new lmt-based delay-dependent criterion for global asymptotic stability of cellular neural networks
TRANSCRIPT
ARTICLE IN PRESS
Neurocomputing 72 (2009) 3331–3336
Contents lists available at ScienceDirect
Neurocomputing
0925-23
doi:10.1
$ Thi
(Grant
for Che
605210
China (G
Univers� Corr
116028,
E-m
laibing3
journal homepage: www.elsevier.com/locate/neucom
New LMT-based delay-dependent criterion for global asymptotic stability ofcellular neural networks$
Cheng-De Zheng a,b,�, Lai-Bing Lu a, Zhan-Shan Wang b
a School of Science, Dalian Jiaotong University, Dalian 116028, PR Chinab School of Information Science and Engineering, Northeastern University, Shenyang 110004, PR China
a r t i c l e i n f o
Article history:
Received 23 July 2008
Received in revised form
3 December 2008
Accepted 13 January 2009
Communicated by T. Heskesinequality are considered. Three illustrative examples are given to demonstrate the effectiveness of the
Available online 5 March 2009
Keywords:
Global asymptotic stability (GAS)
Linear matrix inequality (LMI)
Cellular neural networks
Time-varying delay
Jensen integral inequality
12/$ - see front matter & 2009 Elsevier B.V. A
016/j.neucom.2009.01.013
s work was supported by the National Natura
nos. 60534010, 60572070, 60728307, 607740
ung Kong Scholars and Innovative Research
03), the National High Technology Research a
rant no. 2006AA04Z183) and the Postdoctora
ity (Grant no. 20080314).
esponding author at: School of Science, Dalia
PR China.
ail addresses: [email protected], cdzheng
[email protected] (L.-B. Lu), zhanshan_wang@163
a b s t r a c t
The problem of global asymptotic stability analysis is studied for a class of cellular neural networks with
time-varying delay. By defining a Lyapunov–Krasovskii functional, a new delay-dependent stability
condition is derived in terms of linear matrix inequalities. The obtained criterion is less conser-
vative than some previous literature because free-weighting matrix method and the Jensen integral
proposed results.
& 2009 Elsevier B.V. All rights reserved.
1. Introduction
As neural networks have been successfully applied to variousfields such as image processing, pattern recognition and partialdifferential equations solving, the stability analysis of neuralnetworks has been extensively investigated. However, time delaysare frequently encountered in neural networks due to the finiteswitching speed of amplifiers and the inherent communication ofneurons. It is well known that the existence of time delay maycause divergence, oscillation, and even instability. Therefore,considerable efforts have been devoted to stability analysis ofneural networks with time delays. Among them, delay-dependentstability criteria [1,3–6,9,23] have attracted much attentionrecently because delay-dependent criteria make use of informa-tion on the length of delays, and are less conservative than delay-independent ones [7,8,10–22].
ll rights reserved.
l Science Foundation of China
48, 60774093), the Program
Groups of China (Grant no.
nd Development Program of
l Foundation of Northeastern
n Jiaotong University, Dalian
@djtu.edu.cn (C.-D. Zheng),
.com (Z.-S. Wang).
In order to get a less conservative result, we constructa new Lyapunov–Krasovskii functional and derive newsufficient condition for the globally asymptotic stability of theconcerned delayed neural network, which is delay-dependentand computationally efficient. Since the free-weighting matrixapproach [3,4,6] and the Jensen integral inequality [2] areinvolved, the relationship between the time-varying delay andits lower and upper bounds is considered, the result is lessconservative than some existing ones. Three illustrativeexamples are given to demonstrate the effectiveness of theproposed results.
2. Problem description
Considering the following cellular neural networks withinterval time-varying delays:
_xðtÞ ¼ �CxðtÞ þ Af ðxðtÞÞ þ Bf ðxðt � tðtÞÞÞ þ J, (1)
where xðtÞ ¼ ðx1ðtÞ; x2ðtÞ; . . . ; xnðtÞÞT2Rn is the neural state
vector, C ¼ diagfc1; c2; . . . ; cng is a positive diagonal matrix, A ¼
ðaijÞn�n; B ¼ ðbijÞn�n are known constant matrices, 0pt1ptðtÞpt2
is the time-varying delay, where t1; t2 are constants. J isthe constant external input vector, and f ðxðtÞÞ ¼ ðf 1ðx1ðtÞÞ;
f 2ðx2ðtÞÞ; . . . ; f nðxnðtÞÞÞT2 Rn denotes the neural activation func-
tion. It is assumed that f iðxiðtÞÞ ði ¼ 1;2; . . . ;nÞ are bounded and
ARTICLE IN PRESS
C.-D. Zheng et al. / Neurocomputing 72 (2009) 3331–33363332
there exist constants gj;sj such that (see [9])
gjpf jðs1Þ � f jðs2Þ
s1 � s2psj; j ¼ 1;2; . . . ;n (2)
for any s1; s2 2 R,s1as2:
From the well-known Brouwer’s fixed point theorem, system(1) always has an equilibrium point x�.
