Sie sind auf Seite 1von 5

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2017.2671521, IEEE
Transactions on Circuits and Systems II: Express Briefs
1

Maximum Versoria Criterion-Based Robust


Adaptive Filtering Algorithm
Fuyi Huang, Jiashu Zhang, and Sheng Zhang

AbstractUsing the generalized Gaussian probability density Non-Gaussian interferences (including heavy-tailed
function (GPDF) as the kernel, a generalized correntropy has been non-Gaussian interferences and light-tailed non-Gaussian
proposed. And then a generalized maximum correntropy interferences) often occur in practical applications.
criterion (GMCC) algorithm is developed by maximizing the
Heavy-tailed (e.g., Laplace, -stable, etc. distributions)
generalized correntropy. However, the GMCC algorithm has a
high steady-state misalignment and involves a high calculation non-Gaussian (impulsive) interferences produce more outliers
cost of the exponential term (generalized Gaussian kernel). In this than those assumed by Gaussian models. The characterization
letter, we propose a maximum Versoria criterion (MVC) of a non-Gaussian signal by its second order moment is no
algorithm, which is derived by maximizing the generalized longer optimal and many studies have shown that lower order
Versoria function, to reduce steady-state misalignment and statistic can lead to improved robustness against non-Gaussian
computational effort as compared to the GMCC algorithm. The impulsive interferences. In such situations, the sign algorithm
MVC algorithm is then tested in system identification and (SA) [6], the affine projection sign algorithms (APSAs) [7][9],
acoustic echo cancellation scenarios, which have demonstrated and the continuous mixed p-norm (CMPN) algorithm [10] yield
that the proposed algorithm is robust against non-Gaussian
improved robustness against non-Gaussian impulsive noises.
impulsive noises and performs much better than the LMP and
GMCC algorithms. Using higher order moment algorithms to deal with light-tailed
(e.g., uniform, binary, etc. distributions) non-Gaussian
Index TermsAdaptive filtering, Gaussian probability density interferences are more suitable. For this situation, the least
function, generalized maximum correntropy criterion (GMCC), mean fourth (LMF) algorithm yields improved performance in
maximum Versoria criterion (MVC). terms of the convergence rate and steady-state misalignment,
as compared to the LMS algorithm [11].
Recent years, correntropy has been successfully applied in
I. INTRODUCTION non-Gaussian signal processing [12], [13]. The correntropy is a

T HE minimum mean square error (MMSE) criterion is


widely used in adaptive filtering due to its low
computational complexity and simplicity [1]. Based on the
nonlinear similarity measure between two arbitrary scalar
random variables. Meanwhile it has a close relationship with
M-estimation [12]. Thus, the maximum correntropy criterion
MMSE criterion, the least mean square (LMS) algorithm as a (MCC) algorithm is robust against non-Gaussian impulsive
method optimized for Gaussian models was developed. The noises. Lately, a generalized correntropy has been proposed in
LMS algorithm is the most popular adaptive filtering algorithm. [14]. Then, an adaptive filtering algorithm, called generalized
To obtain improved performance as compared to the LMS maximum correntropy criterion (GMCC) algorithm, was
algorithm, the variable-step-size LMS (VSS-LMS) and developed by maximizing the generalized correntropy [14].
normalized LMS (NLMS) algorithms have been developed in The disadvantages of the GMCC algorithm are high
[2] and [3], respectively. Furthermore, in order to offer steady-state misalignment, and high calculation cost of the
improved convergence performance and reduce computational exponential term (generalized Gaussian kernel) in the update
complexity relative to the NLMS algorithm, the nonparametric equation of the weight vector. The Versoria function has been
VSS-NLMS (NPVSS-NLMS) and set-membership NLMS successfully applied in the particle swarm optimization (PSO)
(SMNLMS) algorithms were proposed in [4] and [5], algorithm for IIR adaptive filtering [15]. The PSO algorithm
respectively. Unfortunately, these algorithms developed under uses a modified Versoria function which is employed to
the MMSE criterion do not perform well in nonlinear and provide an adaptive inertial weight factor instead of the
non-Gaussian interference situations, especially in the presence Sigmoid function for avoiding the exponential computation and
of non-Gaussian impulsive noises. ensuring the low steady-state misalignment. The Versoria
function in the PSO plays a similar role of variable step-size.
Unlike the traditional usage of Versoria, we use the Versoria as
This work was supported in part by the National Natural Science Foundation the cost function in this letter. Thus, a maximum Versoria
of P. R. China under Grants 61271341 and 61671392.
F. Huang and J. Zhang are with Sichuan Province Key Laboratory of Signal
criterion (MVC) algorithm is developed by maximizing the
and Information Processing, Southwest Jiaotong University, Chengdu generalized Versoria function, which reduces the steady-state
611756, P. R. China (e-mail: fukngaiwong@163.com; misalignment and the computational effort as compared to the
jszhang@home.swjtu.edu.cn). GMCC algorithm. The performance of the proposed MVC
S. Zhang is with the School of Computing, Engineering and Mathematics, algorithm is tested in system identification and acoustic echo
University of Western Sydney, Sydney, NSW 2751, Australia (e-mail:
Dr.s.zhang@ieee.org). cancellation scenarios that include non-Gaussian impulsive

