Sheffield UK
April 2005
© 2005 Arturo Molina-Cristobal
This thesis addresses some of the open problems in multiobjective control. The aim of
this thesis is to compare two emerging techniques in multiobjective control: evolutionary
algorithms (EAs) and convex optimisation over linear matrix inequalities (LMIs).
This investigation addresses some of the drawbacks of LMI techniques, such as the
inability to reduce the order of the controller and the LMI technique's inherent
conservatism when tackling multiobjective problems. The truly multiobjective
optimisation using an EA is proposed to overcome these problems. Reduction of the
order of the controller is achieved for problems such as the H∞ controller design with
time-domain specifications. The mixed H2/H∞ control problem is treated as a
multiobjective H2/H∞ control problem, and an improvement of the Pareto optimal set is
achieved.
Both methods are applied to a gas turbine engine controller design problem. Numerical
problems with controllers resulting from using the LMI approach are addressed and a
solution, based on the EA, is shown to design more numerically robust controllers than
the LMI approach.
i
The controller design from this application problem is then extended to a gain scheduling
controller design with the use of linear parametric-varying (LPV) techniques. The issue
of applying LPV modelling and LPV control is addressed, and the EA-based method is
proposed to design multiobjective H2/H∞ controllers for two operating points. The results
give satisfactory, well-conditioned controllers.
ii
Acknowledgments
I would like to thank my supervisors Professor Peter J Fleming and Professor David H
Owens for their advice, guidance, and encouragement throughout the period of this
research work. It has been a privilege working with them.
I would also like to thank Didier Henrion for introducing me to novel areas of LMI
techniques. Thanks also to my colleagues Jorge Escamilla for discussions in related
topics of this thesis, Robin Purshouse for his expert advice in genetic algorithms, Ian
Griffin for his expert advice in gas turbine engines, and thanks to John Watson and Mike
Dewar who proof read this manuscript.
This work has been funded by Consejo Nacional de Ciencia y Tecnología (CONACyT) –
México. I have also received financial support from the Rolls Royce University
Technology Centre (RR-UTC) at the Department of Automatic Control and Systems
Engineering, Sheffield University, which facilitated me to attend technical events related
to this research work.
I am extremely grateful to my parents Eve and Arturo, brothers Felix and Tonatiuh and
especially to my sister Esperanza for their love and support. Thanks and my love goes to
Helen for her understanding and support for the completion of this project.
iii
Statement of Originality
Unless otherwise stated in the text, the work described in this thesis was carried out
solely by the candidate. None of this work has already been accepted for any degree, nor
is it concurrently submitted in candidature for any degree.
Candidate:
________________________________
Arturo Molina-Cristóbal
Supervisors:
_________________________________
Peter J Fleming
_________________________________
David H Owens
iv
A mi Mamá y Papá con Cariño
(To my Parents with love)
v
Contents
Chapter 1 ......................................................................................................................1
Introduction..................................................................................................................1
1.3 Contributions............................................................................................................4
Chapter 2 ......................................................................................................................6
2.1 Introduction..............................................................................................................6
2.3.1 Convexity……………………………………………………………………...7
2.3.2 Geometry of LMI sets .......................................................................................9
2.3.3 Semidefinite Programming (SDP) ..................................................................10
vi
2.8 Multiobjective optimisation ...................................................................................25
2.12 Conclusions..........................................................................................................44
Chapter 3 ....................................................................................................................45
3.1 Introduction............................................................................................................45
vii
3.5. Conclusions...........................................................................................................68
Chapter 4......................................................................................................................69
Mixed H2/H∞ problem with LMIs Approach and MOGA-based method. ..................69
4.1 Introduction............................................................................................................69
4.9. Conclusions...........................................................................................................96
Chapter 5 ....................................................................................................................97
5.1 Introduction............................................................................................................97
viii
5.3.1 Scaling……………………………………………………………………...104
5.3.2 Design Controller Strategy. ..........................................................................104
5.12 Conclusions........................................................................................................136
Chapter 6 ..................................................................................................................137
6.2 Suitability.............................................................................................................138
Appendix A................................................................................................................142
Appendix B ................................................................................................................143
References:.................................................................................................................147
ix
Chapter 1
Introduction
1.1 Motivation of the Research
It is safe to say that any system design involves simultaneous consideration of multiple
criteria. In control systems engineering, the design of a single controller which meets a
collection of performances or objectives is, in theory and in practice, an open problem.
Often some of these objectives will be in conflict, e.g. rise time and overshoot, and might
be limited, e.g., the amount of achievable disturbance attenuation in a certain frequency
region is limited by constraints on the systems bandwidth, where trying to improve one
specification will cause deterioration in another. Thus, a trade-off exists, and is treated by
using multiobjective control optimisation, in which a trade-off is sought between
competing objectives. In such an optimisation, no single optimal solution exists, rather a
set of equally valid solutions, known as the Pareto optimal set.
The multiobjective control problem has been under research for some time and several
techniques have been proposed to solve this problem. However, research in this area has
been initially carried out with single optimisation techniques; this approach makes the
multiobjective problem into a formulation of a single objective problem. Moreover, all
specifications and constraints of a different nature have to be translated into a unique
performance measure (e.g. H2, H∞).
1
Chapter 1 Introduction
Specifically, theoretical results of the mixed H2/H∞ problem have shown the difficulties
that a unified framework involves. Moreover, in practice it is difficult to solve coupled
Riccati equations (Doyle et al. 1994; Zhou et al. 1994). Research into this problem has
shown that it can be reinterpreted as a convex optimisation problem (Khargonekar and
Rotea 1991a; Scherer 1995b). This approach relies on the efficient interior-point
algorithm to solve the optimisation problem. Almost a decade ago, this observation
introduced a whole new area of research known as semidefinite programming (SDP),
which is a convex optimisation problem over linear matrix inequalities (LMIs). The
motivation to use LMIs is that several control specifications are naturally formulated as
convex constraints. Thus, a unified framework can be proposed (Skelton and Iwasaki
1994). Although advances in control systems with the LMI approach were achieved, this
formulation did not overcome open problems, such as conservatism and the low order of
the controller. Moreover, this new formulation leads to single objective optimisation
problems when multiobjective control problems are addressed, such as the multiobjective
H2/H∞ problem. Therefore, this problem is named the mixed H2/H∞ problem.
2
Chapter 1 Introduction
Chapter 4 considers two multiobjective control problems. First, the LQG problem is
studied, which is used to find the exact trade-off between process and sensor noise. For
this LQG problem, a complete LMI formulation is presented, where the equivalence
between LQG and H2 performance is used to formulate the convex optimisation problem
over LMIs. The LQG problem was then presented using MOGA-based method in the
same way as Fonseca and Fleming (1994). Secondly, the mixed H2/H∞ problem is
optimised using the convex optimisation over LMIs, resulting in a trade-off. A discussion
of the conservatism of this result is provided. The mixed H2/H∞ problem is then presented
using the MOGA-based method and comparative results are shown.
Chapter 5 designs a MIMO controller for an operating point of the gas turbine engine
model. Both the mixed H2/H∞ controller design: LMI approach and the multiobjective
H2/H∞ controller design: the MOGA-based method was considered for this design. The
problem of applying gain scheduling control design is studied. Both the linear parametric
varying (LPV) model of the gas turbine engine and the LPV controller design are
3
Chapter 1 Introduction
considered. Finally the LPV gain scheduling: MOGA-based method is proposed and two
operating points are considered.
Chapter 6 presents a conclusion on the research undertaken and offers suggestions for
future research.
1.3 Contributions
• A new approach to optimise controller structures was proposed. This approach
contrasts with the traditional implementation of parameters in state-space
framework. The new controller structure is based on the new developments on
Youla-Kučera parametrization and the optimisation design of weighting functions
by Schroder (1998). The design procedure to optimise controller structure
involved the use of the multiobjective genetic algorithm (MOGA) as an
optimisation tool.