Throughout this paper, let kyk denote the Euclidean normof a vector y 2 Rn; WT ;W�1 denote the transpose, the inverse,respectively. Let W40 ðX0Þ denote a positive (nonnegative)definite symmetric matrix, I denote an identity matrix withcompatible dimension.
In order to prove the global asymptotic stability (GAS) of theequilibrium point x� of system (1), we will first simplify system (1)as follows. Let uð�Þ ¼ xð�Þ � x�; then we have,
_uðtÞ ¼ �CuðtÞ þ AgðuðtÞÞ þ Bgðuðt � tðtÞÞÞ, (3)
where uðtÞ ¼ ðu1ðtÞ;u2ðtÞ; . . . ;unðtÞÞT ; gjðujðtÞÞ ¼ f jðujðtÞ þ x�j Þ � f jðx
�j Þ
with gjð0Þ ¼ 0; j ¼ 1;2; . . . ;n: By assumption (2), we can see that
gjpgjðujðtÞÞ
ujðtÞpsj. (4)
Clearly, the equilibrium point of system (1) is GAS if and only ifthe zero solution of system (3) is GAS.
Through this paper, the Jensen integral inequality will be used,so it is listed as the following lemma.
Lemma 1 (Gu [2]). For any positive symmetric constant matrix
M 2 Rn�n, scalars r1or2 and vector function o : ½r1; r2� !Rn such
that the integrations concerned are well defined, then
Z r2
r1
oðsÞds
� �T
M
Z r2
r1
oðsÞds
� �pðr2 � r1Þ
Z r2
r1
oT ðsÞMoðsÞds.
3. GAS results of delayed cellular neural networks
First, we will present the GAS results for system (3) with_tðtÞpZo1:
Theorem 1. Under assumption (2) and 0pt1ptðtÞpt2; _tðtÞpZo1;suppose that there exist positive definite symmetric matrices Pi ði ¼
1; . . . ;7Þ;Q ¼ ½Qij�2�2;R ¼ ½Rij�2�2; S ¼ ½Sij�2�2; positive diagonal ma-
trices T1; T2;D ¼ diagfd1; d2; . . . ;dng;D ¼ diagfd1; d2; . . . ; dng and
real matrices XT¼ ½XT
1 XT2�;Y
T¼ ½YT
1 YT2�; Z
T¼ ½ZT
1 ZT2� such that the
following linear matrix inequality (LMIs) hold:
X1 ¼R X
XT P3
" #X0, (5)
X2 ¼S Y
YT P5
" #X0, (6)
X3 ¼Rþ S Z
ZT P3 þ P5
" #X0, (7)
O ¼
O11 O12 O13 O14 Y1 �Z1
� O22 0 O24 O25 O26
� � O33 O34 0 0
� � � O44 0 0
� � 0 0 O55 0
� � 0 0 0 O66
26666666664
37777777775o0, (8)
where ‘‘*’’ are entries readily inferred by symmetry, and
O11 ¼ � P1C � CP1 þ CPC þ X1 þ XT1 þ Q11 þ P6 þ P7 þ t2R11
þ ðt2 � t1ÞS11 � 2GT1S� 2ðDS� DGÞC �1
t2P2,
O12 ¼ �X1 þ XT2 þ Z1 � Y1 þ t2R12 þ ðt2 � t1ÞS12 þ
1
t2P2,
O13 ¼ P1A� CðD� DÞ þ Q12 � CPAþ ðDS� DGÞAþ ðSþGÞT1,
O14 ¼ P1B� CPBþ ðDS� DGÞB,
O22 ¼ � ð1� ZÞQ11 � X2 � XT2 þ Z2 þ ZT
2 � Y2 � YT2 þ t2R22
þ ðt2 � t1ÞS22 � 2GT2S�1
t2P2 �
1
t2 � t1ðP2 þ 2P4Þ,
O24 ¼ �ð1� ZÞQ12 þ ðSþGÞT2; O25 ¼ Y2 þ1
t2 � t1P4,
O26 ¼1
t2 � t1ðP2 þ P4Þ � Z2,
O33 ¼ Q22 þ ðD�DÞAþ ATðD� DÞ þ AT PA� 2T1,
O34 ¼ ðD�DÞBþ AT PB,
O44 ¼ �ð1� ZÞQ22 þ BT PB� 2T2; O55 ¼ �1
t2 � t1P4 � P6,
O66 ¼ �1
t2 � t1ðP2 þ P4Þ � P7,
P ¼ t2ðP2 þ P3Þ þ ðt2 � t1ÞðP4 þ P5Þ; G ¼ diagfg1; g2; . . . ; gng,
S ¼ diagfs1;s2; . . . ;sng,
then the equilibrium point of system (3) is unique and GAS.