1549-7747 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2017.2671521, IEEE
Transactions on Circuits and Systems II: Express Briefs
2

interferences. Simulation results have confirmed that the same steady-state error than do the GPDF-basd algorithm; in
proposed MVC algorithm is robust against non-Gaussian other words, the former yields a reduced steady-state error as
impulsive noises and shows the superior performance as compared to the latter under the same convergence rate.
compared to the LMP and GMCC algorithms. Motivated by the generalized GPDF [14], a generalized
Versoria function can be simply regarded as
1 1 (5)
f ( ek ) 2 a p
2a
1 | ek / 2 a | 1 | ek | p

where p>0 is the shape parameter, and = (2a)p. In this


representation, as a special case, it includes the original
Versoria function when p=2.
Fig. 1. Comparison of steepness of the GPDF and the standard Versoria
In order to reduce the steady-state misalignment and the
function. computational effort as compared to the GMCC algorithm, we
here use the generalized Versoria function as cost function
II. PROPOSED MAXIMUM VERSORIA CRITERION ALGORITHM
1 (6)
Define the input vector of an unknown system as xk=[xk, J ( w k 1 ) E p

xk-1,...,xk-L+1]T, where k is the time index, L is the filter length, 1 | ek |


and superscript T denotes vector transpose operation. The where E[] denotes the expectation operator. As can be seen
output signal of the unknown system is modelled as from (6), whenever an impulsive interference-corrupted ek
(1) appears, leads to J(wk1)0 (the gradient of J(wk1) approaches
d k x Tk w o k
zero), which plays the role of suppressing impulsive
where k is the background measurement noise plus impulsive interferences. When the error ek is small, it is easy to obtain
interferences, and wo is an unknown weight vector that needs to J(wk1) E[1|ek|p]. In this case, the proposed Versoria cost
be estimated. The system output error ek is defined as function works like the conventional cost function of the LMP
algorithm (note that the proposed algorithm is based on the
ek d k x Tk w k 1 (2) stochastic gradient ascent method).
A robust adaptive filtering algorithm, called maximum
where wk1 is the estimate of wo at iteration k1.
Versoria criterion (MVC) algorithm, can be developed by
The cost function of the maximum correntropy criterion
maximizing the generalized Versoria cost function (6).
(MCC) algorithm is the Gaussian probability density function
Removing the expectation operator and computing the gradient
(GPDF), which can be simply written as
of (6) with respect to wk1, yields
ek2 (3) 1
f ( ek ) exp( )
2 2 J w k 1 p | e k | p 1 sign( e k ) x k (7)
(1 | e k | p ) 2
where >0 is the kernel width. The GPDF is high calculation
cost when it is used in domains of machine learning and signal Using the stochastic gradient ascent approach, the update
processing, especially in adaptive filtering. For adaptive equation of the weight vector of the MVC algorithm can be
filtering, designing a suitable and high-efficiency cost obtained as
framework is an important issue. In order to draw forth the 1
w k w k 1 | e k | p 1 sign( e k ) x k (8)
proposed cost function, we first introduce the Versoria function, (1 | e k | p ) 2
which is defined as
where is the step size. The Versoria term in (8) approaches
8a 3 1 (4) zero whenever an impulsive interference-corrupted ek appears.
