Page 1

On Interaction between Loss Characterization and

Forward Error Correction in Wireless Multimedia

Communication

Abdelhamid Nafaa, Yassine Hadjadj-Aoul and Ahmed Mehaoua

University of Versailles – CNRS-PRiSM Lab.,

45 avenue des Etats-Unis, 78035 Versailles — France

{anaf, yana, mea}@prism.uvsq.fr

Abstract—With the steadily growing synergy between existing

heterogeneous networks, the wireless LAN appears as the de-facto

wireless access network in the end-to-end multimedia services

distribution chain. Unlike in the traditional wired multi-hop

networks (Internet) where congestions increase persistently both

delays and losses, wireless packet losses are often location- and time-

varying. Particularly, WLAN communication is characterized by

high bit error rates that translates into tight loss dependency. The

loss process may rapidly shift between different loss correlations

levels, resulting in poor forward error correction (FEC) recovery

capabilities. In this paper, we address this issue by providing a

combined loss model to accurately characterize the wireless loss

distribution features. We use control theory guided parameter tuning

in order to urge the convergence of the loss models towards seizing

the instantaneous loss distribution trends. Finally, we derive a new

loss-specific QoS metrics for new FEC block allocation scheme.

I. INTRODUCTION

Among Quality of Service (QoS) metrics, packet loss is an

important metric in shared environments such as the WLAN’s [3].

While existing works focus on capturing the mean loss (long-term

QoS), less emphasis is put on modeling loss distribution (short-term

QoS, cf. [4] [5] and the references therein). In certain real-time

applications, the loss pattern is a key parameter that determines the

performance observed by the users. For the same loss rate, two

different loss distributions could potentially produce widely different

perceptions of performance [6]. Also, many forward error recovery

approaches become less efficient as the loss burstiness increases. It is,

therefore, essential for streaming application to capture and quantify

the loss process with suitable QoS metrics in order to improve the

efficiency of QoS adaptive mechanisms.

The intrinsic wireless link characteristics involve unpredictable

burst errors that are usually uncorrelated with the instantaneous

available bandwidth; this often translates into sporadic and clustered

packet losses. One particular issue to tackle when streaming media is

the FEC efficiency/accuracy. In this work, we investigate video

multicast communications over IEEE 802.11b wireless LAN. Our

target environment is a video server that multicasts video streams for a

large group of clients. In our large-scale network architecture, we

transmit a single multicast stream together with different FEC streams

tailored for different operated WLANs that are connected to the IP

backbone through a broadband metropolitan DVB-T network; the

DVB-T network ensures, among other things, filtering of FEC streams

by transmitting each FEC stream to the appropriate WLAN. Within a

given WLAN, the FEC stream is used at different receivers to recover,

to some extents, possible network losses. Each WLAN may subscribe

to a broad range of QoS levels (Service Level Agreement — SLA),

which correspond to different adaptive FEC responsiveness levels. We

believe that this scenario reasonably represents many of current [2]

and forthcoming streaming services. The emphasis is put on the

WLAN since it is actually considered as the “de-facto” wireless access

network in many Service Providers offers [1].

In this work, we focus on better understanding the WLAN’s

“burstiness” effect in order to characterize the channel behavior with

more accurate QoS metrics. As pointed out in this work, the loss

process in wireless channel shows certain “stationary”1 behavior over

the time [8]. This comes out through a significant stability in the

successive measured loss run lengths distribution. Capturing the loss

distribution trends may be useful, in many manners, for multimedia

applications adaptation, e.g., allocating the optimal FEC block, or

even applying a reasonable interleaved FEC protection [6]. To tackle

the wireless loss uncertainties, we combine a loss run length model

and an inter-loss distance model to accurately capture the channel

burstiness and the clustering between loss runs. This combined loss

model relies on an accurate network feedback (extended RTCP) that

indicates for each transmitted packet whether it was lost or received.

We use a simple fuzzy controller to instantaneously refresh our

combined loss model, and then anticipate wireless channel

fluctuations taking into account the knowledge acquired from past

experience. This achieved by canceling/minimizing each time the

possible model approximations, thereby urging the convergence of our

models towards capturing the current loss correlation features. Finally,

we use the latter prediction, together with new loss-specific metrics, to

effectively protect multimedia streams using a new bandwidth-

efficient FEC allocation scheme.

