Multidimensional MUSIC DOA Estimation Using

Ant Colony Optimization Algorithm

Yameng Jiao, Jianguo Huang, Yunshan Hou
College of Marine, Northwestern Polytechnical University, 710072, Xian, China
jiaoyameng@mail.nwpu.edu.cn
AbstractThe estimation of the direction-of-arrival (DOA) of
signals is an important issue in array signal processing. A
multidimensional version of MUSIC (MD-MUSIC) algorithm is a
well known excellent algorithm for DOA estimation. However,
this algorithm is totally impractical for its prohibitive
computational burden incurred by multi-dimensional grid search.
In order to solve this problem, ant colony optimization (ACO) is
introduced to combine with the MD-MUSIC algorithm, and a
new algorithm with lower computational burden called ACOMD-
MUSIC is proposed. In this proposed algorithm, a kind of
continuous pheromone probability distribution (Gaussian kernel
probability density function) in the sampling process is used. The
global maximum of MD-MUSIC spatial spectrum function could
be reached after reasonable iterations. Simulation results
illustrate that the proposed algorithm provides similar
performance as MD-MUSIC algorithm, while its computational
cost is only 1/13 of MD-MUSIC algorithm.
Keywords-array signal processing; DOA estimation; multiple
MUSIC algorithm; ant colony optimization; computational
complexity
I. INTRODUCTION
In order to solve the classical narrow-band multiple source
location problems, a lot of array signal processing algorithms
have been devised. Most of the researches have focused on the
algorithms for high-resolution DOA estimation. Subspace
based methods are known to have high resolution capabilities
and yield accurate estimates. The introduction of multiple
signal classification (MUSIC) algorithm
[1] [2]
, which requires a
one-dimensional search, was an attempt to more fully exploit
the underlying data model. MUSIC provides a nice geometric
interpretation of the DOA problem and has received much
attention. The asymptotic properties of MUSIC have been well
documented in [3][4]. A multidimensional version of MUSIC
(MD-MUSIC) has been discussed in [1][5] when dealing with
the problems of correlated signals and finite sample bias.
Unfortunately, this results in a highly nonlinear
multidimensional optimization problem. This algorithm
becomes totally impractical because of the complexity and the
computational burden for higher dimension.
Ant colony optimization (ACO) algorithm was initially
developed to be a metaheuristic algorithm for solving
combinatorial optimization problems (COPs) in 1992 by
Dorigo, M.
[6]
. After the initial proof-of-concept application to
the traveling salesman problem (TSP)
[7]
, ACO was applied to
many other COPs such as assignment problems
[8]
, scheduling
problems
[9]
, and vehicle routing problems
[10]
etc. However,
there exists an important class of problems for which this was
not the class of optimization problems that require choosing
values for continuous variables. Dorigo, M. presented an
extension of ACO algorithm in the continuous domain in
2008
[11]
. It has been found that ACO algorithm could be
adapted to solve continuous optimization problem and obtained
the approximate solutions in a reasonable iteration number.
Based on above reasons, this paper suggests the using of
the ACO algorithm to find the optimal parameters of the MD-
MUSIC algorithm. Then, the ACOMD-MUSIC algorithm is
proposed by extending the pheromone probabilistic choice to
pheromone Gaussian kernel probability density function in the
sampling process. This algorithm not only obtains the global
optimum value of highly nonlinear multidimensional problem
but also reduces the computation complexity obviously. The
simulation results show that the algorithm provides similar
performance to that achieved by the MD-MUSIC algorithm,
but its computational cost is only 1/13 of MD-MUSIC.
This paper is organized as follows. In Section the array
signal model and the principle of the MD-MUSIC algorithm is
described. A theoretical development of the ACOMD-MUSIC
DOA estimation is presented in Section . Simulation results
are given in Section that illustrate the efficacy of proposed
algorithm in reducing the computation burden. Finally, the
conclusions are given in Section .
II. SIGNAL MODEL AND MD-MUSIC ALGORITHM
Consider an array consisting M sensors, D far field
narrow-band source signals arrive from
directions , where D<M, is the direction
to the i th source measured relative to the array normal vector.
Background noise is Gaussian white noise with zero-mean and
variance . Noise is uncorrelated with source signals.
Assume that wave propagating medium is isotropic and
homogeneous and a wave field generated by sources in the
presence of noise is sampled temporally and spatially by the
array, then the M1array signal vector can be expressed as
T
D
] [
2 1
" = O
i

