EFFICIENT PLACEMENT OF WIRELESS BASE-STATIONS IN
URB
AN ENVIRONMENT
Mansour A. Aldajani
Center for Communications and Computer Research, KFUPM, Dhahran, Saudi Arabia
K
eywords:
Wireless Communications, Cell planning, Base Station, Placement, Convolution.
Abstract:
This work proposes a novel approach for placing wireless base-stations in urban environments. The new
approach solves the placement problem using 2-D convolution. Convolution searches for the best locations of
base stations based on highest consumption criteria. Convolution theorem is then used to substantially cut the
computation load of the proposed approach. The approach allows simple user-interface and arbitrary demand
and supply patterns of power. Simulations show that the new approach can be used to efficiently model and
place wireless base-stations.
1 INTRODUCTION
Placing wireless base-stations is a complex process
which usually involves many parameters and requires
long time to solve. Modeling the placement problem
using conventional optimization techniques is com-
plex and need expert knowledge of the considered
technique. Moreover, the solution approach usually
take long time for moderate grid-size resolution.
One of the first studies to solve the placement
problem was in (Sherali and Rappaport, 1996). In
this study, the solution of single as well as multiple
transmitters problem was considered. The problem
was modeled as a nonlinear program and then three
nonlinear optimization algorithms were considered to
solve this model. The work in (Hao Q., 1997) for-
mulated the placement problem as a large-scale com-
binatorial optimization model. The model is then
solved using the simulated-annealing approach. The
Hata’s propagation model (Hata, 1980) was used to
determine the transmission loss in this study. Similar
model was developed in (Calegari, 1997). The model
in this work was solved using generic algorithms re-
sulting in sub-optimal solutions. The work in (Park
and Park, 2002) considered the determination of both
BSs placement as well as transmission power. A sim-
ple weighted objective function was established. The
real-coded generic algorithm was then used to obtain
the solution. The solution takes into account the inter-
ference situation to determine the appropriate trans-
mission power. Transmission power in wireless sys-
tems can also be adjusted via well-developed power
control techniques (Aldajani and Sayed, 2003).
In this work, we present a new approach for plac-
ing wireless base stations in urban environment. The
approach uses convolution as a core process to come
up with a minimum number of base-stations such that
the minimum coverage level is achieved. Fast algo-
rithms for computing the convolution are then used to
substantially reduce the computation load.
2 PROBLEM FORMULATION
The objective of the placement problem is to mini-
mize the total number of base-stations N such that the
net power inside a 2-D Euclidian space Γ is at least
equal to the power threshold α at all locations. In
other words,
p(x, y) α x, y Γ (1)
where p(x, y) is the net power at the point with coor-
dinates (x, y).
To find an expression for p(x, y), let us define the
quantity s
n
(x, y) as the power supplied by the n
th
BS
to the mobile station at location (x, y). This quantity
33
A. Aldajani M. (2007).
EFFICIENT PLACEMENT OF WIRELESS BASE-STATIONS IN URBAN ENVIRONMENT.
In Proceedings of the Second International Conference on Wireless Information Networks and Systems, pages 33-38
DOI: 10.5220/0002145200330038
Copyright
c
SciTePress
indicates simply the signal strength at location (x, y)
due to station n. This quantity is dependent mainly on
the radio propagation and path loss model of the trans-
mitter antenna. For example, for omni-directional an-
tenna in free-space, the quantity s
n
(x, y) is given by
the well-known Friis equation (Janaswamy, 2000)
s
n
(x, y) =
p
n
G
t
G
r
λ
4πh
n
(x, y)
2
(2)
where
p
n
, G
t
, G
r
, and λ are the transmission power,
transmitter gain, receiver gain, and wave length re-
spectively. h
n
(x, y) is the distance between the point
(x, y) and the base-station n.
Let us also introduce the term d(x, y) to represent
the demand level at point (x, y). This quantity can
be used to model different priorities of coverage and
extra signal attenuations at location (x, y).
Then we can define the net power p at location
(x, y) as follows
p(x, y) = max
n=[1,N]
{s
n
(x, y)} d(x, y). (3)
We assume here that the mobile station at location
(x, y) will connect to the base station that delivers the
maximum signal power.
The objective of the placement problem is to find
the minimum number of base-stations and their loca-
tions that will satisfy the power constraint (1) where
p(x, y) is given by (3).
