Automatica 43 (2007) 861867

www.elsevier.com/locate/automatica
Brief paper
Decentralized dynamic output feedback for robust stabilization of a class of
nonlinear interconnected systems

Srdjan S. Stankovi c
a,
, Duan M. Stipanovi c
b
, Dragoslav D. iljak
c
a
Faculty of Electrical Engineering, University of Belgrade, Serbia
b
Department of Industrial and Enterprise Systems Engineering and the Coordinated Science Laboratory, University of Illinois at Urbana-Champaign,
Illinois, USA
c
School of Engineering, Santa Clara University, Santa Clara, CA, USA
Received 17 March 2006; received in revised form 21 August 2006; accepted 4 November 2006
Abstract
The objective of this paper is to propose an approach to robust stabilization of systems which are composed of linear subsystems coupled
by nonlinear time-varying interconnections satisfying quadratic constraints. The proposed algorithms, which are formulated within the convex
optimization framework, employ linear dynamic feedback structure involving local Luenberger-type observers. It is also shown how the new
methodology can produce improved results if interconnections have linear parts that are known a priori. Examples of output stabilization of
inverted pendulums and decentralized control of a platoon of vehicles are used to illustrate the applicability of the design method.
2007 Published by Elsevier Ltd.
Keywords: Nonlinear interconnected systems; Decentralized dynamic output feedback; LMIs; Robustness; Control of platoons of vehicles
1. Introduction
With the emergence of the powerful convex optimization
toolboxes involving linear matrix inequalities (LMIs), solving
problems of controller design within the convex optimization
framework became very attractive (e.g. Boyd, El Ghaoui, Feron,
& Balakrishnan, 1994; Dullerud & Paganini, 2000; Gahinet &
Apkarian, 1994; Gahinet, Nemirovski, Laub, & Chilali, 1995;
Garcia, Daafouz, & Bernussou, 2002; Ho & Lu, 2003; Iwasaki
& Skelton, 1994; Scorletti & Duc, 2001). Of wide-spread in-
terest have been the control problems of formulating sufcient
conditions for computing output feedback control laws using
convex optimization methods due to the fact that the neces-
sary and sufcient conditions are known to be nonconvex, in
general. These problems become increasingly more difcult
to solve when decentralized information structure constraints

This paper was not recommended at any IFAC meeting. This paper was
recommended for publication in revised form by Associate Editor Murat
Arcak under the direction of Editor Hassan K. Khalil.

Corresponding author. Tel.: +381 11 3370150; fax: +381 11 3248681

E-mail address: stankovic@etf.bg.ac.yu (S.S. Stankovi c).
0005-1098/$ - see front matter 2007 Published by Elsevier Ltd.
doi:10.1016/j.automatica.2006.11.010
are imposed in the control design (Cao, Sun, & Mao, 1998;
Geromel, Bernussou, &de Oliveira, 1999; Geromel, Bernussou,
& Peres, 1994; Qi, Salapaka, Voulgaris, & Khammash, 2004;
iljak & Ze cevi c, 2004; Yang & Wang, 1999; Ze cevi c &
iljak, 2004). These information structures can be found
in important applications, such as power systems (Ze cevi c,
Nekovi c, & iljak, 2004), control of formations of un-
manned vehicles (Stankovi c, Stanojevi c, & iljak, 2000) and
control of large structures (Li, Kosmatopoulos, Yoannou, &
Ryciotaki-Boussalis, 2000), to name few.
In this paper we propose novel sufcient conditions for
the design of decentralized dynamic output controllers in the
convex optimization context for stabilization of interconnected
systems with linear subsystems and nonlinear time-varying
interconnections. Controllers are designed to guarantee robust
stability of the overall system and, in addition, maximize the
bounds of unknown interconnection terms, starting from the
methodology proposed in iljak and Stipanovi c (2000). In
contrast to the approach in Stankovi c and iljak (2006), where
general dynamic controller structure is treated, we shall adopt
here the controller structure containing local observers of Lu-
enberger type. Several algorithms are proposed in the general
862 S.S. Stankovi c et al. / Automatica 43 (2007) 861867
case of full order observers, differing by complexity and the
degree of interdependence between the observer and the feed-
back gains, where no additional constraints on the parameters
of the system model are imposed (see Pagilla & Zhu, 2005;
iljak & Stipanovi c, 2001). It is also shown how the proposed
scheme can be used to build reduced-order observers. Particular
attention is paid to the case when linear parts of interconnec-
tions are known a priori, and an algorithm is proposed which
takes advantage of this knowledge to come up with improved
results. To illustrate the application of the proposed schemes
we include two examples, the rst dealing with interconnected
pendulums, and the second with the problem of platoons of
vehicles in the case when the velocity and acceleration of the
neighboring vehicles are not accessible.
2. Problem statement
Consider a nonlinear interconnected system S composed of
a nite number N of subsystems represented by
S
i
: x
i
= A
i
x
i
+ B
i
u
i
+ h
i
(t, x), y
i
= C
i
x
i
, (1)
where x
i
R
n
i
, u
i
R
m
i
and y
i
R
p
i
are the subsystem
state, input and output vectors, respectively, x =(x
T
1
, . . . , x
T
N
)
T
(