Proof. Firstly, we show the uniqueness of the equilibrium pointby contradiction. To end this, let u be the equilibrium point of thedelayed recurrent neural network (3), then we have
�Cuþ ðAþ BÞgðuÞ ¼ 0.
Now suppose ua0: It is easy to see that
2fuTðP1 þDS� DGÞ þ gT ðuÞðD� DÞgf�Cuþ ðAþ BÞgðuÞg ¼ 0. (9)
By inequality (4), we get
� 2uTSðT1 þ T2ÞGuþ 2u
TðT1 þ T2ÞðGþ SÞgðuÞ � 2gT ðuÞ
�ðT1 þ T2ÞgðuÞX0.
This together with Eq. (9) gives
uTE11uþ 2u
TE12gðuÞ þ gT ðuÞE22gðuÞX0,
i.e.
½uT
gT ðuÞ�E11 E12
� E22
" #u
gðuÞ
" #X0, (10)
where
E11 ¼ �ðP1 þDS� DGÞC � CðP1 þDS� DGÞ � 2GSðT1 þ T2Þ,
E12 ¼ ðP1 þDS� DGÞðAþ BÞ � CðD� DÞ þ ðT1 þ T2ÞðGþ SÞ,
E22 ¼ ðD� DÞðAþ BÞ þ ðAþ BÞT ðD�DÞ � 2ðT1 þ T2Þ.
On the other hand, let
P ¼I I 0 0 I I
0 0 I I 0 0
� �,
ARTICLE IN PRESS
C.-D. Zheng et al. / Neurocomputing 72 (2009) 3331–3336 3333
multiplying (8) by P and PT on its left and right side, respectively,we have
~E11 E12 þ ZQ12 � CPðAþ BÞ
� E22 þ ZQ22 þ ðAþ BÞT PðAþ BÞ
" #o0,
where ~E11 ¼ E11 þ CPC þ ZQ11 þ t2ðR11 þ R12 þ RT12 þ R22Þ þ ðt2 �
t1ÞðS11 þ S12 þ ST12 þ S22Þ:
That is
E11 E12
� E22
" #þ
I I
0 0
" #ðt2Rþ ðt2 � t1ÞSÞ
I 0
I 0
" #þ ZQ
þ�C
ðAþ BÞT
" #P½�C Aþ B�o0.
Note that P40;QX0;RX0; SX0; we obtain
E11 E12
� E22
" #o0.
Obviously, this contradicts with (10). The contradiction implies
that u ¼ 0: That is, the origin of the delayed recurrent neural
networks (3) is the unique equilibrium point.
Next, we show the unique equilibrium point of (3) is GAS.
