Path Loss Exponent Estimation

in Large Wireless Networks

Sunil Srinivasa and Martin Haenggi
Network Communications and Information Processing (NCIP) Lab
Department of Electrical Engineering
University of Notre Dame
Aug 13, 2009
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 1 / 22
Motivation
Large-scale path loss law: signal strength attenuates with
distance d as d

S
_
d
d
0
_

.
Though it is typically assumed in analysis and design problems
that the path loss exponent (PLE) is known a priori, it is often
not the case.
The PLE has a strong impact on the quality of links, and
therefore needs to be accurately estimated for the ecient
design and operation of systems.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 2 / 22
Motivation (contd.)
Example 1: The information-theoretic capacity of large random
ad hoc networks scales as

n
2/2
for 2 < 3

n for 3.
Depending on the value of , dierent routing strategies are
required to be implemented.

A.

Ozg ur, O. Leveque and D. Tse, Hierarchical Cooperation Achieves
Optimal Capacity Scaling in Ad Hoc Networks, IEEE Trans. Info. Th., 2007.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 3 / 22
Motivation (contd.)
Example 2: Outage probability in a planar Poisson point
process with Rayleigh fading.
2 2.5 3 3.5 4 4.5 5
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
= 1, p = 0.05, m = 1, N
0
= 25 dBm
O
u
t
a
g
e

p
r
o
b
a
b
i
l
i
t
y

= 20 dB
= 10 dB
= 5 dB
= 0 dB
= 5 dB
The system performance critically depends on .
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 4 / 22
System Model
Receiver
Transmitter
Filled circles: transmitters.
Empty circles: receivers.
An innite Poisson point
process (PPP) on R
2
with
density .
Channel access scheme is
ALOHA.
p is the ALOHA contention
parameter. Therefore, the
set of transmitters forms a
PPP with density p.
No synchronization.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 5 / 22
System Model (contd.)
Attenuation in the channel: product of
large-scale path loss, with PLE .
small-scale fading (m-Nakagami).
m= 1: Rayleigh fading ; m : no fading.
Noise is AWGN with variance N
0
.
All the transmit powers are equal to unity (no power control).
Problem: How do you accurately estimate the PLE at each node in
the network in a completely distributed manner?
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 6 / 22
What Makes Estimating the PLE Complicated?
The large-scale path loss is commonly taken to be deterministic
while the small-scale fading is modeled as a stochastic process.
This distinction, however, does not hold when the nodes
themselves are randomly arranged. So, we need to consider
the distance and fading ambiguities jointly.
Moreover, PLE estimation needs to be performed during the
initialization of the network. During this phase, the system is
typically interference-limited due to the presence of
uncoordinated transmissions.
Purely RSS-based estimators cannot be used in these situations.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 7 / 22
Overview
Propose three distributed algorithms for estimating the PLE in
large random wireless networks that explicitly take into account
the uncertainty in the locations of the nodes.
the uncertainty in the fading gains across links.
the interference in the network.
Provide simulation results to demonstrate the performance of
the algorithms and quantify the estimation errors.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 8 / 22
The Big Picture
The PLE estimation problem is essentially tackled by equating
the empirical (observed) values of certain network
characteristics to their theoretically established values.
By obtaining measurements over several time slots, the PLE can
be estimated at each node in a distributed fashion.
The three PLE algorithms are each based on a specic
network characteristic:
the mean interference.
the outage probability.
connectivity properties of a node.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 9 / 22
Simulation Details
We use 50, 000 dierent realizations of the PPP to analyze the
mean error performance of the algorithms, which is characterized
using the relative MSE, dened as E
_
( )
2
_
/.
We used p = 0.05 since it was suitable. Note the tradeos.
Large p: results in few quasi-dierent realizations of the
transmitter PPP.
Small p: takes long for the algorithms to convergence.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 10 / 22
Algo. 1: Using the Mean Interference
This algorithm assumes that the network density is known.
In theory, the mean interference is given by

= 2pA
2
0
/ ( 2) , (1)
where A
0
is the near-eld radius of the antenna.
Implementation
Nodes simply need to record the mean strength of the received
power,

, averaged over several time slots.

Equating to

, and using the known values of p, A

0
, and ,
is found from a look-up table.

J. Venkataraman, M. Haenggi and O. Collins, Shot Noise Models for Outage

and Throughput Analyses in Wireless Ad Hoc Networks, MILCOM, 2006.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 11 / 22
Algo. 1: Using the Mean Interference (contd.)
Relative MSE of versus the number of time slots.
200 400 600 800 1000 1200 1400 1600 1800 2000 100
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0.22
0.24
Number of time slots N
R
e
l
a
t
i
v
e

M
S
E
= 1, p = 0.05, m = 1, N
0
= 25 dBm, A
0
= 1


= 2.5
= 3
= 3.5
= 4
= 4.5
The estimates are fairly accurate over a wide range of parameters.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 12 / 22
Algo 2: Based on (Virtual) Outage Probabilities
This algorithm does not require the knowledge of or m.
In a PPP, when the signal power is exponentially distributed, the
probability of a successful transmission p
s
is
p
s
= P(SIR > ) = exp(c
2/
), (2)
where c = p
_
m+
2