N
i=1
n
i
=n) is the global state vector and h
i
(t, x) : R
n+1

R
n
i
are piecewise continuous vector functions in both argu-
ments, satisfying in their domains of continuity the following
quadratic inequalities:
h
i
(t, x)
T
h
i
(t, x):
2
i
x
T
H
T
i
H
i
x, (2)
where :
i
>0 are bounding parameters and H
i
constant l
i
n
matrices, i = 1, . . . , N (see, e.g. iljak & Stipanovi c, 2000,
2001). The entire interconnected system S can be represented
in a compact form as
S : x = Ax + Bu + h(t, x), y = Cx, (3)
where u=(u
T
1
, . . . , u
T
N
)
T
and y =(y
T
1
, . . . , y
T
N
)
T
are the global
input and output vectors, respectively (

N
i=1
m
i
=m,

N
i=1
p
i
=
p), A = diag{A
1
, . . . , A
N
}, B = diag{B
1
, . . . , B
N
}, C =
diag{C
1
, . . . , C
N
}, and h(t, x) = (h
1
(t, x)
T
, . . . , h
N
(t, x)
T
)
T
is the global interconnection function. Consequently, we can
write
h
T
(t, x)h(t, x)x
T
H
T
I
1
Hx, (4)
where H =diag{H
T
1
, . . . , H
T
N
}
T
is an l n matrix (l =

N
i=1
l
i
),
I=diag{
1
I
l
1
, . . . ,
N
I
l
N
} and
i
=:
2
i
(I
l
i
represents the l
i
l
i
identity matrix).
Our assumptions are: (a) the dynamic controller F for S
is linear; (b) it obeys the decentralized information structure
constraint requiring that each subsystem is controlled using its
own local output; and (c) it is composed of an observer of
Luenberger-type and a static observer state feedback, that is,
F : w = Aw + Bu + L(y Cw), u = Kw, (5)
where w R
n
is the observer state, w = (w
T
1
, . . . , w
T
N
)
T
,
w
i
R
n
i
; K = diag{K
1
, . . . , K
N
} and L = diag{L
1
, . . . , L
N
}
represent the global controller parameter matrices, while pairs
(K
i
, L
i
) determine the local dynamic controllers.
Consequently, the resulting closed-loop system S
f
= (S, F)
can be represented as
S
f
: z = A
f
z + h
f
(t, z), y = C
f
z, (6)
where z is the state vector. Taking z =(z
T
1
, z
T
2
)
T
, z
1
=w, z
2
=
w x, we obtain
(7)
As a consequence of (4), we have now
h
f
(t, z)
T
h
f
(t, z)z
T
H
T
f
I
1
H
f
z, (8)
where .
We shall be looking for a dynamic controller F which robustly
stabilizes S. According to iljak and Stipanovi c (2000, 2001),
S is robustly stabilized with vector degree : = (:
1
, . . . , :
N
)
T
if the equilibrium z = 0 of the closed-loop system S
f
=(S, F)
is globally asymptotically stable for all h
f
(t, z) satisfying (8)
for some H
f
and I. In such a way, the controller stabilizes the
linear part of S and, at the same time, maximizes its tolerance
to uncertain nonlinear interconnections and perturbations. The
following general LMI-based formulation of the robust stabi-
lization problem represents a basis for all our further elabora-
tions in this paper (iljak & Stipanovi c, 2000):
S
f
= (S, F) is robustly stable with vector degree : if the
following problem is feasible:
Minimize Tr I
Subject to X
f
>0,

X
f
A
f
+ A
T
f
X
f
X
f
H
T
f
X
f
I 0
H
f
0 I

<0.
(9)
However, in general, dynamic output feedback design cannot
be done directly using (9), since the second matrix inequality
is not an LMI in both X
f
and the feedback parameter matrix.
In the case of static state feedback the problem can be trans-
formed into an LMI problem by a simple change of variables
(Geromel et al., 1994). In the case of dynamic output feedback
the problem becomes far more complex. A decoupled quadratic
Lyapunov function with block-diagonal weighting matrix has
been used in iljak and Stipanovi c (2001) to determine the
dynamic controller parameters. However, the proposed design
procedure imposes additional constraints on the system model
characteristics. Thus, in the next section we will provide several
modications of (9) obtained by convexifying the constraints.
Solutions to these problems will provide guaranteed feasible
solutions to (9) and the upper bound of the objective function
Tr I. This is implied by the fact that these convex conditions
are more restrictive than the nonconvex conditions in (9).
S.S. Stankovi c et al. / Automatica 43 (2007) 861867 863
3. Robust decentralized controller design
3.1. General casefull order observer
Dene Q=diag{Q
1
, . . . , Q
N
}, P =diag{P
1
, . . . , P
N
}, W =
diag{W
1
, . . . , W
N
} and V = diag{V
1
, . . . , V
N
}, where Q
i
, P
i
,
W
i
and V
i
are n
i
n
i
, n
i
n
i
, m
i
n
i
and n
i
p
i
matrices,
respectively. Also, for compactness of notation let us dene
1(S, L, M, I) =