4
Chapter 1 Introduction
5
Chapter 2
Review of Linear Matrix Inequalities and
Evolutionary Optimisation in Multiobjective
Control
2.1 Introduction
This chapter is devoted to reviewing the two optimisation techniques used in this work
for multiobjective control. Although this thesis is related to the topic of multiobjective
control, the review will start with the definition of linear matrix inequalities (LMIs),
convexity optimisation, and the role of LMIs in control systems. The history of linear
matrix inequalities in control systems aims to explain how convex optimisation theory
was developed from the linear programming optimisation tool to the interior-point
method and to analyse its importance in control systems. Then, the translation of control
problems into a set of LMI constraints will be explained using the LQR optimal problem
and the H∞ norm problem. The multiobjective optimisation problem is stated with
fundamental definitions, as well as the multiobjective control problem. This, in then, will
be followed by illustration of how multiobjective control problems can be tackled with
LMI techniques. Finally, the evolutionary algorithm approach to multiobjective control
problem, known as the multiobjective genetic algorithm (MOGA), is introduced.
6
Chapter 2 Review of LMIs and EAs in Multiobjective Control
m
F ( x) = FO + ∑ xi Fi > 0, (2.1)
i =1
where x ∈ ℜ m is the variable, and Fi = FiT ∈ ℜ nxn , i=0,…,m are given. The inequality
symbol in (2.1) means that F(x) is positive-definite, i.e., u T F ( x)u > 0 for all non-zero
⎛ F1 ( x) 0 0 ⎞
⎜ ⎟
⎜ 0 O 0 ⎟>0. (2.2)
⎜ 0 0 Fn ( x) ⎟⎠
⎝
2.3.1 Convexity
Convexity is one of the main geometric ideas underlying much global optimisation
theory, and the foundation for analytical tools used in LMI techniques.
A set C is convex if the line segment between any two points in C lies in C, i.e., if any x,
y ∈ C and any λ with 0 ≤ λ ≤ 1 , thus the line will be
λx + (1 − λ) y ∈ C . (2.3)
In Figure 2.1a, the set is convex because a line between any two points within the set C
will be contained in the region, whereas in Figure 2.1b not every line between any two
points is contained in the set C.
7
Chapter 2 Review of LMIs and EAs in Multiobjective Control
x
C
x
C
y y
a) b)
if x, y ∈ ℜ n ,0 ≤ λ ≤ 1 then,
f (λx + (1 − λ ) y ) ≤ λf ( x) + (1 − λ ) f ( y ) . (2.4)
A function f is convex if, for every pair of x, y the graph of f (λx + (1 − λ ) y ) lies on or
below a straight line λf ( x) + (1 − λ ) f ( y ) . Geometrically, the inequality means that the line
between (x, f(x)), (y, f(y)) lies above the graph of f. (See Figure 2.2).
f(.)
λf(x)+(1-λ)f(y)
x y
λx+(1-λ)y
8
Chapter 2 Review of LMIs and EAs in Multiobjective Control
C = {x ∈ ℜ n | Fi ( x) ≥ 0, ∀i = 1,..., m} (2.5)
⎧ n
⎫
f = ⎨ x ∈ ℜ n | F ( x) = F0 + ∑ xi Fi > 0⎬ , (2.6)
⎩ i =1 ⎭
the matrix F(x) is positive semidefinite if and only if its diagonal minors fi(x) are non-
negative. Diagonal minors are multivariate polynomials of independent xi. So the LMI set
can be described as
{ }
F ( x) = x ∈ ℜ n | f i ( x) > 0, i = 1,L, n (2.7)
Sets defined by polynomial inequalities and equations are called semialgebraic sets.
Moreover, they are convex sets. In order to illustrate feasibility and the convexity of
LMIs, a two dimensional subspace (x1, x2) is represented in Figure 2.3. For the sake of
illustrating, the geometry of an LMI feasible set, an example was taken from (Parrillo and
Lall 2003),
⎛3 0 1⎞ ⎛ −1 −1 0 ⎞ ⎛ 0 −1 0⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
F ( x) = ⎜ 0 4 0 ⎟ + x1 ⎜ − 1 0 0 ⎟ + x2 ⎜ − 1 − 1 0 ⎟ > 0
⎜1
⎝ 0 0 ⎟⎠ ⎜0
⎝ 0 − 1⎟⎠ ⎜0
⎝ 0 0 ⎟⎠
feasible, if and only if, all principal minors are nonnegative system of polynomial
inequalities f i ( x) > 0 .
9
Chapter 2 Review of LMIs and EAs in Multiobjective Control
The intersection of all these minors of the LMI set (2.8) generates the feasible set, which
is the shaded region in Figure 2.3.
10
Chapter 2 Review of LMIs and EAs in Multiobjective Control
minimize cT x
m
subject to F ( x) = FO + ∑ xi Fi > 0, (2.9)
i =1
where c∈ ℜ m , and F>0 means the matrix F is symmetric and positive semidefinite, also
denoted as F f 0 .
In the 1940’s; the first engineering application appeared when in the Soviet Union Lur’e
and Postnikov, (1944) (cited in (Lur'e 1961; Boyd et al. 1994a)), used a set of three first-
order equations containing nonlinear term to formulate a Lyapunov function. With this
work, they solved the problem of stability of a control system with nonlinearity in the
actuator (Vidsasagar 1993). Although they did not form LMIs, their stability criteria have
the form of LMIs, and were solved by hand for very small systems (Boyd et al. 1994a).
11
Chapter 2 Review of LMIs and EAs in Multiobjective Control
In the 1960’s; Popov (1962), Yakubovich (1962) and Kalman (1963) developed graphical
techniques for solving certain families of LMIs. Nowadays, their results are known as the
KYP lemma. In the late 1960’s; LMIs were intensively studied in control theory by
Yakubovich (1964; 1965a; 1965b), who developed the method of matrix inequalities in
the theory of nonlinear control and found solutions to certain matrix inequalities.
In the 1970’s; Willems (1971) found that analytical treatment of quadratic optimal
control problems leads to a series of matrix relations and frequency domain inequalities,
such as the linear matrix inequality:
⎛ AT K + KA + Q KB + C T ⎞
F ( K ) = ⎜⎜ T
⎟ ≥ 0,
⎟ (2.11)
⎝ B K + C R ⎠
and the algebraic Riccati equation (ARE) was considered for its solution. Anderson and
Vongpanitlerd (1973) found a solution for a simplified version of the general form of the
quadratic matrix inequality and found that solving the general form is difficult. Willems
recognised that LMIs must be solved by computational algorithms, and he quoted:
Although graphical solutions for special forms of LMI systems which appear in
engineering applications were available, the major problem faced was to solve more
general forms of LMI systems.
In the early 1980’s; two separate major advances happened; The first, an interior-point
algorithm was proposed by Karmarkar (1984). This polynomial-time linear programming
method was faster than the simplex method (developed by Dantzing in 1948 (Dantzing
1963)). The second major advance was the observation that LMIs in control and in
systems theory could be formulated as convex optimisation problems (Horisberger and
Belanger 1976). Pyatnitskii and Skorodinskii (1982) reduced the problem of Lur’e to a
convex optimisation problem involving linear matrix inequalities, which they solved
12
Chapter 2 Review of LMIs and EAs in Multiobjective Control
using an ellipsoid algorithm (developed in the 1970’s to prove that linear programs can
be solved in polynomial-time (Boyd et al. 1994b)). At this point, the interior-points
algorithm and convex optimisation involving LMIs in control systems were two separate
areas.
In the late 1980’s; other authors started formulating control systems as a convex
optimisation problem involving LMIs, such as the quadratic stabilisation of uncertain
systems and solved using linear programming (Bernussou et al. 1989). The work of Boyd
et al. (1988) showed that convex optimisation was emerging as a technique for controller
design. In this work, objectives or control specifications, such as overshoot, settling time,
classical single-loop gain/phase margins, were translated into standard convex
programming, using Youla parametrization of all stable controllers. Nesterov and
Nemirovski (1988), in Russia, extended the interior-point method for convex
optimisation over LMI constraints, but it is not until the next decade that their work was
spread and widely recognised. The reader is referred to Tokhomirov (1996) for a more
detailed Russian version of the history of convex optimisation.
13
Chapter 2 Review of LMIs and EAs in Multiobjective Control
research, like linear programming (LP), conic quadratic programming (CQP) and
semidefinite programming (SDP) (Vandenberghe and Boyd 1996; Boyd et al. 1997;
Vandenberghe and Balakrishnan 1997; Forsgren et al. 2002). In control systems, several
problems related to convex optimisation involving linear matrix inequalities can be
solved efficiently using interior-point algorithms. Numerical experiments shows that
these algorithms solve LMI problems with a polynomial-time complexity (Gahinet and
Nemirovski 1997). These algorithm is also known as the Projective method and the first
implementation, as a piece of commercial software, is the LMI Control Toolbox for use
with MATLAB® (Gahinet et al. 1995). Another popular solver is SeDuMi, developed by
Sturm (1999). Both of them can be interfaced with YALMIP, developed by Lofberg
(2004); this software provides a more friendly way to write LMIs in MATLAB. The
interested reader in more theoretical developments of convex optimisation is referred to
the books: “Lectures notes on modern convex optimisation” by Ben-Tal and Nemirovski
(2001) and “Convex optimisation” by Boyd and Vandenberghe (2004).