_
1
2

__
_
(m)m
2/
_
.
Nodes can determine the SIR, and consequently p
s
by
computing the ratio of the power of the signal (which arrives
from a virtual transmitter, and is assumed to be exponentially
distributed) and the total received power.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 13 / 22
Algo 2: Based on (Virtual) Outage Prob. (contd.)
Implementation: A dierential method.
Obtain a histogram of the observed SIR values measured over
several time slots.
The empirical success probabilities (p
s,i
= P(SIR >
i
),
i = 1, 2) are obtained at two dierent threshold values.
Inverting (2), an estimate of is obtained as
=
2 ln(
1
/
2
)
ln (ln p
s,1
/ ln p
s,2
)
. (3)
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 14 / 22
Algo 2: Based on (Virtual) Outage Prob. (contd.)
Relative MSE of versus the number of time slots.
200 400 600 800 1000 1200 1400 1600 1800 2000 100
0
0.05
0.1
0.15
0.2
0.25
= 1, p = 0.05, m = 1, N
0
= 25 dBm,
1
= 10 dB,
2
= 0 dB
R
e
l
a
t
i
v
e

M
S
E
Number of time slots N


= 2.5
= 3
= 3.5
= 4
= 4.5
The estimation error increases with larger .
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 15 / 22
Algo 2: Based on (Virtual) Outage Prob. (contd.)
Relative MSE of versus the Nakagami parameter.
10 100 0.5 1
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Nakagami parameter m
R
e
l
a
t
i
v
e

M
S
E
= 1, p = 0.05, N
0
= 25 dBm,
1
= 10 dB,
2
= 0 dB, N = 10000


= 2.5
= 3
= 3.5
= 4
= 4.5
This algorithm performs more accurately at lower values of m.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 16 / 22
Algo 3: Based on the Cardinality of the Tx Set
This algorithm also does not require to know m or .
Transmitter node y is in receiver node xs transmitting set, T(x)
if they are connected, i.e., the SIR at x due to ys signal is > .
We prove that under the conditions of m N,
E|T(x)| =

N
T
=
(m)
_
1
_
2

_
m
_
(m+
2

)(2
2

)
2/
. (4)
We see that

N
T
is inversely proportional to
2/
, and surmise
that this behavior holds at arbitrary m R
+
.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 17 / 22
Algo 3: Based on the Card. of the Tx Set (contd.)
Implementation
For a known threshold
1
1, at time slot i, 1 i N, set
N
T,1
(i) =

1 if the node can decode a packet

0 otherwise.
Evaluate

N
T,1
and

N
T,2
at two dierent threshold values
1
and

2
respectively.
In theory, we obtain

N
T,1
/

N
T,2
= (
2
/
1
)
2/
.
Inverting this, we have
= (2 ln(
2
/
1
)) / ln(

N
T,1
/

N
T,2
). (5)
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 18 / 22
Algo 3: Based on the Card. of the Tx Set (contd.)
Relative MSE of versus the number of time slots.
200 400 600 800 1000 1200 1400 1600 1800 2000 100
0.12
0.14
0.16
0.18
0.2
0.22
0.24
Number of time slots N
R
e
l
a
t
i
v
e

M
S
E
= 1, p = 0.05, m = 1, N
0
= 25 dBm,
1
= 10 dB,
2
= 0 dB


= 2.5
= 3
= 3.5
= 4
= 4.5
In contrast to Algo. 1 and 2, the relative MSE decreases with .
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 19 / 22
Algo 3: Based on the Card. of the Tx Set (contd.)
Relative MSE of versus the Nakagami parameter.
1 10 100 0.5
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Nakagami parameter m
R
e
l
a
t
i
v
e

M
S
E
= 1, p = 0.05, N
0
= 25 dBm,
1
= 10 dB,
2
= 0 dB, N = 10000


= 2.5
= 3
= 3.5
= 4
= 4.5
The estimates are more accurate at lower m.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 20 / 22
Summary and Discussion
We have addressed the PLE estimation problem in the presence
of node location uncertainties, m-Nakagami fading and most
importantly, interference!
Each of the algorithms are fully distributed and do not require
any information on the location of other nodes or the value of m.
Based on the relative MSE values, we conclude that at low
values of , Algo. 1 performs the best (though it requires the
density to be known), while when is high, Algo. 3 is preferred.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 21 / 22
Summary and Discussion (contd.)
Each of our algorithms work by equating empirical values with
their corresponding theoretically established ones.
The caveat is that the theoretical results are for an average
network while in practice, we have only a single realization of
the node distribution at hand. Thus, in general, the estimates
we obtain are biased.
This also explains the fact that the performance of Algorithms 2
and 3 is better at lower m.
We remark that the bias (and the MSE) can be signicantly
lowered if nodes have access to several independent realizations
of the point process or if they are allowed to communicate.
Sunil Srinivasa and Martin Haenggi () University of Notre Dame IT School 2009 22 / 22