S L M
L
T
I 0
M
T
0 I

, (10)
where S, L, M and I are matrices of appropriate dimensions.
Problem 1.
Miniminize Tr I
Subject to Q>0, P >0, (11)
1(S
1
, I, QH
T
, I) <0, 1(S
2
, P, H
T
, I) <0,
where S
1
=AQ+QA
T
+BW +W
T
B
T
and S
2
=PA+A
T
P
VC C
T
V
T
.
Theorem 1. System S is robustly stabilized by the controller F
if Problem 1 is feasible. The controller parameters are given by
K = WQ
1
, L = P
1
V. (12)
Proof. It will be shown that there exists a real number z >0
such that the matrix X
f
=diag{z
1
Q
1
, P} satises LMIs (9)
for some I>0, where P and Q are solutions of Problem 1 (see
Boyd et al., 1994 for a similar approach to the basic stabilization
problem). Introducing (7) and X
f
into (9), one obtains

zS
1
LC I 0 zQH
T
C
T
L
T
S
2
0 P H
T
I 0 I 0 0
0 P 0 I 0
zHQ H 0 0 I

<0. (13)
Applying the Schurs complement formula after convenient
column and row permutations, we obtain the following condi-
tions equivalent to (13):
E
1
<0, zE
3
(I
z
) E
2
E
1
1
E
T
2
+

I 0
0 0

<0, (14)
where E
1
=

I P
P S
2

, E
2
=

0 LC
0 H

, E
3
(X) =

S
1
QH
T
HQ X

(X denotes an n n matrix), and I

z
= z
1
I.
Let v = z
min
(E
1
1
), a = z
max
(E
2
E
T
2
) and j = z
max
(E
3
(I
0
)),
where I
0
= diag{
0
1
I
l
1
, . . . ,
0
N
I
l
N
} is the optimal I ob-
tained by solving Problem 1. Both eigenvalues j and v are
negative, since E
1
and E
3
(I
0
) represent principal minors
of the negative denite matrices 1(S
1
, I, QH
T
, I
0
) and
1(S
2
, P, H
T
, I
0
). Choose I = z

I
0
, where z

>|0|/|j|,
0 = 1 + av, and assume that 0 <z <z

; then, we have
z
max
{E
3
(I
z
)} =z
max
{E
3
((z

/z)I
0
)}z
max
{E
3
(I
0
)} =j, hav-
ing in mind that z

/z >1. For this choice of I and z, (14) is

implied by
jz 0 <0, (15)
which is true for |0|/|j| <z <z

. Therefore, the required z

exists, and we have the result.
Remark 1. The local robustness degrees :
i
= 1/

0
i
|0|/|j|,
i =1, . . . N, guaranteed on the basis of Theorem 1 are too con-
servative. More realistic values can be obtained by introducing
into (9) the controller parameters obtained by (12), and by solv-
ing the corresponding minimization problem with variables X
f
and I, as will be done in all the examples presented.
Remark 2. Problem 1 implements, in fact, the separation prin-
ciple. The constituent problems Q>0, 1(S
1
, I, QH
T
, I
1
) <0
and P >0, 1(S
2
, P, H
T
, I
2
) <0 can be solved inde-
pendently, the rst providing K as in the state feed-
back design (iljak & Stipanovi c, 2000), and the sec-
ond L, robustly stabilizing the observer, so that I =
diag{max(
1
1
,
2
1
)I
l
1
, . . . , max(
1
N
,
2
N
)I
l
N
}.
Remark 3. The proof of Theorem 1 gives the idea to simplify
LMIs in (11) as follows:
Problem 2.
Minimize Tr I
Subject to Q>0, P >0, E
3
(I) <0, E
1
<0. (16)
Controller parameters are obtained by using (12). The achiev-
able robustness degree is, in general, lower than the one
obtained by solving Problem 1. Namely, it is possible to
show using the methodology of Theorem 1 that if Q
0
, W
0
and I
0
are obtained by solving Problem 2, then there ex-
ist j >0 and [ >1 such that 1(j(AQ
0
+ Q
0
A
T
+ BW
0
+
W
T
0
B
T
), I, jQ
0
H
T
, [I
0
) <0.
The structure of (13) gives an idea to construct more complex
algorithms, trying to get higher robustness degree by taking into
account the interdependence between K and L in the LMIs (9).
Problem 3. Step 1:
Minimize Tr I
Subject to P >0, 1(S
2
, P, H
T
, I) <0, (17)
Step 2:
Minimize Tr A
Subject to Q>0,