f ( ek ) 2
2a
4 a ek2
1 ( ek /2 a ) 2 Thus, this eliminates the likelihood of updating the weight
vector based on wrong information caused by the various
where a>0 is the radius of the generating circle of Versoria. The outliers including the impulsive interferences. When the system
centroid of the generating circle is located at (0, a). A larger output error ek is small, the MVC algorithm works like the
value of the radius a leads to a steeper Versoria. conventional LMP algorithm. We can obtain different orders
As a comparison, Fig. 1 plots the GPDF (2=0.5) and the MVC algorithms for different p values. Particularly, the
standard Versoria function (a=0.5). From this figure, we can generalized MVC algorithm with p=2 becomes
observe that the Versoria function is less steep than the GPDF
for the large error ek while both the GPDF and the Versoria 1 (9)
w k w k 1 ek x k
function exhibit comparable steepness for the small error ek. (1 e k2 ) 2
This means that the error along the direction of the gradient
ascent of the Versoria is faster to reach the optimal point (i.e., which is standard MVC algorithm. When 0, the generalized
zero error) than do the GPDF. From the perspective of the MVC algorithm reverts to the conventional LMP algorithm[16]
adaptive filter, the gradient ascent algorithm based on the
w k w k 1 | e k | p 1 sign( e k ) x k (10)
Versoria function has a faster convergence rate to reach the

1549-7747 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2017.2671521, IEEE
Transactions on Circuits and Systems II: Express Briefs
3

Assume that the impulsive interferences do not appear. The E [|| v k || 2 ] E [|| v k 1 || 2 ]
optimal a ( = (2a)p) can be calculated at each iteration by (19)
maximizing the Versoria term of the error nonlinearity function 2 E [ e a , k F ( e k )] 2 E [|| x k || 2 F 2 ( e k )]
of the MVC algorithm, to make the error with greatest
where ea,k = xkTvk1 is the a priori error, and F(ek) is the error
attenuation along the direction of the gradient ascent, i.e.,
nonlinearity function of the MVC algorithm, i.e.,
1 (11)
max j ( ek ) 1 (20)
ak (1 (2 a k ) p | ek | p ) 2 F ( ek ) | e k | p 1 sign( e k )
(1 | e k | p ) 2
Taking the derivative of (11) with respect to ek, yields
Assuming that ||xk||2 is asymptotically uncorrelated with F2(ek).
4 p (2 a k ) p 1 | e k | p 1 In order to guarantee that the MVC algorithm is stable for
j ( e k ) [ a k | e k | a k sign( e k ) ] (12) convergence, the step size should satisfy
(1 (2 a k ) p | e k | p ) 3
2 E[ea ,k F (ek )] (21)
where j'() and a'k are the first derivative. And then making the inf
E [|| x k || ] a ,k E [ F 2 ( e k )]
2
E [ e 2
]
derivative equal to zero, yields ak = a'kek. After performing
some simple calculation, yields where inf{} denotes infimum, and E [ e a2, k ] : E [ e a2, k ]
a k C | ek | (13) with denoting the Cramer-Rao bound. When the Cramr-Rao
bound is not known, the step size bound is difficult to be
where C is a positive constant. In order to achieve the good obtained. But, one can indirectly study the stability of the MVC
robustness in the presence of impulsive interferences, we adopt algorithm via the probability of divergence (POD) [14], [19].
following method to estimate a robust E[|ek|], instead of |ek| in The divergence means lim || v k || 2 . The POD of the LMF
(13), as k

