Documente Academic
Documente Profesional
Documente Cultură
IMECE2003-43419
Jeongheon Han
Robert E. Skelton
Structural Systems and Control Laboratory
Department of Mechanical and Aeronautical Engineering
University of California, San Diego
La Jolla, California 92093-0411
Email: jehan@ucsd.edu, bobskelton@ucsd.edu
ABSTRACT
This paper presents a new algorithm for the design of
linear controllers with special constraints imposed on the
control gain matrix. This so called SLC (Structured Linear Control) problem can be formulated with linear matrix inequalities (LMIs) with a nonconvex equality constraint. This class of prolems includes fixed order output feedback control, multi-objective controller design,
decentralized controller design, joint plant and controller
design, and other interesting control problems.
Our approach includes two main contributions. One
is that many design specifications such as H performance, generalized H2 performance including H2 performance, ` performance, and upper covariance bounding
controllers are described by a similar matrix inequality. A
new matrix variable is introduced to give more freedom
to design the controller. Indeed this new variable helps to
find the optimal fixed-order output feedback controller.
The second contribution uses a linearization algorithm to search for a solution to the nonconvex SLC
problems. This has the effect of adding a certain potential function to the nonconvex constraints to make them
convex. Although the constraints are added to make
functions convex, those modified matrix inequalities will
not bring significant conservatism because they will ultimately go to zero, guaranteeing the feasibility of the orig-
Address
inal nonconvex problem. Numerical examples demonstrate the performance of the proposed algorithms and
provide a comparison with some of the existing methods.
Introduction
Control problems are usually formulated as optimization problems. Unfortunately, most of them are not
convex [1], and a few of them can be formulated as linear matrix inequalities (LMIs). In this case, powerful
algorithms can be used to find the optimal solution [2, 3].
There are two main approaches to solve a linear control
problem. One is the change of variable technique [4, 13]
and the other is to use the elimination lemma [1]. Indeed, the book [1] was written to show that more than
twenty different control problems can be treated as the
same linear algebra problem. In the LMI framework,
one can solve several linear control problems in the form
minK f (T()), where f () is some suitably defined objective function and T() is the transfer function from
a given input to a given output of interest. Notice that
there is no constraint on the control variable K for this
problem. For this problem, one can find a solution efficiently with the use of any LMI solver. However the
problem becomes difficult when one adds some structural
constraints on K. A typical example is a decentralized
control problem or a fixed-order output feedback control
problem. In this case, the problem can be formulated as
c 2003 by ASME
Copyright
analysis conditions have several advantages over the original performance analysis conditions. First of all, many
design specifications such as general H2 performance including H2 performance, H performance, ` performance, and the upper covariance bounding controllers
can be written in a very similar matrix inequality. We
introduce a new matrix variable Z for several system performance analysis conditions. As a result, we have more
freedom to find the optimal controller. Indeed, this new
variable helps to find the optimal fixed-order output feedback controller.
The paper is organized as follows. Section 2 defines
a framework for SLC problems and describes the new
system performance analysis conditions. We derive the
important class of a matrix inequality which introduces
a new matrix variable. Based on this new system performance specifications, a new linearization algorithm is
proposed in section 3. Two numerical examples illustrate
the performance of the proposed algorithms as compared
with the existing results in section 4 and then draw the
conclusion.
y(k) = Cy xp (k)
+ Dy w(k)
where x <np is the plant state, z <nz is the controlled output, and y <ny is the measured output. We
assume that all matrices have suitable dimensions. Our
goal is to compute an output-feedback controller
K
xc (k + 1) = Ac xc (k) + Bc y(k)
u(k) = Cc xc (k) + Dc y(k)
(2)
x(k + 1)
Acl (K) Bcl (K)
x(k)
=
z(k)
Ccl (K) Dcl (K)
w(k)
(3)
Dc Cc
K=
Bc Ac
4
xp
; x=
xc
4
c 2003 by ASME
Copyright
A =
Cy 0
0 Inc
Cz 0 (4)
Note that (5) describes an upper bound to the observability Gramian X and (6) describes an upper bound to the
controllability Gramian Y. The following results are useful to compute system gains. Twz () denotes the transfer
function from the input w to the output z.
Lemma 1. [1] Let a positive scalar be given and consider the discrete-time system (3). Suppose the system is
asymptotically stable. Then the following statements are
true.
