reliability and efï¬ciency analysis of distributed source
TRANSCRIPT
Reliability and Efficiency Analysis of
Distributed Source Coding in
Wireless Sensor Networks
C. FISCHIONE, S. TENNINA, F. SANTUCCI, L. DI PAOLO, F. GRAZIOSI
Stockholm 2007
Automatic Control Group
School of Electrical Engineering
Kungliga Tekniska Hgskolan
IR-EE-RT 2006:036
1
Reliability and Efficiency Analysis of
Distributed Source Coding in
Wireless Sensor Networks
C. Fischione, S. Tennina, F. Santucci, L. Di Paolo, F. Graziosi
Abstract
The amount of measurements, which can be successfully gathered in tree-based wireless sensor
networks (WSNs) employing distributed source coding (DSC)algorithms, is evaluated in this paper in the
presence of different coding topologies and packet aggregation schemes (PA). The system model includes
a realistic network architecture with multi-hop communication, automatic repeat request protocol (ARQ),
packet losses due to channel impairments and collisions. Four DSC topologies and three alternatives of
PA are considered. The analysis is carried out by first computing the packet loss probability; then the
reliability is evaluated in terms of loss factor and the efficiency in terms of average energy consumption
of the network. The theoretical framework proposed in this paper is exploited to perform a numerical
evaluation of loss factor and energy consumption for some correlation profiles of the measurements.
Keywords: Wireless Sensor Networks (WSNs), Distributed Source Coding (DSC), Packet
Aggregation (PA).
C. Fischione is with the Department of Electrical Engineering and Computer Sciences, University of California at Berkeley. He
was with the School of Electrical Engineering, Royal Institute of Technology, Stockholm, Sweden, when part of this workwas
done. S. Tennina, F. Santucci, and F. Graziosi are all with the University of L’Aquila, Centre of Excellence DEWS and Department
of Electrical and Information Engineering, Poggio di Roio,67040 L’Aquila, Italy. E-mails:[email protected],
stennina, santucci, [email protected].
Work done in the framework of the HYCON Network of Excellence, contract number FP6-IST-511368. C. Fischione wishes
to acknowledge the support of the NSF ITR CHESS and the GSRC.
Part of the topics of this paper were presented at IEEE Globecom 2006, and ICTSM 2006.
2
I. INTRODUCTION
Wireless Sensor Networks (WSNs) consist of sensing devicesoften deployed with limited
energy resources. Distributed Source Coding (DSC) claims to decrease the energy consumption
for lifetime extension of WSNs: to this end it exploits the richness of information usually provided
by spatial correlation of measurements taken by different sensors, to compress data without loss
of information.
Although DSC has recently found in WSNs a relevant application domain, the theoretical
foundations date back to the pioneering work by Slepian and Wolf [1]. A practical code con-
struction to achieve the theoretical bounds devised in [1] has been proposed by Pradhan et al.
in [2] and [3]. Those works have been extended by Chou et al. [4], who have proposed an
algorithm to achieve DSC over WSNs using a single codebook with variable compression rate.
Marco and Neuhoff [5] and Ganesan et al. [6] have pointed out that performance of DSC is
largely influenced by the DSC topology, i.e., the nodes’ position with respect to the spatial
correlation profile of the sensed phenomena. The problem of optimizing the DSC topology has
been investigated by Cristescu et al. [7], [8] for tree-likenetworks.
These contributions assume that DSC acts above the typical protocol stack of a sensor network
platform. However, many other mechanisms for WSNs (e.g., routing and MAC protocols) are
targeted to the same goal of achieving better energy usage. As it was observed in [9], DSC
interactions with lower layers of the protocol stack (network, data link, and physical layer)
deserve particular attention to fully exploit the claimed benefits. For instance, header overheads
that appear in any protocol data unit (PDU) format should be taken into account. A node may
aggregate packets coming from other nodes with its own data in order to reduce the impact
of packet overhead in multi-hop routing. We believe that a crucial concern is represented by a
proper (joint) combination of DSC topologies and packet aggregation (PA) in realistic multi-hop
communication scenarios.
According to our previous considerations, it can be argued that DSC has a relevant impact on
network, data link, and physical layers, and vice versa. Theinteraction between DSC and routing
has been investigated in [10]–[12]. Petrovic et al. [10] have also noted that routing with packet
aggregation allow for decreasing protocol overhead, with reduction of energy consumption. Baek
et al. [13] have dealt with the problem of finding the optimal packet aggregation alternative
3
for WSNs with DSC. Classifications of aggregation protocolswhen nodes collect periodic
measurements have been done by Solis et al. [14], [15] and Roedig et al. [16]. In [17] we have
proposed an analysis of the energy consumption for various cases of DSC coding topologies
with PA. Using the same network scenario in [17], we have addressed in [18] the problem of
characterizing the percentage of measurements that cannotbe correctly decoded at the network
sink.
Although the above mentioned relevant contributions address fundamental topics of DSC, they
neglect some important aspects that relate to network design and operations in realistic scenarios,
namely: packet losses, packet aggregation and fragmentation, and overhead reduction. Indeed,
packet aggregation without packet fragmentation may be prohibitive in real multi-hop WSNs.
It is the purpose of this paper to develop a joint (comprehensive) analysis of the loss factor
and overall energy consumption. Our contribution is related to [5] because we include the four
alternatives of coding topology therein proposed. It is also related to [10], because we consider
three PA techniques. However, we simultaneously consider DSC and PA in a system scenario
including an accurate model of the physical layer, a data link layer, and multi-hop routing. In
this respect we extend considerably the analysis of DSC and PA proposed in [17] and [18].
The remainder of the paper is organized as follows: in Section II, the system model is
described. Section III is dedicated to the analysis of packet loss probability, which is then used
first in Section IV to characterize the network-wide loss factor, and subsequently in Section V to
analyze the energy consumption. Our theoretical frameworkis then used in Section VI to derive
and discuss numerical results in selected scenarios. Finally, in Section VII, we summarize the
main conclusions and evidence future perspectives.
II. SYSTEM MODEL
Consider an area whereinN nodes are deployed: each of them takes some measurements and
attempts to transmit reports towards a sink. The network topology is assumed to be organized
in a tree, where each node belongs to a generic leveli (with i = 1, . . . , L), i being the distance
of a node from the sink in terms of number of hops (see Fig. 1). Tree based architecture is a
relevant network topology for the IEEE 802.15.4 standard [19, pag. 15–16]. Furthermore, the
interest for such a network topology is confirmed by a large number of relevant contributions
that can be found in the literature (see, for instance, [5] – [8], [20] – [22]), which all deal with
4
tree-like networks. Denote withli the number of nodes belonging to leveli. We setl0 = 1,
which is the number of nodes of level0 (only the sink). The equality∑L
i=1 li = N holds true.