Consider the following Lyapunov–Krasovskii functional:
VðuðtÞÞ ¼ uT ðtÞP1uðtÞ þ 2Xn
i¼1
di
Z uiðtÞ
0fgiðsÞ � gisgds
�
þdi
Z uiðtÞ
0fsis� giðsÞgds
�
þ
Z t
t�t2
Z t
y_uTðsÞðP2 þ P3Þ _uðsÞds dy
þ
Z t�t1
t�t2
Z t
y_uTðsÞðP4 þ P5Þ _uðsÞds dy
þX2
i¼1
Z t
t�ti
uT ðsÞPiþ5uðsÞdsþ
Z t
t�tðtÞxTðuðsÞÞQxðuðsÞÞds,
(11)
where xTðuðsÞÞ ¼ ðuT ðsÞ; gT ðuðsÞÞÞ:
For convenience, we denote ut ¼ uðt � tðtÞÞ: The time derivative
of functional (11) along the trajectories of system (3) is obtained
as follows:
_VðuðtÞÞ ¼ 2uT ðtÞP1 _uðtÞ þ 2fgT ðuðtÞÞ � uT ðtÞGgD _uðtÞ þ 2fuT ðtÞS
� gT ðuðtÞÞgD _uðtÞ þ t2 _uTðtÞðP2 þ P3Þ _uðtÞ
�
Z t
t�t2
_uTðsÞðP2 þ P3Þ _uðsÞdsþ ðt2 � t1Þ _u
TðtÞ
�ðP4 þ P5Þ _uðtÞ �
Z t�t1
t�t2
_uTðsÞðP4 þ P5Þ _uðsÞdsþ uT ðtÞ
�ðP6 þ P7ÞuðtÞ �X2
i¼1
uT ðt � tiÞPiþ5uðt � tiÞ
þ xTðuðtÞÞQxðuðtÞÞ � ð1� _tðtÞÞxT
ðutÞQxTðutÞ. (12)
It is clear that the following equations are true:
Z t
t�t2
_uTðsÞP2 _uðsÞds ¼
Z t
t�tðtÞ_uTðsÞP2 _uðsÞdsþ
Z t�tðtÞ
t�t2
_uTðsÞP2 _uðsÞds,
(13)
Z t�t1
t�t2
_uTðsÞP4 _uðsÞds ¼
Z t�t1
t�tðtÞ_uTðsÞP4 _uðsÞdsþ
Z t�tðtÞ
t�t2
_uTðsÞP4 _uðsÞds.
(14)
By using the Jensen integral inequality (Lemma 1), we obtain
�
Z t
t�tðtÞ_uTðsÞP2 _uðsÞdsp�
1
tðtÞ
Z t
t�tðtÞ_uðsÞds
� �T
P2
Z t
t�tðtÞ_uðsÞds
p�1
t2½uðtÞ � ut�
T P2½uðtÞ � ut�, (15)
�
Z t�t1
t�tðtÞ_uTðsÞP4 _uðsÞdsp�
1
tðtÞ � t1
Z t�t1
t�tðtÞ_uðsÞds
� �T
P4
Z t�t1
t�tðtÞ_uðsÞds
p�1
t2 � t1½uðt � t1Þ � ut�
T P4½uðt � t1Þ � ut�,
(16)
�
Z t�tðtÞ
t�t2
_uTðsÞðP2 þ P4Þ _uðsÞds
p�1
t2 � tðtÞ
Z t�tðtÞ
t�t2
_uðsÞds
� �T
ðP2 þ P4Þ
Z t�tðtÞ
t�t2
_uðsÞds
p�1
t2 � t1½ut � uðt � t2Þ�
T ðP2 þ P4Þ½ut � uðt � t2Þ�. (17)
On the other hand, based on Leibniz–Newton formula, for
any real matrix Xi;Yi; Zi ði ¼ 1;2Þ with compatible dimensions,
we get
0 ¼ 2fuT ðtÞX1 þ uTtX2g uðtÞ � ut �
Z t
t�tðtÞ_uðsÞds
� �, (18)
0 ¼ 2fuT ðtÞY1 þ uTtY2g uðt � t1Þ � ut �
Z t�t1
t�tðtÞ_uðsÞds
� �, (19)
0 ¼ 2fuT ðtÞZ1 þ uTtZ2g ut � uðt � t2Þ �
Z t�tðtÞ
t�t2
_uðsÞds
� �. (20)
Furthermore, from inequality (4), the following matrix inequal-
ities hold for any positive diagonal matrices T1; T2 with compa-
tible dimensions
0p� 2uT ðtÞST1GuðtÞ þ 2uT ðtÞT1ðGþSÞgðuðtÞÞ � 2gT ðuðtÞÞT1gðuðtÞÞ,
(21)
0p� 2uTtST2Gut þ 2uT
tT2ðGþSÞgðutÞ � 2gT ðutÞT2gðutÞ. (22)
Moreover, for any real symmetric matrices R; S with compatible
dimensions, we have
0 ¼ t2kT ðtÞRkðtÞ �Z t
t�tðtÞkT ðtÞRkðtÞds�
Z t�tðtÞ
t�t2
kT ðtÞRkðtÞds, (23)
0 ¼ ðt2 � t1ÞkT ðtÞSkðtÞ �Z t�t1
t�tðtÞkT ðtÞSkðtÞds�
Z t�tðtÞ
t�t2
kT ðtÞSkðtÞds,
(24)
where kT ðtÞ ¼ ðuT ðtÞ;uTt Þ:
Therefore it can be seen from (12) to (24) that
_VðuðtÞÞpzTðtÞOzðtÞ �
Z t
t�tðtÞBT ðt; sÞX1Bðt; sÞds
�
Z t�t1
t�tðtÞBT ðt; sÞX2Bðt; sÞds�
Z t�tðtÞ
t�t2
BT ðt; sÞX3Bðt; sÞds,
where BT ðt; sÞ ¼ ½uT ðtÞ;uTt ; _u
TðsÞ� and
zTðtÞ ¼ ½uT ðtÞ uT
t gT ðuðtÞÞ gT ðutÞ uT ðt � t1Þ uT ðt � t2Þ�.