2.1 Discretization of the Model
The variables p(x, y), s
n
(x, y), and d(x, y) are dis-
cretized in 2-D Euclidian space to form the matrices
P, S
n
, and D respectively. Therefore, the optimization
problem can be written in matrix format as
min N (4)
subject to
P
N
= max
n=[1,N]
{S
n
} D α (5)
where P
N
is the power pattern matrix of size (I × J)
after assigning N base-stations, S
n
is the power sup-
ply matrix of the n
th
BS, and D is the demand pattern
matrix. Notice that this constraint states that all the
elements of the matrix P
N
should be greater than the
power threshold α.
The matrix S
n
can be broken down into the convo-
lution of two matrices as follows
S
n
= X
n
A (6)
where the symbol indicates the 2-dimensional con-
volution. The matrix A of size (I
A
× J
A
) is a fixed
propagation pattern matrix of the transmitter radio an-
tenna. The matrix X
n
indicates the location of base-
station n. If we denote this location by the coordinates
(u
n
, v
n
) then X
n
has all its elements equal to zero ex-
cept at (u
n
, v
n
) where it equals to “1”. In other words,
X
n
(i, j) =
(
1 at (u
n
, v
n
)
0 elsewhere.
(7)
Expression (6) means simply shifting the elements of
the A matrix by (u
n
, v
n
).
Notice that minimizing the number of base-
stations N is equivalent to minimizing the summation
norm of the location matrices X
n
for all base-stations.
In view of this fact, the optimization problem can fi-
nally be written as
min k
N
n=1
X
n
k (8)
subject to
P
N
= max
n=[1,N]
{X
n
A} D α. (9)
3 SOLUTION OF THE
PLACEMENT PROBLEM
A flow chart of the proposed algorithm is shown in
Fig. 1. To determine the amount of power consump-
tions associated with placing a BS at a certain grid
point, the antenna propagation matrix A is convolved
with the existing power pattern P
n1
that resulted
from previously assigned BSs, i.e.,
Y
n
= A P
n1
P
0
= D. (10)
The role of the convolution here is as follows. For
each point on the current power pattern P
n1
, the an-
tenna propagation A is centered at that point and dot-
multiplied with the intersecting sector of P
n1
. The
multiplication values are then summed up and the an-
swer is stored at the corresponding point in Y
n
. This
convolution process is repeated for all other points in
P
n1
.
The coordinates that correspond to the minimum
value of the matrix Y
n
indicates the highest consump-
tion. This point is chosen as the location of the n
th
base-station,
(u
n
, v
n
) = argmin
(i, j)
Y
n
. (11)
Once a new base-station location is chosen, the lo-
cation matrix X
n
is constructed from (7). The power
matrix is then updated as follows
P
n
= G
n
D, n = 1, 2, ...., N. (12)
WINSYS 2007 - International Conference on Wireless Information Networks and Systems
34
Given A and D
Set n=1, P
0
= -D , and G
0
=0
Calculate the power contributions
Y
n
=A P
n-1
Place the n
th
BS at the minimum
value of
Y
n
;
(u
n
,v
n
)=
argmin
{Y
n
}
Compute the net power
P
n
= G
n
- D
All area
covered?
p
min
(
n) >
Return N, Locations: { u,v }, and p
min
YES
NO
Construct the location matrix X
n
X
n
(i,j ) = 1 at (u
n
,v
n
) and 0 Elsewhere
n=n+1
N=n
Compute the accumulated power
G
n
=max{ G
n-1
, A X
n
}
Figure 1: The proposed solution algorithm.
where G
n
is the power pattern supplied by the stations
1 to n. This matrix can be computed iteratively from
G
n
= max {G
n1
, S
n
} = max {G
n1
, AX
n
}, G
0
= 0
(13)
3.1 Penalizing Boundaries of the
Demand Grid
Since the design space is always provided as a con-
fined rectangular region, the demand matrix D need to
be surrounded by a negative frame value w as shown
in Fig. 2 to penalize the boundaries of the grid. The
purpose of the penalty w is to push the locations of
the BSs inward and therefore increase the coverage
efficiency. In case of a tie, the algorithm will pick the
value that results in higher p
min
. In this way, not only
the number of stations will be minimized but also the
minimum power will be maximized reflecting an im-
proved over-all coverage.
D =
w w w w w w w
w 5 5 0 0 0 w
w 5 5 0 0 0 w
w 5 0 +10 +10 +5 w
w 0 0 +10 +10 +5 w
w 0 0 +10 +10 +5 w
w w w w w w w
Figure 2: Numerical example of the demand matrix D sur-
rounded by the penalty frame value w.