(i) ep < iff there exist matrices X > 0 and >
0 such that I > and (5) holds.
(ii) ie < iff there exist matrices X and such
that I > and (6) holds
(Bz )i = Li Bz (Cz )i = Li Cz
and zi = Li z.
`2 1
< iff
2.1
(5)
Y > ATcl (K)YAcl (K) + BTcl (K)Bcl (K)
(6)
> CTcl (K)YCcl (K) + DTcl (K)Dcl (K)
T
X Z
Acl (K) Bcl (K)
X 0
Acl (K) Bcl (K)
>
ZT
Ccl (K) Dcl (K)
0 I
Ccl (K) Dcl (K)
(7)
T
Y Z
Acl (K) Bcl (K)
Y0
Acl (K) Bcl (K)
>
ZT
Ccl (K) Dcl (K)
0 I
Ccl (K) Dcl (K)
(8)
kzk` .
c 2003 by ASME
Copyright
design specifications H2 which is given by the statement (v) in Lemma 1. This measure is closely related to
H2 and H measures and can be interpreted by a uncontrained mixed H2 /H problem. In this sense, we label
all problems defined in Lemma 1 by general H problem .
Note that all inequalities given in Lemma 1 have similar
forms which can be parametrized by the matrix inequality ( + K) R ( + K)T < Q. The following
lemma is important to derive alternative analysis conditions.
Q Z
A B
>
.
ZT R
BT C
(9)
3
Linearization Algorithm
In the previous section, we have shown how several control problems can be written as linear matrix inequalities that preserve the affine dependence on the free
variables with the addition of a nonconvex equality constraints. Since analytic condition is nonvex matrix inequality, we need to develop a new algorithm.
Z = B + (Q A) 2 L (R C) 2
where L is an arbitrary matrix such that kLk < 1.
Theorem 2. Let a scalar convex functional f (X), a matrix functional J (X) and H(X) be given and consider
the following nonconvex optimization problem.
(10)
(11)
where
4
T
A B 4 Acl (K) Bcl (K)
X 0
Acl (K) Bcl (K)
=
CD
Ccl (K) Dcl (K)
0 I
Ccl (K) Dcl (K)
c 2003 by ASME
Copyright
is generally a difficult question. Linearization of nonconvex term H(X) may help to find such a convexifying potential function. Now we are ready to develop a new algorithm for .
where LIN X 1 , Xk is given by (12).
3. If kXk+1 Xk k < , go to step 4. Otherwise, set
k k + 1 and go back to Step 2.
LIN X 1 , Xk 0
T
(?)
0
I
LIN Y1 , Yo = Yo1 Yo1 (Y Yo ) Yo1 (12)
>0
We also summarize the algorithm [14] for the comparison purpose. Since the step 1 and 3 are the same as
those in Algorithm 1, we describe the step 2 only.
Algorithm 2. [14]
2. Solve the following convex optimization problem.
Since
Y 1 LIN Y 1 , Yo
ATcl (K)
Y
0 >0
T
B
(K)
0
cl
I
subject to
C
(K)
D
(K)
cl
cl
Ccl (K) Y
0 >0
DTcl (K)
0
I
where LIN Y 1 , Yk is given by (12).
Note that these two algorithm can also be used for the
optimal fixed order control problem. Algorithm 1 and 2
may find an initial feasible solution. If they failed, we
need to find an initial feasible solution in order to use
those algorithms. Here, we also propose a new feasibility
algorithm for the completeness of the proposed algorithm
using the linearization approach. We describe the step 2
only for brevity.
Algorithm 3. Initialization
5
c 2003 by ASME
Copyright
4 Illustrative Examples
4.1
Example 1 : Decentralized H2 Optimal
Static Output Feeback Controller
Consider the following discrete-time plant
I Y
+ Y LIN X 1 , Xk < 0
X Z Acl (K) Bcl (K)
subject to
>0
Y0
(?)
0 I
Ap =
Bp =
Cy =
Feasibility problem is not convex either, since there is the
nonconvex term X 1 . But this nonconvex term is the
same as the previous optimization problem. Hence we
can linearize this term as we explained before.