The remainder of this paper is as follows. Section II investigates

reliable video multicasting over Wireless LAN. Section III describes

our proposal: a combined wireless loss characterization and channel

coding. Section IV is devoted to our models validation as well as to

the definition of an adaptive fuzzy-based channel behavior prediction.

The performance evaluation of our streaming system is presented in

Section V. Finally, we have drawn several key conclusions from this

work; these are stated in Section VI.

II. BACKGROUND AND MOTIVATIONS

Typical communications over WLAN involve a high bit error,

which usually occurs through correlated adjacent packets losses. An

understanding of the effect of packet loss on the reconstructed video

quality is clearly very important for designing and operating video

1 Although the term “stationary” typically refers to stochastic processes that

do not change behavior over time, we use it in this paper to exclusively point

out a cyclic dependency between loss distribution measures.

Page 2

communication systems over error-prone networks. In case of

multicast or broadcast communication over WLAN, the data packets

are not acknowledged, and hence no retransmission is performed at

the MAC/Logical link layer; this mode of communication reduces

transmission delays and data control overheads while making

communications less reliable. Another important implication, with

multicast communication in WLAN’s, is that the AP’s is unable to

change its nominal bit rate (11, 5.5, 2, or 1 Mbps) as the

communication experience an increased BER [7]. Usually, in directed

communication, both the wireless station and the Access Point may

explicitly degrade their nominal bit rate by altering both coding and

physical modulation in order to overcome a channel’s noisy period

(i.e., after repeated unsuccessful frame transmissions). Even though

the AP may be enhanced to sense the receivers’ traffic in order to

degrade its nominal bit rate accordingly, the period during which the

receivers experience severe degradations may last longer time periods.

In this context, the transient time-varying harsh conditions may

steadily persist over the time.

Generally, reliable transmission on wireless channels requires the

use of some type of error control. Efficient error control on time-

varying channels can be performed by implementing an adaptive

control system where the optimum code is selected according to the

actual channel conditions. In [8], the author proposes track the short

time intervals where the channel parameters are stable enough so as to

improve the recovery efficiency. Authors in [9], propose an adaptive

FEC based on redundant audio samples transmission (audio packets),

so the information relative to packet n may be spread over multiple

packets relying on a simple two-states Gilbert model to react to

network fluctuations. This scheme is very efficient for Internet

telephony application since a Gilbert model suffice to capture the quite

“soft” loss process dynamics involved by the persistent congestion of

traditional multi-hop IP networks. Clearly, it is important to capture

the wireless channel dynamics through pertinent QoS metrics to figure

out the appropriate FEC scheme that maximizes recovery

effectiveness.

III. A COMBINED LOSS CHARACTERIZATION AND CHANNEL

CODING

1. Loss Run Length Model

In this section we present an extension of the well-known Gilbert

model. The extended Gilbert model was previously used in modeling

Internet losses. It allows treating the issue of long loss runs by using

models with multiple states. We define the random variable X as

follows: X = 0: “non packet loss”, X = k: “exactly k consecutive

packets lost”, and X ≥ k: “at least k consecutive packets lost”. With

this definition, we establish a loss run-length (loss burst length) model

with n states (see Figure 1). We rely on an extended RTCP feedback

that continuously reports on the measured loss pattern. Each RTCP

report corresponds to the last transmitted 300 packets (i.e., loss pattern

segment).

S1

(1 loss)

S0

(non loss)

P01

S2

(2 losses)

Sn-1

(n-1

losses)

P12

P(n-2)(n-1)

P(n-1)(n-1)

P10

P20

P(n-1)0= 1 - P(n-1)(n-1) = 1

P00 = 1 - P01

. . . .

Fig. 2. Extended Gilbert model with limited states.

The system keeps a counter l, which is the number of

consecutively lost packets; it is reset whenever the next packet is

delivered. The parameter to determine is P[Xi | Xi-1 to Xi-l all lost].

Let mi, i = 1, 2... n-1 denote the number of loss bursts having

length i, where n – 1 is the longest loss bursts. m0 denotes the number

of delivered packets. n represents also the number of model’s states.