algorithm is nonzero for any value of the step size when the
e k e k 1 (1 ) min( Ae , k ) (14) input is infinite. According to the simple example presented in
[14], one can verify that the MVC algorithm is very stable and
where 0<<<1 is the smoothing factor, min() denotes the its POD may be zero, no matter what input distribution is.
sample minimum operation which helps to remove the Consider a scalar filter case (wo and xk are both scalars) as in
impulsive interference-corrupted |ek|, and Ae,k = [|ek| |ek1| [19], the update of the weight-error vector is given as
|ekNw+1|] with Nw being the length of the estimation window.
Thus, (13) can be rewritten as 1
v k v k 1 | e k | p 1 sign( e k ) x k
(1 | e k | p ) 2
a k C ek (15)
1 (22)
v k 1 | ek | p 2 ek x k
In addition, the proposed MVC algorithm will lose robustness (1 | e k | p ) 2
when ak is too large. To circumvent this problem, we shall
1
restrict ak to lie inside an interval [0, C0]. (1 | e k | p | v k 1 | 2 ) v k 1
Remark: Motivated by the NLMS algorithm, it is easy to (1 | e k | p ) 2
develop the normalized MVC (NMVC) algorithm via the
Versoria cost function with respect to p power of the 0
p=2, =0.01, a =1
Misalignment (dB)

normalized output error -5 p=2, =0.01, a =2


p=2, =0.01, a =5
1 (16) -10
J ( w k 1 ) E p
p=2, =0.01, a =10

1 (| e k | || x k ||) -15 p=2, =0.01, ak

-20
We can derive the update equation of the NMVC algorithm by
using the stochastic gradient ascent approach, as 0 2000 4000 6000 8000 10000
Number of iterations
1 | ek | p 1 (a)
w k w k 1 sign( ek ) x k (17) 0
(1 (| ek | || x k ||) p ) 2 || x k || p
Misalignment (dB)

p=4, =0.0001, a =1
-10 p=4, =0.0001, a =2
When 0, the NMVC algorithm reduces to the NLMP p=4, =0.0001, a =5
algorithm [17] -20 p=4, =0.0001, a =10
p=4, =0.0001, ak
p 1
| ek | (18)
-30
w k w k 1 sign( ek ) x k
|| x k || p 0 2 1 3 4 5
Number of iterations x 10
4

(b)
III. CONVERGENCE ANALYSIS OF THE MVC ALGORITHM Fig. 2. Misalignment curves of the MVC with different constant a and adaptive
To begin with, we define the weight-error vector as vk = wo ak, respectively. The background measurement noise b is Gaussian noise with
k
wk. Based on the energy conservation relation [18], the zero-mean and unit variance. (a) =0.99, Nw=20, C=C0=10, 0=|d1|, (b) =0.99,
following relation holds: Nw=10, C=C0=10, 0=|d1|.

1549-7747 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2017.2671521, IEEE
Transactions on Circuits and Systems II: Express Briefs
4

15
SA (=0.007) 10
10 LMS (=0.002) SA (=0.005)
GMCC ( =1, =0.01, =0.007) 5 LMS (=0.0017)
5 GMCC ( =2, =0.05, =0.007) GMCC ( =1, =0.01, =0.005)
Misalignment (dB)