H∞ Control Theory: One of the researchers who intensively studied H∞ optimisation was
Scherer (1990). His work was based on the previous work of Willems (1971), Willems
studied the solvability of Riccati inequalities and the bounded real lemma which are the
regular tools for H∞ optimisation. Other researchers (Lu and Doyle 1992; Packard 1994)
formulated the state-space H∞ control problem as a convex constraint using linear
fractional transformation (LFT). Iwasaki and Skelton (1994a) gave necessary and
sufficient conditions for the existence of an H∞ controller, in terms of LMIs. The first
14
Chapter 2 Review of LMIs and EAs in Multiobjective Control
solvability conditions of H∞ control involving the LMI approach was given by Gahinet
(1994), Gahinet and Apkarian (1994), where the H∞ synthesis is performed in two steps
(find a Lyapunov matrix and obtain the controller); here, the Riccati equations are
replaced by Riccati inequalities. This inequality formulation provided a parametrization
of all H∞ suboptimal controllers, and the computation of the controller reduces to solving
LMIs. In 1997, a one step procedure to overcome the expensive numerical optimisation
was proposed by Gahinet 1996). Other lines of research studied the flexibility of LMI
techniques in combining several constraints with H∞ performance. Chilali and Gahinet
(1996) and Chilali et al. (1999) gave the LMI formulation to combine Pole-Placement
objectives, H2 performance and H∞ performance. This formulation together with the
mixed H2/H∞ formulation by Scherer (1995b) were the first investigations of
multiobjective optimisation with LMI techniques in control systems, a topic which will
be reviewed later in this chapter.
Some researchers have proposed general frameworks in order to obtain a unified LMI
approach. Skelton and Iwasaki (1995) have combined the Covariance Control approach
with the LMI approach, and it has resulted in added computational efficiency. This
approach reduces a large class of single problems to a single problem in linear algebra. It
is also important to mention that all advances of LMI techniques mentioned up to now
have been developed under the framework of state-space techniques. However, LMI
techniques have also been studied under the framework of polynomial systems. A survey
of these developments is found in Henrion (2000); Henrion et al. (2003). Other advances
on robust stability and performance analysis of uncertain systems using LMIs and convex
methods for robust H2 analysis are found in Balakrishnan and Kashyap (1999) and
Paganini (1999a) respectively. For a more complete survey of advances in linear matrix
inequalities up to the year 2000, the reader is referred to the monograph of Ghaoui and
Silviu-lulian (1999) as well as the book: “Robust Control: a convex approach” by
Dullerud and Paganini (2000).
15
Chapter 2 Review of LMIs and EAs in Multiobjective Control
• The mixed H2/H∞ controller design by Scherer et al. (1997) was applied to a space
Launch Vehicle (Clement 2004) and to a satellite controller (Zasadzinski and
Frapard 2004).
• The fly by wire helicopter controller (Prempain and Postlethwaite 2004) was
designed using a static H∞ loop shaping controller, derived from nonlinear
inequalities and the determination of the controller reducing to an optimisation
problem over LMIs. This controller was designed low-order and was successfully
flight-tested and high levels of performance were achieved.
• A LMI approach was used for robustness analysis of fighter aircraft control laws
(Biannic 2004). Because of the need not only for linear time invariant
uncertainties, but also for time varying uncertainties, the integral quadratic
constraint (IQC) approach (Megretski and Rantzer 1997) was also implemented.
16
Chapter 2 Review of LMIs and EAs in Multiobjective Control
As a summary of this workshop, the practical applications showed that LMI techniques
improve results over classical methods, but still conservative, low-order controllers are
required, and in some cases there is a numerical limitation (number of variables) with the
state of the art LMI solvers.
In this workshop, theoretical achievements in the area of LMI techniques were reported.
The problems found that real applications, when LMI techniques are used, have
motivated further investigations on convex optimisation (Nesterov and Nemirovski
1994). In convex optimisation, SDP has been bringing new ways to formulate control
theory as SDP problems (Vandenberghe 2004), i.e. minimization of a linear objective
subject to LMI constraints. Among these formulations, the ones that stand out are: duality
theory, and LMI relaxations.
Duality theory provides systematic and unified proofs of necessary and sufficient
conditions for solvability of LMIs. Recently it has been used to obtain lower bounds of
the optimal value (Balakrishnan and Vandenberghe 2003). In robust control, the system
descriptions are affected by either time-varying parametric or dynamic uncertainties. The
area of LMI relaxations is attempting to deal with these descriptions and give new LMI
formulations, e.g. in gain scheduling design, linear parametric varying (LPV) can be
formulated to handle rate of variations and naturally capture the parameter dependency of
the control system (Scherer 2003; 2004).
17
Chapter 2 Review of LMIs and EAs in Multiobjective Control
⎛0 1 ⎞
Figure 2.5. Lyapunov’s LMI with A = ⎜⎜ ⎟⎟ ,
⎝ − 1 − 2⎠
Lyapunov also showed that this first LMI could be explicitly solved, choosing any value
of Q where Q = QT > 0 . The linear equation AT P + PA = −Q for the matrix P is
guaranteed to be positive-definite if the system x& = Ax is stable. The same problem can
be formulated as a SDP problem or LMI optimisation problem, thus,
where Q = Q T is positive definite. Assume that the system is asymptotically stable, then
all solutions of (2.12) are bounded so that J < ∞ . In order to show this, consider a
quadratic Lyapunov function V ( x(t )) = x T (t ) Px(t ) to establish a bound on J. Using the
dV
dt
( x(t )) =
d
dt
( )
x(t ) T Px(t ) ≤ − x T (t )Qx(t ) dt , (2.14)
18
Chapter 2 Review of LMIs and EAs in Multiobjective Control
which is negative definite, for all trajectories and all t. Integrating this inequality from t=0
to t=T , the equation (2.14) becomes
T
x(T ) Px(T ) − x(0) T Px(0) ≤ − ∫ x(t ) T Qx(t )dt ,
T
d
dt
( )
x(t ) T Px (t ) = x(t ) T ( AT P + PA) x(t ) ≤ − x(t ) T Qx(t ) , (2.16)
Thus, finding a P that provides a bound J can be done by solving an LMI feasibility
problem. Moreover, it can be optimized over the Lyapunov function P by searching for
the best (i.e. smallest) bound, the optimisation problem being written as follows,
Since xoT Pxo is a linear function of the variable P, this is an SDP in the variable P = PT,
and is numerically tractable by using interior-point methods.
19
Chapter 2 Review of LMIs and EAs in Multiobjective Control
In this Section some of the fundamental tools for expressing control problems as LMI
optimisation problem will be presented. This procedure involves changes of variables and
the use of Schur complements, and the use of the KYP lemma.
⎛Q S⎞
⎜⎜ T ⎟ is positive definite, (2.18)
⎝S R ⎟⎠
Proof:
⎛ I 0 ⎞⎛ Q S ⎞⎛ I − Q −1 S ⎞ ⎛ Q 0 ⎞
⎜⎜ T −1 ⎟⎜ ⎟⎟⎜⎜ ⎟ = ⎜⎜ ⎟
T −1 ⎟ . (2.21)
⎝− S Q I ⎟⎠⎜⎝ S T R ⎠⎝ 0 ⎟
I ⎠ ⎝0 R−S Q S⎠
Thus, the set of nonlinear inequalities (2.19) and (2.20) can be represented as the LMI
(2.18). An example to illustrate the use of Schur complement, is the quadratic matrix
inequality
where A, B, Q=QT, R=RT>0 are given matrices of appropriate size, and P=PT is the
variable. Note that this is a quadratic inequality matrix in the variable P. It can be
expressed as the linear matrix inequality using Schur complements, see equation (2.23).