S
1
I LC 0 QH
T
I I 0 0 0
C
T
L
T
0 S
2
P H
T
0 0 P I 0
HQ 0 H 0 IA

<0,
(18)
864 S.S. Stankovi c et al. / Automatica 43 (2007) 861867
where A=diag{o
1
I
l
1
, . . . , o
N
I
l
N
} (o
i
>0), while P, S
2
, I and
L = P
1
V are obtained in Step 1.
Theorem 2. System S is robustly stabilized by the controller
F if Problem 3 is feasible. Controller parameters are given by
(12). The robustness degree bounds are given by :
i
=1/

i
o
i
.
Proof. The proof follows simply from the fact that the sec-
ond inequality in (18) is identical to inequality (9) for X
f
=
diag{Q
1
, P}, with I replaced by IA. Notice that Steps 1 and
2 have to be performed consecutively and not simultaneously,
like in Problems 1 and 2.
Remark 4. Notice that the exposed methodology gives rise
to similar algorithms which could be derived from alter-
native realizations of the closed-loop model (6). One can
take, for example, z = (z
T
1
, z
T
2
)
T
, z
1
= x, z
2
= x w,
, and h
f
(t, z) =
, and come up with a problem similar to
Problem 3, in which K is determined in Step 1, and L in Step 2.
3.2. General casereduced order observer
The results of the preceding paragraph can be directly ex-
tended to the design of controllers with decentralized reduced
order observers.
Assume that , p
i
n
i
; if x
i
is de-
composed as x
i
= ((x

i
)
T
, (x

i
)
T
)
T
, where x

i
R
n
i
p
i
, x

i

R
p
i
, then y
i
= x

i
and the output w
i
of the local reduced or-
der observer is an estimate of x

i
(w
i
R
n
i
p
i
). Assume also
that the local dynamic controllers F
i
have the following form
(Kwakernaak & Sivan, 1972):
w
i
= A
11
i
w
i
+ A
12
i
y
i
+ B
1
i
u
i
(19)
+ L
i
[ y
i
A
21
i
w
i
A
22
i
y
i
B
2
i
u
i
],
u
i
= G
i
w
i
+ J
i
y
i
= K
i

i
, (20)
where A
i
=

A
11
i
A
12
i
A
21
i
A
22
i

, B
i
=

B
1
i
B
2
i

(A
22
i
and B
2
i
are p
i
p
i
and m
i
p
i
matrices, respectively), and
i
=
(w
T
i
, y
T
i
)
T
= (w
T
i
, (x

i
)
T
)
T
(differentiation of y
i
in (19) can
be avoided by standard transformation of variables). Dening
p
i
=w
i
x

i
, =(
T
1
, . . . ,
T
N
)
T
and p=(p
T
1
, . . . , p
T
N
)
T
, we take
z =(
T
, p
T
)
T
as a new state vector for S
f
=(S, F), and obtain
S
f
: z =

A + BK

LA
12
0 A
11
LA
21

z + h
f
(t, z), (21)
where A
11
= diag{A
11
1
, . . . , A
11
N
}, A
12
= diag{A
12
1
,
. . . , A
12
N
}, A
21
= diag{A
21
1
, . . . , A
21
N
}, K = diag{K
1
,
. . . , K
N
}, L = diag{L
1
, . . . , L
N
},

L = diag{

L
1
, . . . ,

L
N
}
where the decomposition h
i
(x) =(h

i
(x)
T
, h

i
(x)
T
)
T
is induced
by the decomposition of x
i
into x

i
and x

i
, so that
h
f
(t, z)
T
h
f
(t, z):
2
z
T

H
T
f

H
f
z, (22)
where while

H
i
is an l
i

(n
i
p
i
) matrix containing the rst n
i
p
i
columns of H
i
,
having in mind that H
i
x = H
i


H
i
p.
The structure of the closed-loop model (21) shows that con-
troller design can be entirely based on the methodology exposed
in the rst paragraph. Problem 1 and Theorem 1 give rise to
Corollary 1. The systemS in which
is ro-
bustly stabilized by the dynamic controller F dened by (19),
(20) if the following problem is feasible:
Minimize Tr I
Subject to Q>0,

