Proceedings of the MC1.

4
2005 IEEE Conference on Control Applications
Toronto, Canada, August 28-31, 2005

Robust pole placement by output feedback using a genetic algorithm

Jérôme Bosche, Riad Riadi, Olivier Bachelier, Driss Mehdi and Laurent Rambault

Abstract— This paper proposes LMI conditions to estimate control law, associated with a matrix K. In other words,
the robustness of uncertain matrices with respect to root- it is aimed at robustly DC -stabilizing the uncertain closed-
clustering regions (any intersection and/or union of half planes, loop state matrix. This class of problems is generally very
discs and outsides of discs). The uncertainty complies with a
polytopic LFT description. This robustness criterion enables, difficult to be solved in a "traditional way". It is known as
as a special case, to derive a non fragility index for a pole N P-hard problems. A sufficient condition allowing to make
placement by output feedback gain. It is then exploited to up for this difficulty is proposed in this paper. This condition
improve the efficiency of such a feedback, using a genetic is expressed in terms of LMI.
algorithm. This paper is organized as follows : Section 2 states the
Index Terms—Robust Control, genetic algorithms, linear problem. Section 3 proposes a bound that estimates the
systems, LMIs, uncertain systems. robustness of pole location with respect to clustering regions
DC . In section 4, it is shown that this bound can be used as
I. INTRODUCTION a non fragility criterion for static output feedback. Models
with a non zero direct transmission matrix are considered.
D-stability of a state matrix is known to be an interesting Section 5 exploits the proposed conditions and considers the
property for linear MIMO model of system. Indeed, the root- possibility of maximizing the non fragility criterion in order
clustering can guarantee satisfactory transient performances. to improve the robustness of a pole placement. Section 6
For this reason many conditions for matrix D-stability were emphasizes the relevance of the notions through a numerical
proposed, some of them in terms of LMI [1], [2]. illustration before to conclude.
Nevertheless, all those formulations involve simple convex
regions. It can be however discerning to enlarge the choice
Notations : M
is the transpose conjugate of M , H(M ) is
of clustering regions especially to unions or combinations
of possibly non-symmetric subregions. This could allow far the Hermitian expression M + M
. The Kronecker product
is denoted by ⊗. ||M ||2 is the 2-norm of matrix M , i.e.
more possibilities in the specification of robust transient
the maximal singular value of M . II n is the identity matrix
performances (seperate fast and slow dynamics, damping
ratio, etc). Based upon the notion of matrix ∂D-regularity of order n, O is a null matrix of suitable dimension.
rCl k denotes the complex Dk -stability radius, Dk being a
[4] (non intersection between the matrix spectrum and some
region of the complex plane. In matrix inequalities “≺ 0”
curve ∂D), this paper aims at proposing this larger choice
(“ 0”) means negative (semi-)definite and ” 0” (“ 0”)
of regions. Unfortunately, this property may not hold in the
positive (semi-)definite. In the Hermitian matrices, notation
presence of uncertainty. For norm-bounded or LFT uncer-
tainty, the result of [3] based upon an LMI formulation (•)
prevent us from repeating the symmetric blocks.
close to the Bounded Real Lemma can be used. If a polytopic II. P ROBLEM S TATEMENT
uncertainty is considered, the result of [2] can be easily used. In this section, the problem is stated. First, the uncertain
Firstly, a robust analysis tool is proposed. It computes a ro- matrix is presented. Then, the description for clustering
bustness for a matrix subject to a polytopic LFT uncertainty. regions is given. At last, the precise goal is formulated.
This robustness bound, as large as possible, ensures that the
spectrum belongs to a specific region DC corresponding to A. The uncertain matrix
some combination of half-planes, discs, outsides of discs. A Consider the uncertain system modelled in a polytopic
matrix is said DC -stable if and only if its spectrum lies inside LFT form such that:
the region DC . The robustness bound is then a lower bound
of the ∂D-stability radius. It is computed owing to LMIs 8
> ω(t) = ∆z(t)
<
involving parameter-dependent Lyapunov matrices (PDLM). ẋ(t) = A(θ)x(t) + J(θ)ω(t) + B(θ)u(t)
The robustness bound is then used as a non fragility index : z(t)
> = L(θ)x(t) + D∆∆ ω(t) + D∆u (θ)u(t)
which can be maximized owing to a genetic algorithm in y(t) = C(θ)x(t) + Dy∆ (θ)ω(t) + D(θ)u(t)
(1)
order to improve the robustness of a static output feedback
This work was not supported by any organization where ∆ ∈ Cl q×r is an unknown matrix, belonging to B(ρ),
J. Bosche, R. Riadi, O. Bachelier, D. Mehdi and L. Rambault are the ball of (q × r) complex matrices checking ||∆||2 ≤ ρ.
with LAII-ESIP (Laboratoire d’Automatique et d’Informatique Industrielle u(t) and y(t) are respectively the control and output signals
- Ecole Supérieure d’Ingénieurs de Poitiers) Bâtiment de Mécanique,
40 Avenue du Recteur Pineau, 86022 POITIERS C EDEX, FRANCE whereas ω(t) and z(t) are fictitious input and output.
[email protected]