Therefore, if Oo0 and XiX0; i ¼ 1;2;3; then _VðuðtÞÞp� �kuðtÞk2
holds for a sufficiently small �40: This implies the GAS of system
(3) with assumption (2), which completes the proof. &
ARTICLE IN PRESS
C.-D. Zheng et al. / Neurocomputing 72 (2009) 3331–33363334
Remark 1. Theorem 1 provides an LMI-based sufficientcondition for the GAS of the delayed neural network in (3). Oneadvantage of the LMI approach is that the LMI condition can bechecked numerically more efficiently than those in [8,21,22]which are not LMIs with respect to the parameters to bedetermined.
Remark 2. Unlike the Lyapunov–Krasovskii functionaland the proof technique in [4], this paper introduced theterms of G;D; P2; P4;Q12; used Jensen integral inequality. Theadvantage is that Theorem 1 can lead to less conservativenessthan [4]. In fact, let G;D;P2; P4;Q12 all be 0; Theorem 1 of thispaper is equal to Theorem 1 of [4]. Furthermore, this paperproved the uniqueness of equilibrium point of above network,while [4] did not.
When tðtÞ is not differentiable or _tðtÞ is unknown, Q
will no longer be helpful to improve the stability condition since�ð1� ZÞQ may be nonnegative definite. Therefore, by settingQ ¼ 0; an easy delay-dependent criterion is derived as follows forunknown Z:
0 5 10 15 20 25 30−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
u1
u2
u3
u4
example 1
Time t
u(t)
Fig. 1. The dynamical behavior of the system (3) in Example 1.
Table 1Calculated maximal upper bounds of t2 for various t1 and Z of Example 1.
t1 Methods Z ¼ 0 Z ¼ 0:1 Z ¼ 0:5 Z ¼ 0:9 Z ¼ 0:95 Unknown Z
0 [6] 3.5841 3.2775 2.1502 1.3164 1.2708 1.2598
[3] 3.5841 3.2793 2.2245 1.5847 1.5514 1.5444
[4] 3.5841 3.3039 2.5376 2.0853 2.0485 2.0389
Corollary 1. Under assumption (2) and 0pt1ptðtÞpt2; _tðtÞpZo1;suppose that there exist positive definite symmetric matrices
Pi ði ¼ 1; . . . ;7Þ;Q ;R; S; positive diagonal matrices T1; T2;D;D and
real matrices X;Y ; Z such that Xl ðl ¼ 1;2;3ÞX0 and the following
LMI holds:
O11 O12 O13 O14 Y1 �Z1
� O22 0 ðSþ GÞT2 O25 O26
� � O33 O34 0 0
� � � O44 0 0
� � 0 0 O55 0
� � 0 0 0 O66
26666666664
37777777775o0,
where
O11 ¼ � P1C � CP1 þ CPC þ X1 þ XT1 þ P6 þ P7 þ t2R11
þ ðt2 � t1ÞS11 � 2ðDS� DGÞC � 2GT1S�1
t2P2,
O13 ¼ P1A� CðD� DÞ � CPAþ ðDS� DGÞAþ ðSþ GÞT1,
O22 ¼ � X2 � XT2 þ Z2 þ ZT
2 � Y2 � YT2 þ t2R22 þ ðt2 � t1ÞS22
� 2GT2S�1
t2P2 �
1
t2 � t1ðP2 þ 2P4Þ,
O33 ¼ ðD� DÞAþ ATðD� DÞ þ AT PA� 2T1,
O44 ¼ BT PB� 2T2,
and the other parameters are determined in Theorem 1, then the
equilibrium point of system (3) is unique and GAS.