⎛ − AT P − PA − Q PB ⎞
⎜ ⎟ > 0. (2.23)
⎜ B T
P R ⎟
⎝ ⎠
20
Chapter 2 Review of LMIs and EAs in Multiobjective Control
and the feedback gain K, i.e., u (t ) = Kx(t ) , so the closed-loop system is described by
x& (t ) = ( A + BK )x (t ) , x ( 0) = x o . (2.25)
The linear quadratic regulator (LQR) can be used to solve a family of regulator design
problems in which the state is accessible and regulation and actuator effort are each
measured by mean-square deviation. The LQR cost function has the form,
∫ (x(t ) )
∞
( )
∞
J = T
Qx (t ) + u (t ) T Ru (t ) dt = ∫ x(t ) T Q + K T RK x(t )dt , (2.26)
0 0
where Q and R are positive definite. This cost depends on the trajectory, of x(t ) , taken,
so the worst trajectory will correspond to the worst cost. The control problem is to find a
state feedback gain K and a quadratic Lyapunov function P that minimizes the bound
xoT Pxo on the worst cost of J . With the LMI formulation from Section 2.6, this problem
can be solved by designing K and finding a Lyapunov function that gives a guaranteed
performance bound. Proceeding in the same way as Section 2.6, the problem can be
translated into an optimisation problem as follows,
minimize xoT Pxo
As a result and contrary to the SDP problem, this optimisation problem cannot be
considered an SDP because the constraint involves a quadratic term K and products
between the variables P and K. However, this problem can be formulated as an SDP with
the following transformations (Boyd et al. 1997). Define new matrices Y and W as,
Y = P −1 , W = KP −1 ,
21
Chapter 2 Review of LMIs and EAs in Multiobjective Control
P = Y −1 , K = WY −1 .
Substituting of the variables Y and W instead of P and K in equation (2.27), and then
multiplying the left and the right hand sides by Y, the inequality above becomes
⎛ − (YAT + W T BT + AY + BW ) Y W T ⎞
⎜ ⎟
⎜ Y Q −1 0 ⎟ ≥ 0 , (2.29)
⎜⎜ ⎟
⎝ W 0 R −1 ⎟⎠
where it is assumed that Q and R are invertible. Once again, using the Schur complement,
the cost,
xoT PxoT = xoT Y −1 xo ≤ γ ,
is expressed as the LMI
⎛γ xoT ⎞
⎜ ⎟≥0. (2.30)
⎜x Y ⎟⎠
⎝ o
Then, the SDP problem can be formulated in terms of LMIs (2.29) and (2.30)
minimize γ
⎛ − (YAT + W T BT + AY + BW ) Y W T ⎞
⎜ ⎟
−1
subject to ⎜ Y Q 0 ⎟≥0, (2.31)
⎜⎜ −1 ⎟
⎝ W 0 R ⎟⎠
⎛γ xoT ⎞
⎜ ⎟≥0.
⎜x Y ⎟⎠
⎝ o
This is an SDP problem that can be solved by interior-point methods. The reader is
referred to Feron et al. (1992) for the stochastic interpretation of the LQR optimal
problem and its LMI formulation. In the stochastic case, the cost function (2.26) is
formulated as the expected value of the quadratic function and the state space equation
(2.24) contains unit intensity white noise.
22
Chapter 2 Review of LMIs and EAs in Multiobjective Control
∀ω ∈ ℜ .
The corresponding equivalence for strict inequalities holds even if (A, B) is not
controllable.
⎛ x& ⎞ ⎛ A B ⎞⎛ x ⎞
⎜⎜ ⎟⎟ = ⎜⎜ ⎟⎟⎜⎜ ⎟⎟ , (2.34)
⎝ y ⎠ ⎝ C D ⎠⎝ u ⎠
23
Chapter 2 Review of LMIs and EAs in Multiobjective Control
where x is the state, u is the input and y the output of the system, with x(0) = 0 . The
transfer function of this system is
T ( s ) = C ( Is − A) −1 B + D . (2.35)
The H∞ norm of the system (2.34) measures the system input-output gain for finite energy
or finite root mean square (RMS) inputs signals and when applied to a transfer function
(2.35) gives the largest gain across frequency in the singular value norm. This constraint
is useful to enforce robust stability in the face of norm-bounded uncertainties.
The system (2.34) is asymptotically stable and the maximum singular value of its transfer
function (2.35) is bounded by γ > 0 . Thus,
y
T ∞
= sup 2
< γ , ∀u ∈ L2 (2.36)
0< u 2 < ∞ u 2
The right hand side of the equality is the input-output relation and is equivalent to the
quadratic condition,
∞
y
2
2
2
(
− γ u 2 := ∫ y T y − γ 2 u T u dt < 0 .
2
) (2.37)
0
Rearranging,
T
∞
⎛ y⎞ ⎛ (γI ) −1 0 ⎞⎛ y ⎞
∫ ⎜⎜ u ⎟⎟ ⎜ ⎟⎜ ⎟dt ,
0⎝ ⎠
⎜ 0
⎝ − γI ⎟⎠⎜⎝ u ⎟⎠
T T
∞
⎛ x⎞ ⎛ C D ⎞ ⎛ (γI ) −1 0 ⎞⎛ C D ⎞⎛ x ⎞
∫ ⎜⎜ u ⎟⎟ ⎜⎜ ⎟⎟ ⎜⎜ ⎟⎜ ⎟⎜ ⎟dt , (2.38)
0⎝ ⎠ ⎝1044
I⎠ ⎝ 0 − γI ⎟⎠⎜⎝ 0 I ⎟⎠⎜⎝ u ⎟⎠
444 42444444 3
M
suppose that (A, B) is controllable and making use of the YKP lemma with the symmetric
matrix M indicated in (2.38),
⎛ AT P + PA PB ⎞ ⎛ γ −1C T C γ −1C T D ⎞
⎜ ⎟+⎜ ⎟
⎜ BT P 0 ⎟ ⎜ γ −1 D T C − γ + γ −1 D T D ⎟ < 0 ,
⎝ ⎠ ⎝ ⎠
24
Chapter 2 Review of LMIs and EAs in Multiobjective Control
The next step is to make use of Schur complements. In summary, the system (2.34) is
asymptotically stable and T ∞
< γ , if and only if, there exists a solution P = P T to the
system of LMIs:
⎛ AT P + PA PB CT ⎞
⎜ ⎟
⎜ BT P − γI DT ⎟ < 0 ,
⎜⎜ ⎟
⎝ C D − γI ⎟⎠
P >0, (2.39)
and the FDI for all ω ∈ ℜ with det(iωI − A) ≠ 0 is
[C (iωI − A) −1
][
∗
]
B + D C (iωI − A) −1 B + D − γ 2 I < 0 ,
which is equivalent to the left hand side of the equality (2.36), i.e.
C ( sI − A) −1 B + D <γ.
∞
hk ( X ) = 0 , k=1, 2, …, K. (2.42)
where X is the vector of n decision variables: X=(x1, x2, …, xn)T and the decision space.
Each decision variable is bound as follows:
25
Chapter 2 Review of LMIs and EAs in Multiobjective Control
f i ( xv ) ≤ f i ( xu ) ∀ i=1, 2, …, n. (2.43)
The Pareto optimal or Pareto front set is also known as the non-inferior or non-dominated
solutions. The set of non-dominated solutions is identified as the boundary of the
feasibility region, illustrated in Figure 2.6.
Feasible region
Objective function 2
Pareto front
Objective function 1
26
Chapter 2 Review of LMIs and EAs in Multiobjective Control
time, overshoot and output error have been observed (Boyd and Barratt 1991; Astrom
2000).
min
K
( f1 ( K ) f 2 (K ) L f n ( K )) ,
T
(2.44)
subject to K ∈ Ω ,
where Ω is the set of all controllers that internally stabilize the plant P.
The multiobjective optimisation problem has been tackled with single objective
optimisation techniques, such as the weighted sum approach, goal programming, goal
attainment, the ε-constraint method, etc. However, these approaches combine objectives
into a single function (Dorato 1991; Coello Coello 1998; Liu et al. 2003). Consequently,
the multiobjective control problem has adopted the same approach, although the
engineering application demands several control specifications e.g. rise time, overshoot,
27
Chapter 2 Review of LMIs and EAs in Multiobjective Control
gain and phase margins (Fleming and Pashkevich 1986). Another approach is the method
of inequalities (MoI) (Zakain and Al-Naib 1973; Maciejowski 1989; Whidborne and Liu
1993). In the method of inequalities, the problem is expressed as a set of algebraic
inequalities which need to be satisfied for a successful design. Based on simulated
annealing, two algorithms to solve a set of inequalities were proposed by Whidborne et
al. (1996). This method and other multiobjective approaches are well explained in the
book “Multiobjective Optimisation and Control” by Liu et al. (2003).