Each node is connected to a parent node and takesm measurements ofm different physical
phenomena per sensing interval. We assume that the measurements of different phenomena are
continuous i.i.d. random variables, and that the measurements of a given phenomenon by different
sensor nodes are spatially correlated. The measurements are quantized by an analog-to-digital
converter, which is assumed to have the same resolution in each sensor. Denote withYi,j, for
i = 1, . . . , L, and for j = 1, . . . , li, the quantized measurement taken from the generic nodej
of level i at a given time instant for one of them phenomena. Coherently with the purpose of
this paper, we retainYi,j as a discrete random variable taken from random process, which has
a countable discrete alphabet [4], [6], [7]. Furthermore, aset of variablesYi,js at a given time
instant are generically spatially correlated, the correlation pattern being constant in time under
the assumption of stationarity [5]. In this context we suppose that the number of bits produced
by DSC encoding of measurements is determined by the entropy(this is a common assumption
in literature, see for instance [6] – [8], [11], [23], [24], and references therein). We denote by
H(Yij) the entropy of the considered source.
Each node performs the following tasks: it encodes the measurements according to the DSC
algorithm proposed in [4], receives PDUs from the child nodes, aggregates them, and then
transmits them towards its parent node. When an ARQ protocolis employed, retransmissions
are attempted until either successful reception is attained or a maximum number of tries is
reached.
For defining and computing performance indexes, in the following we refer to a packet as a
PDU, which results from application of major headers at the various layers of the protocol stack.
Hence, a packet is composed by a frame length indicator plus apreamble and a synchronization
symbol ofP bits; a data link header ofO bits, a network header ofQ bits; a payload incorporating
coded measurements taken by nodes, andID bits for node identification;CRC bits for the
forward error detection [25]. The payload may have variablesize as a consequence of the DSC
topology and aggregation mechanism, as it will be clear in the following subsections. The packet
structure is adherent to specifications by the IEEE 802.15.4communication standard [19] for the
data link layer and physical layer of the protocol stack, while we refer to Tmote Sky sensors [26]
for the network layer.
5
In the next subsections, we first describe the DSC topologiesand the aggregation mechanisms
we intend to analyze; next we define the performance indexes we intend to adopt.
A. DSC Topology and Packet Aggregation
Four prominent DSC coding schemes are here studied: No-Slepian-Wolf (NOSW), Sequential
Slepian-Wolf (SEQ), Clustered Slepian-Wolf (CL), and Slepian-Wolf Master Slave (MS). They
were firstly proposed in [5]. In the following, we briefly summarize their characteristics.
In the NOSW topology, each sensor encodes measurements independently from other nodes.
In the SEQ scheme (see Fig. 2), node(1, 1) performs encoding ofY1,1 with H1,1 = H(Y1,1)
bits; node(1, 2) encodesY1,2 with H1,2 = H(Y1,2|Y1,1) bits, provided that the decoder knows
Y1,1; node (1, j) encodesY1,j with H1,j = H(Y1,j|Y1,1 . . . Y1,j−1) bits, provided that the de-
coder knowsY1,1 . . . Y1,j−1; in general, node(i, j) encodesYi,j using the following number of
bits Hi,j = H(Yi,j|Y1,1 . . . Y1,l1 Y2,1 . . . Y2,l2 . . . Yi,1 . . . Yi,j−1), provided that the decoder knows
Y1,1 . . . Y1,l1Y2,1 . . . Y2,l2 . . . Yi,1 . . . Yi,j−1.
In the CL topology (see Fig. 3), nodes are grouped inK = l1 clusters, where each cluster
consists of a sub-tree having a node of level1 as a root. Each cluster includes a number of
nodes, which is denoted byNk. For each node of a cluster, a SEQ coding topology is adopted
independently from other clusters and starting from the root node(1, k), for k = 1, . . . , l1, of
clusterk.
In the MS topology (see Fig. 4), nodes are grouped in clustersas in the case of CL, but each
root (1, k), for k = 1, . . . , l1, of a sub-tree is elected as master node. However, in contrast to
CL, for each node of a clusterk, the DSC is performed only with respect to the master(1, k)
of the cluster the node belongs to.
We consider three alternatives for aggregation: the classical multi-hop (CMH), where nodes
relay packets without any aggregation; the aggregated multi-hop (AMH), where nodes collect
received packets, aggregate them at the MAC layer into a single frame and relay it; the Frag-
mented Aggregated Multihop (FAMH), which differs from AMH as it allows fragmentation:
when an aggregated packet reaches a size that exceeds a maximum threshold, it is fragmented
in multiple packets.
6
B. Performance Indexes
Two performance indexes are considered in this paper, namely: the loss factor and the energy
consumption of the network, which we denote withΛ andEN, respectively. The loss factor is
defined as the fraction of measurements generated by the network that cannot be reconstructed
at the sink. The energy consumption of the network is insteadexpressed as the average number
of bits (per reporting interval) transmitted across the network. In computingEN, we will neglect
the impact of acknowledgement packets, which is a rather common assumption in literature, and
is motivated by the fact that they have marginal impact on overall performance.1
Characterization of bothΛ and EN is founded on the packet error probability. In the next
sections we derive this probability for the various combinations of DSC topologies and PA
mechanisms.
III. PACKET LOSSPROBABILITY
Consider a wireless channel which exhibits non-selective fading behavior both in frequency and
in time, and a O-QPSK (offset quadrature phase shift keying)modulation. These assumptions are
consistent with transceivers operating in the ISM frequency band according to the IEEE 802.15.4
communication standard [19]. Therefore, the bit error probability can be expressed as [25]
Pe(d) =1
2
(
1 −
√
γ(d)
1 + γ(d)
)
, (III.1)
where γ(d) is the average value of the Signal-to-Noise Ratio (SNR) computed at distanced
from the signal source. We adopt the following model forγ(d) in dB [27]: γ(d)dB = Pt dB −
P (d)dB − Pn dB, wherePt is the transmit power,P (d) is the path loss at distanced from the
transmitter, andPn is the noise floor at the receiver. The path loss is expressed as P (d) =
P (d0) + 10 · α · log10(d/d0), whereP (d0) is the path loss computed at the reference distance
d0 (see e.g. [28]); whileα is the path-loss decay constant with typical values belonging to
the interval[2, 4]. Under the assumption that the CRC code is always able to detect perfectly
1Note that the performance indexes do not include the packet delay caused by the coding topology. Indeed, under the
assumptions that the correlation properties of the sensed phenomena are stationary, that the number of measurements per
sampling rate is the typical one provided by sensors off-the-shelf [26] (i.e. less than ten), and that the number of levels is not
large, then the delays are negligible.