3.2 Verification of the Solution
Since analytical solution for the placement problem
given by (8) and (9) is not available, we follow two
numerical approaches to verify the proposed solution.
First, the algorithm is implemented on simple models
where solutions are known and the results are then
compared (Park and Park, 2002). Second, solution
is verified by performing an exhaustive search on all
possible locations.
4 SIMULATIONS
Matlab was used to implement the algorithm on a
2.1GHz personal computer with 256MB of memory.
The Matlab program provides a friendly User Inter-
face (UI). It inputs a color-coded map, similar to that
of Fig. 3, in a common image format (JPEG) and then
constructs the corresponding demand pattern matrix
D. It also inputs the propagation pattern matrix A. It
then computes the number of base-stations and their
locations and then shows them on the color-coded im-
age. The program also returns the final minimum
power p
min
, and the percentage coverage (PC) of each
assigned base-station.
In our simulations, the size of the matrices D and
A is fixed to 41 × 61 for each (corresponds to 2501
possible locations). Furthermore, the power threshold
is arbitrarily fixed in all simulations to the normalized
value α = 1%.
To test the algorithm, we considered the place-
ment problem in (Park and Park, 2002). In this case,
a configuration of seven hexagonal cells is to be cov-
ered with omni-directional antennas having the same
radius as that of the cells. The solution for this prob-
lem is obvious: exactly seven BS are needed which
EFFICIENT PLACEMENT OF WIRELESS BASE-STATIONS IN URBAN ENVIRONMENT
35
Red: No-Demand
(Avoid)
White: Normal Demand
(Urban area, Campuses, schools, etc.)
Green: Highest Demand
(Malls, Business district,
Large population, etc.)
Blue: High Demand
(Highways, roads, etc.)
Figure 3: Example of designing the demand levels on a real
map using color codes.
Table 1: Example of color codes and their corresponding
values inside the demand matrix D .
Color Demand Value inside D
Green Highest +20
Blue High +10
White Normal 0
Red No Demand 20
should be located at the centers of the cells. To imple-
ment the proposed approach, the edges of the seven
cells are drawn using popular drawing software and
then feeded directly to the algorithm. The edges are
mapped as negative values in D. The results are
shown in Fig. 4. The algorithm achieved 99.7% cov-
erage in seven iterations. Solving the same problem
with Genetic Algorithm (GA), for example, would
need more than 1000 generations to get the same cov-
erage (Park and Park, 2002). Furthermore, the fact
that the model can be built by simply drawing the cell
boundaries and feeding the drawing to the algorithm
makes the proposed approach much more attractive
when compared to the cumbrous modeling process
demanded by the GA.
In another experiment, the color-coded map of
Fig. 3 is used in simulation to build a different non-
trivial demand matrix D. The numerical weights as-
signed to the four colors in this example are listed in
Table 1. An omni-directional propagation matrix A
is used here. Fig. 5 shows the resulting placement
0 10 20 30 40 50 60 70 80
0
10
20
30
40
50
60
70
Figure 4: Solution of the hexagonal cells problem.
of the base-stations. In this case, six base-stations
were sufficient to meet the coverage requirement. No-
tice that as expected, the first base-station was lo-
cated at the green region (corresponds to very high
demand). Also, the algorithm avoided the placement
of any base-station at the red region (correspond to
no-demand region). The minimum power returned by
the algorithm is p
min
= 1.0724 which is just above the
required power threshold α = 1. The optimal frame
value w
in this example is -86. The percentage cover-
age and accumulated percentage coverage for this ex-
ample are shown in Fig. 6. This information can play
very useful role for cell planners. Through this infor-
mation, they have a choice to eliminate those stations
on the map that have negligible coverage. The second
base-station covered about 40% of the area while the
6th one covered about 1% only. This means that if
99% total coverage is sufficient, then the 6th station
can simply be removed. The algorithm returned the
results in less than 2 minutes.
5 COMPUTATION COMPLEXITY
From the discussions above, the proposed scheme has
an outer loop as well as an inner loop. The outer loop
searches for the optimal frame value w
while the in-
ner loop implements Fig. 1 to find the location of the
base-stations.
For the outer loop, a simple line search was found
sufficient to find w
. The search is limited to the inte-
ger values in the range [w
min
, 0]. Still, more efficient
search algorithm could be adopted to find this value.
In the inner loop represented by Fig 1, the only
WINSYS 2007 - International Conference on Wireless Information Networks and Systems
36
1
2
3
4
5
6
Figure 5: Result of the placement problem with omni-
directional antenna and a given demand map.