Note that a model can be entirely described by its burst loss length

occurrences vector M (i.e. the mi coefficient vector, M = (m0, m1, mn-

1)-1. The formula to calculate the parameters of the extended Gilbert

model is given by:

∑

=

∑

=

n

−

−

−

=

−≥

≥

=−≥≥=−

1

1

1

)(

)1(

)(

)1() )(1(

ki

n

ki

mi

mi

kXP

kXP

kXkXPkkP

At this point, the mean burst loss length (MBL) is easily deduced:

∑

=

=

n

MBL

∑

=

i

−

−

⋅

1

1

1

1

)(

n

i

i

mi

mi

(1)

MBL gives the expected mean loss run length based on the

previously observed loss distribution.

2. Inter-loss Distance Model

The inter-loss distance metric was recently proposed to describe

the distance between packet losses in terms of sequence number. The

ILD (Inter-Loss Distance) metric is useful in two respects. First, an

accurate loss model is able to model loss run distributions, but it does

not model distances between loss runs. Second, small ILDs may also

degrade the performance of FEC codes.

As with the loss model, we derive a model to characterize the

ILD’s distribution features. This is useful to understand and foretell

the spacing between loss events. Let di, i=1,.. n-1 denote the number

of ILDs having length i. The ILD model is completely described by its

ILD’s occurrence vector given by: D=(d1, d2, …., dn-1)-1. The mean

inter-loss distance (MILD) is given by:

∑

=

i

∑

=

i

−

−

⋅

=

1

1

1

1

)(

n

n

i

di

di

MILD

(2)

The inter-loss distance metric allows one to study the separation

between packet losses. This is particularly useful to complement the

loss model for an enhanced loss pattern prediction and multimedia

application adaptation.

3. Efficient FEC Protocol for Video Distribution

Packet level FEC consists of producing h redundant packets from

k original ones, providing a resiliency against a maximum of h losses

out of n packets constituting the FEC block (n=h+k ). In order to

reflect at application level the wireless channel dynamics, we use both

MBL and MILD as parameters in FEC blocks. Using an appropriate

amount of redundancy h = MBL makes communications robust

against the most likely expected clustered losses. On the other hand,

we take as the number of original data k = MILD in order to maximize

the number of protected media packets (see Fig. 3). At this point, the

FEC block is constituted of n=MBL+MILD packets. The above

described FEC block allocation scheme is efficient as long as the

wireless channel still exhibits stable behavior over the time.

FEC

RTP media

Packets

FEC

Packets

Burst

error (MBL)

FEC

MILD

Fig. 3. Model-based FEC block allocation scheme.

Since we are restricted by delay constraints, choosing a large k >

MILD may lead to intra-media synchronization and timeouts

problems. In realistic streaming systems it is, indeed, much suitable to

apply FEC to no more than one frame (see [6]). Consequently, we

generalize the previous analysis in order to apply our redundancy

allocation scheme to any block size (k). In our case, k stands for the

Page 3

number of packets constituting the current video Frame to be

transmitted. The amount of FEC redundancy h for k original video

packets is then obtained as follows:

⋅

=

MILD

MBLk

h

(3)

IV. ADAPTIVE MODEL-BASED LOSS CHARACTERIZATION

In this section, we aim at better predicting the wireless loss

pattern, focusing on burst loss distribution. One important issue that

should be overcame, when streaming media with FEC protection, is to

capture the dependency between losses. Especially, both burst loss

length and inter-losses distance (ILD) have devastating consequences

on FEC efficiency, and thus the video quality at receiver. In this

context, it is much suitable to use FEC in order to evaluate a particular

loss model and its capability to predict the short-term loss features

(burst loss length and ILD); this consists in comparing the

experimented loss pattern and the model predicted loss pattern after

recovering with FEC. We use trace-based simulation. This evaluation

process is usually used to estimate the overall loss model performance

in respect to multimedia service quality (see [11], [5]).

The trace-based simulation consists in collecting loss patterns

resulting from real traffic transmission over the network. In our case,

we use a WLAN network to collect the loss pattern resulting from real

H.264 video multicasting. Each transmitted RTP packet contains a

sequence number that is intended for intra-media synchronization. If

the packet arrives, the receiver will write the sequence number into the

trace file. Afterwards, during an off-line analysis, we calculate the

final loss pattern file and then divide it into several segments

(corresponding to the extended RTCP reports). The simulation of

streaming server behavior is obtained after a multi-pass processing.