) ( ) ( ) ( ) ( t t t n s A x + = (1)
where is the MD array manifold which is consisting of
D steering vectors , describe the transfer function
between DOA and the output of the array. is the D1
source signal vector. is the M1 additive noise vector.
) ( A
) (
i
a
) (t s
) (t n
lCSP2010 Proceedings

291
___________________________________
978-1-4244-5900-1/10/$26.00 2010 IEEE

The array output is sampled at K time instances and these
snapshots are collected in the columns of an MK data matrix
) (k X
) ( ) ( ) ( ) ( k k k N S A X + = (2)
Therefore, the array output is complex Gaussian with zero
mean and covariance matrix R. In practical engineering the
estimated covariance matrix R

is defined as
_
=
=
K
k
H
k k
K
1
) ( ) (
1
X X R

(3)
For the asymptotic analysis, it is assumed that the array
manifold vectors are twice continuously differentiable with
bounded second derivatives in a neighborhood of the true
DOAs. To enable unique identification of the signal
parameters, the following condition is imposed on the array
response:
s
U T A = O) ( (4)
where is a full rank matrix, is the signal subspace.
Moreover, it is clear that the signal subspace is a subset of the
range space of . The MD-MUSIC algorithm fits the
subspace spanned by to the measurements in a
least squares sense. We define the MD-MUSIC problem by the
following equation:
T
s
U
) ( A
) ( A ) (k X
2
) ( min arg ) (
F
S

MUSIC MD
T A U T ,
T ,

O =

(5)
where
2
F
- denotes Frobenius norm.
By substituting the pseudoinverse solution,
, into (5), we can obtain the
following equivalent problem:
S S
H H
U A U A A A T
+
= =
1
) (

} tr{ max arg

H
S S
U U P
A

MUSIC MD
=

(6)
where is a projection matrix that
projects into the column space of A.
+
= = AA A A A A P
H 1 H
) (
A
III. ACOMD-MUSIC ALGORITHM
A. ACO Algorithm for Continuous Domain
The inspiring source of ACO is the foraging behavior of the
real ants. The capability of the real ant colonies has inspired the
definition of the artificial ant colonies that can find
approximate solutions to hard combinatorial optimization
problems. The principal component of ACO algorithms is the
pheromone model, which is used to probabilistically sample
the search space
[6]
. The idea that is central to the way ACO
works is the incremental construction of solutions based on the
biased (by pheromone) probabilistic choice of solution
components.
The fundamental idea of ACO for the continuous problem
is the shift from using a discrete probability distribution to
using a Gaussian kernel probability density function (PDF). In
ACO for the continuous problem, instead of choosing a
component, an ant samples a Gaussian kernel PDF. We define
a Gaussian kernel as a weighted sum of several one-
dimensional Gaussian functions , and denote it as : ) ( g
i
l
) ( G
i
2
2
2
1 1
2
1
) ( ) ( ) (
i
l
i
l

i
l
L
l
l
i
l
L
l
l
i
e

w g w G
) (
) (

= =
t
= =
_ _
(7)
Since we use as many Gaussian kernel PDFs as the number of
the targets of the problem, i=1,, D identifies a single such
PDF. The is parameterized with three vectors of
parameters: w, and . L is the number of sampling of the
search space.
) (O
i
G
i

B. The Steps of the ACOMD-MUSIC Algorithm

The steps of the MD-MUSIC algorithm based on ACO are
given as follows:
Step1: Global sampling of the search space
Sampling the search space with some solutions uniformly
distributed in a given direction search range . For each
solution to a D-dimensional problem, we use the values of
its D variables to calculate the value of the objective function
according to (6). Each solution has an associated
weight, which is associated with the individual Gaussian
functions. It is calculated according to the following formula:
] , [
h l

l

) (
l
f
2 2
2
2
1
2
1
L q
l
l
e
qL
w
) (
t
=
(8)
where the parameter q adjusts the balance between the
iteration-best and the best-so-far pheromone update, Then, the
solutions are sorted in descending order in the archive
according to . Finally, we store all of these in an archive
presented in Figure 1.
) (
l
f
Figure 1. Archive
Step2: Updating the chosen solution
The probability for each ant of choosing the l th
Gaussian function is given by:
l
p
_
=
=
L
r
r
l
l
w
w
p
1
(9)
The choice of the l th Gaussian function is done only once per
ant, per iteration. An ant uses the Gaussian function associated
with the chosen solution . is the means, that is, the values
i
l

