Вы находитесь на странице: 1из 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/311532272

Quality of Experience analysis for VoLTE services through Artificial Neural


Network fitting

Conference Paper · October 2016


DOI: 10.1109/WINCOM.2016.7777208

CITATION READS

1 176

2 authors:

Alessandro Vizzarri Fabrizio A. M. Davide


University of Rome Tor Vergata Università Telematica Guglielmo Marconi
13 PUBLICATIONS   41 CITATIONS    111 PUBLICATIONS   2,522 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

CROSS crossprogramme.eu View project

PASION View project

All content following this page was uploaded by Alessandro Vizzarri on 31 August 2017.

The user has requested enhancement of the downloaded file.


Quality of Experience Analysis for VoLTE Services
through Artificial Neural Network Fitting
Alessandro Vizzarri Fabrizio Davide
Department of Enterprise Engineering Department of Engineering of Innovation
University of Rome Tor Vergata University of G. Marconi
Rome, Italy Rome, Italy
alessandro.vizzarri@uniroma2.it f.davide@unimarconi.it

Abstract —We consider Quality of Experience (QoE), experiments, reported in Sect. 4. In Sect. 5 QoS metrics have
measured in terms of Mean Opinion Score Listening Quality been used as predictors, together with a feed-forward Neural
Subjective (MOS_LQS) for a Voice over LTE (VoLTE) Network compared with a linear multi-regression technique.
application in realistic situations. A set of service scenarios have This is motivated by a shared evidence of nonlinear behavior
been identified and the network performance simulated. We for MOS versus network parameters in realistic scenarios. Sect.
organized a listening panel to measure MOS according to the 6 will draft the most relevant conclusions out of the present
standard procedures. MOS-LQS results and QoS metrics have approach.
been correlated using a set of Artificial Neural Network (ANN)
models. As a clear result the ANN accurately models the II. QOE ANALYSIS FOR A VOLTE APPLICATION
relationship. This confirms again the need for an approach that is
non linear and capable of generalization, as made by the human Since LTE is fully IP-based wireless standard it only
judgment. Future research is necessary to generalize the ANN enables entire wireless transmission over Packet Switching (PS)
model to a far larger scale while controlling the combinatorial paths using the Internet Protocol (IP) protocol as the network
explosion of scenarios and technical parameters. protocol. All applications delivered over LTE systems are IP-
based, including voice applications as Voice Over LTE
Keywords—LTE; VoIP; VoLTE; end-to-end QoS; QoE; LTE (VoLTE) [3], [4], [5]. VoLTE is delivered in best effort
KPIs; IP cloud; ANN; MOS; MOS-LQS. modality. This implies VoLTE, as a best effort service, requires
All the authors contributed equally to this work.
an effective service management in order to guarantee both an
acceptable end-to-end QoS at network layer and an acceptable
QoE perceived by end-users. Vizzarri in [6] presents a review
of the most important papers focused on the end-to-end QoS in
I. INTRODUCTION case of applications delivered over LTE networks. Reference
Voice over LTE (VoLTE) is a key service for Long Term [7] analyses the QoS for a VoLTE service through and end-to-
Evolution (LTE) networks, and needs specific attention for end approach. Reference [8] studies the impact of network
quality of service (QoS). End-to-end approach for QoS is congestions on the VoLTE end-to-end performance. Major
strongly recommended by the standard not only for data-based standardization entities already treated the QoS issue in LTE