7
corrupted packets (see [29] for an experimental support), the average packet loss probability at
level i is defined as
Ψ(d, si) = 1 − (1 − φi) · [1 − Pe(d)]2·(P+O+Q+CRC+si) , (III.2)
whereφi is the collision probability, and the factor2 accounts for Manchester encoding. We
remark that the collision probability is computed on a bit basis in order to account for the effects
of the packet size. The termsi is defined as the average payload size among the nodes of leveli.
The average packet loss probability across one hop of leveli is defined as the average of (III.2)
with respect to the hop length:
Ψ(si) =1
dmax − dmin
∫ dmax
dmin
Ψ(ρ, si)dρ , (III.3)
wheredmin anddmax are, respectively, the minimum and maximum distances between any pair
of child-parent nodes. We call (III.3) single-hop packet loss probability. When an ARQ protocol
is implemented at levelj, with Mj maximum number of retransmissions, we can easily express
the packet loss probability of a packet generated at leveli as follows:
Ψ(si, Mj) = Ψ(si)Mj . (III.4)
The packet loss probability over a multi-hop path from nodei up to the sink can be computed
with
Ψ(si) = 1 −
i∏
j=1
[1 − Ψ(si, Mj)] . (III.5)
Remark 3.1: We can observe that in (III.3) and (III.4) the dependence on the aggregation
scheme and coding topology is included inΨ(si, Mj) through the payload size, which will be
characterized in the next subsections for the cases of aggregation CMH, AMH, and FAMH, and
coding NOSW, SEQ, CL, and MS.
We define the average connectivityCi = li+1/li, i = 0, . . . , L − 1, of a node belonging to
level i as the average number of children of that node, where we impose CL = 0. Furthermore,
for the sake of the performance analysis, we define the average coding rate of the leveli as
Hi =1
li
li∑
l=1
Hi,l .
8
A. CMH
The contribution to the average payload size for the CMH scheme at leveli, as given from
measurements generated by a node of levelj, can be expressed as
si = ID + m · Hj i = 1, . . . , L andj ≥ i , (III.6)
Therefore, the single-hop packet loss probability is givenby (III.3), wheresi is given by (III.6).
B. AMH
The following claim holds:
Claim 3.2: In the AMH scheme, the average payload size at theith level is
si =
ID ·(
1 +∑L−i
j=1
∏j
k=1 ni+k
)
+ m ·(
Hi +∑L−i
j=1 Hi+j ·∏j
k=1 ni+k
)
i = 1, . . . , L − 1 ,
ID + m · HL i = L .
(III.7)
whereni is the average number of packets of leveli that are successfully received by the parent
node of leveli − 1:
ni = [1 − Ψ(si, Mi)] · Ci−1, i = 2, . . . , L , (III.8)
Proof: In order to prove (III.7) and (III.8), consider the average number of packets that are
successfully transmitted to each parent-node of levelL − 1: nL = [1 − Ψ(sL, ML)] · CL−1 . At
level L − 1, each node aggregates the received packets and relays a unique packet having an
average payloadsL−1 = ID + m · HL−1 + sL · nL = ID + ID · nL + m · (HL−1 + HL · nL) .
At level L − 2, the average number of received packets isnL−1 = [1 − Ψ(sL−1, ML−1)] · CL−2.
Hence, the average payload size and the average number of packets coming to this level are as
follows
sL−2 =ID + ID · nL−1 · nL + ID · nL−1 + m · (HL−2 + HL−1 · nL−1 + HL · nL · nL−1) ,
nL−2 =[
1 − Ψ(sL−2, ML−2)]
· CL−3 .
Generalization of previous expressions to obtain (III.7) and (III.8) turns out from iteration.
According to the previous Claim, the single-hop packet lossprobability is given by (III.3)
and (III.7).
9
C. FAMH
In the FAMH case, the payload size is limited to a maximum length smax. When the limit is
exceeded, the payload is split intoFi fragments, withFi = ⌈ si
smax⌉, where⌈·⌉ denotes the upper
integer (ceiling). Each fragment is then mapped onto a packet and let si = (Q + si)/Fi denote
the length of the related PDU at the data link layer. Since next hop node is known at the moment
of the fragmentation, it is not necessary to transmitFi times the same network header overhead,
which can be split intoFi fragments. The fragmented packets, having payloadsi, need to have
all other headers (preamble, MAC and CRC) to be transmitted,however, they do not need to
have the network header (for next hop is known at each fragment). The loss probability of a
fragmented packet, averaged over a single-hop distance, isgiven as follows:
ΨF (si) =1
dmax − dmin
∫ dmax
dmin
1 − (1 − φi) · [1 − Pe(ρ)]2·(P+O+CRC+si) dρ . (III.9)
The packet loss probability over a single-hop can be expressed as
Ψ(si) = 1 − [1 − ΨF (si)]Fi . (III.10)
Finally, expression (III.10) can be used in (III.4) to obtain the single-hop packet loss probability.
IV. L OSSFACTOR
Here we build on the packet loss probability analysis of previous section to derive the loss
factor. First, let us introduce some definitions. Denote with Di the average number of descendants
of the generic node of leveli whose measurements are successfully received at that node.Denote
with Di,j the average number of descendants at levelj whose measurements are successfully
received at the generic node of leveli. We also imposeDi,i = 1, and thatDi,j = 0 if i > j.
Then,
Claim 4.1: At level i of the network, the average number of descendants is
Di =
L∑
j=i+1
Di,j , i = 0, . . . , (L − 1) , (IV.1)
where
Di,j =
j∏
k=i+1
[1 − Ψ(sj, Mk)] · Ck−1. (IV.2)
10
Proof: The proof is by induction. At levelL − 1, DL−1 is easily obtained byDL−1 =
DL−1,L = [1− Ψ(sL, ML)] ·CL−1 . At level L− 2, DL−2 depends on the number of packets that
are generated from levelL and levelL − 1 and that are correctly received at levelL − 2:
DL−2 = DL−2,L−1 + DL−2,L =[
1 − Ψ(sL−1, ML−1)]
· CL−2 +[
1 − Ψ(sL, ML−1)]
·[
1 − Ψ(sL, ML)]
CL−2 · CL−1 . (IV.3)
For the generic leveli, (IV.3) can be extended in a straightforward way to obtain (IV.1).