The number of processing passes is tightly dependent on the number

of loss pattern segments. We use a more precise model’s evaluation

process by measuring the accuracy gap between the experienced and

the model-predicted results. In other words, each time we measure a

new loss pattern segment (receiving RTCP feedback), we calculate the

distance between the variations of both measured and model-predicted

loss patterns; then, by making a summation of these distances, we

obtain a surface that roughly represents the accuracy gap.

1. Sliding Number of Model’s States (SNMS)

To evaluate the performance of a network with respect to real-

time video application, a model with a limited number of states is

sufficient (for example, authors in [11] show that their Internet packet

traces typically match with a 6 states model), which optimize memory

and computational capabilities of the system that performs modeling.

Meanwhile, it was previously shown that models with more states

capture better the long burst loss effect (see [5]). In order to face this

issue, we propose an adaptive number (n) of model’s states to

accurately capture the fluctuating channel conditions while

minimizing the computation and memory cost.

We conduct experimentations where the number of models’ states

is fixed to a maximum of 50 states for both models. The model with

static number of states uses 50 states to capture the channel

fluctuation, while SNMS varies the number (n) of model’s states

according to the RTCP feedback (i.e. n values is fixed according to the

maximum loss run length experienced by the receiver during the last

transmitted N packets — see formula (4)).

( )

] 50 , 2 [

∈=

andii Maxn

Here, mi denotes the number of loss bursts having length i.

{}

0

>

i

m

(4)

Tab. 1: Percentage of lost packets unrecoverable by FEC.

Experimented

5.07%

Adaptive States Model

4.04%

Static States Model

4.12%

We observe that SNMS behaves globally like the model with

static number of states (the overall performance is depicted in Tab. 1).

SNMS still provides loss pattern predictions very close to the

experienced loss pattern for a reduced memory and computation cost.

As an example, in our experimentation, the model with static number

of states used an average of 10 kb memory while SNMS used only 2.5

Kb. Note that the higher the memory used in modeling the higher the

computation cost is. In the remaining of the paper we systematically

use the SNMS in both loss and ILD models.

2. Adaptive Loss Pattern Prediction

This section presents a design of fuzzy system that predicts the

next loss pattern segment. Consider a variable vector M

(corresponding to burst loss length occurrences vector) that assumes

the sequence of the following values:

{

=

M M

(i-1) is the number of received feedback reports.

We use the exponential weight mean average (EWMA) of the

precedent sequence {

()

)1(

1

+−=

ii

MαM

The predictor, illustrated in equation (5), is controlled by a vector

parameter α, where α is the weight given to past history (α is also

considered as the smoothing factor). The larger it is, the more weight

past history has in relation to the last observation. It is much suitable

to keep certain influence of past history in order to smoothly react to

ephemeral channel fluctuations. Although statistical characteristics of

real channels can significantly vary with time, propagation

experiments for various types of channels [8], indicates that the basic

system parameters remain constant over short time intervals.

A major problem with the exponential averaging predictor is in

the choice of α. It would be useful to automatically determine a ‘good’

value of the smoothing factor α, and to be able to change this value

online if the loss distribution fluctuates. Our approach uses fuzzy logic

control to achieve this tuning. The fuzzy controller is effective and

costless in term of computation latencies, which is much suitable for

stringent network services such as multimedia streaming applications.

Our proposed fuzzy system enables run-time adaptation based on

iterative feedback control and knowledge acquired from past

experience. In our case, the fuzzy controller behaves like a

minimization function. Thus, we vary the vector α in order to

minimize/cancel the error vector between the last measured loss

pattern segment

i

{ }

/ M - M min

j

i

i

α

Hence, we simply replace

i

M by

appropriate {α}j vector elements. In other words, we correct the error

produced in the last step through finding out the appropriate α vector

that should have been employed – see equations (6) and (7).

}

1211

−−

ii

M ... M ,

}

1i

M − to predict the next value of M (Mi).

)1(

−−

i

Μα.

(5)

Mand the last predicted loss pattern segment

[

0,

∈

M in (5), and then figure out the

i

M .