services but also for delay sensitive services VoLTE [1]. through an end-to-end approach. In [9] ETSI provides end-to-
MNOs analyze and monitor continuously the main network end QoS reference architecture for LTE and a description of
parameters (Key Performance Indicators, KPI) at network level relevant management functions. Implementation of QoS
and try to estimate Quality of Experience (QoE) perceived by policies and strategies are left to MNOs. QoE of a voice
end user at the application level. application is usually represented by the Mean Opinion Score
The present work considers QoE measured in terms of (MOS). MOS is a scalar variable, whose values are limited in
Mean Opinion Score Listening Quality Subjective for a VoLTE the range from 1 (worst case) to 5 (best case) [10], which
application in realistic situations. We are interested in studying indicates an average level of service acceptance for the end
the behavior for MOS (as QoE parameter) versus the network user. The MOS value is strictly related to the R factor provided
KPIs (i.e. QoS parameters). Research presents extensive by the ITU E-Model [11]. The E-Model combines a number of
samples of LTE networks modelled in realistic usage different impairments to be considered for an overall quality
conditions through scenario making. We propose the approach measure.
of modelling multiuser and multiservice scenarios according to
There are various different MOS measures, as shown in Fig.
reported conditions, and simulate the network performance
1. We can distinguish objective measurement that rely on
using the OPNET simulation software (as described in Sect. 2).
estimates of conversational and listening quality, based on
Many authors applied polynomial, exponential and logarithmic
technical measures and subjective measurements that rely on
functions to fit the objective MOS value (see references in [2]).
collaboration of human subjects. In case measures are
Then we attacked the subjective measurements for MOS. This
performed before the receiver we speak of MOS
required to invest a certain amount of resources to train and
Conversational Quality Estimated (MOS-CQE) and MOS
employ a panel of listeners for an extensive series of
Listening Quality Estimated (MOS-LQE). In case measures are
performed after the receiver we speak of MOS Listening the callee might start to talk at the same moment or interrupt
Quality Objective (MOS-LQO) and MOS Listening Quality each other. Further an excess of delay might generate Packet
Subjective (MOS-LQS), respectively objective and subjective. Loss because the delayed voice packet could be dropped. Jitter
In the present paper we are interested in subjective is the variability in the arrival time between packets, caused by
measurement of MOS. network congestion or route changes [17]. Negative effects of
Delay and Jitter are voice echoes, while high value of Packet
Loss can produce overlapping of words with a strong negative
impact on voice intelligibility. The effects of these QoS metrics
on QoE are well known, though we are interested in correlating
the QoE with the measured QoS metrics via neural networks.
There are two set of reasons. First, neural networks are a well-
known tool to fit non-linear input-output functions, and may be
easily used in our problem. Second, neural networks were
demonstrated to have a great potential in modeling human
behavior. On consequence we are interested in exploiting this
subtler property to find a way to reduce the expensive process
of full MOS-LQS measurement.
Fig. 1. MOS and voice quality estimation schemes for VoIP Services.

