Automatica 45 (2009) 684–691

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Robust H∞ control for a class of nonlinear discrete time-delay stochastic systems

with missing measurementsI
Zidong Wang a,∗ , Daniel W.C. Ho b , Yurong Liu c , Xiaohui Liu a
a
Department of Information Systems and Computing, Brunel University, Uxbridge, Middlesex, UB8 3PH, UK
b
Department of Mathematics, City University of Hong Kong, Kowloon, Hong Kong
c
Department of Mathematics, Yangzhou University, Yangzhou 225002, PR China

article info a b s t r a c t

Article history: This paper is concerned with the problem of robust H∞ output feedback control for a class of uncertain
Received 16 January 2008 discrete-time delayed nonlinear stochastic systems with missing measurements. The parameter
Received in revised form uncertainties enter into all the system matrices, the time-varying delay is unknown with given low
25 September 2008
and upper bounds, the nonlinearities satisfy the sector conditions, and the missing measurements are
Accepted 17 October 2008
Available online 8 January 2009
described by a binary switching sequence that obeys a conditional probability distribution. The problem
addressed is the design of an output feedback controller such that, for all admissible uncertainties, the
Keywords:
resulting closed-loop system is exponentially stable in the mean square for the zero disturbance input and
Discrete nonlinear stochastic system also achieves a prescribed H∞ performance level. By using the Lyapunov method and stochastic analysis
H∞ output feedback control techniques, sufficient conditions are first derived to guarantee the existence of the desired controllers, and
Time-varying delays then the controller parameters are characterized in terms of linear matrix inequalities (LMIs). A numerical
Missing measurements example is exploited to show the usefulness of the results obtained.
Lyapunov–Krasovskii functional © 2008 Elsevier Ltd. All rights reserved.
Linear matrix inequality

1. Introduction Wang (2005), Huang and Ho (2007), Niu, Ho, and Lam (2005),
Shi, Mahmoud, Nguang, and Ismail (2006) and Shi, Mahmoud,
Time delay, stochasticity and nonlinearity are arguably three Yi, and Ismail (2005). Third, it is well known that nonlinearities
of the main sources that contribute to the complexity of a exist universally in practical systems, and therefore nonlinear
system. First, various engineering systems (e.g. electrical networks, control has been an ever hot topic in the past few decades.
turbojet engines, microwave oscillators, nuclear reactors and It is worth mentioning that, among different descriptions of
hydraulic systems) have the characteristics of time delay in signal the nonlinearities, the so-called sector nonlinearity Khalil (1996)
transmissions. So far, the stability analysis and robust control has gained much attention for deterministic systems, and both
for dynamic time-delay systems have attracted a number of the control analysis and model reduction problems have been
researchers over the past years. Second, as stochastic modeling investigated, see Han (2005) and Lam, Gao, Xu, and Wang (2005).
has had extensive applications in control and communication Recently, the control problem of nonlinear stochastic systems
problems, the stability analysis problem for linear stochastic time- has stirred renewed research interests, and a variety of nonlin-
delay systems has been studied by many authors, see e.g. Chen ear stochastic systems have been investigated by different ap-
and Zhang (2004), Gao, Lam, and Wang (2006), Gao, Wang, and proaches Deng and Krstic (2000), Deng, Krstic, and Williams
(2001), Liu, Pan, and Shi (2003) and Xie and Xie (2000). Most re-
cently, in Berman and Shaked (2006), Shu and Wei (2005) and
I This paper was not presented at any IFAC meeting. This paper was
Zhang, Chen, and Tseng (2005), an H∞ -type theory has been de-
veloped for a large class of continuous- and discrete-time non-
recommended for publication in revised form by Associate Editor Dragan Nešic
under the direction of Editor Hassan K. Khalil. This work was supported in part by linear stochastic systems. As for nonlinear stochastic time-delay
the Engineering and Physical Sciences Research Council (EPSRC) of the U.K. under systems, the relevant literature has not been very much. Most of
Grant GR/S27658/01, the City University of Hong Kong under Grant 7001992, the the existing results have been only concerned with the stability
Royal Society of the U.K. under an International Joint Project, the Natural Science analysis issue, see e.g. Fu and Liao (2003) and Yuan and Mao (2004).
Foundation of Jiangsu Province of China under Grant BK2007075, the National
Natural Science Foundation of China under Grant 60774073, and the Alexander von
The performance indices of the controlled systems, such as the
Humboldt Foundation of Germany. robustness and disturbance rejection attenuation level, have not
∗ Corresponding author. Tel.: +44 1895 266021; fax: +44 1895 251686. received enough attention despite their importance in practical
E-mail address: [email protected] (Z. Wang). applications.
0005-1098/$ – see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2008.10.025
 Z. Wang et al. / Automatica 45 (2009) 684–691 685

In almost all the literature mentioned above, the assumption ỹ(k) = Cx(k) + Cd x(k − d(k)) + φ(Sx(k)),
of consecutive measurements has been made implicitly. Unfortu- z (k) = Lx(k) + B2 u(k),
nately, in many practical applications, such an assumption does not
hold. For example, due to sensor temporal failure or network trans- x(j) = ψ(j), j = −dM , −dM + 1, . . . , −1, 0,
mission delay/loss, at certain time points, the system measurement where x(k) ∈ R is the state vector, u(k) ∈ Rm is the control input,
n

may contain noise only, indicating that the real signal is missing. z (k) ∈ Rp is the controlled output, and w(k) ∈ Rq is the exogenous
Such a phenomenon is also called dropout or intermittence in net- disturbance signal which is assumed to belong to l2 [0, ∞).
worked control systems, see e.g. Sinopoli et al. (2004) and Wang, For the system (Σ 0 ), the positive integer d(k) denotes the time-
Yang, Ho, and Liu (2006). One of the most popular ways to de- varying delay satisfying
scribe the missing measurement is to view it as a Bernoulli dis-
tributed (binary switching) white sequence specified by a condi- dm ≤ d(k) ≤ dM , k ∈ N+ (1)
tional probability distribution in the output equation. The Bernoulli where dm and dM are known positive integers. ψ(j) (j =
distribution description was first proposed in Nahi (1969) to deal −dM , −dM + 1, . . . , −1, 0) are the initial conditions. B1 , B2 , C , Cd ,
with the optimal recursive filtering problem, and has then been D1 , D2 , S and L are known real constant matrices, and the matrices
used in Hounkpevi and Yaz (2007), Sinopoli et al. (2004), Wang, A(k), Ad (k), E (k) and Ed (k) are time-varying matrices of the form
Ho, and Liu (2003), Wang et al. (2006) and Wang, Yang, Ho, and Liu A(k) = A + ∆A(k), Ad (k) = Ad + ∆Ad (k), E (k) = E + ∆E (k),
(2007) for various control and filtering problems of linear systems Ed (k) = Ed + ∆Ed (k). Here, the constant matrices A, Ad , E and
with probabilistic measurement losses. However, so far, the robust Ed are known while ∆A(k), ∆Ad (k), ∆A(k), ∆Ad (k) are unknown
H∞ output feedback control problem for nonlinear stochastic time- matrices representing time-varying parameter uncertainties that
delayed systems with missing measurements has not been fully in- are assumed to satisfy the following admissible condition:
vestigated, especially for discrete-time cases, which motivates us
∆A(k) ∆Ad (k) ∆E (k) ∆Ed (k)
 
to shorten such a gap in the present investigation.
In this paper, we consider the robust H∞ control problem = MF (k) N1 N2 N3 N4 ,
 