0-7803-9354-6/05/$20.00 ©2005 IEEE 400

Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.
 The generalized model is represented by: Clearly, Dk denotes either one side or the other side of the
⎡ ⎤ boundary ∂Dk . It can then be a half plane, a disc or the
A(θ) J(θ) B(θ) outside of a disc.
N (θ) = ⎣ L(θ) D∆∆ D∆u (θ) ⎦ . (2) Also define the region DC as a combination, i.e. any union
C(θ) Dy∆ (θ) D(θ) and/or intersection of the various subregions Dk . Such a
N (θ) is assumed to belong to a polytope of matrices: formulation of DC enables a very large choice of clustering
regions. A matrix is DC -stable when all its eigenvalues lie
( ) in DC . It is said ∂DC -singular when at least one of the
X
N
N = N = N (θ) ∈ Cl | N = (θi Ni ) ; θ ∈ Θ (3) eigenvalues belongs to ∂DC . Otherwise, it is DC -regular.
i=1
C. Problem statement
where Θ is the set of all barycentric coordinates: The purpose of this work is:
 1) to derive LMI conditions that enable to compute a
 
N
 robustness bound ρC , as large as possible, such that
Θ = θ = θ1 . . . θN ∈ {IR+ }N | θi = 1 . the uncertain matrix A defined in (8) is DC -stable
i=1 whatever ∆ in B(ρC ) is and for any θ in Θ,
(4)
2) to use the previous robust analysis tools in order to
Extreme matrices Ni are the vertices of N :
estimate the fragility of a static output feedback which
⎡ ⎤ DC -stabilizes a system,
Ai Ji Bi 3) to exploit this fragility criterion in a robust control
Ni = ⎣ L i D∆∆ D∆ui ⎦ , i = 1, ..., N . (5) context.
Ci Dy∆ i Di
III. ROBUST DC - STABILITY ANALYSIS
Then, the uncertain model can be written in the following
form: In this section, an LMI condition for A to be DC -stable,
» – » – leading to the computation of a bound ρC , is given.
A(θ) B(θ) J(θ) ˆ ˜
M (∆, θ) = + ∆

D∆u (θ)
C(θ) D(θ) Dy∆ (θ)
L(θ)
Theorem 1: Let A ∈ Cl n×n , DC and ∂DC be respectively
(6) an uncertain matrix as defined in (8) with (7) and (2),
a clustering region as defined in subsection II-B and its
with boundary. A is robustly DC -stable with respect to B(ρ)
∆ = ∆(II r − D∆∆ ∆)−1

(7) and N if there exist N sets Pi , each one made up by k̄
In this work, we are interested in the expression of the Hermitian matrices Pk,i ∈ Cl n×n , k ∈ {1, .., k̄} , and k̄
uncertain state matrix A = A(∆, θ), i.e.: matrices Gk ∈ Cl (2n+q+r)×n , k = 1, ..., k̄, such that:

∆L(θ) ∈ Cl n×n .
A = A(θ) + J(θ)
(8) Gk (Pi , Ni ) = Gk ≺ 0 ∀k, i ∈ {1, .., k̄} × {1, .., N } (12)

It is well known that this formulation encompasses the where 2 3

L
i
case where N linearly depends on some parameter deflec- Rk ⊗ Pki O
6 O 7
Gk = 4
5+
tions around a nominal value. O −II q D∆∆
Li O D∆∆ −γII r
B. Clustering regions DC ` ˆ ˜´
In this part, the reader is reminded of the definition of a H Gk Ai −II n Ji O ≺ 0 (13)
class of regions : "combined regions" (or DC -regions).
Definition 1: [5] : Let R be a set of k̄ Hermitian matrices with γ = ρ−2 .
Rk as in (9)
⎧   Proof:
⎪
⎪ rk00 rk10
⎨ Rk = Rk =