Misalignment (dB)
GMCC ( =4, =0.03, =0.006) 0 GMCC ( =2, =0.05, =0.005)
0 GMCC ( =6, =0.01, =0.002) GMCC ( =4, =0.005, =0.0013)
MVC (p =1, =0.0025, =0.0065) GMCC ( =6, =0.01, =0.001)
-5
-5 MVC (p =2, =0.0031, =0.002) MVC (p =1, =0.0025, =0.005)
MVC (p =4, =0.0039, =0.001) MVC (p =2, =0.0031, =0.0017)
-10
-10 MVC (p =6, =0.0006, =0.0003) MVC (p =4, =0.0039, =0.0008)
MVC (p =6, =0.0004, =0.00015)
-15 -15

-20 -20

-25 -25
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Number of iterations x 10
4 Number of iterations x 10
4

(a) (b)

10
10 SA (=0.005) SA (=0.0035)
LMS (=0.0015) LMS (=0.0012)
GMCC ( =1, =0.01, =0.005) GMCC ( =1, =0.01, =0.0035)
GMCC ( =2, =0.05, =0.005) 0 GMCC ( =2, =0.05, =0.0035)
0
Misalignment (dB)

Misalignment (dB)
GMCC ( =4, =0.005, =0.0013) GMCC ( =4, =0.01, =0.002)
GMCC ( =6, =0.01, =0.001) GMCC ( =6, =0.01, =0.0015)
MVC (p =1, =0.0025, =0.005) MVC (p =1, =0.0025, =0.0035)
-10 MVC (p =2, =0.0031, =0.0012)
-10 MVC (p =2, =0.0031, =0.0015)
MVC (p =4, =0.0039, =0.0008) MVC (p =4, =0.0039, =0.0008)
MVC (p =6, =0.0004, =0.00015) MVC (p =6, =0.0006, =0.0003)
-20
-20

-30
-30
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Number of iterations x 10
4 Number of iterations x 10
4

(c) (d)
Fig. 3. Misalignment curves of the SA, LMS, GMCC and MVC algorithms for different background measurement noise bk : (a) Laplace noise with zero-mean and


unit variance; (b) Gaussian noise with zero-mean and unit variance; (c) Uniform noise over 3, 3 ; (d) Binary noise over {1, 1} with probability mass P(x=
1)=P(x=1)=0.5. The impulsive interferences occur with probability Pr=0.1.

where the noise k is assumed to be zero. It is not difficult to get IV. SIMULATION RESULTS
1 1 Monte Carlo simulations are carried out for system
0 | ek | p (23)
p 2
(1 | e k | ) 4 identification and acoustic echo cancellation scenarios. The
performance of the adaptive algorithms is evaluated by the
normalized misalignment defined as:
If | v k 1 | 2 , we have Misalignment(k)=10log10(||w(k)wo||2/||wo||2). The misalignment
8
learning curves are obtained from the ensemble averages of 500
1 (24)
0 | e k | p | v k 1 | 2 2 independent runs.
(1 | e k | p ) 2
A. System Identification Scenario
Then, we obtain In the system identification scenario, the length of the
2
1 unknown system generated randomly and adaptive filters is 128
| v k |2 1 | ek | p | v k 1 | 2 | v k 1 |2 | v k 1 |2 (25) taps (L=128). The input is a white Gaussian random signal with
(1 | ek | p ) 2
zero-mean and unit variance. The system output is
Thus, the limit lim | v k | , if exists, is always bounded. This contaminated by a noise model as k bk k n k , where b k is
k

means that in this case the MVC algorithm will never diverge the background measurement noise, and k n k models the
(or the MVC has a zero POD), while the POD of the LMF strong impulsive interferences, wherein k is an i.i.d. Bernoulli
algorithm will be non-zero [19]. Comparing the convergent
condition of the proposed MVC algorithm with that (|vk1|2 process with a probability of success P ( k 1) Pr and
exp(1)/(2) [14]) of the GMCC algorithm, we obtain /(4) = P ( k 0 ) 1 Pr (Pr is the probability of the impulsive
exp(1)/. As = , yields =4exp(1)>. That is, in this interferences occurring), and n k is a white Gaussian random
simple example, the step-size bound of the MVC algorithm for
converging is larger than that of the GMCC algorithm. sequence with zero-mean and variance n2 1000 y2 . For all