In robust control, the problem of multiobjective control has been treated with the use of
performance indices, e.g. linear quadratic index, H∞ norm and H2 norm, etc, and then the
approach of reducing the multiobjective control problem into a single objective. These
techniques are known as mixed techniques, e.g. mixed sensitivity H∞ approach
(Jonckheere and Verma 1986; Baeyens and Khargonekar 1994), H∞ control and pole-
placement constraints (Yedavalli and Liu 1995), multiple objective optimal control of
linear systems: the quadratic case (Khargonekar and Rotea 1991b).
The mixed H2/H∞ control problem, which has been intensely studied (Bernstein and
Haddad 1989; Khargonekar and Rotea 1991a; Kaminer et al. 1993; Doyle et al. 1994;
Sznaier 1994; Zhou et al. 1994; Scherer 1995a), guarantees robust stability while
minimizing a nominal H2 norm performance measure. Since the solvability of Riccati
equations have been shown to be fairly numerically reliable for the H∞ control theory
(Glover and Doyle 1988; Doyle et al. 1989), the solution to the mixed H2/H∞ control
problem has been tackled by coupling Riccati equations from H2 theory and H∞ theory
(Bernstein and Haddad 1989; Doyle et al. 1989; Doyle et al. 1994). However, solving the
coupled Riccati equations is a major problem and the controllers resulting from this
method have a sub-optimal performance, along with many assumptions on the plant.
During the early 1990’s, Khargonekar and Rotea (1991c) showed that the mixed H2/H∞
control problem might be cast into a format of a constrained convex optimisation
problems over a bounded set of quadratic matrix inequalities, expecting that well
established techniques from convex programming literature will solve the computational
problem. Notice, from Section 2.4.1, that at this time, it was relatively early for interior-
point methods to become recognized as efficient algorithms for solving convex problems
28
Chapter 2 Review of LMIs and EAs in Multiobjective Control
over LMIs. Other algorithms such as the method of centers by Boyd and El Ghaoui
(1993) were still used to solve the mixed H2/H∞ control (Baeyens and Khargonekar
1994).
Then, after the consolidation of the interior-point method to solve convex problems by
Nesterov and Nemirovski (1994), research in robust control started formulating the mixed
H2/H∞ control problem as a convex problem over LMIs (Scherer 1995a). Due to the
“flexibility” of LMI techniques to include multiple constraints on the control design
(Boyd et al. 1994b), the LMI control formulation was combined with constraints, such as
root-clustering, the state-feedback case was investigated by Chilali and Gahinet (1996).
Further investigations showed that a general quadratic synthesis procedure could lead to
formulation of multiobjective output-feedback control via LMI optimisation. This
synthesis method gives a framework that can include several constraints, such as H2
performance, H∞ performance, peak to peak gain and regional pole constraints (Scherer et
al. 1997). In this approach, all LMI characterizations have the following common origin:
Let A and xcl denote the closed-loop state matrix and state vector, respectively. Since the
controller has to be internally stabilizing, the closed-loop system must admit a quadratic
Lyapunov function
V ( x cl ) = x clT Pxcl , P > 0,
This LMI characterization relies on the assumption that it can be solved using the same
Lyapunov function, at the expenses of conservatism. Otherwise this approach becomes a
multiobjective controller design problem written as a system of bilinear matrix
inequalities (BMIs). Bilinear matrix inequalities are affine in two different sets of
variables but not in all together. Unfortunately, no reliable numerical algorithms are
available to solve BMIs (Arzelier and Peaucelle 2002; Dettori and Scherer 2002).
In Chapter 4 of this thesis, the mixed H2/H∞ control problem is studied using the general
quadratic synthesis procedure developed by Scherer et al. (1997). The numerical results
of this LMI technique are compared with numerical results resulting from the MOGA-
base method (in Section 2.11).
29
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Another approach resulted from the fact that the multiobjective H2/H∞ control problem is
equivalent to the generally infinite-dimensional convex optimisation problem (Scherer
1995b; Elia and Dahleh 1997). The infinite-dimensional formulation can be shown by
using the Youla parametrization of all internally stabilizing controllers,
where Tj is obtained from an arbitrary initial stable controller and H ∞ℜqxp denotes the
algebra of real-rational proper and stable transfer matrices of dimension q x p. Thus, the
multiobjective H2/H∞ control problem can be stated as follows:
This problem is then translated into a convex optimisation problem in the parameter Q,
which varies in the infinite-dimensional space of all transfer functions which are in the
space H ∞ℜqxp . An approximation sketched by Scherer and Weiland (1999) showed that a
Ritz-Galerkin approximation scheme leads to finite-dimensional problem by using the
following sequence of finite-dimensional subspaces, and a fixed real a > 0 ,
⎧ s−a ( s − a) 2 (s − a) υ qxp ⎫
ξ υ = ⎨Q0 + Q1 + Q2 + L + Qυ | Qυ , L , Qυ ∈ ℜ ⎬. (2.46)
⎩ s+a (s + a) 2 ( s + a) υ ⎭
For a more comprehensive discussion, the reader is referred to Scherer (1995b). The
advantage over the single Lyapunov approach method by Scherer et al. (1997) is that the
Youla parametrization gives a better approximation to the optimal result at the expense of
increasing the order of the controller (Scherer 1995a; Hindi et al. 1998; Scherer 2000).
Another approximation to the mixed H2/H∞ problem was proposed by Arzelier and
Peaucelle (2002). They proposed an iterative method which relies on the “smart” choice
of an initial seed.
30
Chapter 2 Review of LMIs and EAs in Multiobjective Control
The remaining sections of this chapter are devoted to introducing a novel approach to the
multiobjective control problem; this is the multiobjective genetic algorithm. This
approach is fundamentally different to traditional optimisation techniques because it is
not deterministic and it can be applied to a wide range of applications in engineering,
medicine, social science, etc. First, the basis of this approach will be presented with the
standard genetic algorithm.
GAs are part of a broad class of search techniques within the field of evolutionary
algorithms (EAs) or evolutionary computing (EC) (Fonseca 1994; Fleming and
Purshouse 2002). The genetic algorithm concept was developed by Holland (1975) and
the computational implementation and mathematical foundations were developed by
Goldberg, who first investigated the application of GAs to the control of gas pipeline
transmission for his PhD dissertation (Investigation co-chaired by J. Holland). In his
book, “Genetic Algorithms in Search, Optimisation, and Machine Learning” (Goldberg
1989), the computational code of the so-called standard genetic algorithm (SGA) is
introduced, and it constitutes of three genetic operators: reproduction, crossover, and
mutation. These operators perform the evolution using probabilistic rules, which
manipulate the “genetic code”. Goldberg identifies fundamental differences with
traditional search methods (deterministic):
• GAs work from a population; many other methods work from a single point.
• GAs only require an evaluation function or objective function to assess the fitness
of the individuals, this objective function is the only information from the
31
Chapter 2 Review of LMIs and EAs in Multiobjective Control
32
Chapter 2 Review of LMIs and EAs in Multiobjective Control
For the simple optimisation problem, the roulette is constructed with slots sized
according to their fitness. With the percentage of fitness taken from Table 2.1, the
roulettes, in Figure 2.7, are divided in four slots each of one is proportional to their
fitness in percentage. If the roulette in Figure 2.7a spins, individual two (Table 2.1) has a
probability 0.279 of been selected by the pointer. In this way, individuals with height
fitness have more probability of being selected for the succeeding generation, by making
an exact replica of the selected individual. This new individual is then entered into a
mating pool for further manipulations of its code. This kind of selection is known as
roulette wheel selection (RWS). Unfortunately, the RWS approach allows large selection
errors to occur (Fonseca 1994), and introduces a large variance in its realizations (Deb
2001). An alternative selection approach is the stochastic universal sampling (SUS),
which (Fonseca 1994) visualizes this selection as follows: SUS has the same principia of
RWS but with multiple, equally spaced pointers.
5.3% 5.3%
8.8% 8.8%
57.8% 57.8%
27.9% 27.9%
a) (RWS) b) (SUS)
Figure 2.7 Reproduction using a roulette wheel with slots sized according to fitness.
33
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Notice that in the roulette Figure 2.7b, individual one will have four copies (see Table
2.1), the individual two will have two copies and the remainder only one copy each.