(2)
for a class of uncertain discrete stochastic time-delay systems
where M and Ni (i = 1, 2, 3, 4) are known real constant matrices,
involving sector nonlinearities and missing measurements. The
and F (k) is the unknown time-varying matrix-valued function
parameter uncertainties are assumed to be norm-bounded, the
delays are time-varying, and the sector nonlinearities appear subject to F T (k)F (k) ≤ I, ∀k ∈ N+ .
In this paper, without loss of generality, we always assume that
in the system states and delayed states. Similar to Hounkpevi
and Yaz (2007), Sinopoli et al. (2004) and Wang et al. (2003, f (0) = 0, fd (0) = 0 and φ(0) = 0. For vector-valued functions
2006, 2007), the missing measurements are characterized as a f , fd , φ , we assume:
binary switching sequence satisfying a conditional probability [f (x) − f (y) − U1 (x − y)]T
distribution. An effective linear matrix inequality (LMI) approach
× [f (x) − f (y) − U2 (x − y)] ≤ 0, ∀ x, y ∈ R n , (3)
is proposed to design the output feedback controllers such that,
for all admissible nonlinearities, time-delays and probabilistic [fd (x) − fd (y) − V1 (x − y)]T
measurement missing, the overall uncertain closed-loop system is × [fd (x) − fd (y) − V2 (x − y)] ≤ 0, ∀x, y ∈ Rn , (4)
robustly asymptotically stable in the mean square and a prescribed
H∞ disturbance rejection attenuation level is guaranteed. We
[φ(x) − φ(y) − W1 (x − y)]T
first establish the sufficient conditions for the uncertain nonlinear × [φ(x) − φ(y) − W2 (x − y)] ≤ 0, ∀x, y ∈ Rm , (5)
stochastic time-delay systems to be stable in the mean square, where U1 , U2 , V1 , V2 ∈ Rn×n
, W1 , W2 ∈ Rm×m
are known real
and then derive the explicit expression of the desired controller constant matrices, and U1 − U2 , V1 − V2 , W1 − W2 are positive
gains. A numerical example is provided to show the usefulness and definite matrices.
effectiveness of the proposed design method.
Notations: Throughout this paper, N is the set of natural Remark 1. It is customary that the nonlinear functions f , fd , φ
numbers and N+ stands for the set of nonnegative integers; Rn and are said to belong to sectors [U1 , U2 ], [V1 , V2 ] and [W1 , W2 ],
Rn×m denote, respectively, the n dimensional Euclidean space and respectively (see Khalil (1996)). The nonlinear descriptions in (3)–
the set of all n × m real matrices. The superscript ‘‘T’’ denotes the (5) are quite general that include the usual Lipschitz conditions
transpose and the notation X ≥ Y (respectively, X > Y ) where X as a special case. Note that both the control analysis and model
and Y are symmetric matrices, means that X − Y is positive semi- reduction problems for systems with sector nonlinearities have
definite (respectively, positive definite). I is the identity matrix been intensively studied, see e.g. Han (2005) and Lam et al. (2005).
with compatible dimension. l2 [0, ∞) denotes the space of square
summable infinite vector sequences. The notation k · k stands for In the system (Σ 0 ), ỹ(k) ∈ Rm is the ideal system output
the usual l2 [0, ∞) norm while | · | refers to the Euclidean vector that always contains the real signal. However, as discussed in
norm. Moreover, let (Ω , F , {Ft }t ≥0 , P ) be a complete probability the introduction, in practical engineering systems (e.g., networked
space with a filtration {Ft }t ≥0 satisfying the usual conditions control systems) the system output usually contains probabilistic
(i.e., the filtration contains all P-null sets and is right continuous). missing data. In such a case, the actual system output can be
E{·} stands for the mathematical expectation operator with respect described preferably by
to the given probability measure P. The asterisk ∗ in a matrix is y(k) = γk ỹ(k) + D2 w(k)
used to denote term that is induced by symmetry. Sometimes, the = γk [Cx(k) + Cd x(k − d(k)) + φ(Sx(k))]
arguments of a function will be omitted in the analysis when no
confusion can arise. + D2 w(k), (6)
where the stochastic variable γk ∈ R is a Bernoulli distributed
2. Problem formulation white sequence specified by the following distribution law:

Consider the following uncertain discrete nonlinear system Prob{γk = 1} = E{γk } = β, (7)
with time delays of the form: Prob{γk = 0} = 1 − E{γk } = 1 − β (8)
(Σ ) : x(k + 1) = A(k)x(k) + Ad (k)x(k − d(k))
0
with β > 0 a known constant. Obviously, for the stochastic
+ E (k)f (x(k)) + Ed (k)fd (x(k − d(k))) variable γk , we have the mean value E{γk } = β and variance
+ B1 u(k) + D1 w(k), σ 2 = β(1 − β).
686 Z. Wang et al. / Automatica 45 (2009) 684–691
   
Remark 2. The system measurement mode (6)–(8), which can be A B 1 CK Ad
Ā = , Ād = , (9)
used to represent missing measurements or uncertain observa- β BK C AK β BK Cd
tions, was first introduced in Nahi (1969), and has been subse-    
E Ed
quently studied in many papers, see e.g. Hounkpevi and Yaz (2007), Ē = , Ēd = ,
0 0
Sinopoli et al. (2004) and Wang et al. (2003, 2006, 2007). Note that
when the real signal is missing (i.e., γk = 0), the system measure- ∆A(k) ∆Ad (k)
   