] 1

(6)

i

{ }

α

[] 1,0

11

1

- MM ) M Mα.(

j

/ i-i i-

i

∈=−

−

(7)

The resolution of the previous equations will permit to find out an

adequate vector α; each element {α}j in the vector α varies

independently according to the observed occurrence of burst length j

during past measurements. By casting the problem into a fuzzy-based

control theory framework, we obtain a simple SISO (Single Input

Single Output) system. The input represents the elements of the error

vector (

i

i

M - M

) and the output represents the {α}i vector elements.

This means that when the error (i.e. difference between the

predicted and the measured loss segments) is high we should increase

α in order to give further weight to recently measured loss segments.

Otherwise, if the predicted loss segment is close to the measured loss

segment, we decrease α to keep a high influence of the past history.

This way, α elements fall around zero when the channel is in

Page 4

stationary period, which comes out through stable and quite correlated

successive loss segments. The shape and the range of the fuzzy

membership functions should be tuned according to the max burst

error. We recommend using a Gaussian membership function that

provides a broad choice in setting the coarseness/smoothness of α

adaptation. One may tune the function until reaching an acceptable

behavior in respect to the wireless environment.

Tab. 2: Accuracy gap between measured and estimated loss patterns (x103).

Simple model

0.1607

Fuzzy-based adaptive model

0.1414

We evaluate the performance of the combined-model loss

prediction through extensive trace-driven simulation. In order to better

quantify the achieved performance, we give in Tab. 2 the accuracy gap

between the fuzzy-based adaptive model and the simple model (i.e.,

that does not consider past history). This is a good indictor of the

overall “short-term” model predictions accuracy. The presented values

are obtained after averaging the results of 10 trace-based simulations;

each time, we use a different loss pattern that is collected from a real

WLAN communication. For each simulation, the accuracy gap

between the experienced and predicted loss pattern is evaluated using

the runge-kutta numeric integration method; it basically makes a

summation of the distances between two functions. The accuracy gain

of the fuzzy-based adaptive loss distribution prediction is around 12%.

V. PERFORMANCE EVALUATION

In this section, we evaluate the implementation of our streaming

system through trace-based simulations. We emphasize the ability of

our adaptive combined model to foreseen wireless LAN fluctuations.

The simulation is performed for a streaming scenario where a

multicast server distributes video together with FEC. To this end, we

use video loss traces collected in real WLAN communication.

1. Experimental Results Analysis

Fig. 4 depicts the instantaneous application-level loss rate (in

percentage) measured when using our proposal versus the

conventional method. We include the loss rate curve when no FEC is

used in order to better discern the possible loss recovery gain. The No-

FEC curve represents, indeed, the instantaneous loss pattern segments

perceived by both adaptive streaming systems. This loss rate (green

curve with square marker) will actually steer the streaming adaptation

as it basically represents the successive reported loss pattern segments.

The measurements are depicted for 12 min of CBR Soccer sequence

streaming, while the reported loss rate values are each time averaged

over 300 packets. Note that the loss rate curves represent final packet

loss rates (FLP) measured after recovering with FEC.

050100150200250300350400450

0

10

20

30

40

50

60

70

80

90

Loss pattern segment index (corresponding to receiver report for N = 300 packets)

Fig. 4. Instantaneous receiver-perceived loss rate.

application-level perceived loss rate (%)

Witout FEC transmission

Conventional adaptive FEC

Our proposal

We observe a high loss resiliency with our proposal when the

channel burstiness passes through stationary period (for instance, see

Fig. 4 during the periods delimited by the receiver reports: “8” to

“16”, and “396” to “404”). This is particularly noticeable when both

the measured ILD and loss-run-length distribution are stable over the

time, which increase the FEC efficiency (see Fig. 5 and Fig. 6 between

the receiver reports “8” to “16”). The mean measured loss rate during

the multicast streaming session is around 6.1% (i.e., 7607 lost

packets). When using the conventional adaptive streaming server, the

mean perceived loss rate reached over 5% (6210 lost packets after

FEC recovery). At the same time with our proposal, we notice a

perceived loss rate of about 4% (4968 lost packets after FEC

recovery). This difference in mean loss rate (1%) can have devastating

consequence on video streams. As the wireless channel may exhibits

too long periods of perturbation (unavailability), our model don’t

account for the excessive losses that are brief spikes. Thus MBL

values are limited by 15, so larger values are filtered by the fuzzy-

controller, and thus not considered in predicting the next loss pattern

segment. MILD values are limited by the maximum number states of

the ILD model, i.e., 50.