292
of the i th variable of the chosen solution, and is the
standard deviations, that is, the average distance from the
chosen solution to other solution in the search space, and we
multiply it by the parameter :
i
l

) 0 ( >
_
=

=
L
m
i
l
i
m
i
l
L


1
1
(10)
where is the i th variable of l th solution. Then we update
the chosen solution according to (7).
i
l

Step3: Updating the search space according to the chosen

solution
For the chosen solution , we calculate the value of
objective function . If , then replace the
last solution
L
with the chosen solution . Finally, the
solutions are sorted in descending order again in the new
archive according to the new . Otherwise, there is no
change in the archive.

) ( f ) ( ) (
1
f f >
) (
l
f
Step4: Judging the condition of iteration and obtaining the
DOA estimation
If

b a
I b a
<
e
} {
,
1 1
max
, I is the iteration number. Then,
the mathematical expectation of the solutions of last five
iterations is calculated to be the final DOA estimation.
Otherwise, go to step 2.
IV. SIMULATION RESULTS
A. Simulation Model
To illustrate the performance of the proposed algorithm we
consider a twelve-sensor ULA with sensor spacing of half-
wavelength. 100 Monte Carlo experiments are carried out for
each scenario. Two uncorrelated narrow-band source in far-
field impinge on the receiving array from -3 and 3
respectively. The sampling frequency is 120 kHz.
B. Performance Analysis
Figure 2 shows the convergence of ACOMD_MUSIC
algorithm when SNR is 10dB. 2 = T 0, 10 = L 0.1 = q
are chosen in
this example. It is obviously shown that the proposed algorithm
reaches the global maximum of MD-MUSIC spatial spectrum
function after reasonable iterations.
0.01 = 0.001 = ] ,60 60 [ ] , [
D D
=
h l

Figure 3 shows the spatial spectrum of MD-MUSIC
algorithm. Figure 4 shows the resolution probability of ML,
MD-MUSIC, and ACOMD-MUSIC versus SNR. The SNR
range is from -30dB to 30dB. The searching step of ML
algorithm is 0.5e. MSE of ML, MD-MUSIC and ACOMD-
MUSIC for the sources at -3 and 3 versus SNR are compared
in Figure 5 and Figure 6. It is obviously shown that the ML
algorithm has the highest resolution probability and lowest
MSE when SNR is blew -5dB. The proposed algorithm
provides similar performance to that achieved by the MD-
MUSIC algorithm.
200 400 600 800 1000 1200 1400 1600 1800
-10
-8
-6
-4
-2
0
2
4
6
8
10
iteration times
D
O
A
(

)
DOA of signal1
DOA of signal2
Figure 2. Result of optimization
Figure 3. The spatial spectrum of MD-MUSIC algorithm
-30 -20 -10 0 10 20 30
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR(dB)
R
e
s
o
lu
t
io
n

p
r
o
b
a
b
ilit
y
ML
MD-MUSIC
ACOMD-MUSIC
Figure 4. Resolution probability
-30 -20 -10 0 10 20 30
0
2
4
6
8
10
12
14
SNR(dB)
M
S
E
(