∈ Cl 2×2 (12) leads to (by convex combination)
rk
10 rk11
(9) Gk (Pk (θ), N (θ), D∆∆ ) ≺ 0 ∀k ∈ {1, .., k̄} , ∀ θ ∈ Θ.
⎪
⎪
(14)
⎩
rk11 = 0, rk11 ∈ Cl ∀k ∈ {1, ..., k̄} .
Then, the outline of the proof is given in [4] where the
and consider the following curves: authors propose a sufficient condition for robust ∂DC -
    regularity against B(ρ). In fact, this condition is the same as
  1
∂Dk = z ∈ Cl | fk (z) = 1 z̃ Rk = 0 . the one of theorem 1 where DC -stability is considered (and
z
(10) not ∂DC -regularity) due to the fact that nominal matrix A is
Relevant curves are lines or circles. Each curve ∂Dk enables supposed to be DC -stable. In [4], the considered uncertain
to define an associated region: matrix is A with D∆∆ = 0 what does not lead to major
modifications of the proof. 2
Dk = {z ∈ Cl | fk (z) ≺ 0} ∀k ∈ {1, ..k̄} . (11)

401
Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.
 IV. F RAGILITY ANALYSIS  = ∆(II p − D∆)−1 . The polytopic structure of
uncertainty ∆
In this section, a system with uncertain polytopic realiza- Ac is characterized by
tion (A(θ); B(θ); C(θ); D = D(O)) is subject to an output  
Ao B
feedback gain K. First, the uncertain closed-loop state matrix M = (22)
C D
is expressed under a polytopic LFT form, due to uncertainty
on K. Then an LMI condition leading to a fragility criterion that belongs to
of such a model is proposed. (
X
N
M = M = M(θ) ∈ Cl | M(θ) = (θi Mi ) =
A. The uncertain closed-loop state matrix i=1
Consider the open-loop continuous-time model given by: » – )
X
N
Ai + Bi K̂Ci Bi (II m + K̂D)
 (θi ); θ ∈ Θ
ẋ(t) = A(θ)x(t) + B(θ)u(t) (II p − DK)−1 Ci D
i=1
(15)
y(t) = C(θ)x(t) + Du(t) (23)

where A(θ) ∈ Cl n×n , B(θ) ∈ Cl n×m , C(θ) ∈ Cl p×n and where Θ is defined as in (4). The aim of this work, in this
D ∈ Cl p×m . It matters to note that matrix D is not assumed section, is then to compute a bound on ||∆||2 , as large as
zero. possible, such that, once the clustering region DC is spec-
In this work, it is assumed that B(θ) and C(θ) cannot ified, DC -stability of matrix Ac defined in (20) is ensured.
belong both to a polytope, i.e. the two uncertain polytopic We now propose a fragility analysis condition allowing the
realizations considered are (A(θ); B(θ); C = C(O)) and computation of this bound.
(A(θ); B = B(O); C(θ)).
Then, suppose that there exists a static output feedback K B. Fragility analysis and DC -stability
which DC -stabilizes the closed-loop system. The correspond- Theorem 2: Let Ac ∈ Cl n×n be defined as in (20) with
ing control writes (18) and (19), DC and ∂DC be respectively a region as
u(t) = Ky(t) (16) defined in subsection II-B and its boundary. Assume that
and we are interested in analyzing its fragility. Indeed, it A0 is DC -stable, then A is robustly DC -stable with respect
is impossible in practice to precisely implement this control to B(ρ) and M if there exist N sets Pi , each one made up
law so it is reasonnable to assume that the feedback matrix by k̄ Hermitian matrices Pk,i ∈ Cl n×n , k ∈ {1, .., k̄} , and
actually reads K = K + ∆ where ∆ is an uncertain term. k̄ matrices Gk ∈ Cl d(2n+q+r)×dn , k = 1, ..., k̄, such that:
The fact that D = 0 is not so easy to take into account so
we define ŷ(t) = y(t) − Du(t) = C(θ)x(t). Hence, we are Gk (Pi , Mi ) ≺ 0 ∀k, i ∈ {1, .., k̄} × {1, .., N } (24)
interested in:
ˆ
u(t) = K̂ŷ(t) = (K̂ + ∆)ŷ(t) = (II m + K̂D)−1 K̂y(t) (17)
with γ = ρ−2 , Gk and Mi defined in (13) and (23).
From (17) it is easy to deduce that: Proof:
This proof is given for all sub-boundaries ∂Dk , k = 1, .., k̄.
K̂ = (II m + K̂D)K and (18) For any θ ∈ Θ, it is easy to deduce that
ˆ = (II m + K̂D)∆(II p − DK − D∆)−1
∆ (19)
Gk (Pk (θ), M(θ), D) ≺ 0 ∀k ∈ {1, .., k̄} (25)
The closed-loop state matrix is then given by
Define matrice S by:
⎡ ⎤
ˆ ˆ II n A
0 O O
Ac = A(θ)+B(θ)K̂ C(θ)+B(θ)∆C(θ) = A0 +B(θ)∆C(θ)
S=⎣ O B
II m O ⎦ . (26)
(20)
O O O II p
With (19), equality (20) becomes:
Left and right multiply (25) respectively by S and S
yields
−1
Ac = A0 + B(θ)(II m + K̂D)∆(II p − DK − D∆) C(θ) , 2
Uk (A0 , Pk (θ)) Rk10 Pk (θ)B + Rk11 A
0 Pk (θ)B C