050100150200 250300350400450

0

5

10

15

20

25

Loss pattern segment index (corresponding to receiver report for N = 300 packets)

Fig. 5. Instantaneous MBL.

mean burst length (MBL)

0 50 100150200250300350400450

0

5

10

15

20

25

30

35

40

45

50

Loss pattern segment index (corresponding to receiver report for N = 300 packets)

Fig. 6. Instantaneous MILD.

mean inter loss distance (MILD)

It is clear that our streaming system provides better

communication robustness through increasing the loss resiliency,

which certainly mitigates multimedia quality degradation in harsh

wireless conditions. However, in most of existing wireless multimedia

services deployment the bandwidth consumption can be a matter of

concern as well. Consequently, the tradeoff between bandwidth

consumption and perceived quality must be carefully investigated. Fig.

7 illustrates the instantaneous measured throughput during the

multicast streaming session with our proposal (and respectively the

conventional streaming system). The mean bandwidth consumption

achieved with the conventional streaming system is around 495 Kbps,

while the bandwidth consumption is about 510.5 Kbps with our

proposal. This excess bandwidth usage (3.1%) can be afforded by

Service Providers (SP) if the video quality continuity is sustained.

Additionally, in almost all service-level-agreement networks, the

WLAN resources are usually dedicated to sensitive multimedia

services.

050100150200250300350400450

450

500

550

600

650

700

750

800

850

900

Loss pattern segment index (corresponding to receiver report for N = 300 packets)

Fig. 7. Instantaneous bandwidth consumption.

It is commonly accepted, from the Service Provider point of view,

that additional bandwidth consumption is tolerated only for enhanced

loss resiliency and video quality at the receiver side. For the purpose

Bandwidth consumption (xKbps)

Conventional adaptive FEC

Our proposal

Page 5

of measuring the tradeoff between the useful received data and the

transmitted data, we define the FEC efficiency factor as:

ed FEC Transmitt a ceived DatRe

Total Data

+

EF FEC

=−

Here, the total data represents the obtained video packets after

recovering with FEC (i.e., received packets and recovered packets). So

ideally FEC-EF = 1 when (i) no FEC is transmitted and the

communication doesn’t suffer from losses or (ii) all transmitted FEC

redundancy is used to recover from packet losses. Fig. 8 illustrates the

measured FEC-EF throughout the multicast streaming session.

0 50 100150200250300350400450

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Loss pattern segment index (corresponding to receiver report for N = 300 packets)

Fig. 8. The instantaneous measured FEC-EF.

The mean measured FEC-EF for the conventional streaming

system is around 0.8874, while we measured 0.91 as a mean value

with our proposal during the multicast streaming session. The FEC

efficiency is particularly noticeable when the loss pattern segment

predictions are accurate, which usually occurs during channel’s

stationary periods (for instance, see periods between 8-16 in Fig. 4,

Fig. 5, Fig. 6, and Figure 9). An important observation is that FEC-EF

some how depends on the channel burstiness forecasting. The more

accurate the loss pattern prediction is, the higher the observed FEC

efficiency, and the more the service quality is. As an example, during

the period 396-404, both measured MBL and MILD show confined

variances (see Fig. 5 and Fig. 6), which translate to an increased loss

resiliency for a reduced bandwidth consumption. Besides being

strongly dependent on the loss distribution, the FEC efficiency

improvement is significantly sensitive to video traffic characteristics.

In fact, when using high bitrate video streams, our FEC block

allocation scheme provides better robustness. This is especially due to

the increased mean video frame size that improves the precision with

which the FEC redundancy is transmitted (see formula (3)).

FEC-Efficiency Factor (<=1)

Conventional adaptive FEC

Our proposal

VI. CONCLUSION

As pointed out in this paper, multimedia streaming applications

depend rather on the loss pattern than the averaged loss rate.