)
ML
MD-MUSIC
ACOMD-MUSIC
Figure 5. MSE of signal1
293
-30 -20 -10 0 10 20 30
0
2
4
6
8
10
12
14
SNR(dB)
M
S
E
(
)
ML
MD-MUSIC
ACOMD-MUSIC
Figure 6. MSE of signal2
C. Computational complexities analysis
To compare the computational complexities we consider a
twelve-sensor ULA, and the parameter values are the same as
that of the previous simulations. 100 Monte Carlo experiments
are carried out for each scenario. TABLE shows the mean
iteration number of ACOMD-MUSIC associated with the SNR.
TABLE I. THE MEAN NUMBER OF ITERATION
SNR(dB)
-10 0 10
Mean iteration number 2066.8 1608.7 1210.3
Assume the computational complexity of one spectral peak
searching is . Then we can obtain the computational
complexities of MD-MUSIC algorithm and ACOMD-MUSIC
algorithm denoted by (11) and (12).
A
A =
D
l h
a J ) / ) ((
MD_MUSIC
(11)
A + ~ ) (
C ACOMD_MUSI
L I T J (12)
where a is the step of searching, D is the number of signals,
T, I and L is the number of ants, mean iteration number and
sampling number of the searching space respectively.
Form TABLE, we can calculate the computational
complexity by (11) and (12) respectively. For example,
and when
SNR is -10dB. It is obviously shown that the computational
complexity of ACOMD-MUSIC is only 1/13 of MD-MUSIC.
From (11) (12), we can see that it may lead to exponential
growth of the computational complexity of the MD-MUSIC
algorithm with increasing the signal number, the proposed
algorithm has obvious advantage in the computational
complexity. From TABLE, we can also see that the
proposed algorithm has more advantage in the computational
complexity than the MD-MUSIC algorithm because its
iteration number decreases with the increase of the SNR.
A = 57600
MD_MUSIC
J A ~ 4233.6
C ACOMD_MUSI
J
V. CONCLUSIONS
In view of the MD-MUSIC algorithm is totally impractical
for its prohibitive computational burden incurred by multi-
dimensional grid search, a new algorithm with lower
computational burden called ACOMD-MUSIC is proposed in
this paper to solve this problem. Simulation results illustrate
that the proposed algorithm is more suitable to engineering
application because it provides similar performance to that of
MD-MUSIC algorithm, while its computational cost is only
1/13 of MD-MUSIC algorithm.
ACKNOWLEDGMENT
This work is partially supported by National Natural
Science Foundation of China (60972152), Aviation Science
Fund (2009ZC53031), National Laboratory Foundation of
China (9140C2304080607), and NPU Foundation for
Fundamental Research (NPU-FFR-W018102).
REFERENCES
[1] R. O. Schmidt, A signal subspace approach to multiple emitter location
and spectral estimation, Ph.D. dissertation, Stanford Univ., Stanford,
CA, Nov. 1981.
[2] G. Bienvenu and L. Kopp, Adaptivity to background noise spatial
coherence for high resolution passive methods, in Proc. IEEE ICASSP
(Denver, CO), 1980, pp. 307-310.
[3] P. Stoica and A. Nehorai, MUSIC, maximum likelihood and Cramer-
Rao bound, in Proc. ICASSP 88 Conf. (New York, NY), Apr.1988, pp.
2296-2299.
[4] H. Clergeot, S. Tressens, and A. Ouamri, Performance of high
resolution frequencies estimation methods compared to the Cramer-Rao
bounds, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-37,
pp. 1703-1720, Nov. 1989.
[5] J. A. Cadzow, A high resolution direction-of-arrival algorithm for
narrow-band coherent and incoherent sources, IEEE Trans. Acoust.,
Speech, Signal Processing, vol. ASSP-36, pp. 965-979, July 1988.
[6] M.Dorigo, Optimization, Learning and Natural Algorithms (in Italian),
Ph.D. thesis, Dipartimento di Elettronica, Politccnico di Milano, Italy,
1992.
[7] M.Dorigo, L.M.Gambardella, Ant Colony System: A cooperative
learning approach to the traveling salesman problem, IEEE
Transactions on Evolutionary Computation, 1(1), pp.53-66,1997.
[8] D.Costa, A.Hertz, , Ants can colour graphs, Journal of the Operational
Research Society 48, pp.295-305, 1997.
[9] C.Gagne, W.L.Price, M.Gravel, Comparing an ACO algorithm with
other heuristics for the single machine scheduling problem with
sequence-dependent setup times, Journal of the Operational Research
Society 53, pp. 895-906,2002.
[10] L.M.Gambardclla, E.Taillard, G. Agazzi, MACSVRPTW: A multiple
ant colony system for vehicle routing problems with time windows, In:
Corne, D., Dorigo, M., Glover, F. (Eds.), New Ideas in Optimization.
McGraw-Hill, UK, pp. 63-76, 1999.
[11] K. Socha, M. Dorigo. Ant colony optimization for continuous
domains, European Journal of Operational Research 185, 2008,
pp.1155-1173
294