In the next subsections, we characterize the loss factor foreach case of coding topology and
aggregation mechanism.
A. NOSW
In the CMH scenario, the probability that a packet generatedat leveli does not reach the sink
is given by (III.5). Hence, after averaging over all levels and number of nodes, the loss factor
can be expressed as
Λ(a)NOSW =
1
N
L∑
i=1
1 −
i∏
j=1
[
1 − Ψ(si, Mj)]
· li , (IV.4)
where, for the sake of clarity of notation, we have denoted with the superscript(a) the CMH
case.
When aggregation procedures are adopted, derivation of theloss factor has to take into account
that loosing a packet generated from a node of leveli determines the avalanche effect of losing
measurements successfully received by that node and comingfrom lower levels of the network. It
is not difficult to see that a packet loss at leveli causes an average loss of1+Di measurements.
Therefore, the loss factor can be expressed as
Λ(b)NOSW =
1
N
L∑
i=1
1 −i
∏
j=1
[
1 − Ψ(si, Mj)]
· (1 + Di) · li , (IV.5)
where the computation ofΨ(si, Mj), which appears also in theDi, is performed by using (III.3),
(III.4), and (III.7). Note that the superscript(b) denotes AMH.
When the FAMH scheme is considered, the loss factor can be derived in the same way as in
the case of AMH:
Λ(c)NOSW =
1
N
L∑
i=1
1 −i
∏
j=1
[
1 − Ψ(si, Mj)]
· (1 + Di) · li , (IV.6)
11
where the computation ofΨ(si, Mj) is accomplished using (III.4), (III.9), and (III.10), respec-
tively. Note thatΨ(si, Mj) is also implicit in computation ofDi, although, as done before, it is
not evidenced for the sake of simple notation. The superscript (c) denotes the FAMH case.
B. SEQ
Let us firstly consider the CMH (a) case:
Claim 4.2: The loss factor can be expressed by
Λ(a)SEQ =
1
N
L∑
i=1
A(a)SEQ,i ·
li∑
j=1
BSEQ,i,j · C(a)SEQ,i,j , (IV.7)
where
A(a)SEQ,i =
1 −
i∏
m=1
[
1 − Ψ(si, Mm)]
,
BSEQ,i,j = N + 1 − j −i−1∑
m=1
lm ,
C(a)SEQ,i,j =
i−1∏
k=1
k∏
m=1
[
1 − Ψ(sk, Mm)]lk
·
i∏
m=1
[
1 − Ψ(si, Mm)]j−1
.
and whereΨ(si, Mk) is computed with (III.4).
Proof: We prove (IV.7) by induction. Denote withΦ1,1 the event of losingY1,1, so that
measurements encoded with respect toY1,1 cannot be decoded. Then, in the case of the event
Φ1,1, it causes the avalanche effect of losingN measurements, sinceY1,1 is lost along with
all measurements taken by the otherN − 1 nodes that were encoded with respect toY1,1.
Therefore, the amount of measurements lost when only the event Φ1,1 happens, is given by
ΛΦ1,1= Ψ(s1, M1) · N .
Denote withΦ1,2 the event of successfully receivingY1,1 while losingY1,2. Then, whenΦ1,2
occurs,N − 1 measurements are lost, sinceY1,2 is lost along with theN − 2 measurements that
were coded with respect toY1,2. Therefore, the amount of measurements lost when onlyΦ1,2
takes place is given byΛΦ1,2= Ψ(s1, M1)· (N − 1) · [1 − Ψ(s1, M1)] .
Denote withΦ1,l1 the event of losingY1,l1 , while Y1,1, Y1,2, . . . , Y1,l1−1 were successfully re-
ceived. IfΦ1,l1 happens, thenN−l1+1 measurements are lost, since measurementsY2,1, . . . , Y2,l2,
12
Y3,1, . . . , Yl,lL were coded with respect toY1,l1. Therefore, the amount of measurements lost
in correspondence of the only eventΦ1,l1 is given by ΛΦ1,l1= Ψ(s1, M1) · (N + 1 − l1) ·
[
1 − Ψ(s1, M1)]l1−1
.
Consider now the general case of the eventΦi,j , when Yi,j is lost, while Y1,1, Y1,2, . . . ,
Y1,l1, Y2,1 . . . , Yi,1 . . . , Yi,j−1 are successfully received. WhenΦi,j occurs, thenYi,j+1, . . . , Yi,li,
Yi+1,1, . . . , YL,1, . . . , YL,ll are lost, since they were coded with respect toYi,j. Hence, the total
number of lost measurements isN + 1 − j −∑i−1
k=1 lk , and the amount of measurements lost
when the only eventΦi,j takes place isΛΦi,j= A
(a)SEQ,i · BSEQ,i,j · C
(a)SEQ,i,j . By averagingΛΦi,j
with respect to all measurements, Eq. (IV.7) is obtained.
In the cases of AMH and FAMH, expressions for the loss factorsare similar to those given
by Claim 4.2. In fact, the loss of a packet coming from node(i, j) induces the loss of all
measurements aggregated from lower layers of the network until (i, j): its average value is
given by Di in (IV.1). However, this number of losses is already included in the number of
measurements that, being coded with respect toYi,j, cannot be reconstructed. It is easy to see
from Fig. 1 that this metric can be expressed asN +1− j −∑i−1
k=1 lk. Therefore, we can readily
express the loss factor for the AMH (b) case as
Λ(b)SEQ =
1
N
L∑
i=1
A(b)SEQ,i ·
li∑
j=1
BSEQ,i,j · C(b)SEQ,i,j , (IV.8)
where the superscript(b) denotes that the probabilityΨ(si, Mj) has to be computed using (III.3),
(III.4), and (III.7).
For the case of FAMH(c), we have
Λ(c)SEQ =
1
N
L∑
i=1
A(c)SEQ,i ·
li∑
j=1
BSEQ,i,j · C(c)SEQ,i,j , (IV.9)
where the superscript(c) denotes that the computation ofΨ(si, Mj) is accomplished using (III.4),
(III.9), and (III.10).