These proportionate selection methods have a scaling problem (Deb 2001), for example
in the simple problem, Table 2.1, individual one has a probability tending to 1. This
might dominate the mating pool with its copies, and it might not necessarily be the
optimal solution. The scaling problem can also be avoided by using a ranking selection
method. A third selection operation known as the tournament selection works by
comparing a pair of individuals, where the better of those two is declared winner of the
tournament.
34
Chapter 2 Review of LMIs and EAs in Multiobjective Control
0 1 0 0 1 0 1 0 1 1 0 1 0 1 1
0 0 1 1 1 0 0 1 0 1 0 0 1 0 1
Crossover point
Each string is split in two by the crossover point, which is chosen at random. The second
part of each string is in a crossover position with the second. As a result, two new
offspring are created, see Figure 2.8. Other typical recombination operators are:
2.10.5 Mutation
This genetic operation causes individuals’ genotypes to be changed according to some
probabilistic rule. In other words it is a random alteration of some bits in individuals. In
artificial genetic algorithms, the mutation operator protects against an irrecoverable loss
35
Chapter 2 Review of LMIs and EAs in Multiobjective Control
of some potentially useful genetic material (Goldberg 1989). The probability is calculated
in different ways, the probability pm of mutation according to Fonseca (1994) is
p m = 1 − σ −1 / l (2.47)
where l = length of the chromosome, σ is the selective pressure, recommended value 1.8.
Initialize population
gen = 0
False
stop gen < maxgen
True
Evaluation
Assign Fitness
gen =gen +1
Replication
Crossover
Mutation
36
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Some applications have found difficulties with the binary representation of the
population. Research in GAs has brought continuous search space representation. This
has advantages, such as achieving any arbitrary precision in the optimal solution. The
replication operator remains the same, but the crossover and mutation operator are
applied in a different way; the approaches used in this work are, simulated binary
crossover (SBX) (Deb and Agrawal 1995) and polynomial mutation (Deb and Goyal
1996) respectively.
The SGA consists of replication, crossover and mutation, and it works as follows (see
Figure 2.9): it starts with a random population, while a condition is true that the following
will be executed; each individual will be evaluated and the assignment of fitness will be
performed, and then replication crossover and mutation is executed as explained above.
Traditionally by practitioners of GAs, a simulation of a SGA is conditioned to be
executed for a finite number of generations. This number changes according to the
application; in this work, the problems treated had a range of generations between 100
and 250.
Recall that the aim of this section and the next one is to introduce an alternative approach
to the multiobjective control problem. Before completing this section, two observations
by researchers in the field of multiobjective control and in the field of genetic algorithms
will be presented.
The field of multiobjective control has been developed though techniques with the aims
of finding a unified framework to include a variety of control engineering specifications
and overcome conservatism, high order controllers, nonlinearities, etc. Moreover, solving
a truly multiobjective control problem is desired instead of mixed or aggregating
approaches. The benefits will be to get a better approximation to the optimal Pareto-front,
and consequently a deeper understanding of the trade-offs that a particular problem might
have. The reason for applying GAs to the multiobjective control problem is linked with
the following two quotes:
37
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Quote one: first, consider the population, strings and fitness, generated for the simple
optimisation problem. In the book by Goldberg (1989, page 18), a similar population was
presented with the following explanation of ideas:
String Fitness
10111 529
10000 256
01001 81
00111 49
Quote two: Let us think that the strings are controllers and the fitness is their
performance in controlling a plant. The following quote was written in a paper by
Khargonekar and Rotea (1991c), where they explain the challenges of the multiobjective
control problem.
Although research on multiobjective control has been carried out up to today, this
challenge of the multiobjective control problem has not been entirely solved. The next
38
Chapter 2 Review of LMIs and EAs in Multiobjective Control
There exist numerous MOEAs available in literature, the interested reader is referred to
(Coello Coello 1998; Veldhuizen and Lamont 2000; Deb 2001) for a survey of the most
popular approaches. In this work the one used is the Multiobjective Genetic Algorithm
(MOGA) by Fonseca and Fleming (1993). MOGA can be applied to multiobjective
control problems by encoding the control problem, defining the vector spaces and the
objectives. If a successful design is easily located, further improvements such as reduced
controller order can be incorporated. Alternatively, if acceptable solutions are not readily
obtained, the designer may choose to relax the design goals.
MOGA has been successfully applied to control systems problems, e.g. the MIMO
controller structure and parameter optimisation for a gas turbine engine (Chipperfield
and Fleming 1996), the H∞ design of an electromagnetic suspension (EMS) control
system for maglev vehicle (Dakev et al. 1997), the optimisation of the low-pressure spool
speed governor of a Rolls-Royce Pegasus gas turbine engine (Fonseca and Fleming
1998b). MOGA was also used in conjunction with an H∞ loop-shaping design procedure
(LSDP) for the ALSTOM benchmark problem (Griffin et al. 2000). MOGA has been
applied in combination with other approaches, such as the Fuzzy Logic approach to the
scheduling controller design of a gas turbine engine (Chipperfield et al. 2002), and a
multiobjective optimisation approach with DK-Iteration (Griffin and Fleming 2003).
39
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Initialize population
gen = 0
False
stop gen < maxgen
True
Multiobjective Evaluation
Fitness
Assignment Kernel density Estimation
Fitness Sharing
gen = gen+1
σ share = σ mate Replication
Mating Restriction
Crossover
Mutation
40
Chapter 2 Review of LMIs and EAs in Multiobjective Control
Within MOGA, the initial population is randomly generated within a defined range and
then decoded (in case of a non-real chromosome) to produce the corresponding vectors of
decision variables. A set of objective function values is then evaluated for each individual
within the population (see Figure 2.11a). The sequence of genetic operators is then
applied, resulting in the subsequent generation of further potential solutions. Note that
fitness assignment is a more elaborate process (see shaded region in Figure 2.10).
Dominated solutions
0 5
2
2
0 1
F2 F2
0
nondominated
0 solutions
attainment
surface
F1 F1
a) b)
approximation
to optimality
F2 F2
diversity
density estimation
F1 F1
c) d)
41
Chapter 2 Review of LMIs and EAs in Multiobjective Control
In the Pareto set, interpolating between the solutions to obtain a smooth representation of
the Pareto set is not generally correct. Solutions corresponding to intermediate non-
dominated individuals, even if they exist, are unknown. A more convenient
representation will be to draw a boundary separating those points in objective space or
the attainment surface, see Figure 2.11b. This concept was introduced by Fonseca and
Fleming (1993).
42
Chapter 2 Review of LMIs and EAs in Multiobjective Control
The genetic operators presented in this section might be implemented with the SGA as in
Figure 2.10. The MOGA has been recognized as a good approach, efficient and easy to
implement (Coello Coello 1998). The approximation to the Pareto optimal is particularly
difficult because the Pareto optimal is unknown. The traditional way is to experiment and
decide either by using prior knowledge or by comparing with the results of other methods
if the current Pareto-front is improving towards the Pareto optimal. In general, caution
must be taken when using evolutionary algorithms. Even if a good choice of parameters
is found for a particular application; this set will be sub-optimal for many other problems.
Making the wrong choice of parameters can produce exceedingly poor results (Fleming
and Purshouse 2002).
43
Chapter 2 Review of LMIs and EAs in Multiobjective Control
2.12 Conclusions
LMI formulations can mix various specifications and objectives, which is an
improvement over classical LQG or H∞ techniques. Semidefinite program (SDP) has an
efficient numerical solution using recent interior-point methods for convex optimisation.
This brings a numerical solution to problems when no analytical solution is known. Some
problems in control systems will reduce to convex problems, in general, due to the
existence of convex constraints.
According to researches Ghaoui and Silviu-lulian (1999), in the future, LMIs will play
the same central role in post-modern theory as the Lyapunov function and Riccati
equations played in modern, theory and in turn, the role various graphical techniques
such as Bode, Nyquist and Nichols plots played in the classical.
There is not a general method that converts a control specification into LMI constraints
and even if this is achieved it might be non-convex as in the case of the reduced order
controller. Alternatively, MOGA can deal with non-convex constraints and provide truly
multiobjective optimisation treatment to multiobjective control problems.
The following chapters will present multiobjective control problems and their solution
with both LMI techniques and the MOGA-based method.