P >0,
1(S
1
, I, QH
T
, I) <0,
1(

S
2
,

P, H
T
, I) <0, (23)
where

S
2
=

PA
11
+ (A
11
)
T

P

VA
21
(A
21
)
T

V
T
,

P =
diag{

P
1
, . . . ,

P
N
} and

V = diag{

V
1
, . . . ,

V
N
} (

P
i
and

V
i
are
(n
i
p
i
) (n
i
p
i
) and (n
i
p
i
) p
i
matrices, respectively).
Controller parameters are obtained by using (12).
3.3. A linear part of interconnections known
Known linear interconnections between the subsystems S
i
in
S can be represented by a full n n matrix A
o
containing off-
diagonal interconnection blocks, so that A + A
o
becomes the
new state matrix in the linear part of S in (2); function h(t, x)
still represents the unknown part of interconnections. The de-
sign methodology proposed in Sections 3.1 and 3.2 can be ex-
tended to this case, trying to take advantage of the additional
a priori information. Notice, however, that by replacing A by
A + A
o
in the observer equation for F in (5) one violates the
adopted information structure constraint, i.e. the dynamic con-
troller ceases to be decentralized. Inserting A+A
o
only in the
state model (3), we obtain
This fact indicates that the design scheme could now be based
on modifying the problems described in Sections 3.1 and 3.2
by inserting the new information in the form of A
o
at the cor-
responding places in the related LMIs. However, robust stabi-
lization is achievable when the interconnections are sufciently
weak. For example, Problem 1 gives rise to:
Problem 4.
Minimize Tr I
Subject to P >0, Q>0,
1(S
1
, I, QH
T
, I) <0,
1(S
2o
, P, H
T
, I) <0, (24)
where S
2o
= P(A + A
o
) + (A + A
o
)
T
P VC C
T
V
T
.
S.S. Stankovi c et al. / Automatica 43 (2007) 861867 865
Theorem 3. The system S with known linear interconnections
(modelled by adding A
o
to A in (3)) is robustly stabilized by
the decentralized dynamic controller F in (5) if Problem 4 is
feasible and
o <
j
2
80
o
v
o
z
P
z
Q
, (25)
where o =z
max
(A
T
o
A
o
), z
P
=z
max
(P
2
), z
Q
=z
max
(Q
2
), v
o
=
z
min
(E
1
1o
), matrix E
1o
is obtained from E
1
in (14) by replacing
S
2
with S
2o
, and 0
o
= 1 + 2av
o
.
Proof. The proof is based on a line of thought similar to that
applied in Theorem 1. Inserting X
f
= diag{z
1
Q
1
, P} and
A
f
=

A + BK LC
A
o
A + A
o
LC

into (9) we obtain

zS
1
L
o
I 0 zQH
T
L
T
o
S
2o
0 P H
T
I 0 I 0 0
0 P 0 I 0
zHQ H 0 0 I

<0, (26)
where L
o
=LC +zQA
T
o
P. The last inequality is equivalent to
E
1o
<0 and
zE
3
(I
z
) (E
2
+ zE
2o
)E
1
1o
(E
2
+ zE
2o
)
T
+

I 0
0 0

<0,
(27)
where E
2o
=

0 QA
T
o
P
0 0

. Assume, similarly as in Theorem

1, that I=z

I
0
for some z

>0, where I
0
is the optimal value
obtained by solving Problem 4. Assume that 0 <z <z

. Then,
(27) is implied by
2ov
o
z
P
z
Q
z
2
+ jz 0
o
<0, (28)
having in mind that z
max
{E
3
(I
z
)}z
max
{E
3
(I
0
)} = j; notice
that v
o
<0 by assumption, as a consequence of the feasibility of
Problem 4. The existence of z >0 satisfying (28) is guaranteed
if (25) holds, since then we have D = j
2
8o0
o
v
o
z
P
z
Q
>0.
Consequently, we choose
j

D
4ov
o
z
P
z
Q
= z
1
<z

z
2
=
j +

D
4ov
o
z
P
z
Q
,
where 0 <z
1
<z
2
since j <0 and

D|j|, so that z can take
any value in the interval [z
1
, z

]. Thus, the result.

The local guaranteed robustness degree bounds are now :
i
=
1/

0
i
z
1
, i = 1, . . . , N.
4. Examples
4.1. Inverted pendulums
The rst example illustrates how the proposed methodology
can be used to obtain connective stability (iljak, 1991) and
Table 1
Robustness degree : for different algorithms
Problem 1 Problem 2 Problem 3
Case A 5.6450 0.3813 21.7304
Case B 4.3840 0.5787 15.2214
Case C 0.6564 0.3191 0.7003
robustness with respect to modelling errors in the case when a
system is not stabilizable using static output feedback.
Consider the motion of two inverted pendulums connected
by a spring which can slide up and down the rods of the pendu-
lums in jumps of unpredictable size and direction between the
support and the height equal to 1 (iljak & Stipanovi c, 2000).
A linearized and normalized model is given by
(29)
where h(t, x) = e(t, x)Gx, and e(t, x) :
R
5
[0, 1] represents a normalized interconnection param-
eter. We want to compute a decentralized control law which
would connectively stabilize the system for all values of
e(t, x) [0, 1].
The decentralized static state feedback provides : = 4.4950
with the local controller gain matrix K = [725.9085
40.4346] and the corresponding closed-loop poles {20
j17.8093}. Clearly, the system is not stabilizable by static
output feedback, since two coefcients of the characteristic
equation remain xed to zero irrespective of the controller
parameters.
Table 1 gives robustness degree : provided by dynamic
output feedback obtained by the proposed algorithms. Case
A corresponds to the situation in which H = G in the
three algorithms from Section 3.1, Case B to H = G with
when the algorithms from Section 3.3
are applied (derived from Problems 13 in accordance with the
methodology of Problem 4 and Theorem 3), and Case C to the
situation with no a priori knowledge, when H =I and A
o
=0,
and the algorithms from Section 3.1 are applied. It is possible to
conclude that the best results are obtained by solving Problem
3; the worst case corresponds to Problem 2, as expected. No-
tice that in Case C none of the algorithms ensures connective
stability. For Problem 2, connective stability is achieved only
in Case B, when the information about the interconnections is
included (in Case B we have, in fact, e(t, x) = 0.5 + e