Available methods can be grouped in three classes. The III. RELATED WORK
“Conversational Opinion Test” requires a preparation in the The approach of the authors is to correlate the objective
laboratory in terms of creation of impairments in the VoLTE QoS KPIs at network level with MOS as QoE subjective
testbed system such as loss, delay and echo. Each test requires indicator. We will apply in the following different statistical
two subjects who have to sit in separate sound-proof rooms and regression techniques: Multiple Linear Regression (MLR) and
then talk to each other before finally rating the individual score, Artificial Neural Network (ANN) with different training
using 5-point scale. Also, each condition requires participants algorithms. In [18] the authors propose fuzzy-logic for
at least 24-32 subjects to test. The final result is the MOS QoS/QoE mapping. QoS/QoE mapping is carried out
Conversational Quality Subjective (a variant of MOS-LQS correlating Mean Opinion Score (MOS) as QoE parameter to
shown on Fig. 1). The “LQS Test” requires a laboratory too network KPIs as QoS parameter. In particular Delay, Jitter and
[12]. Preparation is somehow more complicate, because Packet Loss Rate are considered as KPIs. A comparison of
participants are taken individually and provided each with different approaches for the best fitting curve of the simulation
audio files that is the result of a high fidelity recording results is in [19] [20].
transmitted along the VoLTE service chain, including the
receiver. Each participant has to give the individual score, In [21] the authors conducted experimental tests in order to
using the usual 5-point scale. Measurement of each scenario evaluate the QoS of a generic VoIP service. On the basis of the
should be conducted with at least 16 subjects [13]. obtained test results, the authors tried to correlate the metrics
measured at network level (Packet Loss and Packet Delay) to
In the following we will focus on the listening test MOS- the corresponding MOS-CQS resulting from the subjective test.
LQS. We based on a transmission test carried out with pre- They proceeded to fit the obtained values through a MLR and
recorded phrases; the goal of this test is to obtain the absolute identified a polynomial relationship. Finally they calculated the
quality of the voice sample after transmission, through the related Mean Absolute Percent Error (MAPE).
direct hearing of the sample, without a reference sample. There
are various listening tests, i.e. Absolute Category Rating QoE/QoS mapping can be also analyzed and modelled
(ACR), Degradation Category Rating (DCR) and Comparison through the involvement of the ANNs [22] [23], since a
Category Rating (CCR) [14]. The quality must be evaluated desiderable output (QoE in our case) can be predicted on the
through different opinion scales: Listening quality scale, basis of multiple inputs (network QoS metrics at network
Listening-effort scale and Loudness-preference scale. After level). In [24] the authors try to correlate the QoE to QoS in
collection of each participant’s opinion of quality (expressed in case of a mobile data service through an ANN. After training,
the usual range), the MOS is computed as the mean of the test the ANN model gives two different types of relationship for the
panel. The ITU-T Rec. P.800 (1996) defined the requirements a QoE: direct if QoE is estimated against the bandwidth; inverse
MOS test has to comply with [15]. in case of estimation of Delay. In [25] V. A. Machado et alii
defined an ANN-based network model for QoS/QoE correlation
The final objective of our research is to correlate the MOS- in case of a video service delivered over a WiMAX network.
LQS with the QoS metrics of the VoLTE applications. Here we Test results confirm the goodness of the ANN model, which
adopt as the QoS metrics relevant for a VoLTE service three exibits an acceptable error. In [26] the authors apply the ANN
Key Performance Indicators (KPIs): end-to-end Delay, Packet to the estimation of the volume traffic matrix in a large scale [P
Loss rate (PLR) and Jitter (other quality of service measures for network. They compare different training algorithms (as
a generic VoIP application can be found in [16]). Delay is Levenberg-Marquardt and Bayesian Regularization) in terms of
expressed as the amount of time a packet sent by the source robustness and accuracy in order to perform a good prediction
(caller) takes to reach destination (callee). An excess of delay of traffic. Results confirm the Levenberg-Marquardt training
can make an audio conversation very difficult: both caller and algorithm to have the best error robustness.
IV. EXPERIMENTAL QCI 1 (GBR) and Allocation and Retention Priority (ARP 1).
The VoLTE application has been launched with a start offset of
A. Methodology 20s till the end of simulation period. In the HTTP application,
Our rationale is as follows: concerning scenarios from n.25 to n.48, UE_2 can download 1
∑ Define realistic scenarios for VoLTE application, to KB web page, n. 5 medium images with dimension up to 2 KB
be simulated for computing KPIs. and two short videos with dimension up to 350 KB. HTTP
∑ Perform MOS LQS testing per each scenario. application is launched with a start offset of 40s. Since modeled
∑ Attempt multi parameter fitting of QoE vs QoS KPIs profiles add a start offset around 40s, VoLTE and HTTP
through the ANN. applications start after 80s from simulation initiation.