C. CL
When the network is partitioned into clusters, for each cluster the loss factor can be computed
as in the SEQ case. Hence, the loss factor for CMH (a), AMH (b),and FAMH (c) is provided
13
by the following expression:
Λ(a,b,c)CL =
1
K
K∑
k=1
Λ(a,b,c)SEQ,k , (IV.10)
where the subscriptk denotes the loss factor of cluster numberk, and it is computed using (IV.7)
for the case(a), (IV.8) for the case(b), and (IV.9) for the case(c). When computingΛ(a,b,c)SEQ,k,
N and li have to be replaced (whenever they appear in (IV.7), (IV.8),and (IV.9)) with Nk and
l(k)i respectively, wherel(k)
i is defined as the number of nodes of leveli of clusterk. Notice that
l(k)0 = 1 is the number of nodes of level0 (only the master node) of clusterk.
D. MS
Let us consider first the CMH (a) scheme. We have the followingclaim:
Claim 4.3: The loss factor for the CMH scheme is
Λ(a)MS =
1
K
K∑
k=1
Λ(a)MS,k ,
whereΛ(a)MS,k is the loss factor for clusterk, and it is given by
Λ(a)MS,k = Ψ(s1, M1) +
1
Nk
L∑
i=2
l(k)i ·
[
1 − Ψ(s1, M1)]
·
1 −i
∏
m=1
[
1 − Ψ(si, Mm)]
,
and whereΨ(si, Mm) is computed using (III.4).
Proof: Denote withΩ1,k, for k = 1, . . . , l1, the event of losingY1,k. When Ω1,k takes
place, allNk measurements of clustersk are lost, since they were coded with respect toY1,k.
Therefore, the amount of lost measurements induced by the only eventΩ1,k is given byΛΩ1,k=
NkΨ(s1, M1) .
Consider now the eventΩi,j , for i = 2, . . . , L and j = jk, . . . , jk + l(k)i − 1, where jk is
the index of the first node of leveli of cluster k, and jk + l(k)i − 1 is the index of the last
node of the same level of the same cluster. The eventΩi,j occurs whenY1,k is successfully
received andYi,j is lost. Therefore, the loss factor induced by the only eventΩi,j is given as
ΛΩi,j=
[
1 − Ψ(s1, M1)]
·
1 −∏i
m=1
[
1 − Ψ(si, Mm)]
. By averaging with respect toΛΩi,j,
we get expression (IV.11).
In the case of packet aggregation, the procedure to derive the loss factor is the same as in the
CMH scheme. However, the loss of a measurement taken by a node(i, j) in clusterk, induces
to lose furtherD(k)i measurements on the average, whereD
(k)i is defined as in (IV.1), but using
14
C(k)i instead ofCi, with C
(k)i defined asC(k)
i = l(k)i+1/l
(k)i , i = 0, . . . , L− 1 , and thatC(k)
L = 0.
Therefore, the loss factor is given by
Λ(b)MS =
1
K
K∑
k=1
Λ(b)MS,k
where
Λ(b)MS,k = Ψ(s1, M1) + (IV.11)
1
Nk
L∑
i=2
l(k)i · (1 + D
(k)i ) ·
[
1 − Ψ(s1, M1)]
·
1 −
i∏
m=1
[
1 − Ψ(si, Mm)]
.
In (IV.11), Ψ(si, Mm) and D(k)i are computed by resorting to (III.3), (III.4), and (III.7).Ob-
viously, (III.7) is computed by replacingN , li, andCi with Nk, l(k)i , andC
(k)i whenever they
appear.
The loss factor in the FAMH (c) case is derived by applying thesame steps described for the
(b) case. Hence,
Λ(c)MS =
1
K
K∑
k=1
Λ(c)MS,k
where
Λ(c)MS,k = Ψ(s1, M1) +
1
Nk
L∑
i=2
l(k)i · (1 + D
(k)i ) ·
[
1 − Ψ(s1, M1)]
·
1 −i
∏
m=1
[
1 − Ψ(si, Mm)]
.
In previous expression,Ψ(si, Mm) andD(k)i are given by (III.4), (III.9), and (III.10).
V. AVERAGE ENERGY CONSUMPTION
Expressing the average number of bits transmitted by the network requires characterization
of the average number of transmitted packets at the generic level i. Recall that such a number
depends on the coding topology by the payload size. In the following, we derive the energy
consumption for the CMH, AMH and FAMH schemes.
The number of retransmissions of a packet sent from a node of level i over the hop between
level i and i − 1 is denoted asΘi. It is expressed as [25]
Θi =1 + Mi · Ψ(si)
Mi+1 − (Mi + 1) · Ψ(si)Mi
1 − Ψ(si).
15
Therefore, the average number of transmitted bits from the generic node of leveli, when an
ARQ protocol is employed, is
Bi =
L∑
j=i
(P + O + Q + sj + CRC) · Di,j · Θj ,
whereDi,j has been defined in (IV.2).
Consider the CMH (a) case. The average number of transmittedbits is
E(a)N =
L∑
i=1
Bi · li . (V.1)
Let us consider now the AMH (b) case. Each node of leveli transmits only one packet
containing both the measurements taken from local sensing and the measurements coming from
lower levels of the network and encapsulated therein. Therefore, EN is readily expressed as
follows:
E(b)N =
L∑
i=1
(P + O + Q + si + CRC) · Θi · li . (V.2)
Finally, consider the FAMH (c) case. It is easy to see that theaverage number of transmitted
bits is expressed as
E(c)N =
L∑
i=1
(P + O + si + CRC) · Fi · Θi · li , (V.3)
whereΘi is computed using the payload sizesi for the FAMH case.
VI. NUMERICAL EXAMPLES
In this section we report numerical results obtained by a Matlab implementation of the
analytical framework developed in previous sections. We consider two representative scenarios:
the first one refers to large spatial correlations among measurements, while the second one refers
to low correlation patterns. The parameters setting adopted for simulations is representative of the
Tmote Sky sensors [26] and the communication standard IEEE 802.15.4 [19], and are introduced
in the sequel.
We consider a network deployed in a square area having facet of 11m with the sink located in
the middle andN = 64 nodes distributed inL = 4 levels, where each node is randomly located
within a sub-square of1.2m. The values set for the number of nodes for each level, alongwith
average connectivity, are reported in Tab. II. Each node sensesm = 8 measurements per sensing
event. The packet frame format is as follows:P = 48, O = 184, Q = 56, ID = 32, CRC = 16,
16
andsmax = 760, where the unit is intended in bit. Transmission power has been set to−10 dBm.