0
∆Ā(k) = , Ād (k) = ,
ment contains noise only. Such a case does happen in practice. For 0 0 0
example, in target tracking control, due to high maneuverability
∆E (k) ∆Ed (k)
   
of the tracked target, there may be a nonzero probability that any ∆Ē (k) = , ∆Ēd (k) = ,
0 0
observation consists of noise alone if the target is absent, i.e., the
measurements are not consecutive but contain missing observa- M̄ = [M T 0]T , N̄1 = [N1 0].
tions. On the other hand, there are still other ways to model the
missing phenomenon, such as those using randomly delayed sen- It is clear that
sor output and probabilistic jumps. Ā(k) = Ā + ∆Ā(k), Ād (k) = Ād + ∆Ād (k),
Replacing ỹ(k) in (Σ ) by y(k), we have the following system to
0
Ē (k) = Ē + ∆Ē (k), Ēd (k) = Ēd + ∆Ēd (k).
be investigated in this paper:
In this paper, we shall focus on the robust stabilization problem
(Σ ) : x(k + 1) = A(k)x(k) + Ad (k)x(k − d(k)) whose purpose is to design a full-order exponentially stable
+ E (k)f (x(k)) + Ed (k)fd (x(k − d(k))) H∞ controller for the system (Σ ) via output feedback. More
+ B1 u(k) + D1 w(k), specifically, an output feedback controller Σo is to be designed
such that (i) the closed-loop system (Σc ) is exponentially stable
y(k) = γk [Cx(k) + Cd x(k − d(k)) + φ(Sx(k))]
in the mean square; and (ii) Under zero initial condition, the
+ D2 w(k),
controlled output z satisfies kz k2 ≤ γ kwk2 for any nonzero w ∈
z (k) = Lx(k) + B2 u(k), l2 , where γ > 0 is a prescribed constant.
x(j) = ψ(j), j = −dM , −dM + 1, . . . , 0. The following lemma will be used in establishing our main
results.
In this paper, for the system (Σ ), we consider the following full-
order dynamic output feedback controller: Lemma 1. Let D , S and F be real matrices of appropriate dimensions
with F T F ≤ I . Then, for any scalar ε > 0, we have D F S + (D F S )T ≤
(Σo ) : x̂(k + 1) = AK x̂(k) + BK y(k),
ε −1 DD T + ε S T S .
u(k) = CK x̂(k),
where x̂(k) ∈ Rn is the controller state, AK , BK and Ck are the 3. Main results
controller parameters to be determined. Due to the existence of the
stochastic variable γk ∈ R, we will need the following definition for First of all, let us deal with the robust stability analysis problem
the exponential stability in the mean square. and derive a sufficient condition under which the closed-loop
system (Σc ) with given controller and w(k) = 0 is robustly
Definition 1. The controlled system (Σ ) is said to be exponentially exponentially stable in the mean-square.
stable in the mean square if, in case of w(k) = 0 and for all admissi-
ble uncertainties, there exist constants µ > 0 and 0 < α < 1 such Theorem 1. Let the controller parameters AK , BK and CK be given and
that the admissible condition hold. Then, the closed-loop system (Σc ) with
w(t ) = 0 is robustly exponentially stable in the mean square if there
E|x(k)|2 ≤ µα k sup E|ψ(i)|2 . exist two matrices P > 0, Q > 0 and two positive constant scalars ε ,
−dM ≤i≤0
λ such that the following matrix inequality holds:
The combination of the controller (Σo ) and the system (Σ )  Π ∗ ∗ ∗ ∗ ∗ ∗ ∗ 
yields the following closed-loop system:
 0 Πa ∗ ∗ ∗ ∗ ∗ ∗ 
(Σc ) : η(k + 1) = Ã(k)η(k) + Ãd (k)H η(k − d(k))  −Ŭ T H
 2 0 −I ∗ ∗ ∗ ∗ ∗  
+ Ē (k)f (x(k)) + Ēd (k)fd (x(k − d(k)))  0 −λV̆2T 0 −λI ∗ ∗ ∗ ∗ 
 
Ψ =
+ γk B̄K φ(Sx(k)) + Dw(k), −W2T H 0 0 0 −I ∗ ∗ ∗ 

 Ā Ād Ē Ēd 0 Πb ∗ ∗ 
 
z (k) = L̄η(k),  σ C̄ σ C̄d 0 0 σ B̄K 0 −P − 1 ∗ 
where N̄1 N2 N3 N4 0 0 0 −ε I
T
x(k) A(k) < 0,
    
I B1 CK (10)
ηk = , H = n×n , Ã(k) = ,
x̂(k) 0n×n γk BK C AK
with Πa = −Q − λV̆1 and Πb = −P −1
+ ε M̄ M̄ , where T

Ad (k) E (k) E (k)

     
Ãd (k) = , Ē (k) = , Ēd (k) = d ,
γk BK Cd 0 0 Ŭ1 = (U1T U2 + U2T U1 )/2; Ŭ2 = −(U1T + U2T )/2; (11)

V̆1 = ( )/2; )/2;

 T T
V1T V2 V2T V1 V1T V2T
   
0 D1 L + V̆2 = −( + (12)
B̄K = , D= , L̄ = T T .
BK BK D2 CK B2 W1 = (S T W1T W2 S + S T W2T W1 S )/2; (13)
For convenience, we denote
W2 = −(S T
W1T +S T
W2T )/2; (14)
A(k) Ad (k)
   
B1 CK
Ā(k) = , Ād (k) = , Π = −P + (dM − dm + 1)H QH − H Ŭ1 H T T
β BK C AK β BK Cd
    − H T W1 H , (15)
0 0 0 √
C̄ = , C̄d = , and σ = β(1 − β) with β = E{γk }.
BK C 0 BK Cd
 Z. Wang et al. / Automatica 45 (2009) 684–691 687

Proof. For the stability analysis of the system (Σc ), we construct where Ŭ1 , Ŭ2 , V̆1 , V̆2 , W1 and W2 are defined in (11) and (13). Then,
the following Lyapunov–Krasovskii functional: it implies from (23), (24) and (26) that
E{∆V (k)} ≤ E ξ0T (k)Ψ1 (k)ξ0 (k) + ξ0T (k)AT0 (k)P A0 (k)ξ0 (k)

V (k) = V1 (k) + V2 (k) + V3 (k), (16)

where V1 (k) = ηT (k)P η(k), V2 (k) =

Pk−1
ηT (i)H T QH η(i) and + σ 2 ξ0T (k)B0T P B0 ξ0 (k)
i=k−d(k) ( T  T
V3 (k) = i=j η (i)H QH η(i).
Pk−d Pk−1
T T
η(k) H Ŭ1 H H T Ŭ2 η(k)
m
 
j=k−dM +1
−E
Calculating the difference of V (k) along the system (Σc ) with f (x(k)) Ŭ2T H I f (x(k))
w(k) = 0 and taking the mathematical expectation, we have T 
x(k − d(k)) x(k − d(k))
  
V̆1 V̆2
+λ
3
X fd (x(k − d(k))) V̆2T I fd (x(k − d(k)))
E{∆V (k)} = E{∆Vi (k)}, (17)
T  T )
η(k) H W1 H H T W2 η(k)
  
i=1
+
where φ(Sx(k)) W2T H I φ(Sx(k))

E{∆V1 (k)} = E{V1 (k + 1) − V1 (k)} = E ξ0T (k) Ψ2 (k) + AT0 (k)P A0 (k) + σ 2 B0T P B0 ξ0 (k) ,
  
(27)
ÂT0 (k)P Â0 (k)

=E where
+ σ B̂0T (k)P B̂0 (k) −ηT (k)P η(k)
2
Π −H T Ŭ2 −H T W 2
 
(18) 0 0
with
 0 −Q − λV̆1 0 −λV̆2 0 
Ψ2 (k) =  .
 
