoE-Oriented Scheduling in optimization of mobile radio access networks based
OFDMA Networks,” in 24th International Teletraffic Congress (ITC), on quality of experience.
no. 15, 2012.
[25] J. Chen, R. Mahindra, M. Khojastepour, S. Rangarajan, and M. Chiang,
“A scheduling framework for adaptive video delivery over cellular
networks,” Sep. 2013, pp. 389–400.
[26] V. Joseph and G. de Veciana, “NOVA: QoE-driven optimization of
DASH-based video delivery in networks,” in IEEE INFOCOM 2014 -
IEEE Conference on Computer Communications, April 2014, pp. 82–90.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2019 13

Salvador Luna-Ramı́rez received his M.S in Tele-

communication Engineering and the Ph.D degrees
from the University of Málaga, Spain, in 2000
and 2010, respectively. Since 2000, he has been
with the department of Communications Engineer-
ing, University of Málaga, where he is currently
Associate Professor. His research interests include
self-optimization of mobile radio access networks
and radio resource management.

Matı́as Toril received his M.S in Telecommunica-

tion Engineering and the Ph.D degrees from the
University of Málaga, Spain, in 1995 and 2007
respectively. Since 1997, he is Lecturer in the Com-
munications Engineering Department, University of
Málaga, where he is currently Full Professor. He
has authored more than 110 publications in leading
conferences and journals and 3 patents owned by
Nokia Corporation. His current research interests
include self-organizing networks, radio resource ma-
nagement and data analytics.

Carolina Gijón received her B.S. degree in Tele-

communication Systems Engineering from the Uni-
versity of Málaga, Spain, in 2016. Currently, she
is working towards the Ph.D. degree. Her research
interests include self-organizing networks and radio
resource management.