44
Chapter 3
Multiobjective Optimisation of Controller
Structure via YK parametrization using GAs and
LMIs
3.1 Introduction
This chapter studies the open problem of reduced- and fixed-order synthesis, which is
known to be a rank minimization problem subject to a linear matrix inequality (LMI)
constraint in H∞ state-space techniques. The optimisation involved is difficult to solve
due to the non-convexity of the objectives. In order to address this problem, the
application of an evolutionary approach is considered. A controller of optimal order is
selected from a set of feasible controller structures. Two controller design problems are
presented. A comparison of time and frequency domain responses shows that low order
controllers resulting from the evolutionary optimisation perform as well as those low
order controllers designed using the latest achievements in LMI techniques.
Within the field of embedded control systems, it is desirable to implement a low order
controller. When a state-space H∞ technique is used, the order of the controller must be at
least the same as the order of the plant. There are three alternatives to reduce the order of
such controllers:
45
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
1. A reduced order approximation of the plant can be generated before designing the
controller.
This work examines the third alternative. Using LMIs, the intention is to include a
constraint dim (Ac) ≤ k for some k that is less than or equal to the dimension of A where A
and Ac are the state matrices of the plant and controller respectively. The corresponding
LMIs can be derived.
⎧⎪⎡ X I ⎤ ⎫⎪
rank ⎨⎢ ⎬ ≤ n + nk
⎪⎩⎣ I Y ⎥⎦ ⎪⎭
where n is the order of matrix Ac and nk is the order of the matrix Ac, X>0 and Y>0 are n x
n matrices. However, the rank constraints are non-convex and difficult, if not impossible,
to treat by current optimisation techniques (Scherer et al. 1997). However, recent
research has developed some solutions to this problem; e.g. an algorithm in state space
realization (Iwasaki and Skelton 1994b; Apkarian et al. 2003; Xin 2004) and a sufficient
LMI condition in polynomial systems (Henrion 2003a) have been proposed to overcome
this non-convexity and can be used to fix the order of the controller.
Alternatively, the multiobjective genetic algorithm (MOGA) has been shown capable of
optimizing controller structure and controller parameters (Chipperfield and Fleming
1996; Schroder 1998). A controller is designed by optimising the controller structure
(order of the controller) and multiple design objectives in both the time domain
(overshoot, undershoot and settling time) and the frequency domain. In Section 3.4, two
examples of flexible structures have been solved numerically to illustrate the proposed
procedure. The results and the procedure are compared with those obtained by LMI
techniques.
46
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
v
u +
r + K P z
- +
with coprime polynomials an and bn, and where K is the controller of order m with
coprime polynomials yn and xn,
yn
K=
xn , (3.2)
r is the reference signal and v is the disturbance and z is the control system output. Given
a plant, P, the design of the feedback control system consists of designing the controller,
K, such that the resulting feedback system exhibits the desired performance.
and all the controllers that stabilize the given plant are generated by all pairs of x̂ , ŷ that
solve the Bézout equation (3.3).
47
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
This particular Diophantine equation (Bézout equation) has been useful in the area of
control systems because it provides the parametrization of a family of all stabilizing
controllers for P and it was proposed by Kučera (1975); Youla et al. (1976a); Youla et al.
(1976b) (see equation 3.4):
yˆ − aq
K= (3.4)
xˆ + bq
where q is an arbitrary proper stable rational parameter.
an xˆn + bn yˆ n = ad xd (3.5)
where the degree of the polynomials ad and xd are n and m=n-1 respectively and x̂ n and
ŷ n define the initial controller. Defining
an bn xˆ n yˆ n
a= , b= , xˆ = , yˆ = , (3.6)
ad ad xd xd
equation (3.5) can be transformed into the Bézout equation (3.3). Defining the Youla-
ad qn
Kučera polynomial q = , and substituting (3.6) in (3.4), the parametrization can be
xd qd
⎡ x n ⎤ ⎡ bn x d a d xˆ n ⎤ ⎡ a d q n ⎤
(a d x d )2 ⎢ ⎥=⎢ ⎥ , (3.7)
⎣ y n ⎦ ⎣− a n x d a d yˆ n ⎦ ⎢⎣ x d q d ⎥⎦
48
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
where xdqd is an arbitrary stable polynomial and adqn is an arbitrary polynomial of equal
or lower degree. Equation (3.7) can be simplified in equation (3.8) indicating that the
A,
⎡ xn ⎤
⎢
⎡a d x d 0 − bn − xˆ n ⎤ ⎢ y n ⎥⎥
⎢ ⎥ =0 (3.8)
⎣14 0 ad xd a n − yˆ n ⎦ ⎢ q n ⎥
4444244444 3⎢ ⎥
A ⎣q d ⎦
Assuming that q n , q d are polynomials and noting also that polynomial matrix A has full
row rank, the dimension of the null-space will be equal to two. Let
⎡ x1n x 2n ⎤
⎢y y 2 n ⎥⎥
N = ⎢ 1n (3.9)
⎢ q1n q 2n ⎥
⎢ ⎥
⎣q1d q 2d ⎦
be a minimal basis for the null-space of matrix A, i.e. such that AN = 0 and the column
degree of N are minimal among all possible null-space bases. By extracting a minimal
polynomial basis N for the null-space of the polynomial matrix A defined in (3.8), all the
⎡ xn ⎤ ⎡ x1n x2 n ⎤ ⎡ λ1 ⎤
⎢y ⎥ = ⎢y y 2 n ⎥⎦ ⎢⎣ λ2 ⎥⎦
, (3.10)
⎣ n ⎦ ⎣ 1n
where λ1 , λ2 are polynomials such that the YK denominator polynomial, qd, is stable and
the corresponding YK numerator polynomial is given by
⎡q d ⎤ ⎡q1d q 2 d ⎤ ⎡ λ1 ⎤
⎢q ⎥ = ⎢q q 2 n ⎥⎦ ⎢⎣ λ2 ⎥⎦
. (3.11)
⎣ n ⎦ ⎣ 1n
Then a controller, K, of fixed order m can be found if there exist polynomials λ1 and λ2 of
order m. The remainder of the controller design procedure using LMI techniques is fully
49
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
explained in Henrion et al. (2004) and used later in this chapter to find the LMI
controllers.
Given a set of polynomials ni (s) , d i (s) for i = 1,2,... Ni as well as the positive real
∑ ni ( s ) x i ( s )
i
< γ, (3.12)
∑ d i (s) xi ( s)
i ∞
denotes the peak value of the magnitude of rational transfer function S evaluated along
the imaginary axis. If c(s) is defined as a central polynomial, the degree of this
polynomial is equal to the highest degree of polynomials ni (s) , d i (s) plus the degree of
the desired polynomial xi ( s ) , which is the desired controller. Similarly to standard H∞
techniques, this LMI design technique is iterative, and a trial-and-error approach cannot
be avoided in choosing appropriate the central polynomial. A general rule-of-thumb is
that open-loop stable poles must be mirrored in the central polynomial, completed by
sufficiently fast additional dynamics.
⎛ ⎞
n ( s ) = γ⎜⎜ ∑ d i ( s ) xi ( s ) ⎟⎟ + ∑ ni ( s ) xi ( s ) ,
⎝ i ⎠ i
⎛ ⎞
d ( s ) = γ⎜⎜ ∑ d i ( s ) xi ( s ) ⎟⎟ − ∑ ni ( s ) x i ( s ) .
⎝ i ⎠ i
50
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
Let us define the 2 by 2 polynomial matrices
⎡ n( s ) 0 ⎤ ⎡d ( s ) 0 ⎤ ⎡ c( s ) c( s ) ⎤
N (s) = ⎢ , D( s ) = ⎢ , C ( s) = ⎢ ⎥,
⎣ 0 n( s )⎥⎦ ⎣ 0 d ( s )⎥⎦ ⎣ − c( s ) c( s ) ⎦
N = N0[ N1 ... ]
Nδ , [
D = D0 D1 ... Dδ , ] C = [C 0 C1 ... Cδ ] .
the highest degree arising in polynomials ni (s) , d i (s) and c(s). With these notations,
results on positive polynomial matrices are invoked in Henrion (2003a) to derive the
following solution.