3
4 (•)
Rk11 B
Pk (θ)B − II m D
5≺ 0
with following definitions C D −γII p
⎧ (27)
⎨ B = B(θ)(II m + K̂D)
D = (II p − DK)−1 D with
⎩
C = (II p − DK)−1 C(θ) .
Uk (A0 , Pk (θ)) = Rk00 ⊗ Pk (θ) + H(Rk10 ⊗ Pk (θ)A0 )+
Ac complies to
Rk11 ⊗ A
0 Pk (θ)A0 . (28)
−1
Ac = A0 + B∆(II p − D∆) C. (21)
One can note that the expression of Ac given by (21) is Applying Schur’s lemma yields:
similar to the one given by (8) with (7) and (2). Closed- » –
Uk (A0 , Pk (θ)) + C
C Rk10 Pk (θ)B + Rk11 A
0 Pk (θ)B + C
D
loop state matrix Ac is subject to both polytopic and LFT ≺ 0.
(•)
Rk11 B
Pk (θ)B − γII m + D
D

402
Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.
 Remark 1: This proof could be carried out by left and
The previous inequality can be written: right multiplying (29) by [II n | C
∆
] and its transpose.
Nevertheless, it would only proved sufficiency whereas the
» –
» –
II n O II n O proof proposed in this paper also proves the necessity in the
(Rk ⊗ Pk (θ)) +Φ≺ 0 (29)
A0 B A0 B special case where N = 1 and the considered regions are
disjoined.
with
» –
C
C C
D V. A PPLICATION TO ROBUST DC - STABILIZATION
Φ = .
D
C

D D − γII m This section proposes a method to compute a robust
control law by static output feedback for any LTI system
By virtue of the KYP lemma, it is equivalent to: defined as in (15) with either B(θ) and C(θ) not uncertain.
» –
» – The purpose is to compute the matrix K which leads to the
(sII n − A0 )−1 B (sII n − A0 )−1 B
Φ ≺ 0 ∀s ∈ ∂Dk largest bound ρC . This method is based on the resolution
II m II m
of LMI conditions. First, the model (15) is considered
that can be written
with D = O and a DC -stabilization condition of the
√
||C(sII n − A0 )−1 B + D||2 ≺ γ = ρ−1 ∀s ∈ ∂Dk triplet (A(θ), B(θ), C(θ)), leading to K̂, is presented. In
the second part, D is a non zero matrix and a heuristic
⇔{ sup σ k ( C(sII n − A0 )−1 B + D )}−1  ρ . technique for robust placement design is proposed. This
s∈∂Dk    technique leads to a matrix K 3 associated with the more
Q robust static output feedback control law in the sense of the
Simple singular value arguments show that σ k (Q) = non fragility criterion.
µCl k (Q). Hence
{ sup µCl k (Q)}−1  ρ A. DC -stabilization of a polytope
s∈∂Dk
This part proposes a method to compute a robust control
Using classical arguments on singular values [6], it yields law by static output feedback for any MIMO linear system.
" #−1 This method is based on the resolution of tractable LMI
µCl k (Q) = inf {σ k (∆)|det(II p − Q∆) = 0} ≺ ρ−1 conditions and relies on an idea originally proposed in [7],
l m×p
∆∈C
[8] and then reformulated in [9]. This idea is here extended
to the concept of DC -stability and consists in finding a state
⇔ rCl (Q) = inf { inf {σ k (∆)|det(II p − Q∆) = 0}}  ρ
k s∈∂D k l m×p
∆∈C feedback and an output feedback checking the same robust
(30) DC -stability property through the existence of the same set
with of Lyapunov matrices. It amounts to solving the following
det(II p − Q∆) = 0 LMI system