T
 −Ŭ2 H 0 −I 0 0
Â0 (k) = Ā(k)η(k) + Ād (k)H η(k − d(k)) + Ē (k)f (x(k))

 0 −λV̆2T 0 −λI 0 
+ Ēd (k)fd (x(k − d(k))) + β B̄K φ(Sx(k)), (19) −W2T H 0 0 0 −I
B̂0 (k) = C̄ η(k) + C̄d H η(k − d(k)) + B̄K φ(Sx(k)). (20) It is clear from Ψ < 0 that there exists a sufficiently small scalar
ε0 > 0 such that
Similarly, it is not difficult to check that
( Ψ + ε0 diag{I2n×2n , 0} < 0. (28)

E{∆V2 (k)} ≤ E ηT (k)H T QH η(k) In order to deal with the exponential stability, we shall first
prove that (28) indicates
− ηT (k − d(k))H T QH η(k −)d(k))
−dm
kX
Ψ2 (k) + ε0 diag{I2n×2n , 0} + AT0 (k)P A0 (k) + σ 2 B0T P B0 < 0. (29)
+ ηT (i)H T QH η(i) , (21) In fact, by the Schur complement, the above inequality is
i=k−dM +1 equivalent to
(
E{∆V3 (k)} = E (dM − dm )ηT (k)H T QH η(k) Ψ3 (k)
Π + ε I ∗ ∗ ∗ ∗ ∗ ∗ 
) 0
k−dm
X  0 −Q − λV̆1 ∗ ∗ ∗ ∗ ∗
ηT (i)H T Q η(i) .

− (22)  −Ŭ2T H

0 −I ∗ ∗ ∗ ∗


i=k−dM +1 = 0 −λV̆2T 0 −λI ∗ ∗ ∗
 

 −W2T H 0 0 0 −I ∗ ∗
 
Substituting (18)–(22) into (17) leads to 
Ā(k) Ād (k) Ē (k) Ēd (k) 0 −P −1 ∗
 
E{∆V (k)} ≤ E ξ0T (k)Ψ1 (k)ξ0 (k) + ξ0T (k)AT0 (k)P σ C̄ σ C̄d σ B̄K −P −1

0 0 0
< 0,
× A0 (k)ξ0 (k) + σ 2 ξ0T (k)B0T P B0 ξ0 (k) , (23)
which can be decomposed as follows:
where
Ψ3 (k) = Ψ3 + ∆Ψ3 (k), (30)
ξ0 (k) = [ηT (k) xT (k − d(k)) f T (x(k))
where
fdT (x(k − d(k))) φ T (Sx(k))]T ,
Ψ3
A0 (k) = [Ā(k) Ād (k) Ē (k) Ēd (k) 0], Π + ε I
0 ∗ ∗ ∗ ∗ ∗ ∗ 
B0 = [C̄ C̄d 0 0 B̄K ],  0 −Q − λV̆1 ∗ ∗ ∗ ∗ ∗ 
Ψ1 (k) = diag {Π1 , −Q , 0, 0, 0}  −Ŭ2T H 0 −I ∗ ∗ ∗ ∗
 

:=  0 −λV̆2T 0 −λI ∗ ∗ ∗ ,
 
with Π1 = −P + (dM − dm + 1)H T QH .  −W2T H

0 0 0 −I ∗ ∗


It follows readily from (3)–(5) that 
Ā Ād Ē Ēd 0 −P −1 ∗

T  T σ C̄ σ C̄d 0 0 σ B̄K 0 −P −1
η(k) H Ŭ1 H H T Ŭ2 η(k)
  
≤ 0, (24) (31)
f (x(k)) Ŭ2T H I f (x(k))  0 ∗ ∗ ∗ ∗ ∗ ∗
T  ∗ ∗ ∗ ∗ ∗
x(k − d(k)) x(k − d(k))  0 0
  
V̆1 V̆2
≤ 0, (25)  0 0 0 ∗ ∗ ∗ ∗
fd (x(k − d(k))) V̆2T I fd (x(k − d(k))) ∆Ψ3 (k) :=  .
 
 0 0 0 0 ∗ ∗ ∗ (32)
T  T  0 0 0 0 0 ∗ ∗
η(k) H W1 H H T W2 η(k)
  
∆Ā(k) ∆Ād (k) ∆Ē (k) ∆Ēd (k) 0 0 ∗
 
≤ 0, (26)
φ(Sx(k)) W2 HT
I φ(Sx(k)) 0 0 0 0 0 0 0
688 Z. Wang et al. / Automatica 45 (2009) 684–691

From (2) and Lemma 1, it follows that N −1

X
≤ dM µdM max E{|ψ(i)|2 } + dM µdM µi E |ηi |2 .

(40)
∆Ψ3 (k) = M̃F (k)Ñ + (M̃F (k)Ñ ) T −dM ≤i≤0
i=0

≤ ε M̃ M̃ T + ε −1 Ñ T Ñ , (33) Then, from (39) and (40), one has

µN E{V (N )} ≤ E{V (0)} + ω1 (µ) + dM µdM ω2 (µ)
 
where M̃ = [0 0 0 0 0 M̄ T 0]T and Ñ = [N̄1 N2 N3 N4 0 0 0]. With
(31) and (33), one has N −1
X
× µk E{|ηk |2 } + dM µdM ω2 (µ)
Ψ3 (k) ≤ Ψ4 + ε −1 Ñ T Ñ , (34) k=0

where × max E{|ψ(i)|2 }. (41)

−dM ≤i≤0
Π + ε I ∗ ∗ ∗ ∗ ∗ ∗ 
0
Πa ∗ ∗ ∗ ∗ ∗  Let ρ0 = λmin (P ) and ρ = max{ρ1 , ρ2 }. It is obvious that
 0
 −Ŭ T H 0 −I ∗ ∗ ∗ ∗ 
 2  E{V (N )} ≥ ρ0 E{|ηN |2 }. (42)
Ψ4 :=  0 −λV̆2T 0 −λI ∗ ∗ ∗  (35)
 