This paper 3.8363 3.5209 2.7176 2.2151 2.1339 2.0771
1 [4] 3.5841 3.3068 2.5802 2.2736 2.2510 2.2393
This paper 3.8363 3.5234 2.7280 2.3510 2.3084 2.2739
2 [4] 3.5841 3.3125 2.7500 2.6468 2.6361 2.6299
This paper 3.8363 3.5289 2.8392 2.7064 2.6884 2.6735
4. Illustrative examples
In this section, we provide three numerical examples todemonstrate the effectiveness and less conservativeness of ourdelay-dependent stability criteria.
Example 1. Consider system (3) with
C ¼ diagf1:2769;0:6231;0:9230;0:4480g;
A ¼
�0:0373 0:4852 �0:3351 0:2336
�1:6033 0:5988 �0:3224 1:2352
0:3394 �0:0860 �0:3824 �0:5785
�0:1311 0:3253 �0:9534 �0:5015
26664
37775;
B ¼
0:8674 �1:2405 �0:5325 0:0220
0:0474 �0:9164 0:0360 0:9816
1:8495 2:6117 �0:3788 0:8428
�2:0413 0:5179 1:1734 �0:2775
26664
37775;
f 1ðsÞ ¼ 0:05685ðjsþ 1j � js� 1jÞ; f 2ðsÞ ¼ 0:06395ðjsþ 1j � js� 1jÞ;
f 3ðsÞ ¼ 0:3997ðjsþ 1j � js� 1jÞ; f 4ðsÞ ¼ 0:1184ðjsþ 1j � js� 1jÞ:
Obviously, the activation functions are bounded and satisfyassumption (2) with
s1 ¼ 0:1137; s2 ¼ 0:1279; s3 ¼ 0:7994; s4 ¼ 0:2368,
g1 ¼ g2 ¼ g3 ¼ g4 ¼ 0.
For this model, from Theorem 1 we can verify that ð0;0;0;0ÞT isthe unique equilibrium point and is GAS. The GAS with the initialstate ð�0:2;0:5;�0:4;0:2ÞT is shown in Fig. 1.
The same model has been studied in [3,4,6]. The comparison
among Theorem 1 and Corollary 1 in this paper and those in
[3,4,6] are listed in Table 1, where ‘‘unknown Z’’ means that Z can
be arbitrary value or tðtÞ can be not differentiable.
However, for any t1 and Z; all of the criteria given in
[5,8,9,13,14,21,22] fail to conclude whether the model is GAS or
ARTICLE IN PRESS
Table 3Calculated maximal upper bounds of t2 for various t1 and Z of Example 3.
t1 Methods Z ¼ 0:4 Z ¼ 0:8 Z ¼ 0:9 Z ¼ 0:95 Z ¼ 0:99 Unknown Z
0 [5] 0.4067 0.3928 0.3920 0.3920 0.3920 0.3920
This paper 1.7654 1.4270 1.3888 1.3660 1.3465 1.3418
1 [5] 1.3901 1.3794 1.3782 1.3777 1.3774 1.3774
This paper 2.6190 2.3514 2.3461 2.3435 2.3420 2.3418
2 [5] 2.3900 2.3794 2.3782 2.3777 2.3773 2.3773
This paper 3.6172 3.3506 3.3456 3.3432 3.3420 3.3418
100 [5] 100.3898 100.3794 100.3782 100.3777 100.3773 100.3772
This paper 101.6141 101.3506 101.3456 101.3432 101.3420 101.3418
Table 2Calculated maximal upper bounds of t2 for various t1 and Z of Example 2.