B. Simulated scenarios
Realistic scenarios are simulated on the basis of network
impairments that disturb VoLTE calls. Impairments are
represented by:
∑ mixed traffic: VoLTE application is delivered over the
LTE network together with HTTP web browsing
application.
∑ insertion of IP cloud: additive delay and IP packet
discard ratio are produced across entire end-to-end
transmission chain.
This framework is typical when a backbone network
section is involved in service delivery. The HTTP application
is the application considered for modeling a set of multiuser a) b)
and multiservice scenarios. We made the exercise of
considering different LTE network topology (with or without
IP cloud) and traffic flows (single or mixed), as shown in Fig.
2, and identified a series of 48 scenarios.
Scenarios from n.1 to n.12 are characterized by two UEs:
UE_1 (caller) is performing a VoLTE call to UE_2 (callee)
using a direct link. Traffic flow is single. Topology is shown in
Fig. 2.a.
Scenarios from n.13 to n.24 are characterized by three UEs
and one HTTP web server: UE_1 and UE_2 are performing a
VoLTE call while UE_3 is performing an HTTP web browsing
session. Traffic flow is mixed: VoLTE and HTTP web
browsing services are performed by UEs (see Fig. 2.b). c) d)
Fig. 2. LTE network topology simulated for VoLTE application. One UE for
Scenarios from n.25 to n.36 are similar to those in the group VoLTE without IP cloud (a), two UEs for VoLTE and HTTP browsing
1-12: UE_1 is performing a VoLTE call to UE_2, but link without IP cloud (b), One UE for VoLTE with IP cloud (c), two UEs for
among them is affected by insertion of IP cloud. It adds 1% VoLTE and HTTP browsing with IP cloud (d).
packet discard ratio and 0.1s delay between caller (UE_1) and
callee (UE_2) (refer to Fig. 2.c). The simulation period is equal to 5 (3mins plus offset) for
each scenario.
Scenarios from n.37 to n.48 are similar to those in the group
13-24: UE_1 and UE_2 are performing a VoLTE call using a TABLE III. SCENARIO CONFIGURATION
direct link interrupted by IP cloud, UE_3 is performing HTTP
web session (refer to Fig. 2.d). Table III resumes the main LTE LTE S1 eNB
Scen. IP User Hop
Serv. Serv. Capacity BW
characteristics for each simulated scenario. No.
Type No.
cloud
[%]
No. No.
[MHz]
C. Simulation settings 1-12 VoLTE 1 NO
100; 75;
2 2
5; 10;
50;30 20
Scenarios have been simulated using the LTE network
model provided by OPNET 17.5 PL6. The UE’s antenna gain 100; 75; 5; 10;
13-24 VoLTE 1 YES 2 3
is -1 dBi with a receiver sensitivity of -200 dBm. eNodeB uses 50;30 20
10 MHz LTE bandwidths and FDD Duplex Mode. Link among
LTE network nodes is of type PPP D3, with a data rate of VoLTE 100; 75;
44.736 Mbps. The simulation area is a typical campus area 25-36 2 NO 3 2 5; 10; 20
+ HTTP 50;30
(100 Km-square wide). We employed the Voice Codec GSM
EFR and one voice frame per packet. As per the LTE standard,
the VoLTE application is carried out over the EPS bearer with
scenario number) versus each of the KPIs (which show abrupt
VoLTE 100; 75; 5; 10; changes among contiguous scenarios).
37-48 2 YES 3 3
+ HTTP 50;30 20