⇔ det(II p − C(sII n − A0 )−1 B∆ − D∆) = 0 Uk (As (θ), Pk (θ)) ≺ 0
(32)
Uk (Âo (θ), Pk (θ)) ≺ 0 ∀k ∈ {1, .., k̄}
⇔ det(II p − C(sII n − A0 )−1 B∆(II p − D∆)−1 )det
with Uk (•) as defined in (28) and state matrices As (θ) and
(II p − D∆) = 0 . Âo (θ) defined by:
Now assume that the uncertain model is well-posed, it comes
As (θ) = A(θ) + B(θ)Ks and Âo (θ) = A(θ) + B(θ)K̂C(θ) .
det(II p − C(sII n − A0 )−1 B∆) = 0
(33)
⇔ det(II p − (sII n − A0 )−1 B∆C) = 0 Ks is associated with state feedback whereas K̂ DC -
stabilizes the triplet (A(θ), B(θ), C(θ)). Theorem 3
proposes a NSC for LMI system (32) to hold.
⇔ det(sII n − A0 )−1 det(sII n − A0 − B∆C) = 0 .
If A0 is assumed to be DC -stable then the previous equality Theorem 3: The LMI system (32) is satisfied if there
is equivalent to exist N sets Pi , each one made up by k̄ Hermitian
det(sII n − Ac ) = 0 matrices Pk,i ∈ Cl n×n , k ∈ {1, .., k̄} , a nonsingular matrix
According to the above equality, (30) can be written as G ∈ IRm×m as well as k̄ matrices Fk ∈ Cl (2n+m)×n
follows and H ∈ IRm×p such that the LMI (34) is satisfied
∀ k, i ∈ {1, ..., k̄} × {1, ..., N } for a given state feedback
⇔ rCl k (Q) = inf { inf {σ k (∆)|det(sII n − Ac ) = 0}}  ρ
s∈∂Dk ∆∈C
l m×p Ks that DC -stabilizes the pair (A(θ), B(θ)).
(31)
» –
K(Pi ) O ˘ ˆ ˜¯
with rCl k (Q) ≺ rCl D (Q) ∀k = 1, .., k̄.2 H ki = + H Fk Asi −II n Bi
O O

403
Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.
 82 3 9
< O ˜=
+H 4
ˆ
O 5 G −Ks O −II m
Step 2 : A static output feedback matrix K̂ γ which DC -
: ; stabilizes the triplet (A(θ), B(θ), C(θ)) is computed, if pos-
II m
82 3 9 sible, for each matrix Ksγ thanks to the tool presented by
< O ˜=
ˆ theorem 3. When possible, we obtain matrix K γ from (18).
+ H 4 O 5 H Ci O O ≺0. (34)
: II ; Step 3 : A robustness bound is then computed for each
m
matrix K γ thanks to the analysis tool proposed by theorem
with 2. When step 2 fails (i.e. when (34) is not feasible), the
k̄

K(Pi ) = (Rk ⊗ Pki ). (35) robustness bound is imposed zero (ργ = 0).
k=1
Test : Then we can define a convergence criterion which, if
satisfied, a priori picks up the "best" K (denoted by K 3 )
An output feedback matrix is then given by:
in the sense of the criterion, and stops the process. If the
K̂ = G−1 H (36) convergence test does not hold, an evolution process based
Proof : This proof is given for all subregions DCk , k = on genetic algorithms (not detailed in this paper) is used
1, .., k̄. For any θ ∈ Θ, it comes: to improve the robustness of the control law, iteration after
iteration, until the convergence criterion is finally satisfied.
N
 The robustness bound corresponding to K 3 is denoted by
HK (P(θ), As (θ)) = (θi Hki ) ≺ 0 . (37) ρ3 .
i=1

Then, the reasoning of the proof is quite analogous to the Step 1 :