 −W2T H

0 0 0 −I ∗ ∗ 
 It also follows from (37) that
 Ā Ād Ē Ēd 0 Πb ∗ 
E{V (0)} ≤ ρ max E{|ψ(i)|2 }. (43)
σ C̄ σ C̄d 0 0 σ B̄K 0 −P − 1 −dM ≤i≤0

with Πa and Πb defined in Theorem 1. In addition, it can be verified that there exists a scalar µ0 > 1 such
By Schur complement, the inequality (28) is true if and only if that
the right-hand side of (34) is negative definite, which implies that d
Ψ3 (k) < 0. Therefore, (29) holds, and it follows from (27) and (29) ω1 (µ0 ) + dM µ0M ω2 (µ0 ) = 0. (44)
that Substituting (42)–(44) into (41), we obtain
E {∆V (k)} ≤ −ε0 E{|ηk |2 }. (36) 
1
N
2
E{|ηN | } ≤ c0 max E{|ψ(i)|2 },
We are now in a position to proceed with the exponential µ0 −dM ≤i≤0
stability analysis of the system (Σc ). According to the definition  
d
of V (k), we have where c0 = ρ + dM µ0M ω2 (µ0 ) . This indicates that the
ρ0
1

k−1
closed-loop system (Σc ) with w(k) = 0 is robustly exponentially
stable in the mean square. The proof is complete.
X
E{V (k)} ≤ ρ1 E{|ηk |2 } + ρ2 E{|ηi |2 }, (37) 
i=k−dM Next, we will analyze the H∞ performance of the closed-loop
where ρ1 = λmax (P ) and ρ2 = (dM − dm + 1)λmax (Q ). For any system (Σc ).
scalar µ > 1, the above inequality, together with (36), implies that
Theorem 2. Let the controller parameters AK , BK and CK be given and
µk+1 E{V (k + 1)} − µk E{V (k)} γ be a prespecified positive constant. Then, the closed-loop system
= µk+1 E{∆V (k)} + µk (µ − 1)E{V (k)} (Σc ) is robustly exponentially stable in the mean square for w(k) = 0
and satisfies kz k2 ≤ γ kwk2 under the zero initial condition for any
k−1
X nonzero w ∈ l2 [0, +∞), if there exist two positive definite matrices
≤ ω1 (µ)µk E{|ηk |2 } + ω2 (µ) µk E |ηi |2 ,

(38)
P, Q and two positive scalars ε, λ such that the following matrix
i=k−dM
inequality holds:
where ω1 (µ) = −µε0 + (µ − 1)ρ1 and ω2 (µ) = (µ − 1)ρ2 .
Furthermore, for N ≥ dM + 1, summing up both sides of (38) Φ
from 0 to N − 1 with respect to k, we have  Υ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 
N −1  0 Πa ∗ ∗ ∗ ∗ ∗ ∗ ∗ 
 −Ŭ2T H
X 
0 −I ∗ ∗ ∗ ∗ ∗ ∗ 

µN E{V (N )} − E{V (0)} ≤ ω1 (µ) µk E{|ηk |2 }
−λV̆2T
 
 0 0 −λI ∗ ∗ ∗ ∗ ∗ 
k=0  T

=
−W2 H 0 0 0 −I ∗ ∗ ∗ ∗ 
N −1 X
X k−1  0 0 0 0 0 −γ 2 I ∗ ∗ ∗ 
+ ω2 (µ) µk E |ηi |2 .
  
(39) Πb ∗ ∗ 
 Ā Ād Ē Ēd 0 D
 
σ C̄ σ C̄d 0 0 σ B̄K 0 0 −P −1 ∗
 
k=0 i=k−dM
N̄1 N2 N3 N4 0 0 0 0 −εI
Note that for dM ≥ 1,
<0 (45)
N −1 X
k−1
where Πa , Πb , L̆1 , L̆2 , Ŭ1 , Ŭ2 , W1 and W2 are defined as in Theorem 1,
X
µk E |ηi |2

k=0 i=k−dM and
Υ = −P + (dM − dm + 1)H T QH − H T Ŭ1 H − H T W1 H + L̄T L̄.
+dM N −1−dM i+dM
!
−1 iX
X X X N −1
X N −1
X
µk E |ηi |2

≤ + +
i=−dM k=0 i =0 k=i+1 i=N −1−dM k=i+1 Proof. It is not difficult to verify that Φ < 0 implies Ψ < 0.
−1
X N −1−dM
X According to Theorem 1, the closed-loop system (Σc ) is robustly
µi+dM E |ηi |2 + dM µi+dM E |ηi |2
 
≤ dM exponentially stable in the mean square.
i=−dM i=0 We now deal with the H∞ performance of the closed-
N −1 loop system. Construct the same Lyapunov–Krasovskii functional
candidate V (k) as in Theorem 1. A similar calculation as in
X
µi+dM E |ηi |2

+ dM
i=N −1−dM Theorem 1 leads to
 Z. Wang et al. / Automatica 45 (2009) 684–691 689

E{∆V (k)} ≤ E{ξ T (k)Φ1 (k)ξ (k) + ξ T (k)AT (k)P A(k)ξ (k) Π ∗ ∗ ∗ ∗ ∗
 
0 −Q − λV̆1 ∗ ∗ ∗ ∗ 
+σ 2 ξ T (k)B T P B ξ (k)}, (46)
Ξ

0 −I ∗ ∗ ∗ 

Φ011 = 1 ,
where 0 −λV̆2T 0 −λI ∗ ∗ 
Ξ 0 0 0 −I ∗ 
2
ξ (k) = [η (k) x (k − d(k)) f (x(k))
T T T
0 0 0 0 0 −γ 2 I
fdT (x(k − d(k))) φ T (Sx(k)) w T (k)]T , Θ
1 Θ2 Θ3 Θ4 0 Θ5 
A(k) = [Ā(k) Ād (k) Ē (k) Ēd (k) 0 D], Θ6 Θ7 0 0 Θ8 0
B = [C̄ C̄d 0 0 B̄K 0], Ξ3 N2 N3 N4 0 0
Φ021 = ,
Ξ4 0 0 0 0 0
Φ1 (k) = diag{−P + (dM − dm + 1)H QH , −Q , 0, 0, 0, 0}. T
Ξ5 0 0 0 0 0
 
In order to deal with the H∞ performance of the system (Σc ), 0 0 0 0 0 0
we introduce
Π ∗ ∗ ∗ ∗ ∗
 
n
X  0 Π ∗ ∗ ∗ ∗ 
J (n) = E z T (k)z (k) − γ 2 w T (k)w(k) ,
 
(47)  0 0 −ε I ∗ ∗ ∗ 
Φ022 = ,
 
k=0
 0 0 0 −I ∗ ∗  
where n is non-negative integer.
 0 0 0 0 −Q̂ −1 ∗ 
Under the zero initial condition, from (24)–(26), (46) and (47), Ξ6 0 0 0 0 −ε −1 I
one has 
−X −I AX + B1 Λ
 