D. Experimental Setup
The experimental setting is based on the OPNET simulation
tool configured in the System-in-the-loop (SITL) modality.

Fig. 4. MOS-LQS (leftmost) and KPIs (rightmost) represented along the


series of scenarios. On the x-axis there is the scenario order number.

V. ANALYSIS OF RESULTS
A. Multi Linear Regression
In [2] we studied the relationship between QoS KPIs and a
MOS, intended as a QoE metrics, fitting the linear model with
bias shown in (1) :
Fig. 3. The experimental testbed.

OPNET SITL module is installed on a fixed PC, shown in = + ∗ DEL + ∗ + ∗ + (1)


the middle of Fig. 3. This provides the emulation of the LTE
network with two different UEs, e.g. the Sender UE (caller, on where β0, β1, β2 and β3 are the unknown regression
the left) and the Receiver UE (callee, on the right), to perform coefficients and εi ~ (N,σi2) is the error for each observation.
an unidirectional VoLTE call. The two UEs are emulated by In [2] we used the same data as in Table IV and got results
two laptops equipped with a single Network Interface Cards through MLR as shown in Table V.
(NICs). In this setting the Sender transmits a high quality voice
TABLE V. MULTI LINEAR REGRESSION
track, three minutes long, to the Receiver through the OPNET
SITL. The resulting voice track (RVT) is recorded at the Multi Linear Regression’s Statistical Indicators
Receiver and stored. For each scenario we prepared three
Factors β Std. Error tStat pValue
RVTs using a GSM-EFR audio codec. Finally, the authors estimated
performed an ACR listening test, employing a panel of 25 Intercept 5.7371 0.48221 11.898 2.4177e-15
participants, chosen in the age range from 20 to 50. Each panel
participant was provided with a RVT per scenario randomly DEL -15.802 4.8986 -3.2258 0.0023718
chosen out of three. At the end of each run the participant gave JIT -1.7824 3.1521 -0.56547 0.57463
a numeric score (from 1 to 5) to evaluate the quality (in terms PLR 1.5226 2.4779 0.61447 0.54207
of subjective MOS) of the listened RVT.
E. Results Statistical indicators for MLR are: Root Mean Squared
Table IV shows results we have got, in terms of the Error (RMSE) equal to 0.489; Mean Squared Error (MSE)
network QoS metrics (end-to-end Delay, Jitter and Packet Loss equal to 0.699285. Then R-squared is 0.798 and Adjusted R-
Rate) and the QoE metrics MOS-LQS. Squared 0.784. Looking at the p-values for JIT e PLR we see
the null hypothesis cannot be rejected for these variables.
TABLE IV. SIMULATION RESULTS Patterns for residuals, as shown in Fig. 5, reveal a poor fitting
quality, both for size of residuals and the asymmetric
QoS and QoE results [min; max] distribution around their median.
Packet Loss
Scen. Delay Jitter MOS-LQS B. Neural Network Fitting
Rate
No. [s] [s] [1-5]
[%]
We selected a 3-10-1 feed-forward network: 3 linear input
[0.11; [0.06; [0.20; [3.10;
1-12
0.12] 0.11] 0.50] 4.55] neurons, 10 sigmoid neurons in the hidden layer and a linear
[0.13; [0.09; [1.05; [3.39; output neuron. For the training process, we adopted 34 samples
13-24
0.13] 0.13] 1.50] 3.96] (70% of the data set) as the training set, 7 samples (15% of the
25-36
[0.19; [0.15; [7.10; [1.95; whole data set) as the validation set, 7 samples (15% of the
0.21] 0.19] 9.23] 2.71] whole data set) as the test set. The choice of only ten hidden
[0.26; [0.21; [19.92; [1.03;
37-48
0.27] 0.27] 26.06] 1.80]
neurons is a guarantee against overfitting. Then we employed
some training strategies as in the following.
Fig. 4 makes evident at first inspection a fairly non-linear
behavior of MOS-LQS (which varies smoothly with the
validation performance at epoch 61. Coefficient of correlation
R ends at 0.95 putting this network at the same performance
level (on the test set) as the NNLM, as shown in Table VII but
farer from unwanted effects, such as overfitting.

Fig. 6. Error histogram for the ANN trained with the Levenberg-Marquardt
Algorithm.
Fig. 5. Statistical indicators of the MLR approach.