Cz =
0.0045 0.0044
0.0953 0
0.1001 0.0100
0.0145 0
0
0
0
0
which is borrowed from [14]. Our objective is to minimize kTwz ()kH2 using two decentralized static output
feedback controllers. Hence the structure of a controller
Dc is diagonal. This is a typical SLC problem. In order
to use initialization algorithm 3, we set X0 = I + RRT
and Y0 = X01 , where R is a random matrix. After using
the initialization Algorithm 3, we have run the algorithm
1 and 2. The precision has been set to 103 .
Notice that the proposed algorithm is very similar to the one proposed in [5], which adopts conecomplementarity linearization algorithm. The new proposed algorithm minimized trace[Y+Xk1 X Xk1 ], while
the cone complementarity linearization algorithm minimizes trace[YXk + X Yk + Yk X + Xk Y]. It is clear
that the cone complementarity algorithm linearized at
a matrix pair (Xk , Yk ) and our algorithm linearized
only
at Xk
. Another feature is that we minimized
Y X 1
and cone complementarity algorithm minimized kX Y + YX k. Clearly we minimized the controllability Gramian Y and maximized the observability
Gramian X . This implies that our algorithm is suitable for
solving the optimal H2 control problem, while the conecomplementarity algorithm is suitable for solving optimal
H control problem. Because there always exists a positive scalar such that X Y I [1]. Note that we can
establish another feasibility algorithm since
6
ALG 1
ALG 2
5
4
3
2
1
0
10
15
20
25
20
25
= (X + Y) (X + Y) + Xo2 + Yo2
X Y + YX = (X + Y) (X + Y) X Y
X Xo Xo X YYo Yo Y
(X Xo ) (X Xo ) (Y Yo ) (Y Yo )
10
15
iterations
Hence we can apply the linearization approach by ignoring the nonconvex term (X Xo ) (X Xo )
(Y Yo ) (Y Yo ). This approach also linearizes at a
matrix pair (Xo , Yo ).
Figure 4.1 shows the performance of the decentralized static output feedback controller for both algorithms.
6
c 2003 by ASME
Copyright
6
ALG 3
ALG 1
ALG 2
5.5
4.2
4.5
10
12
14
16
1.9
1.8
1.7
1.6
1.5
1.4
20
1
00
xp (k + 1) =
xp (k) +
u(k) +
w(k)
1 1
0
10
2
01
0
z1 (k) =
x (k) +
u(k)
00 p
1
11
0
z2 (k) =
xp (k) +
u(k)
00
1
y(k) = 1 0 xp (k) + 0 1 w(k)
Figure 2.
8
iterations
10
12
14
ation, a certain potential function is added to the nonconvex constraints to enforce them convex. Although
we solved those sufficient conditions iteratively, this approach will not bring significant conservatism because
they will converge to zero. Local optimality is guaranteed. Our approach can be applied to other linear synthesis problems as long as the dependence on the augmented
plant on the synthesis parameters are affine. Two control
design problems have been illustrated and compared with
the existing method numerically.
REFERENCES
[1] R.E. Skelton, T. Iwasaki and K. Grigoriadis, A Unified Algebraic Approach to Linear Control Design,
Taylor & Francis, 1994.
[2] S. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan Linear matrix Inequalities in System and Control Theory, SIAM Studies in Applied Mathematics ,
1994
[3] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control Toolbox, Natick, MA : The MathWorks, 1995
[4] Scherer C., Gahinet P. and Chilali M., Multiobjective output-feedback control via LMI optimization ,
IEEE Trans. on Automatic Control , Vol. 42, No. 7,
pp. 896-911, 1997
[5] L. El Ghaoui, F. Oustry, and M. A. Rami A cone
complementarity linearization algorithm for static
output-feedback and related problems, IEEE Trans.
on Automatic Control , 42(8), 1997
[6] J.C. Geromel, C.C. de Souza, and R.E. Skelton,
Static output feedback controllers : Stability and con-
5 Conclusion
We have addressed the SLC (Structured Linear Control) problem for linear discrete-time systems. New system performance conditions have been derived. This
new results introduce the augmented matrix variable Z.
Hence the number of variables of the synthesis condition
is maximized. It turns out that the behavior of new synthesis conditions is better than the original system synthesis.
In the SLC framework, these objectives are characterized as a set of LMIs with an additional nonconvex
equality constraint. To overcome this nonconvex constraint, we adopt a linearization method. At each iter7
c 2003 by ASME
Copyright
16
c 2003 by ASME
Copyright