1549-7747 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2017.2671521, IEEE
Transactions on Circuits and Systems II: Express Briefs
5

simulations of this subsection, the parameter settings are: than the GMCC algorithm, has been developed by maximizing
Pr=0.1; other parameters are reported in the figures. As two the generalized Versoria cost function. Simulations in system
examples, Fig. 2 shows the misalignment curves of the identification and acoustic echo cancellation scenarios have
proposed MVC algorithm with different constant a and confirmed that the proposed MVC algorithm is robust and
adaptive ak, respectively. As we can see, a smaller value of a performs much better than the LMP and GMCC algorithms in
results in a lower steady-state misalignment but a slower the presence of non-Gaussian impulsive noises. In the MVC
convergence rate, and vice versa. In other words, a larger value algorithm, a larger value of the radius of the generating circle of
of ( = (2a)p) yields a lower steady-state misalignment but a Versoria yields a faster convergence rate but a higher
slower convergence rate, and vice versa. One can also see that steady-state misalignment, and vice versa.
the proposed MVC algorithm with adaptive ak not only
achieves a fast convergence rate, but also achieves a low REFERENCES
steady-state misalignment. In the following simulations, the [1] A. H. Sayed, Fundamentals of Adaptive Filtering. Hoboken, NJ,USA:
performance of the proposed MVC algorithm is compared with Wiley, 2003.
that of the SA, LMS and GMCC algorithms. The computer [2] R. H. Kwong and E. W. Johnston, A variable step size LMS
simulation results are shown in Fig. 3, where the background algorithm, IEEE Trans. Signal Process., vol. 40, no. 7, pp.
16331642, Jul. 1992.
measurement noise b k is (a) Laplace noise; (b) Gaussian noise;
[3] D. T. M. Slock, On the convergence behavior of the LMS and the
(c) Uniform noise; (d) Binary noise, respectively. As shown in normalized LMS algorithms, IEEE Trans. Signal Process., vol. 41,
Fig. 3, it is clearly seen that the proposed MVC (p>1) algorithm no. 9, pp. 28112825, Sep. 1993.
[4] J. Benesty, H. Rey, L. R. Vega, and S. Tressens, A nonparametric
outclasses the SA, LMS and GMCC algorithms in terms of the VSS NLMS algorithm, IEEE Signal Process. Lett., vol. 13, no. 10, pp.
convergence rate and/or the steady-state misalignment. Note 581584, Oct. 2006.
that we do not show the misalignment curves of the LMF since [5] S. Gollamudi, S. Nagaraj, S. Kapoor, and Y.-F. Huang,
the LMF is not stable for any value of the step-size when the Set-membership filtering and a set-membership normalized LMS
noises include the strong impulsive interferences. algorithm with an adaptive step size, IEEE Signal Process. Lett., vol.
5, no. 5, pp.111114, May 1998.
[6] S. Zhang, J. Zhang, and H. Han, Robust shrinkage normalized sign
SA (=0.0005) algorithm in an impulsive noise environment, IEEE Trans. Circuits
10 LMS (=0.00017) Syst. II, Exp. Briefs, vol. 64, no. 1, pp. 9195, Jan. 2017.
GMCC ( =1, =0.01, =0.0005) [7] T. Shao, Y. R. Zheng, and J. Benesty, An affine projection sign
GMCC ( =2, =0.05, =0.0005) algorithm robust against impulsive interference, IEEE Signal Process.
Misalignment (dB)