1) Levenberg-Marquardt Training Algorithm

TABLE VI. NNLM REGRESSION RESULTS

Statistical Indicators
Dataset
Samples MSE R-factor

Training 34 0.0689156 0.94922


Validation 7 0.279969 0.96953
Testing 7 0.0945887 0.97147

Let us start with the Levenberg-Marquardt algorithm for


training of the NNLM. We repeated the training ten times and
picked up the best network, which reached the best validation
performance after eight training epochs. Table VII shows the
mean sum-squared error that is fairly good and largely lower
than the MSE for MLR. Fig. 6 the error histogram for NNLM:
we learn that errors are more in the negative range, as for MLR Fig. 7. Performance of the ANN trained with the the Levenberg-Marquardt
residuals, though their size is lower. This issue improves Algorithm.
substantially for testing and validation samples (a view that we
cannot have in the usual MLR approach that uses all the data 3) Scaled Conjugate Gradient Training Algorithm
for fitting). Fig. 7 shows the regression plots of the network We also tried the Scaled Conjugate Gradient training and
outputs with respect to targets for training, validation, and test achieved the best validation performance for the NNSCG at
sets, and finally the whole data set. Coefficient of correlation R epoch 34, and R still equal to at 0.95.
indicates the proportionate amount of variation in the target
value explained by the network response, considered as the C. Discussion
independent variables in the linear regression model between The statistical indicators of performance for the four
the two ones. Nearer the R is to 1, better the network fits the adopted methods are reported in Table VII. The mean squared
MOS-LQS. Indeed, the NNLM has a fairly good RALL=0.95. error gets the minimum with the neural network trained by the
Levenberg-Marquardt algorithm. The MLR exhibits greater
2) Bayesian Regularization Training Algorithm
error, even if it has benefitted of 30% more data points for
We tried the Bayesian Regularization for training because
fitting. A specific remark is due on the three learning
our dataset is small. This Bayesian regularization takes place
approaches for the neural network. The most interesting
within the Levenberg-Marquardt algorithm modifying the
capability of a model for MOS is that the model output, though
optimization function. It utilizes all the data for training and
affected by error, still remains within the thresholds for the
requires longer training periods. In our case it gets the best
MOS target (that are discrete). This is generally the case for
NNLM as can be seen in the right lowermost plot of Fig. 7, with [5] 3GPP TS 23.203, “PCC – Policy and Charging Control Architecture”.
an exception for the errors occurring around the target value [6] A. Vizzarri, S. Forconi “Review of Studies on End-to-End QoS in LTE
3.3 that is confused as 4 by NNLM. This is an issue solved by Networks”, in Proc. AEIT Congress, Mondello, Italy, 2013, pp. 1-6.
the NNBR (comparing both the plots: here one is not reported [7] A. Vizzarri, “Analysis of VoLTE end-to-end quality of service using
OPNET” in Proc. European Symposium on Computer Modeling and
for brevity). The fact that NNBR has a slightly higher MSE is Simulation, Pisa, Italy, 2014, pag. 452 – 457.
meaningless. Further it requires more intense training, and this
[8] A. Vizzarri, “Analysis of VoIP Over LTE End-To-End Performances in
is tolerable because training is a negligible cost if compared Congested Scenarios” in Proc. Artificial, Intelligence, Modelling and
with the cost of the listening test. Simulation, Madrid, Spain, 2014, pp. 393-343.
[9] 3GPP TS 123.207 “Digital cellular telecommunications system (Phase
TABLE VII. MODEL COMPARISON 2+); Universal Mobile Telecommunications System (UMTS); LTE; End-
to-end Quality of Service (QoS) concept and architecture”.
Main Statistical Indicators [10] ITU Recommendation P.800.2, “Series P: Terminals and Subjective and
Model objective assessment methods”.
MSE RMSE RAll R2All [11] ITU Recommendation G.109 “Definition of categories of speech
transmission quality”
MLR 0.23912 0.48900 0.6371 0.7982 [12] A. Lakaniemi, J. Rosti, V. I. Raisanen, “Subjective VoIP speech quality
evaluation based on network measurements”, IEEE ICC 2001;
NNLM 0.10442 0.32314 0.9534 0.9764
[13] Telchemy. Voice Quality Measurement.
NNBR 0.17862 0.42263 0.9524 0.9759 http://www.telchemy.com/appnotes/TelchemyVoiceQualityMeasureme
NNSCG 0.18338 0.42823 0.9527 0.9760 nt.pdf, Mar. 2008 (accessed: May 12th, 2016).
[14] F. D. Rango, M. Tropea, P. Fazio, and S. Marano, “Overview on VoIP:
VI. CONCLUSIONS Subjective and Objective Measurement Methods”, IJCSNS, vol. 6, pp.
140-153, Jan. 2006.