Particularly, the loss distribution process has a determinant impact on

multimedia codec resiliency as well as Forward Error Correction

efficiency. Besides, typical communications over WLAN pass

frequently through stationary periods that manifest their self as highly

correlated loss pattern segments. We address this issue by proposing a

combined loss-run-length model and inter-loss model for efficient loss

dynamics characterization. We focus on tracking stationary periods by

each time reducing the possible modeling approximations. A fuzzy

controller is used as a generic minimization function, providing fast

convergence towards capturing the transient stationary periods. Our

proposed loss-effect characterization is validated for multimedia

streaming scenario. Thus, we derive a simple yet accurate QoS metrics

for a new FEC-block allocation schema that provides efficiency and

responsiveness.

Based on the performed set of achieved experiments, it appears

that our proposal provide better bandwidth efficiency than the

conventional method in highly changing network conditions. Still, the

performance can drastically change depending on the Fuzzy-specific

parameters (member functions). Using these parameters, the tradeoff

between perceived video quality and bandwidth consumption can be

adjusted for every network configuration. In fact, we observed that

when these parameters are chosen in such a way that there is a high

models’ convergence, the error recovery capability is particularly

efficient during repetitive brief channel spikes. Whereas, when the

channel shows persistent loss pattern correlation, bandwidth efficiency

is somehow limited. In addition, the user’s experience in a wireless

LAN is dependent on the radio propagation environment in which the

wireless LAN operates. The radio propagation environment may

change from time to time, affecting connection speeds and error rates.

In a manufacturing environment, for example, where the multipath

environment changes as equipment is moved about, it is quite possible

for a link to fail completely even if the mobile is stationary. This may

entail a significant accuracy loss in our fuzzy-based loss

characterization, and thus a poor FEC efficiency. Therefore, the fuzzy-

specific parameters must be carefully adjusted, based on the WLAN

deployment scenario.

We are pursuing this work in a way to adjust the basic fuzzy

tuning according to a service level agreement (SLA) that may involve

the service provider and the network access operator [1]. The idea

there would be to provide different levels of FEC responsiveness

according to a-priory service-level commitments that explicitly

specify the desired service quality. This would allow the service

provider to provide a broad range of QoS offers for different

ratifications.

REFERENCES

[1] IST Integrated Project, “ENTHRONE: End-to-end QoS through

Integrated Management of Content, Networks and Terminals”, Sixth

EU’s Framework Program for Research and Development, 2003-2007.

[2] IST Project, “ATHENA Digital Switchover: Developing Infrastructures

for Broadband Access”, Sixth EU’s FP for Research and

Development,2003-2007.

[3] Information technology, Wireless LAN medium access control (MAC)

and physical layer (PHY) specifications, IEEE 802.11, August 1999.

[4] H. Sanneck, et. al., “A framework model for packet loss metrics based

on loss run length”, in SPIE/ACM SIGMM MMCN’00, January 2000.

[5] W. Jiang and H. Schulzrinne, “Modeling of Packet Loss and Delay and

Their Effect on Real-Time Multimedia Service Quality”, In Proc. of

ACM International Workshp NOSSDAV’00, June 26th 2000, Chapel

Hill, USA.

[6] A. Nafaa, et al., “Unequal and interleaved FEC protocol for robust

MPEG-4 multicasting over wireless LANs”, in Proc. of IEEE ICC’04,

Volume: 3, Pages:1431–1435l, Paris, France, 20-24 June 2004.

[7] M. Heusse, et al., “Performance Anomaly of 802.11b”, in Proc. of IEEE

INFOCOM’03, vol. 2, pp. 836-843, San Francisco, CA, Mar. 30th 2003.

[8] B. Vucetic, “An adaptive coding scheme for time-varying channels”, in

IEEE Transaction on Communications, vol. 39, no. 5, May 1991.

[9] J-C. Bolot, et al., “Adaptive FEC-based error control for Internet

Telephony”, Proc. INFOCOM’99, vol.3, pp.1453-1460, 21 Mar. 1999,

New York, USA.

[10] Y. J. Liang, et al., "Analysis of Packet Loss for Compressed Video:

Does Burst-Length Matter?", in Proc. IEEE ICASSP’03, Hong Kong,

Apr. 2003.

[11] M. Yajnik, et al., “Measurement and modeling of the temporal

dependence in packet loss”, in Proc. of IEEE INFOCOM’99, New York,

March 1999.