t1 Methods Z ¼ 0:76 Z ¼ 0:8 Z ¼ 0:85 Z ¼ 0:9 Z ¼ 0:95 Unknown Z
0 [5] 11.4405 5.6225 4.5075 4.2448 4.2426 4.2426
[3] 14.5340 7.2083 5.7608 5.3274 5.2769 5.2769
[4] 20.5135 10.1217 8.0686 7.4622 7.3977 7.3977
This paper 25.8290 13.0158 10.4913 9.6516 9.3993 9.3711
1 [5] 12.4401 5.6226 5.5076 5.2449 5.2426 5.2426
[4] 21.3875 10.8601 8.7276 8.0760 7.9985 7.9985
This paper 26.5007 13.3606 10.6876 9.8375 9.6602 9.6602
2 [5] 13.4401 6.6226 6.5076 6.2449 6.2426 6.2426
[4] 22.3880 11.8602 9.7276 9.0760 8.9985 8.9985
This paper 27.5005 14.3606 11.6876 10.8275 10.6602 10.6602
100 [5] – – – – – –
[4] 120.3883 109.8602 107.7275 107.0760 106.9985 106.9985
This paper 125.4997 111.9289 109.6877 108.8274 108.6602 108.6602
C.-D. Zheng et al. / Neurocomputing 72 (2009) 3331–3336 3335
not. Therefore, we can say that for this system the results in
this paper are effective and less conservative than those in
[3–6,8,9,13,14,21,22].
Example 2. Consider system (1) with
C ¼1 0
0 1
" #; A ¼
0:3 0
0 0:2
" #; B ¼
0:1 �0:2
0 0:1
" #,
f 1ðsÞ ¼ ðjsþ 1j � js� 1jÞ; f 2ðsÞ ¼ ðjsþ 1j � js� 1jÞ.
Obviously, the activation functions are continuous and satisfyassumption (2) with S ¼ 2I;G ¼ 0:
For this model, from Theorem 1 we can verify that ð0;0ÞT is the
unique equilibrium point and is GAS.
The comparison among Theorem 1 and Corollary 1 in this paper
and those in [3–5] are listed in Table 2, where ‘‘–’’ denotes the
result failing to verify the GAS of the equilibrium point.
However, for any t1 and given Z; all the criteria given in
[8,9,21,22] fail to conclude whether the model is GAS or not.
Therefore, we can say that for this system the results in this
paper are much effective and less conservative than those in
[3–5,8,9,21,22].
Example 3. Consider system (3) with
C ¼ diagf1:34;1:08;1:46g,
A ¼
1:08 �0:52 0:05
�0:08 0:68 0:19
0:14 0:29 �0:58
264
375; B ¼
2:45 �0:64 1:2
0:45 0:88 0:47
0:07 �1:56 0:97
264
375,
s1 ¼ 0:31 s2 ¼ 0:40; s3 ¼ 0:23,
g1 ¼ �0:31; g2 ¼ �0:40; g3 ¼ �0:23.
Obviously, none of the results in [1,3,4,6,8,21,22] are applicableto ascertain the stability of this model. The comparison amongTheorem 1 and Corollary 1 in this letter and those in [5] are listedin Table 3.
Therefore, we can say that for this system the results in this
paper are more effective and less conservative than those in
[1,3–6,8,13,21,22].
5. Conclusion
Several sufficient conditions are derived to guarantee the GASof CNN with time-varying delays, which are different from theexisting ones and have wider application fields than some results
ARTICLE IN PRESS
C.-D. Zheng et al. / Neurocomputing 72 (2009) 3331–33363336
in literature. The obtained results can be expressed in the form ofLMI and are easy to be verified. Three illustrative examples arealso given to demonstrate the effectiveness of the proposedresults.
Acknowledgment
The authors would like to thank all of the referees and theAssociate Editor for their constructive comments, which led to asignificant improvement in the paper.
References
[1] H.J. Cho, J.H. Park, Novel delay-dependent robust stability criterion ofdelayed cellular neural networks, Chaos, Solitons and Fractals 32 (2007)1194–1200.
[2] K. Gu, An integral inequality in the stability problem of time-delay systems,in: Proceedings of the 39th IEEE Conference on Decision and Control, 2000,pp. 2805–2810.
[3] Y. He, G.P. Liu, D. Rees, New delay-dependent stability criteria for neuralnetworks with time-varying delay, IEEE Transactions on Neural Networks 18(1) (2007) 310–314.
[4] Y. He, G.P. Liu, D. Rees, M. Wu, Stability analysis for neural networks withtime-varying interval delay, IEEE Transactions on Neural Networks 18 (6)(2007) 1850–1854.
[5] C.D. Li, G. Feng, Delay-interval-dependent stability of recurrent neuralnetworks with time-varying delay, Neurocomputing 72 (2009) 1179–1183.
[6] H. Liu, G. Chen, Delay-dependent stability for neural networks with time-varying delay, Chaos, Solitons and Fractals 33 (2007) 171–177.
[7] T. Shen, Y. Zhang, Improved global robust stability criteria for delayedneural networks, IEEE Transactions on Circuits Systems II 54 (8) (2007)715–719.