Calculation of γ̄ Ksγ
the one given in [10] where the authors propose a necessary
and sufficient condition of DU -stabilization. 2
Step 2 :
Remark 2: Condition (34) depends on Ks associated with
Calculation of γ̄ K̂ γ
a static output feedback. In fact, this matrix is computed in Evolution Process

the nominal case. According to this, one of the two reasons

for which (34) is not satisfied is that KS does not DC - Step 3 :
Calculation of the γ̄ ργ
stabilize the polytope As (θ) = A(θ) + B(θ)Ks . The other
reason is that there is no matrix K̂ that shares the same sets
of Lyapunov matrices with Ks .
no satisfi ed CONVERGENCE
Remark 3: It is important to note that condition (34)
TEST
becomes also necessary for (32) when A(θ), B(θ) and C(θ)
do not belong to a polytope of matrices. satisfi ed

B. Robust pole placement K 3 = (II m + K̂D)−1 K̂

This work aims at computing a robust control law by

static output feedback. The considered clustering regions Fig. 1. Heuristic resolution
are the ones defined in (1).
The method consist in DC -stabilizing the triplet VI. N UMERICAL ILLUSTRATION
(A(θ), B(θ), C(θ)) leading to K̂ according to theorem
The computations are performed with M ATLAB on a PC
3. It is then easy to deduce, from theorem 2, the matrix
Pentium 1.2 GHz.
K associated with the static feedback control law and its
The example corresponds to the reduced model of the lateral
corresponding non fragility criterion. In fact, a genetic
automatic pilot of an aircraft. The open-loop continuous-time
algorithm begins with a set of parameters (chromosomes),
model is given by (15) where
associated with γ̄ static state feedback Ksγ and constituting  
the initial population. This population is used to form A(θ) B
M=
many others populations at least better than the former C(θ) D
ones by selecting the best chromosomes (i.e. the best set
describes a polytope with four vertices:
of parameters) generation after generation. After repeated 2 3
−0.0351 0.1016 −0.0094 −0.3426 −0.8625
reproductions, recombinations, mutations and selections 6 −0.0999 −0.1614 −0.8894 0.0860 −1.0510 7
processes, the algorithm leads a suitable solution K 3 . M1,2,3,4
6 0.0245
=6 0.9320 −0.0248 −0.1397 0.3705 7
7
4 0.0209 −0.4270 −0.3928 −0.0401 0.0181 5
The resolution method that we propose, illustrated by the 0.9278 −0.9638 0.0494 0.2789 −0.1230
flow chart of the figure 1, is as follows:
2 3 2 3
0 0.02 0
O O
Step 1 : Several (γ̄) state feedbacks Ksγ , each one that 6
±4
0 0 0 O 7 4
5± 0.002 0 0 5 .
0 0 0 O
DC -stabilizing the pair (A, B), are computed by using O O
0.0093 0 0
a beforehand specified eigenstructure assignment technique
and exploiting the dof (degree of f reedom) on the eigenvalues It is aimed at determining matrix K 3 associated with
and eigenvectors [11]. the static output feedback, placing the poles of the above

404
Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.
uncertain model in a region DC of the complex plane, while
maximizing the robustness bound ρ corresponding to the
radius of the uncertainty domain.

DC here results from the intersection of three subregions :