The noise floor isPn = −120 dBm [26]. The collision probability has been set toφi = 10−5 (this
corresponds to a packet collision probability of about0.005, which is quite large if related to the
sampling rate, the packet duration, and the number of nodes). Smaller values for the collision
probability basically induce similar effects. Without loss of generality, and coherently with [19],
we set the maximum retransmission iterations of the ARQ protocol for levels1, ..., L = 4,
to 3, 3, 1, 1, respectively. This choice is motivated by the fact that nodes closer to the sink are
subject to larger packet losses as a consequence of larger traffic load. Other choices for the ARQ
protocol provide basically the same trends in the results asthose presented later in this section.
Without loss of generality, we make use of the differential entropy instead of the usual entropy,
as proposed in [6]– [8], and [30, pagg. 227–229]). By adopting this model, measurements are
characterized with aN dimensional multi-variate normal distributionGN(µ,K) having average
µ and covariance matrixK = [Ki,j ], whereKii = σ2i and Ki,j = σ2
i e−rdij for i 6= j, where
r is the spatial correlation decay parameter, anddij is the distance between nodesi and j.
The correlation decay parameter is defined such that it tendsto 0 for highly correlated sources,
whereas tends to1 for low correlated sources, as discussed in [7].
In Fig. 5, the loss factor is plotted for various cases of the coding topology (NOSW, SEQ,
CL, and MS) and aggregation scheme (CMH, AMH, and FAMH) as obtained in Section IV,
for a highly correlated measurements scenario, wherer = 10−5. As a first general remark, it
can be observed that the adoption of ARQ significantly decreases the amount of measurements
that cannot be decoded at the sink node. This is particularlyevident in the MS case, where
the decrease of loss factor due to ARQ is relevant. When looking at the SEQ case, the loss
factor takes on the largest value among DSC topologies. Indeed, recalling the behavior of the
SEQ topology summarized in Section II.A, losing a measurement at node(i, j) determines the
avalanche effect of losing all measurementst, q, with t = i+1, . . . , L andq = j +1, . . . , li. The
same effect appears also in the CL topology, but clustering reduces the amount of performance
loss. Indeed, no relevant avalanche effect is present in this topology. When looking at aggregation
mechanisms, we can observe that they cause a slight increaseof the loss factor, since packets
have larger sizes if compared to the CMH case. A relevant conclusion that can be drawn from
Fig. 5 is that MS seems to be the most appealing coding topology in terms of reliability, and
that aggregation mechanisms exhibit similar performance within the same coding topology.
17
In Fig. 6, values of the energy consumptionEN for the various DSC topologies and aggregation
schemes are plotted as obtained for highly correlated sources. They have been computed through
the analysis presented in Section V. The right column is referred to the ARQ protocol, while
the left column is referred to the case where ARQ is not adopted. Although ARQ introduces
a larger number of packets per node to be relayed, we can clearly observe that it determines a
small energy increase. Nevertheless, by considering that the ARQ protocol is beneficial in terms
of loss factor, it can be concluded that the adoption of ARQ induces significant advantages. The
NOSW topology exhibits the worst performance, as no compression is adopted. It is interesting
to observe that SEQ, CL and MS topologies do not exhibit largedifferences. This is mainly due
to the large spatial correlation, so that measurements froma few nodes are enough to achieve
good compression rates. Considering the aggregation mechanisms, it is evident that they provide
significant performance improvement. Specifically, CMH exhibits the worst performance among
aggregation mechanisms, whatever DSC scheme is adopted. This is due to the fact that packet
overhead introduces extra bits in the packet transmission.
For a better understanding of energy consumption, define thefollowing coding efficiency
metric:
ηCT = 1 −EN,CT
EN,NOSW
, (VI.1)
where the subscript CT denotes one of the coding topologies SEQ, CL, or MS. Large values of
ηCT mean that the corresponding coding topology exhibits reduced energy consumptions when
compared to the NOSW case.
In Fig. 7, the coding efficiency is plotted versus the different coding topologies (x axis)
and aggregation algorithms for the case of highly correlated sources. The coding efficiency
increases remarkably when we move from CMH to FAMH, with AMH providing intermediate
performance. We can observe that the performance provided by the three topologies, SEQ, CL
and MS, are quite different for the cases of AMH and FAMH, while they are less evident in
CMH. It is interesting to remark that the MS coding topology shows the worst performance.
Indeed we recall that the compression is performed with respect to only one measurement for
each cluster.
In order to compare both efficiency and reliability, in Fig. 8the coding efficiency is reported
as a function of the loss factor, for the case of highly correlated sources. It turns out that the
18
SEQ alternative often offers the largest coding efficiency,but also induces larger loss factors. On
the contrary, the MS case exhibits poor energy savings, but good loss factor. The CL topology
presents fair joint performance. Finally, it is confirmed that aggregation mechanisms significantly
impact on the energy consumption.
Let us now consider the case of low correlated measurements,in which we setr = 10−3.
In Fig. 9, we plot the coding efficiency. As expected, we can observe that the maximum
achievable coding efficiency is smaller than the maximum values obtained for highly correlated
measurements in Fig. 7. Notice that once again the ARQ protocol does not induce significant
degradation of the energy efficiency. In Fig. 10, we reportedthe coding efficiency as a function
of the loss factor. The remarks done for Fig. 8 still hold true, confirming the good behavior of
the MS in terms of loss factor, and its poor performance in terms of energy efficiency.
VII. CONCLUSIONS AND FUTURE PERSPECTIVES
A theoretical framework to evaluate expressions for the loss factor and energy efficiency in the
presence of three alternatives of packet aggregation mechanisms (classic multi-hop, aggregated
multi-hop and aggregated and fragmented multi-hop), and four alternatives of distributed source
coding topology (no Slepian-Wolf coding, sequential, clustered and master slave) has been
proposed.
Numerical results show that, under the correlation model wehave used, SEQ coding topology
performs better in terms of energy consumption, yet showingpoor reliability performance.
On the contrary, the MS case offers better performance in terms of measurements that are
successfully decoded, but it is poorer in terms of energy efficiency. Furthermore, the ARQ
protocol significantly decreases the amount of measurements which cannot be decoded at the
sink; moreover, retransmissions do not seem to remarkably affect energy efficiency. We have
instead shown that packet overheads and packet loss probability may play a significant role in
network energy consumption.