[8] Y. Shen, J. Wang, An improved algebraic criterion for global exponentialstability of recurrent neural networks with time-varying delays, IEEETransactions on Neural Networks 19 (3) (2008) 528–831.
[9] Q. Song, Exponential stability of recurrent neural networks with both time-varying delays and general activation functions via LMI approach, Neuro-computing 71 (2008) 2823–2830.
[10] Z. Wang, H. Zhang, W. Yu, Robust stability of Cohen–Grossberg neuralnetworks via state transmission matrix, IEEE Transactions on NeuralNetworks 20 (1) (2009) 169–174.
[11] Z. Wang, H. Zhang, W. Yu, Robust stability criteria for interval Cohen–Grossberg neural networks with time varying delay, Neurocomputing 72(2009) 1105–1110.
[12] Z. Wang, H. Zhang, W. Yu, Robust exponential stability analysis of neuralnetworks with multiple time delays, Neurocomputing 70 (2007) 2534–2543.
[13] S. Xu, Y. Chu, J. Lu, New results on global exponential stability of recurrentneural networks with time-varying delays, Physics Letters A 352 (2006)371–379.
[14] S. Xu, J. Lam, A new approach to exponential stability analysis ofneural networks with time-varying delays, Neural Networks 19 (2006)76–83.
[15] H. Zhang, G. Wang, New criteria of global exponential stability for a class ofgeneralized neural networks with time-varying delays, Neurocomputing 70(13–15) (2007) 2486–2494.
[16] H. Zhang, Z. Wang, Global asymptotic stability of delayed cellularneural networks, IEEE Transactions on Neural Networks 18 (3) (2007)947–950.
[17] H.G. Zhang, Y.C. Wang, Stability analysis of Markovian jumping stochasticCohen–Grossberg neural networks with mixed time delays, IEEE Transactionson Neural Networks 19 (2) (2008) 366–370.
[18] H.G. Zhang, Z. Wang, D. Liu, Robust stability analysis for interval Cohen–Grossberg neural networks with unknown time-varying delays, IEEETransactions on Neural Networks 19 (11) (2008) 1942–1955.
[19] H. Zhang, Z. Wang, D. Liu, Global asymptotic stability of recurrent neuralnetworks with multiple time-varying delays, IEEE Transactions on NeuralNetworks 19 (5) (2008) 855–873.
[20] H. Zhang, Z. Wang, D. Liu, Robust exponential stability of cellular neuralnetworks with multiple time varying delays, IEEE Transactions on Circuitsand Systems II 54 (8) (2007) 730–734.
[21] Z. Zeng, J. Wang, Global exponential stability of recurrent neural networkswith time-varying delays in the presence of strong external stimuli, NeuralNetworks 19 (2006) 1528–1537.
[22] Z. Zeng, J. Wang, Improved conditions for global exponential stability ofrecurrent neural networks with time-varying delays, IEEE Transactions onNeural Networks 17 (3) (2006) 623–635.
[23] C.D. Zheng, H. Zhang, Z. Wang, Novel delay-dependent criteria for globalrobust exponential stability of delayed cellular neural networks with norm-bounded uncertainties, Neurocomputing 72 (2009) 1744–1754.
Cheng-De Zheng was born in Shandong, China, in1966. He received the B.S. degree in mathematics fromJilin University of China in 1987, and the M.S. and Ph.D.degrees in computational mathematics from DalianUniversity of Technology of China in 1998 and 2004,respectively. He joined the Department of Mathe-matics, Dalian Jiaotong University, China, in 1987.Since 2004, he has been a Professor. His main researchinterests are stability of recurrent neural networks,numerical approximation, fuzzy set and its application.
Lai-Bing Lu was born in Shandong, China, in 1982. Hereceived the B.S. degree in mathematics from QufuNormal University, Shandong, China, in 2005. Now heis pursuing the M.S. degree in applied mathematics inDalian Jiaotong University.
Zhan-Shan Wang was born in Liaoning, China, in 1971.He received the Master Degree in control theory andcontrol engineering from Fushun Petroleum Institute,Fushun, China, in 2001. He received the Ph.D. degree incontrol theory and control engineering from North-eastern University, Shenyang, China, in 2006. He is nowan associate professor in Northeastern University. Hisresearch interests include stability analysis of recur-rent neural networks, fault diagnosis, fault tolerantcontrol and nonlinear control.