(t, x),
where e

(t, x) [0.5, 0.5], so that any value of : >0.5 is

sufcient for connective stability). All values of K and L and
the corresponding modes are not presented because of the
lack of space. For example, for Problem 1 and Case A we
866 S.S. Stankovi c et al. / Automatica 43 (2007) 861867
have K
i
=[ 79.1666 11.2883], L
T
i
=[ 27.7711 15.7991],
with local closed-loop poles {27.2275, 0.5435, 0.5441
j6.8052}. Notice also that the algorithm in Problem 3 gives
higher gains and the pole positions farther in the left half plane.
4.2. Decentralized control of a platoon of vehicles
A feedback-linearized state space model of a platoon of N au-
tomotive vehicles is based, according to Stankovi c et al. (2000),
on the following feedback linearized individual vehicle model:

d
i
= v
i1
v
i
, v
i
= a
i
, a
i
= t
1
i
a
i
+ t
1
i
u
i
, (30)
where d
i
= x
i1
x
i
is the distance between two consecutive
vehicles, x
i1
and x
i
being their positions, v
i
and a
i
are the ve-
locity and acceleration of ith vehicle, respectively, u
i
the input
signal chosen to make the closed-loop system satisfy certain
performance criteria, and t
i
the time constant of the engine. Af-
ter obtaining the overall platoon state space model with the state
X=(d
1
d
r
, v
1
v
r
, a
1
a
r
, . . . , d
N
d
r
, v
N
v
r
, a
N
a
r
)
T
and input u = (u
1
, u
2
, . . . , u
N
)
T
, where d
r
, v
r
and a
r
are the
reference values for inter-vehicle distance, velocity and accel-
eration, respectively, and applying the state and input expan-
sion by using convenient full-rank linear transformations, the
following model in the expanded space is obtained (Stankovi c
et al., 2000):

S :

=

A +

B (31)
where =(
T
1
, . . . ,
T
N
)
T
, =(
T
1
, . . . ,
T
N
)
T
,

A=diag{A
1
, . . . ,
A
N
} and

B=diag{B
1
, . . . , B
N
}; vectors
i
and
i
and matrices
A
i
and B
i
are dened within the formally dened subsystem
models connected to each vehicle:
S
i
:

i
= A
i

i
+ B
i

i
=

A
l
i
0

A
d
A
v
i

i
+

B
l
i
0
0 B
v
i

i
, (32)
where
i
=(v
i1
v
r
, a
i1
a
r
, d
i
d
r
, v
i
v
r
, a
i
a
r
)
T
is
the state vector of ith subsystem,
i
=(u
i1
, u
i
)
T
represents its
control vector, while A
l
i
=

0 1
0 t
1
i

,

A
T
d
=

1 0 0
0 0 0

, B
l
i
=

0
t
1
i

, A
v
i
=

0 1 0
0 0 1
0 0 t
1
i

and B
v
i
=

0
0
t
1
i

(Stankovi c
et al., 2000). A decentralized dynamic control law can now be
designed for the expanded system using the methodology from
Section 3.2, supposing that only the subsystem states d
i
d
r
,
v
i
v
r
and a
i
a
r
are exactly known in ith vehicle (subsystem),
that is, v
i1
and a
i1
are not accessible in ith vehicle. According
to Section 3.2, the reduced-order Luenberger observer for
1
i
=
(v
i1
v
r
, a
i1
a
r
)
T
is given by
w
i
= A
l
i
w
i
+ B
l
i
u
i1
+ L
i
[

2
i


A
d
w
i
A
v
i

2
i
], (33)
where
2
i
= (d
i
d
r
, v
i
v
r
, a
i
a
r
)
T
. The local control law
has the following specic structure:
u
i1
= G
1
i
w
i
, u
i
= G
2
i
w
i
+ J
2
i

2
i
, (34)
having in mind that (i 1)st vehicle does not have any
information about ith vehicle. Matrices K
i
=