A

Π= , Θ1 = ,
n
X −I −Y Ω YA + β Γ C
J (n) = E z T (k)z (k) − γ 2 w T (k)w(k) + ∆V (k)
 
     
Ad E Ed
k=0
Θ2 = , Θ3 = , Θ4 = ,
− EV ( n + 1 ) YAd + β Γ Cd YE YEd
n
   
D1 0 0
Θ5 = , Θ6 = σ ,
X
η (k)L̄T L̄η(k) − γ 2 w T (k)w(k) + ∆V (k)
 T 
≤E YD1 + Γ D2 Γ CX Γ C
k=0    
n 0 0
Θ7 = σ , Θ8 = σ , Ξ1 = −Ŭ2T [X I ],
Γ Cd Γ
X
≤E [ξ T (k)Φ2 (k)ξ (k) + ξ T (k)AT (k)P A(k)ξ (k)
k=0
Ξ2 = −W2T [X I ], Ξ3 = N1 [X I ], Ξ4 = [LX + B2 Λ L] ,
+ σ ξ (k)B P B ξ (k)],
2 T T
(48)
Ξ5 = [X I] , Ξ6 = ε M T [I Y ] ,
where Q̂ = (dM − dm + 1)Q − Ŭ1 − W1 ,
Φ2 (k) and L̆1 , L̆2 , Ŭ1 , Ŭ2 , W1 and W2 are defined as in Theorem 1.
Υ −H T Ŭ2 −H T W 2 In this case, the parameters of the output feedback controller can
 
0 0 0
 0 −Q − λV̆1 0 −λV̆2 0 0  be designed as
 
 −Ŭ T H 0 −I 0 0 0 . AK = R−1 (Ω − YAX − β Γ CX + YB1 Λ)G−T ,

= 2 (50)
 0 −λV̆2T 0 −λI 0 0 
BK = R−1 Γ , CK = ΛG−T , (51)
 
 −W T H 0 0 0 −I 0 
2
2
0 0 0 0 0 −γ I where G and R are any matrices satisfying

By Schur complement and along the same line as in the proof of RGT = I − YX . (52)
Theorem 1, we can show that J (n) < 0. The details are omitted
here for space saving. Letting n → ∞, we obtain kz k2 ≤ γ kwk2 . Proof. First, it follows readily from (49) that
This completes the proof of the theorem.   
−X −I
Finally, we are ready to deal with the design of H∞ output Π= < 0,
−I −Y
feedback controller for the system (Σ ) and give the main result
of this paper in the following theorem where the technique of which, by Schur complement, implies that Y − X −1 > 0 and I − YX
linearizing change of variables Gao and Wang (2003) and Gao and is nonsingular. Therefore, there always exist nonsingular matrices
Chen (2007) is used. G and R such that Eq. (52) holds. Now, adopting the similar method
as in Gahinet and Apkarian (1994), we define
Theorem 3. Let γ > 0 be a given positive constant. Then, for    
X I I Y
the nonlinear stochastic system (Σ ) with missing measurements, an Π1 = , Π2 = , P = Π2 Π1−1 , (53)
output feedback controller can be designed such that the closed-loop GT 0 0 RT
system (Σc ) is robustly exponentially stable in the mean-square for and then
w(k) = 0 and also satisfies kz k2 ≤ γ kwk2 under the zero initial  
condition for any nonzero w , if there exist four real constant matrices P =
Y R
, (54)
Y > 0, Ω , Γ and Λ, and a scalar λ > 0 such that, for given matrices RT Z
X > 0, Q > 0 and a scalar ε > 0, the following LMI holds:
where Z = G−1 X (Y − X −1 )XG−T > 0.
It is clear that
 
Φ011 Φ21
T
Φ0 := <0 (49)
Φ21 Φ22 Z − RT Y −1 R = RT (XY − I )−1 (Y − X −1 )(YX − I )−1 R > 0,
where which, by Schur complement, implies that P > 0.
690 Z. Wang et al. / Automatica 45 (2009) 684–691

By a tedious calculation, one can rewrite LMI (49) as are more restrictive than the sector conditions, see e.g. Gao and
Wang (2003) and Lu and Feng (2007); and (2) most existing
− Π1T P Π1 ∗ ∗ ∗ ∗ ∗

literature has been concerned with continuous-time systems, see
 0 −Q − λV̆1 ∗ ∗ ∗ ∗ e.g. Han (2005) and Lam et al. (2005), and the corresponding
 −Ŭ2T H Π1 results on discrete-time systems are relatively few. Moreover, it

0 −I ∗ ∗ ∗
−λV̆2T would be interesting to further investigate the systems with Itô-

 0 0 −λI ∗ ∗
−W2T H Π1
 type stochastic disturbances Wang, Liu, and Liu (2008), and the
0 0 0 −I ∗
corresponding results will appear in the near future.
−γ 2 I

 T0 0 0 0 0

 Π2 ĀΠ1 Π2T Ād Π2T Ē Π2T Ēd 0 Π2T D 4. Numerical example
 σ Π2T C̄ Π1 σ Π2T C̄d σ Π2T B̄K

0 0 0
 N̄1 Π1 Consider the system (Σ ) with the following parameters:

N2 N3 N4 0 0
 L̄Π1

0 0 0 0 0 0.2 0 0.1 0.1 −0.1 0
" # " #
H Π1 0.1 −0.3 0.1 , 0.1 −0.2 ,

0 0 0 0 0 A= Ad = 0
0 0 0 0 0 0 0.1 0 −0.2 0 −0.2 −0.1
∗ ∗ ∗ ∗ ∗ ∗  0.1 0.1 0 0.1 0 0.1
" # " #
∗ ∗ ∗ ∗ ∗ ∗  E= 0.1 0.2 0 , Ed = 0.1 0.2 0 ,
∗ ∗ ∗ ∗ ∗ ∗   0.1 0.2 0.1 0.1 0 0.1
∗ ∗ ∗ ∗ ∗ ∗ 
1 0.1 −0.2 0
 " # " #
∗ ∗ ∗ ∗ ∗ ∗ 
B1 = 0.2 1 , D1 = −0.1 0.1 ,
∗ ∗ ∗ ∗ ∗ ∗ 