Future study will be focused on the extension of our theoretical analysis to evaluate the
interplay of main communication system parameters, e.g. including radio frequency power
control. Furthermore, while in the present paper we have considered a tree-like network, we
plan to extend our analysis to a joint optimization of the routing, aggregation scheme, source
coding with respect to the spatio-temporal correlation pattern of the distributed source.
19
REFERENCES
[1] D. Slepian and J. Wolf, “Noiseless Coding of Correlated Information Sources,”IEEE Transactions on Information Theory,
vol. 19, no. 4, pp. 471 – 480, July 1973.
[2] S. Pradhan, J. Kusuma, and K. Ramchandran, “DistributedCompression in a Dense Microsensor Network,”IEEE Signal
Processing Magazine, vol. 19, no. 2, pp. 51 – 60, March 2002.
[3] S. Pradhan and K. Ramchandran, “Distributed Source Coding Using Syndromes (Discus): Design and Construction,”IEEE
Transactions on Information Theory, vol. 49, no. 3, pp. 626 – 643, March 2003.
[4] J. Chou, D. Petrovic, and K. Ramchandran, “A Distributedand Adaptive Signal Processing Approach to Reducing Energy
Consumption in Sensor Networks,” inTwenty-Second Annual Joint Conference of the IEEE Computerand Communications
Societies (INFOCOM), vol. 2, March 2003, pp. 1054 – 1062.
[5] D. Marco and D. Neuhoff, “Reliability vs. Efficiency in Distributed Source Coding for Field-Gathering Sensor Networks,”
in Third International Symposium on Information Processing in Sensor Networks, 2004, pp. 161 – 168.
[6] D. Ganesan, R. Cristescu, and B. Beferull-Lozano, “Power-Efficient Sensor Placement and Transmission Structure for Data
Gathering under Transmission Constraints,”ACM Transaction on Sensor Networks, vol. 2, pp. 155 – 181, May 2006.
[7] R. Cristescu, B. Beferull-Lozano, and M. Vetterli, “Networked Slepian-Wolf: Theory, Algorithms and Scaling Laws,” IEEE
Transactions on Information Theory, vol. 51, no. 12, pp. 4057–4073, December 2005.
[8] R. Cristescu, B. Beferull-Lozano, M. Vetterli, and R. Wattenhofer, “Network Correlated Data Gathering with Explicit
Communication: NP-Completeness and Algorithms,”IEEE/ACM Transactions on Networking, vol. 14, no. 1, December
2006.
[9] Z.-Q. Luo, M. Gatspar, J. Liu, and A. Swami, Eds.,Signal Processing Magazine: Special Issue on Distributed Signal
Processing in Sensor Networks. IEEE, 2006.
[10] D. Petrovic, R. Shah, K. Ramchandran, and J. Rabaey, “Data Funneling: Routing with Aggregation and Compression for
Wireless Sensor Networks,” inFirst IEEE International Workshop on Sensor Network Protocols and Applications, May
2003, pp. 156 – 162.
[11] A. Scaglione, “Routing and Data Compression in Sensor Networks: Stochastic Models for Sensor Data that guarantee
Scalability,” in IEEE International Symposium on Information Theory, June 2003.
[12] S. Pattem, B. Krishnamachari, and R. Govindan, “The Impact of Spatial Correlation on Routing with Compression in
Wireless Sensor Networks,” inThird International Symposium on Information Processing in Sensor Networks, 2004.
[13] S. Baek, G. de Veciana, and X. Su, “Minimizing Energy Consumption in Large-Scale Sensor Networks through Distributed
Data Compression and Hierarchical Aggregation,”IEEE Journal on Selected Areas in Communications, vol. 22, no. 6, pp.
1130 – 1140, August 2004.
[14] I. Solis and K. Obraczka, “The Impact of Timing in Data Aggregation for Sensor Networks,” inIEEE International
Conference on Communications, June 2004.
[15] ——, “In-Network Aggregation Trade-Offs for Data Collection in Wireless Sensor Networks,”International Journal of
Sensor Networks, vol. 1, no. 3/4, pp. 200 – 212, 2006.
[16] U. Roedig, A. Barroso, and C. Sreenan, “Determination of Aggregation Points in Wireless Sensor Networks,” in30th
Euromicro Conference, September 2004.
[17] L. Di Paolo, C. Fischione, F. Graziosi, F. Santucci, andS. Tennina, “Performance Analysis of Distributed Source Coding
and Packet Aggregation in Wireless Sensor Networks,” inIEEE Global Telecommunications Conference, November 2006.
20
[18] ——, “Loss Factor of Distributed Source Coding with Packet Aggregation in Wireless Sensor Networks,” in14th
International Conference on Telecommunication Systems - Modeling and Analysis (ICTSMA), October 2006.
[19] IEEE Std 802.15.4: Wireless Medium Access Control (MAC) andPhysical Layer (PHY) Specifications for Low-Rate Wireless
Personal Area Networks (WPANs). IEEE Computer Society, 2003.
[20] Simultaneous Optimization of Concave Costs: Single Sink Aggregation or Single Source Buy-at-Bulk, 2003.
[21] S. Draper and G. Wornell, “Side Information Aware Coding Strategies for Sensor Networks,”IEEE Journal on Selected
Areas in Communications, vol. 22, no. 6, pp. 966 – 976, August 2004.
[22] A. Koubaa, M. Alves, and E. Tovar, “Modelling and Worst-Case Dimensioning of Cluster-Tree Wireless Sensor Networks,”
in 27th IEEE International Real-Time Systems Symposium, 2006.
[23] A. Scaglione and S. D. Servetto, “On the Interdependance of Routing and Data Compression in Multi-hop Sensor Network,”
in IEEE MOBICOM, September 2002.
[24] D. Marco, E. J. Duarte-Melo, M. Liu, and D. Neuhoff, “On the Many-To-One Transport Capacity of a Dense Wirless
Sensor Network and the Compressibility of its Data,” inIEEE Information Processing in Sensor Networks, 2003.
[25] S. Haykin,Communications System, 4th ed. John Wiley, May 2000.
[26] Tmote Sky Data Sheet, Moteiv, San Francisco, CA, 2006. [Online]. Available: http://www.moteiv.com/products/docs/
tmote-sky-datasheet.pdf
[27] M. Zuniga and B. Krishnamachari, “Analyzing the Transitional Region in Low Power Wireless Links,” inIEEE First
Annual Conference on Sensor and Ad Hoc Communications and Networks (SECON), October 2004, pp. 517 – 526.