VoLTE is a key service for LTE networks, and still needs [15] ITU-T Recommendation P.800 (08/96), "Methods for subjective
improvement and operalization of quality measures. Research determination of transmission quality".
is strongly engaged in modelling LTE networks in realistic [16] L. Monacelli, R. Francescangeli, “Fault Management for VoIP
usage conditions and modelling the quality perceived by the Applications over wireless and wired NGN Networks: an operational
users as a consequence of measurable technical parameters. prospective” in Proc. 7th IEEE International Workshop on Performance
and Management of Wireless and Mobile Networks, Bonn, Germany,
Here we modelled the network through scenario making, 2011, pp. 711-718.
produced a limited number of cases, simulated the network [17] J. Yu and I. Al-Ajarmeh, “Call admission control and traffic engineering
performances. Then we attacked the subjective measurements of VoIP. The Second International Conference on Digital
for MOS. This required to invest a certain amount of resources Telecommunications”, San Jose, California, USA, July 2007, pp. 11–16.
to train and employ a panel of listeners. Thanks to a good [18] M. Alreshoodi, J. Woods, “An empirical study based on a Fuzzy Logic
organization testing was performed at sustainable costs. System to Assess the QoS/QoE Corelation for Layered Video Streaming”
in Proc. 2013 IEEE International Conference on Computational
We studied how MOS-LQS results can be modelled using a Intelligence and Virtual Environments for Measurement Systems and
set of Artificial Neural Network models. The ANN model Applications, Milan, Italy, 2013.
outperforms the MLR, and this result is crystal clear, due to the [19] M. Alreshoodi, J. Woods, “Survey on QoE/QoS correlation models for
inner nonlinearities in the data (as per Fig. 4) and the filtering multimedia services”, International Journal of Distributed and Parallel
Systems (IJDPS) Vol.4, No.3, May 2013.
action made by the human judgment. This confirms again the
need for an approach that is non linear and capable of [20] S. Khorsandroo, R. Md Noor “Stimulus-Centric versus Perception-
Centric Relations in Quality of Experience Assessment”, in Proc.
generalization. Wireless Telecommunications Symposium, London, UK, 2012, pp. 1-6.
However panel testing cannot be extensively repeated, as [21] N. Khitmoh, P. Wuttidittachotti, T. Daengsi “A Subjective - VoIP
Quality Estimation Model for G.729 Based on Native Thai Users” in
would be necessary for a MNO that might face a incredible Proc. 16th International Conference on Advanced Communication
number of operational situation. Our future research is aimed at Technology, Pyeongchang, South Corea, 2014, pp. 48-53.
generalising the ANN model to a far larger scale (three orders [22] R. E. Uhrig, “Introduction to artificial neural networks” in Proc. 1995
of magnitude at least) thanks to a fully distributed smart IEEE IECON 21st Int. Conference on Industrial Electronics, Control,
approach. We have at stake the search for a viable way to make and Instrumentation, Orlando, Florida, 1995, pp. 33-37.
quality nearer to user perception while controlling the [23] G. Dreyfus, Neural Networks: methodology and applications, Ed. Berlin,
combinatorial explosion of scenarios and technical parameters. Springer, 2005.
[24] S. Rivera et alii, “QoS-QoE Correlation Neural Network Modeling for
REFERENCES Mobile Internet Services” in Proc. 2013 Int. Conference on Computing,
Management and Telecommunications, Ho Chi Minh City, Vietnam,
[1] 3GPP, Release 8 V0.0.3, “Overview of 3GPP Release 8: Summary of all 2013, pp. 75-80.
Release 8 Features” November 2008.
[25] V. A. Machado et alii, “A New Proposal to Provide Estimation of QoS
[2] A. Vizzarri, F. Davide, “Quality of Experience Model for VoLTE and QoE over WiMAX Networks” in Proc. 2011 IEEE Third Latin-
Services through Statistical Analysis”, submitted to GLOBECOM 2016. American Conference on Communications, Belem do Para, Brasile, 2011,
[3] 3GPP TS 23.401, “3rd Generation Partnership Project; Technical pp. 1-6.
Specification Group Services and System Aspects; General Packet Radio [26] C. Benhamed “A large scale IP Network Traffic matrix Estimation based
Service (GPRS) enhancements for Evolved Universal Terrestrial Radio on ANN: A comparison study on training algorithms” in Proc. 2015 4th
Access Network (E-UTRAN) access (Release 8)”. Int. Conf. on Electrical Engineering, Boumerdes, Algery, 2015, pp. 1-5.
[4] 3GPP TS 23.228, “3rd Generation Partnership Project; Technical
Specification Group Services and System Aspects; IP Multimedia (IM)
Subsystem - Stage 2 (3G TS 23.228 version 2.0.0)”.

View publication stats

Вам также может понравиться