5 GMCC ( =4, =0.005, =0.00013) Lett., vol. 17, no. 4, pp. 327330, Apr. 2010.
GMCC ( =6, =0.01, =0.0001) [8] S. Wang, J. Feng, and C. K. Tse, Kernel affine projection sign
MVC (p =1, =0.0025, =0.0005)
algorithms for combating impulse interference, IEEE Trans. Circuits
MVC (p =2, =0.0031, =0.00017)
0 Syst. II, Exp. Briefs, vol. 60, no. 11, pp. 811815, Nov. 2013.
MVC (p =4, =0.0039, =0.00008)
MVC (p =6, =0.0004, =0.000015)
[9] F. Huang, J. Zhang, and S. Zhang, Combined-step-size affine
projection sign algorithm for robust adaptive filtering in impulsive
-5 interference environments, IEEE Trans. Circuits Syst. II, Exp. Briefs,
vol. 63, no. 5, pp. 493497, May 2016.
[10] H. Zayyani, Continuous mixed p-norm adaptive algorithm for system
-10 identification, IEEE Signal Process. Lett., vol. 21, no. 9, pp.
0 2 4 6 8 10 11081110, Sep. 2014.
Time (sec) [11] E. Walach and B. Widrow, The least mean fourth (LMF) adaptive
Fig. 4. Misalignment curves of the SA, LMS, GMCC and MVC algorithms for algorithm and its family, IEEE Trans. Inf. Theory, vol. 30, no. 2, pp.
acoustic echo cancellation application. 275283, Mar. 1984.
[12] W. Liu, P. P. Pokharel, and J. C. Principe, Correntropy: Properties
and applications in non-Gaussian signal processing, IEEE Trans.
B. Acoustic Echo Cancellation Scenario Signal Process., vol. 55, no. 11, pp. 52865298, Nov. 2007.
Here, the performance of the SA, LMS, GMCC, and [13] T. Ogunfunmi and T. Paul, The quarternion maximum correntropy
proposed MVC algorithms is compared in an acoustic echo algorithm, IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 62, no. 6,
pp. 598602, Jun. 2015.
cancellation scenario, under the same initial convergence rate.
[14] B. Chen, L. Xing, H. Zhao, N. Zheng, and J. C. Principe, Generalized
The length of the impulse response of the acoustic channel is correntropy for robust adaptive filtering, IEEE Trans. Signal
512. The input signal is speech signal sampled at 8 kHz. The Process., vol. 64, no. 13, pp. 33763387, Jul. 2016.
output signal is contaminated by k bk k nk , where b k is [15] X. Yu, J. Liu, and H. Li, An adaptive inertia weight particle swarm
optimization algorithm for IIR digital filter, in Proc. Inter. Conf. on
Gaussian noise with zero-mean and unit variance, and k n k is Artificial Intelligence and Computational Intelligence, 2009,
pp.114-118.
the same as that in the system identification scenario. The
[16] S.-C. Pei and C.-C. Tseng, Least mean p-power error criterion for
impulsive interferences occur with probability Pr=0.1. The adaptive fir filter, IEEE J. Sel. Areas Commun., vol. 12, no. 9, pp.
parameter settings are presented in the figure. As can be seen 15401547, Dec. 1994.
from Fig. 4, the proposed MVC (p>1) algorithm performs much [17] O. Arikan, A. Enis Cetin, E. Erzin, Adaptive filtering for
better than the SA, LMS and corresponding orders GMCC non-Gaussian stable processes, IEEE Signal Processing Letters, vol.
1, no. 11, pp.163-165, Nov. 1994.
algorithms.
[18] T. Y. Al-Naffouri and A. H. Sayed, Transient analysis of adaptive
filters with error nonlinearities, IEEE Trans. Signal Process., vol. 51,
V. CONCLUSION no. 3, pp. 653663, Mar. 2003.
[19] V. H. Nascimento and J. C. M. Bermudez, Probability of divergence
A maximum Versoria criterion (MVC) algorithm, which is for the least-mean fourth algorithm, IEEE Trans. Signal Process., vol.
lower steady-state misalignment and less computational effort 54, no. 4, pp. 13761385, 2006.

1549-7747 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Das könnte Ihnen auch gefallen