1 2 3 4 5 6
0
10
20
30
40
50
PC %
1 2 3 4 5 6
20
40
60
80
100
APC %
Base−stations
Figure 6: Percentage coverage (PC) and accumulative per-
centage coverage (APC) for the base-stations in Fig. 5.
computationally expensive operation is the convolu-
tion A P
n1
. As mentioned in the motivation sec-
tion, the cost of the convolution can be reduced sub-
stantially from m
2
to mlog(m) using available fast
convolution techniques (Bernardini and Mian, 1996;
Chiang and Chew, 2005). This feature makes the pro-
posed solution feasible even for large grid-sizes.
Table 2 summarizes the order of complexity for
the proposed approach. In Fig. 7, we show the
simulated and theoretical time needed to assign one
base-station using fast convolution, averaged over 100
runs.
Furthermore, the sparsity of the matrix P
n
can be
utilized to reduce the computations of the convolu-
tion. This matrix is usually full of zeros. Finally,
the search space reduces as new stations are assigned.
This can also reduce the computation load signifi-
cantly.
0 0.5 1 1.5 2
x 10
4
0
0.5
1
1.5
2
2.5
3
Map grid size m=I × J
Time to assign single station (Sec.)
Actual
Theoritical fit
Figure 7: Simulated and theoretical time needed to assign
single station as a function of the grid size m = I × J, us-
ing fast convolution. The theoretical plot is obtained from
βmlog(m), where β is a processing constant.
Table 2: Order of complexity for the proposed approach.
Method used Order of Complexity
for convolution
Direct m
2
Fast mlog(m)
6 CONCLUSION
In this work, we proposed a new approach for placing
the wireless base-stations. The new approach simpli-
fies modeling and solution of the placement problem.
Modeling the problem is performed by drawing color-
codes on the map. Solution is obtained through the
convolution process which searches for the highest
consumption areas. Convolution theorem is then used
to substantially reduce the computation load. Simula-
tions of the proposed approach showed its efficiency
and flexibility in solving the placement problem.
ACKNOWLEDGEMENTS
The author would like to acknowledge the support
of King Fahd University of Petroleum and Minerals,
Dhahran.
REFERENCES
Aldajani, M. A. and Sayed, A. H. (Nov. 2003). Adaptive
predictive power control for the uplink channel in ds-
cdma cellular systems. In IEEE Transactions on Ve-
hicular Technology. Vol. 52, No. 6, pp. 1447 –1462.
EFFICIENT PLACEMENT OF WIRELESS BASE-STATIONS IN URBAN ENVIRONMENT
37
Bernardini, R., C. G. and Mian, G. A. (Nov. 1996). A new
multidimensional fast convolution algorithm. In IEEE
Transactions on Signal Processing I. Vol. 44, No. 11,
pp. 2853 -2864.
Calegari, P., et al.. (May 1997). Genetic approach to radio
network optimization for mobile systems. In Vehicu-
lar Technology Conference (97). Vol. 2, pp. 755 - 759.
Chiang, I. and Chew, W. C. (July 2005). Fast real-time con-
volution algorithm for microwave multiport networks
with nonlinear terminations. In IEEE Transactions on
Circuits and Systems II: Express Briefs. Vol. 52, No.
7, pp. 370 - 375.
Hao Q., et al.. (Sep. 1997). A low-cost cellular mobile com-
munication system: a hierarchical optimization net-
work resource planning approach. In IEEE Journal
on Selected Areas in Communications. Vol. 15, No. 7,
pp. 1315 - 1326.
Hata, M. (Aug. 1980). Emperical formula for propagation
loss in land mobile radio service. In IEEE Trans. Ve-
hicular Technology. Vol. VT-29, pp. 317-325.
Janaswamy, R. (2000). Radiowave propagation and smart
antennas for wireless communications. Kluwer Aca-
demic Publishing, Massachusetts, USA.
Park, B., Y. J. and Park, H. (Sep. 2002). The determination
of base-station placement and transmit power in an
inhomogeneous traffic distribution for radio network
planning. In Vehicular Technology Conference (02).
Vol. 4, pp. 2051 - 2055.
Sherali, H., P. C. and Rappaport, T. (May 1996). Optimal
location of transmitters for micro-cellular radio com-
munication system design. In IEEE Journal on Se-
lected Areas in Communications. Vol. 14, No. 4, pp.
662 - 673.
WINSYS 2007 - International Conference on Wireless Information Networks and Systems
38