0 −0.1 0 0.2
−Π2T P −1 Π2 ∗ ∗ ∗ ∗ ∗ 

0.8 0.7 −0.6 0.8 0.9
   
− Π2T P −1 Π2 ∗ ∗ ∗ ∗  1

0 C = , Cd = ,
0 0 −ε I ∗ ∗ ∗ 
 −0.6 0.9 0.6 0.8 0.7 0.5
0 0 0 −I ∗ ∗  
0.6 0.7 0.2
 
0.9 −0.1

, D2 = ,

−Q̂ −1 S=
0 0 0 0 ∗  0.5 0.8 0.3 0.5 0.8
ε M̄ T Π2 0 0 0 0 −ε −1 I
−0.1 0.1 0.2 0.1
   
0
< 0. L= , B2 = ,
(55) −0.1 −0.1 0 0.1 0.2
Pre- and post-multiplying (55) by −0.2 0.1 0.2 0.1 0
" # " #
D2 = 0.1 0.2 , U1 = V1 = 0.1 0.3 0 ,
diag{Π1−T , I , I , I , I , I , Π2−T , Π2−T , I , I , I , I }
0 0.3 −0.1 0.1 0.3
and diag{Π1−1 , I , I , I , I , I , Π2−1 , Π2−1 , I , I , I , I }, respectively, we −0.2 0.1 0
" #
0.3 0.1
 
can show by Schur complement that the inequality (45) holds. W1 = , U2 = V2 = 0.1 −0.3 −0.1 ,
0 0.4
Therefore, from Theorem 2, the desired result follows and the proof −0.1 0 −0.3
is complete.  " #T
0.1 0.1
" #
−0.3
 
0
W2 = , M = 0.1 , Ni = 0.1 ,
Remark 3. The robust H∞ controller design problem is solved −0.1 −0.4
in Theorem 3 for the addressed uncertain nonlinear stochastic 0.1 0.1
time-delay systems with missing measurements. An LMI-based (i = 1, ..., 4), β = 0.8, dm = 2, dM = 3.
sufficient condition is derived for the existence of state feedback In this example, the H∞ performance level is taken as γ = 0.9.
controllers, which ensures the robust exponential stability in the In order to design output feedback controller, we first choose ε =
mean square for the resulting closed-loop system and also reduces 0.4, Q = diag{0.4, 0.5, 0.4}, X = diag{0.5, 0.6, 0.5}.
the effect of the disturbance input on the controlled output to With the above parameters and by using the Matlab LMI
a prescribed level for all admissible uncertainties. The feasibility Toolbox, we can solve the LMI (49). Furthermore, we choose
of the controller design problem can be readily checked by the "
4 1 0
# "
−0.5412 0.2249 0.3215
#
solvability of an LMI, and this can be done by resorting to the G= 0 2 1 , R = 0.4105 −0.9993 0.2696 .
Matlab LMI toolbox. 1 0 3 −0.1454 1.0047 −0.5463
Remark 4. Note that Lemma 1 has been used in the proof of Therefore, by (50) and (51), the parameters of the desired output
Theorem 1 to tackle the parameter uncertainties, hence certain feedback controller can be designed as
3.7847
−2.3238 −1.5778
" #
conservatism might be introduced. Such conservatism can be
significantly reduced by imposing more constraints on the LMI AK = 2.9212
−2.5302 −1.0455 ,
parameters ε . A selection of more general Lyapunov functional 3.8859
−3.0151 −1.6630
in latest literature (e.g. Gao and Chen (2007)) could reduce the "
2.4653 −2.8886
#
possible conservatism as well. The conservatism reduction leaves BK = 1.5387 −2.2687 ,
an interesting topic for our future research. 1.8090 −3.7140
−0.0197 −0.0302 0.0152
 
Remark 5. It is noticed that the only stochastic term involved in CK = ,
the system (Σ ) is from the stochastic variable γk in (6). If γk = 1 −0.0462 0.0770 0.0113
all the time, our main results will reduce to the case of robust with which, according to Theorem 3, the addressed uncertain non-
H∞ control for a class of deterministic nonlinear discrete time- linear stochastic time-delay systems with missing measurements
delay systems. Such specialized results are still believed to be is robustly exponentially stable in the mean square and the effect
new for two reasons: (1) most existing LMI-based results have of the disturbance input on the controlled output is constrained to
dealt with the nonlinearities satisfying Lipschitz conditions which the prescribed level.
 Z. Wang et al. / Automatica 45 (2009) 684–691 691