[28] K. Sohrabi, B. Manriquez, and G. Pottie, “Near Ground Wideband Channel Measurement in 800-1000 Mhz,” inIEEE
49th Vehicular Technology Conference, July 1999.
[29] J. Jeong and C. Ee, “Forward Error Correction in Sensor Networks,” University of California at Berkeley, http://webs.cs.
berkeley.edu/papers/FECreport.pdf, Tech. Rep., 2003.
[30] T. Cover and J. Thomas,Elements of Information Theory, D. Schilling, Ed. Wiley & Sons, 1991.
21
Symbol Meaning
N Number of nodes
Nk Number of nodes in the clusterk
L Number of levels of the network
li Number of nodes of leveli of the network
Ci Average connectivity of a node of leveli
Di Average number of descendants of a node of leveli
Di,j Average number of descendants at levelj whose measurements are successfully received at a node of level i
K Number of clusters of nodes
dmin Minimum distance among nodes
dmax Maximum distance among nodes
Yi,j Data measured by nodej of level i
m Number of sensed phenomena per node per sampling time
H(·) Entropy function of the source
P Frame length indicator + Preamble+Sync size (bits)
O DataLink section size (bits)
Q Network section size (bits)
si Average payload size (bits)
si Average fragmented payload size (bits)
smax Maximum payload size (bits)
ni Average number of packets of leveli that are successfully received by a parent node of leveli − 1
CRC Correction code section size (bits)
ID Identification code size (bits)
Fi Number of packet fragments at leveli
Mi Maximum number of retrasmissions at leveli
γ(·) SNR function
φi Collision probability at leveli
Pe(·) Bit error probability function
Ψ(d, si) Single-hop packet loss probability function
Ψ(si) Average single-hop packet loss probability function
Ψ(si, Mj) Single-hop packet loss probability function with ARQ
Ψ(si) Multi-hop packet loss probability function with ARQ
Θi Average number of retransmissions for a packet of leveli
Bi Average number of transmitted bits for a node of leveli
Λ Loss Factor of the network
EN Average number of transmitted bits of the network
ηCT Coding efficiency
TABLE I
SYMBOLS.
22
i 0 1 2 3 4
li 1 4 12 20 28
Ci 4 3 20/12 28/20 0
TABLE II
NUMBER OF NODES FOR EACH LEVEL AND AVERAGE CONNECTIVITY.
Fig. 1. Network topology. Nodes are organized as a tree withL levels. The root node acts as sink.
Fig. 2. Sequential coding topology (SEQ).
23
Fig. 3. Clustered coding topology (CL). In each cluster, a SEQ coding is employed. The superscript in the indexes corresponds
to the cluster the node belongs to. Notice that the level 1 of each cluster contains only a node.
Fig. 4. Master Slave coding topology (MS). It is similar to the CL, but the coding is done only with respect to a cluster’
master. The superscript in the indexes corresponds to the cluster the node belongs to. Notice that the level 1 of each cluster
contains only a node.
24
CMH AMH FAMH0
0.2
0.4
0.6
0.8
LOS
S F
AC
TO
R
NO SLEPIAN−WOLF
CMH AMH FAMH0
0.2
0.4
0.6
0.8
LOS
S F
AC
TO
R
SEQUENTIAL SLEPIAN−WOLF
CMH AMH FAMH0
0.2
0.4
0.6
0.8
LOS
S F
AC
TO
R
CLUSTERED SLEPIAN−WOLF
CMH AMH FAMH0
0.2
0.4
0.6
0.8
LOS
S F
AC
TO
R
MASTER−SLAVE
NO ARQ
ARQ
Fig. 5. Loss factor for CMH, AMH, and FAMH with and without ARQ, for the case of highly correlated measurements. The
right bars are referred to ARQ case, whereas the left ones arereferred to the case without ARQ.
CMH AMH FAMH0
2
4
6
8x 10
4
BIT
TX
/ S
NA
PS
HO
T
NO SLEPIAN−WOLF
CMH AMH FAMH0
2
4
6
8x 10
4
BIT
TX
/ S
NA
PS
HO
T
SEQUENTIAL SLEPIAN−WOLF
CMH AMH FAMH0
2
4
6
8x 10
4
BIT
TX
/ S
NA
PS
HO
T
CLUSTERED SLEPIAN−WOLF
CMH AMH FAMH0
2
4
6
8x 10
4
BIT
TX
/ S
NA
PS
HO
T
MASTER−SLAVE
NO ARQ
ARQ
Fig. 6. Average number of byte transmitted by the network forthe case of highly correlated measurements. The right bars are
referred to ARQ case, whereas the left ones are referred to the case without ARQ.
25
SEQ CL MS5
10
15
20
25
CO
DIN
G E
FF
ICIE
NC
Y (
%)
CMHCMH−ARQAMHAMH−ARQFAMHFAMH−ARQ
Fig. 7. Coding Efficiency for CMH, AMH, and FAMH with and without ARQ for the case of highly correlated measurements.
On the abscissa the cases SEQ, CL, and MS are reported.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.96
8
10
12
14
16
18
20
22
24
26
LOSS FACTOR
CO
DIN
G E
FF
ICIE
NC
Y (
%)
SEQ
CL
MS
CMHCMH−ARQAMHAMH−ARQFAMHFAMH−ARQ
Fig. 8. Coding Efficiency vs Loss factor for CMH, AMH, and FAMHwith and without ARQ, for the case of highly correlated
measurements. On each curve, the markers on the left are referred to MS, in the middle to CL, and on the right to SEQ.
26
SEQ CL MS2
4
6
8
10
12
14
16
CO
DIN
G E
FF
ICIE
NC
Y (
%)
CMHCMH−ARQAMHAMH−ARQFAMHFAMH−ARQ
Fig. 9. Coding Efficiency for CMH, AMH, and FAMH with and without ARQ, for the case of weakly correlated measurements.
On the abscissa the cases SEQ, CL, and MS are reported.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.92
4
6
8
10
12
14
16
LOSS FACTOR
CO
DIN
G E
FF
ICIE
NC
Y (
%)
SEQ
CL
MS
CMHCMH−ARQAMHAMH−ARQFAMHFAMH−ARQ
Fig. 10. Coding Efficiency vs Loss factor for CMH, AMH, and FAMH with and without ARQ, for the case of weakly correlated
measurements. On each curve, the markers on the left are referred to MS, in the middle to CL, and on the right to SEQ.