G
1
i
0
G
2
i
J
2
i

and L
i
, i = 1, . . . , N, can now be obtained by using the
algorithm from Corollary 1, taking care of the specic
lower-block-triangular structure of K
i
. For t
i
= t = 0.1,
one obtains: G
i
= G = [ 38.6940 2.1224], G
1
i
= G
1
=
[ 38.6940 2.1224], G
2
i
= G
2
= [ 0.0095 0.0005],
J
2
i
= J
2
= [ 351.4028 319.3970 13.2356], L
i
= L =
10
4

0.0001 0 0
3.2068 0 0

and :
i
= : = 1/4.080, with closed loop
poles 10
2
{1.1480, 0.0116, 0.1561+j0.1197, 0.1561
j0.1197, 0.2640, 320.68, 0}. Obviously, it is also possible
to apply the alternative design schemes from Section 3.1.
The obtained controller has to be nally contracted to
the original space for implementation by using the expan-
sion/contraction matrices as in iljak (1991) and Stankovi c
et al. (2000).
5. Conclusion
This paper proposes several schemes for the design of
decentralized robust dynamic output controllers for linear inter-
connected systems with unknown nonlinear interconnections
satisfying quadratic constraints. The procedures assume the
controller structure containing observers of Luenberger-type,
and are in the form of convex optimization problems with
LMI constraints giving both feedback and observer gains of
local linear dynamic controllers with no explicit constraint on
the system model. It is shown that the proposed methodology
can be also applied to the case of known linear interconnec-
tions, taking advantage of the additional a priori knowledge.
Two numerical examples illustrate applicability aspects of the
proposed procedures. The rst example, dealing with inverted
pendulums, shows how connective stability can be achieved
for large unpredictable variations of system parameters. The
second example is especially important, since it shows the ap-
plicability of the method to decentralized control of platoons
of vehicles, in the case when measurements of the velocity
and acceleration of the neighboring vehicles are not available.
Our future goal is to apply the proposed procedures to the
design of overlapping decentralized dynamic output controllers
for general large-scale systems, within the framework of the
Inclusion Principle (iljak, 1991; Stankovi c et al., 2000; Chu
& iljak, 2005).
References
Boyd, S., El Ghaoui, L., Feron, E., & Balakrishnan, V. (1994). Linear matrix
inequalities in system and control theory. Philadelphia: SIAM.
Cao, Y. Y., Sun, Y. X., & Mao, W. J. (1998). Output feedback decentralized
stabilization: ILMI approach. Systems & Control Letters, 35, 183194.
Chu, D., & iljak, D. D. (2005). A canonical form for the inclusion principle
of dynamic systems. SIAM Journal on Control and Optimization, 44,
969990.
Dullerud, G. E., & Paganini, F. (2000). A course in robust control theory
a convex approach. New York: Springer.
S.S. Stankovi c et al. / Automatica 43 (2007) 861867 867
Gahinet, P., & Apkarian, P. (1994). A linear matrix inequality approach to
H

control. International Journal of Robust and Nonlinear Control, 4,

421448.
Gahinet, P., Nemirovski, A., Laub, A.J., & Chilali, M. (1995). LMI control
toolbox, Natick, MA: The Math Works.
Garcia, G., Daafouz, J., & Bernussou, J. (2002). The innite time near optimal
decentralized regulator problem for singularly perturbed systems: a convex
optimization approach. Automatica, 38, 13971406.
Geromel, J. C., Bernussou, J., & de Oliveira, M. C. (1999). H
2
-norm
optimization with constrained dynamic output feedback controllers:
decentralized and reliable control. IEEE Transactions on Automatic
Control, 44, 14491454.
Geromel, J. C., Bernussou, J., & Peres, P. L. D. (1994). Decentralized control
through parameter space optimization. Automatica, 30, 15651578.
Ho, D. W. C., & Lu, G. (2003). Robust stabilization for a class of discrete-
time non-linear systems via output feedback: The unied LMI approach.
International Journal of Control, 76, 105115.
Iwasaki, T., & Skelton, R. E. (1994). All controllers for the general H