5. Conclusions Wang, Z., Yang, F., Ho, D. W. C., & Liu, X. (2006). Robust H∞ filtering for stochastic
time-delay systems with missing measurements. IEEE Transactions on Signal
Processing, 54(7), 2579–2587.
In this paper, we have studied the problem of robust H∞ Wang, Z., Yang, F., Ho, D. W. C., & Liu, X. (2007). Robust H∞ control for networked
output feedback control for a class of uncertain discrete-time systems with random packet losses. IEEE Transactions on Systems, Man, and
delayed nonlinear stochastic systems with missing measurements. Cybernetics - Part B, 37(4), 916–924.
Wang, Z., Liu, Y., & Liu, X. (2008). H∞ filtering for uncertain stochastic time-delay
An output feedback controller has been designed for all admissible systems with sector-bounded nonlinearities. Automatica, 44(5), 1268–1277.
uncertainties such that the resulting closed-loop system is Xie, S., & Xie, L. (2000). Decentralized stabilization of a class of interconnected
exponentially stable in the mean square for the zero disturbance stochastic nonlinear systems. IEEE Transactions on Automatic Control, 45(1),
132–137.
input and also achieves a prescribed H∞ performance level. A Yuan, C., & Mao, X. (2004). Robust stability and controllability of stochastic
numerical example has been exploited to show the usefulness of differential delay equations with Markovian switching. Automatica, 40(3),
the results obtained. Further research directions would include the 343–354.
Zhang, W., Chen, B.-S., & Tseng, C.-S. (2005). Robust H∞ filtering for nonlinear
investigation on more general nonlinear systems and the extension
stochastic systems. IEEE Transactions on Signal Processing, 53, 589–598.
of our main results to more complex discrete-time systems with
distributed delays.
Zidong Wang was born in Jiangsu, China, in 1966. He
received the B.Sc. degree in mathematics in 1986 from
References Suzhou University, Suzhou, China, and the M.Sc. degree
in applied mathematics in 1990 and the Ph.D. degree
Berman, N., & Shaked, U. (2006). H∞ control for discrete-time nonlinear stochastic in electrical engineering in 1994, both from Nanjing
systems. IEEE Transactions on Automatic Control, 51(6), 1041–1046. University of Science and Technology, Nanjing, China.
Chen, B.-S., & Zhang, W. H. (2004). Stochastic H2 /H∞ control with state-dependent He is currently Professor of Dynamical Systems and
noise. IEEE Transactions on Automatic Control, 49(1), 45–57. Computing in the Department of Information Systems and
Deng, H., & Krstic, M. (2000). Output-feedback stabilization of stochastic nonlinear Computing, Brunel University, UK. From 1990 to 2002, he
systems driven by noise of unknown covariance. Systems and Control Letters, 39, held teaching and research appointments in universities
173–182. in China, Germany and the UK. Prof. Wang’s research
Deng, H., Krstic, M., & Williams, R. (2001). Stabilization of stochastic nonlinear interests include dynamical systems, signal processing, bioinformatics, control
systems driven by noise of unknown covariance. IEEE Transactions on Automatic theory and applications. He has published more than 100 papers in refereed
Control, 46, 1237–1253. international journals. He is a holder of the Alexander von Humboldt Research
Fu, Y., & Liao, X. (2003). BIBO stabilization of stochastic delay systems with Fellowship of Germany, the JSPS Research Fellowship of Japan, William Mong
uncertainty. IEEE Transactions on Automatic Control, 48(1), 133–138. Visiting Research Fellowship of Hong Kong.
Gahinet, P., & Apkarian, P. (1994). A linear matrix inequality approach to H∞ control. Prof. Wang serves as an Associate Editor for 11 international journals, including
International Journal of Robust and Nonlinear Control, 4, 421–448. IEEE Transactions on Automatic Control, IEEE Transactions on Control Systems
Gao, H., Wang, C., & Wang, J. (2005). On H∞ performance analysis for continuous- Technology, IEEE Transactions on Neural Networks, IEEE Transactions on Signal
time stochastic systems with polytopic uncertainties. Circuits, Systems and Processing, and IEEE Transactions on Systems, Man, and Cybernetics—Part C. He is
Signal Processing, 24(4), 415–429. a Senior Member of the IEEE, a Fellow of the Royal Statistical Society and a member
Gao, H., Lam, J., & Wang, C. (2006). Robust energy-to-peak filter design for stochastic of program committee for many international conferences.
time-delay systems. Systems & Control Letters, 55(2), 101–111.
Gao, H., & Wang, C. (2003). Delay-dependent robust H∞ and L2 /L∞ filtering for a
Daniel W. C. Ho received the B.Sc., M.Sc., and Ph.D. degrees
class of uncertain nonlinear time-delay systems. IEEE Transactions on Automatic
in mathematics from the University of Salford, Salford, UK,
Control, 48(9), 1661–1666.
in 1980, 1982, and 1986, respectively.
Gao, H., & Chen, T. (2007). New results on stability of discrete-time systems with
From 1985 to 1988, he was with the Industrial
time-varying state delay. IEEE Transactions on Automatic Control, 52(2), 328–334.
Control Unit, University of Strathclyde, Glasgow, UK, as
Han, Q.-L. (2005). Absolute stability of time-delay systems with sector-bounded
a Research Fellow. Since 1989, he has been with the
nonlinearity. Automatica, 41(12), 2171–2176.
Department of Mathematics, City University of Hong
Hounkpevi, F. O., & Yaz, E. E. (2007). Robust minimum variance linear state
Kong, Hong Kong, where he is currently a Professor.
estimators for multiple sensors with different failure rates. Automatica, 43(7),
His current research interests include H-infinity control
1274–1280.
theory, robust pole assignment problem, adaptive neural
Huang, H., & Ho, D. W. C. (2007). Delay-dependent robust control of uncertain
wavelet identification, and nonlinear control theory. Dr.
stochastic fuzzy systems with time-varying delay. IET Control Theory and
Ho is an Associate Editor of Asian Journal of Control.
Applications, 1(4), 1075–1085.
Khalil, H. K. (1996). Nonlinear systems. Upper Saddle River, NJ: Prentice-Hall.
Lam, J., Gao, H., Xu, S., & Wang, C. (2005). H∞ and L2 /L∞ infinity model reduction Yurong Liu received the B.Sc. degree in mathematics
for system input with sector nonlinearities. Journal of Optimization Theory and from Suzhou University, Suzhou, China, in 1986, the M.Sc.
Applications, 125(1), 137–155. degree in applied mathematics from Nanjing University
Liu, Y., Pan, Z., & Shi, S. (2003). Output feedback control design for strict-feedback of Science and Technology, Nanjing, China, in 1989, and
stochastic nonlinear systems under a risk-sensitive cost. IEEE Transactions on the Ph.D. degree in applied mathematics from Suzhou
Automatic Control, 48(3), 509–513. University in 2000. He is currently a Professor with
Lu, G., & Feng, G. (2007). Robust H∞ observers for Lipschitz nonlinear discrete-time the Department of Mathematics, Yangzhou University,
systems with time delay. IET Control Theory and Applications, 1(3), 810–816. Yangzhou, China. His current interests include neural
Nahi, N. (1969). Optimal recursive estimation with uncertain observation. IEEE networks, nonlinear dynamics, time-delay systems, and
Transactions on Information Theory, 15, 457–462. chaotic dynamics.
Niu, Y., Ho, D. W. C., & Lam, J. (2005). Robust integral sliding mode control
for uncertain stochastic systems with time-varying delay. Automatica, 41(5),
873–880. Xiaohui Liu received the B.Eng. degree in computing from
Shi, P., Mahmoud, M., Nguang, S. K., & Ismail, A. (2006). Robust filtering for jumping Hohai University, Nanjing, China, in 1982 and the Ph.D. de-
systems with mode-dependent delays. Signal Processing, 86(1), 140–152. gree in computer science from Heriot-Watt University, Ed-
Shi, P., Mahmoud, M., Yi, J., & Ismail, A. (2005). Worst case control of uncertain inburg, UK, in 1988.
jumping systems with multi-state and input delay information. Information He is currently a Professor of computing with Brunel
Sciences, 176(2), 186–200. University, Uxbridge, UK. He leads the Intelligent Data
Shu, H., & Wei, G. (2005). H∞ analysis of nonlinear stochastic time-delay systems. Analysis (IDA) Group, performing interdisciplinary re-
Chaos, Solitons & Fractals, 26(2), 637–647. search involving artificial intelligence, dynamic systems,
Sinopoli, B., Schenato, L., Franceschetti, M., Poolla, K., Jordan, M. I., & Sastry, S. S. image and signal processing, and statistics, particularly
(2004). Kalman filtering with intermittent observations. IEEE Transactions on for applications in biology, engineering, and medicine.
Automatic Control, 49, 1453–1464. Prof. Liu serves on editorial boards of four computing jour-
Wang, Z., Ho, D. W. C., & Liu, X. (2003). Variance-constrained filtering for uncertain nals, founded the biennial international conference series on IDA in 1995, and has
stochastic systems with missing measurements. IEEE Transactions on Automatic given numerous invited talks in bioinformatics, data mining, and statistics confer-
Control, 48, 1254–1258. ences.