D1 , D2 and D3 . D1 corresponding to the outside of a disc
of radius 1.7 centered around −2. D2 is a disc also centered
around −2 of radius 4. D3 is a vertical half-plane defined
by x < −1.5. This algorithm was initialized with a popu-
lation of 50 chromosomes, i.e. with 50 randomly generated
state feedback matrices. After 9 generations (so nearly 450
matrices Ksγ ), heuristic method proposed in subsection V-B Fig. 2. Spectrum of Ac with ||∆K  ||2 ≤ ρ3
leads to the following first static output feedback respectively
associated with matrices K̂0 and K0 .
» – » – output feedback gains. Finally, a genetic algorithm was
7.0683 24.1834 −3.3297 3.1581
K̂0 = , K0 = .
−7.1951 7.8093 −11.2104 −0.2178 used to maximize such a criterion while designing a pole
placement by output feedback that enables to improve the
The corresponding robust bound is then: robust transient performances of the control law.
R EFERENCES
ρ0 = 0.0148 .
[1] M. Chilali and P. Gahinet, “H ∞ Design with Pole Placement Con-
Although this control law DC -stabilizes the uncertain model, straints : An LMI Approach,” IEEE Transactions on Automatic Con-
trol, vol. 41, no. 3, pp. 358–367, march 1996.
the evolution process enables to improve the robustness of [2] D. Peaucelle, D. Arzelier, O. Bachelier, and J. Bernussou, “A new
the closed-loop system. After 82 generations, it finally leads robust D-stability condition for real convex polytopic uncertainty,”
to : Systems and Control Letters, vol. 40, no. 1, pp. 21–30, may 2000.
» – » – [3] M. Chilali, P. Gahinet, and P. Apkarian, “Robust pole placement
4.4274 19.0775 −3.3186 2.9648
K̂ 3 = , K3 = . in LMI regions,” IEEE Transactions on Automatic Control, vol. 44,
−8.2306 5.1818 −10.9486 −0.3813 no. 12, pp. 2257–2269, december 1999.
[4] O. Bachelier, D. Henrion, B. Pradin, and D. Mehdi, “Robust matrix
root-clustering of a matrix in intersections or unions of subregions,”
with the following robustness bound To appear in SIAM Journal of Control and Optimization, pp. xxx–xxx,
x 2004.
ρ3 = 0.0787 . [5] O. Bachelier and D. Mehdi, “Robust D U -stability analysis,” Interna-
tional Journal of Robust and Non-linear Control, vol. 13, no. 6, pp.
The average computation time was about 560 seconds by 533–558, 2003.
generation. [6] L. Qiu, B. Bernhardsson, A. Rantzer, E. J. Davison, P. M. Young,
and J. Doyle, “A formula for computation of the real stability radius,”
On figure 2, the uncertain closed-loop state matrix roots Automatica, vol. 31, no. 6, pp. 879–890, 1995.
are plotted for many random values of ∆K  such that [7] D. Arzelier, D. Peaucelle, and S. Salhi, “Robust static output feedback
||∆K  ||2 ≤ ρ3 . Such a plot is a way to appreciate the stabilization for polytopic uncertain systems : improving the guaran-
teed performance bound,” in Proceedings of ROCOND, Milan-Italy,
shape of the spectral value set [12], represented by black June 2003.
points on the figure. It was also studied under the name [8] D. Peaucelle and D. Arzelier, “An effi cient numerical solution for
of pseudospectrum in [13], [14]. The closed-loop output H2 static output feedback synthesis,” Europeen Control Conference,
September 2001.
matrices roots corresponding to the four vertices of the [9] D. Mehdi, E. Boukas, and O. Bachelier, “Static output feedback
polytope are also plotted on this figure, represented by design for uncertain linear discrete time systems,” IMA Journal of
yellow ’+’. Mathematical Control and Information, vol. 21, no. 1, pp. 1–13, 2004.
[10] J. Bosche, O. Bachelier, and D. Mehdi, “Robust pole placement
It shows the efficiency of this approach and the conservatism by static output feedback,” Conference on Decision and Control,
induced by the technique appears relatively weak even if this December 2004.
example is not trivial. [11] B. C. Moore, “On the flexibility offered by state feedback in mul-
tivariable systems beyond closed loop eigenvalue assignment,” IEEE
Transactions on Automatic Control, vol. 21, pp. 689–692, 1976.
VII. C ONCLUSION [12] D. Hinrichsen and B. Kelb, “Spectral value sets: a graphical tool for
In this paper, a method for robust pole placement by robustness analysis,” Systems and Control Letters, vol. 21, pp. 127–
136, 1993.
static output feedback using a genetic algorithm is proposed. [13] F. Tisseur and N. J. Higham, “Structured pseudospectra for polynomial
Firstly, a condition for robust DC -stability of a matrix with eigenvalue problems, with applications,” SIAM Journal of Matrix
respect to polytopic LFT uncertainty is proposed. Through Analysis and Application, vol. 23, no. 1, pp. 187–208, 2001.
[14] L. . N. Trefethen, “Pseudospectra of linear operators,” SIAM Review,
this condition, a robustness bound is derived. It is expected to vol. 39, no. 3, pp. 383–406, 1997.
be good lower bound of the complex D-stability radius. The
main originality of the paper stands in the large choice of
clustering regions that corresponds to some combination of
half-planes, discs and outsides of discs. Then, the robustness
bound turned to be useful as a non-fragility criterion for

405
Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:16:55 UTC from IEEE Xplore. Restrictions apply.