control
problem: LMI existence conditions and state space formulas. Automatica,
30, 13071317.
Kwakernaak, H., & Sivan, R. (1972). Linear optimal control systems. New
York: Wiley.
Li, K., Kosmatopoulos, E. B., Yoannou, P. A., & Ryciotaki-Boussalis,
H. (2000). Large segmented telescopes: Centralized decentralized and
overlapping control designs. IEEE Control Systems Magazine, 20, 5972.
Pagilla, P. R., & Zhu, Y. (2005). A decentralized output feedback controller
for a class of large-scale interconnected nonlinear systems. Journal of
Dynamic Systems Measurement and ControlTransactions of the ASME,
127, 167172.
Qi, X., Salapaka, M. V., Voulgaris, P. G., & Khammash, M. (2004). Structured
optimal and robust control with multiple criteria: A convex solution. IEEE
Transactions on Automatic Control, 49, 16231640.
Scorletti, G., & Duc, G. (2001). An LMI approach to decentralized control.
International Journal of Control, 74, 211224.
iljak, D. D. (1991). Decentralized control of complex systems. New York:
Academic Press.
iljak, D. D., & Stipanovi c, D. (2000). Robust stabilization of nonlinear
systems. Mathematical Problems in Engineering, 6, 461493.
iljak, D. D., & Stipanovi c, D. (2001). Autonomous decentralized control.
Proceedings of ASME International Mechanical Engineering Congress,
761765.
iljak, D. D., & Ze cevi c, A. I. (2004). Control of large-scale systems: Beyond
decentralized feedback. Annual Revision in Control, 20, 169179.
Stankovi c, S.S., & iljak, D.D., (2006). Robust stabilization of nonlinear
interconnected systems by decentralized dynamic output feedback. IEEE
Transactions on Automatic Control, submitted.
Stankovi c, S. S., Stanojevi c, M. J., & iljak, D. D. (2000). Decentralized
overlapping control of a platoon of vehicles. IEEE Transactions on Control
Systems Technology, 8, 816832.
Yang, G. H., & Wang, J. L. (1999). Decentralized controller design for
composite systems: Linear case. International Journal of Control, 72,
815825.
Ze cevi c, A. I., Nekovi c, G., & iljak, D. D. (2004). Robust decentralized
exciter control with linear feedback. IEEE Transactions on Power Systems,
19, 10961103.
Ze cevi c, A. I., & iljak, D. D. (2004). Design of robust static output feedback
for large-scale systems. IEEE Transactions on Automatic Control, 49,
20402044.
Srdjan S. Stankovi c got his Dipl. Ing. Degree
from the Faculty of Electrical Engineering, Uni-
versity of Belgrade, Yugoslavia, in 1968. He
got his M.Sc. and Ph.D. degrees from the same
Faculty in 1972 and 1975, respectively. He was
with the Institute for Nuclear Sciences, Vinca,
Belgrade, Yugoslavia, from 1968 to 1972. Since
1973 he has been with the Faculty of Electri-
cal Engineering, University of Belgrade, where
he is currently Professor of Automatic Control
and Head of the Department for Signals and
Systems. He held visiting positions at the Eindhoven University of Technology,
Eindhoven, the Netherlands and at the Santa Clara University, Santa Clara,
CA. He published numerous scientic papers from diverse elds, including
estimation and identication, adaptive systems, digital signal processing, pro-
cessing of medical images, large-scale systems and neural networks. He was
also leader of numerous scientic projects for government and industry. His
research interests include currently large-scale systems, multiagent systems,
vehicle formation control and statistical learning applied to biomedicine.
Duan M. Stipanovi c received the B.S. de-
gree in electrical engineering from the Univer-
sity of Belgrade, Belgrade, Serbia, in 1994, and
the M.S.E.E. and Ph.D. degrees (under super-
vision of Professor Dragoslav iljak) in elec-
trical engineering from Santa Clara University,
Santa Clara, CA, in 1996 and 2000, respectively.
Dr. Stipanovi c was an Adjunct Lecturer and Re-
search Associate with the Department of Electri-
cal Engineering at Santa Clara University in the
period between 1998 and 2001, and a Research
Associate in Professor Claire Tomlins Hybrid Systems Laboratory of the
Department of Aeronautics and Astronautics at Stanford University, Stanford,
CA, in the period between 2001 and 2004. In 2004 he joined the University
of Illinois at Urbana-Champaign as an Assistant Professor in the Department
of Industrial and Enterprise Systems Engineering and a Research Assistant
Professor in the Control and Decision Group of the Coordinated Science Lab-
oratory. His research interests include decentralized control of interconnected
systems with application to control of formations of vehicles and sensor net-
works, hybrid and discontinuous dynamic systems, differential game theory,
and optimization with application to multiple vehicle coordination and sys-
tems safety verication.
Dragoslav D. iljak received the Ph.D. degree
in 1963 from the University of Belgrade, Bel-
grade, Serbia. Since 1964 he has been with
Santa Clara University, Santa Clara, CA, where
he is the B&M Swig University Professor in
the School of Engineering, and teaches courses
in control theory and applications. His research
interests are in the theory of large complex
systems and its applications to problems in
control engineering, power systems, economics,
multi-agent systems, aerospace, and model
ecosystems. He is the author of monographs Nonlinear Systems: The Pa-
rameter Analysis and Design (Wiley, 1969), Large-Scale Dynamic Systems:
Stability and Structure (North-Holland, 1978), and Decentralized Control of
Complex Systems (Academic Press, 1991). He is an Honorary Member of
The Serbian Academy of Sciences and Arts, and is a Life Fellow of IEEE.