Theorem 3.1 (Henrion 2003a) Given polynomials ni (s) , d i (s) , positive scalars γ for and
central polynomial c( s ) , there exist polynomials xi ( s ) solving H ∞ specifications (3.12)
if matrix inequalities
( N ) T C + C T N − H ( Pn ) > 0 , (3.13)
( D) T C + C T D − H ( Pd ) > 0 , (3.14)
H ( P) = Π T (H ⊗ P )Π
where,
⎡I 2 0⎤
⎢
⎢ O M ⎥⎥
⎢ I2 0⎥
Π=⎢ ⎥
⎢0 I2 ⎥
⎢M O ⎥
⎢ ⎥
⎢⎣ 0 I 2 ⎥⎦
51
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
is a matrix of size {4δ x 2(δ+1)},
⎡0 1⎤ ⎡− 1 0⎤
P=⎢ , P=⎢ ⎥.
⎣1 0⎥⎦ ⎣ 0 1⎦
for the left half-plane and the unit disk respectively.
1 an xn a n xn
S= = , then the H∞ norm is: S ∞
= (3.15)
b y a n x x + bn y n a n x n + bn y n
1+ n n ∞
an xn
bn y n bn y n
T =1− S = , then the H∞ norm is: T ∞
= (3.16)
a n x n + bn y n a n x n + bn y n ∞
If the controller design requires limiting the actuator effort, the transfer function KS
should be included as another objective. In this work this objective was not required and
hence not studied.
52
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
Because of the nature of the problems in this Chapter (flexible structures), a third
objective called Total Variation (TV) (Owens and Chotai 1981) was included. TV
provides a measure of the oscillations in the time response to a step change in reference
signal. The TV of a function f, is defined by
N
TV ( f ) = max ∑
0≤ti ≤t N i =1
f (t i +1 ) − f (t i ) (3.17)
The function f used in this chapter is the step response of the complementary transfer
function T. Another way to compute the total variation is by using the truncated integral
of absolute area (L1 norm) of the derivative of f(t) Boyd and Barratt (1991, page 98).
tN
TV = ∫ | f& (t ) |dt.
0
Another suitable function would be the integral of the step response of the sensitivity
function S (Whidborne and Liu 1993, Chapter 8), as this would give a bound on the
maximum tracking error.
An additional objective was included that measured the order of the controller. The
chromosome structure shown in Figure 3.2 was implemented along the lines of Schroder
(1998), where he used for the optimisation of the weighting function structure.
C α0 α1 β0 α2 α3 α4 β1
β2 α5 α6 α7 α8 β3 β4 β5
53
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
The variable C shown in Figure 3.2 selects the controller structure, and then the controller
K(λ1,λ2) is decoded according to equations (3.18). Finally, the controller K=yn/xn can be
obtained using equation (3.10).
⎧ λ1 = α 0 + α1 s
⎪ 0 ≤ C < 1,
⎪ λ2 = β 0 + s
⎪⎪ λ1 = α 2 + α 3 s + α 4 s 2
K ( λ1 , λ2 ) = ⎨1 ≤ C < 2,
⎪ λ2 = β1 + β 2 s + s 2 (3.18)
⎪ λ1 = α 5 + α 6 s + α 7 s 2 + α8 s 3
⎪ 2 ≤ C ≤ 3,
⎪⎩ λ2 = β 3 + β 4 s + β 5 s 2 + s 3
bn 1
P= = 2 (3.19)
an s ( s + s + 10)
The control problem is to design a controller, K, so that the resulting feedback system
exhibits a step response with no undershoot and the minimal possible overshoot and
settling time as well as a high damping signal response. The first step is to obtain an
initial controller, for example by solving the Diophantine equation (3.5) with arbitrary
placement of the poles at s = -1, giving
a d x d = ( s + 1) 5 .
yˆ n 1 + 45s − 26s 2
= .
xˆ n − 4 + 4s + s 2
54
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
Step Response
10
6
Amplitude
-2
0 5 10 15 20 25 30
Time (sec)
By solving equations (3.10) and (3.11), all the stabilizing controllers can be computed
from polynomials λ1 and λ 2 in equation. (3.20) and it may be possible to find
polynomials x n and y n (controllers K) that have low degree with a stable YK polynomial
qd (3.11),
⎧ x n = λ2
K =⎨ (3.20)
⎩ y n = − λ1 + 26λ2
55
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
LMI Approach
The H∞ design procedure was used to find polynomials x n and y n (controllers K). The
formulation to solve the control problem is as follows: Given the polynomials an and bn
from plant (14) and the bound γ s , the sensitivity function with YK controller (3.4) is:
1
S= = axˆ + abq
b ⎛ yˆ − aq ⎞
1+ n ⎜⎜ ⎟⎟
an ⎝ xˆ + bq ⎠
ad qn
This transfer function can be simplified by substituting q = , equation (3.8), and
xd qd
a n xn a n xn
S ∞
= = < γs (3.21)
a n xn + bn y n ∞
qd ∞
a d x d x n = bn q n + xˆ n q d (3.22)
This frequency domain specification (3.21) can be expressed as an LMI constraint, using
specification 3.12 and theorem 3.1, and then solved simultaneously with the linear
constraint. The H ∞ design algorithm was set up to solve the optimisation control
problem, and then the polynomials xn, qd and qn of a given degree were sought such that
(3.21) and (3.22) are satisfied.
In each single run, a controller was found by using the concept of the central polynomial,
c(s), which is the key design step. The H∞ design procedure consists of iteratively
adjusting roots of c(s), while lowering the upper bound γ s , see Table 3.1. After a series
of trials a set of controllers of orders m=1, 2 and 3 were obtained and the closed-loop step
responses of the best controllers are shown in Figure 3.3.
56
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
Table 3.1 Family of controllers resulting from the H∞ design procedure: LMI Approach.
Step Response
1.2
K (2nd)
1
0.8
Amplitude
0.6 K (1st)
0.4 K (3rd)
0.2
0
0 5 10 15
Time (sec)
Figure 3.4 Step response of: first-order controllers (‘dotted line‘), second-order
controllers (‘dashed line’) and third order controllers (‘solid line’): LMI approach.
MOGA-based method
The controller design problem was then tackled using the MOGA-based method. The
decision variables used were the controller parameters defined in (3.18) in the
chromosome structure from Figure 3.2. An evaluation function was encoded with seven
objectives defined in Section 3.3.1. The MOGA parameters used were:
57
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
• The probability of applying mutation is 7e-4.
• A population of 100 individuals.
The algorithm was iterated for 100 generations. The results are illustrated in Figures 3.5
and 3.6. It can be seen that one single solution does not exist, rather a family of solutions.
The trade-off of the objectives is shown in Figure 3.5, which is a graphical format known
as the method of parallel coordinates. Each line represents a Pareto optimal solution. The
x-axis represents the objective and the y-axis shows the normalized performance in the
interval [0, 1]. Table 3.2 shows objective functions values (the second and the third row)
of the cost for each objective in Figure 3.5 as well as the goal vector (the fourth row).
MOGA Trade-off
Cost
1 2 3 4 5 6 7
Objective no.
1 2 3 4 5 6 7
Objective ||T||∞ ||S||∞ TV Overshoot Settling time K Undershoot
(amplitude) (sec) nth-order (amplitude)
Normalized 1.1 2 1.8 1.04 22 4 0.1
Cost [1]
Normalized 0.9 1 0.9 0.94 0 0 -0.5
Cost [0]
Goal vector 2 2 4 1.001 25 4 0
Table 3.2 Normalization of the design objectives in Figure 3.5 and goal vector.
For comparison, the LMI controller, shown in equation (3.24) using trial number 4 as
shown in Table 3.1, was included with bold lines and marked (o) in Figure 3.5. The
objective 2 H∞ norm of the sensitivity function, the objective 3 TV, the objective 4
overshoot and objective 5 settling time appear to compete quite heavily, since the
improvement in one objective is the detriment in the other. Here is where the control
58
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
engineer can select candidate solutions depending on the specifications of the
applications. Figure 3.6 illustrates the most representative step responses. Most of the
step responses that have large settling times are first-order controllers. The second-order
and third order controllers show better settling time, and this is confirmed by looking at
the Trade-off graph in Figure 3.5.
Step Response
1.2
0.8
Amplitude
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20
Time (Sec)
By changing the values of the goals, the search is forced to examine other areas of the
trade-off surface. Note for example, different values of TV produce a range of responses
from smooth to oscillatory. Taking advantage of the initial information, some of the goals
(||T||∞, ||S||∞, TV and settling time) were tightened. The new goal vector is shown in Table
3.3. The new results are displayed in Figure 3.7, which shows the performance of the
controllers that meet the new design requirements. The order of the objectives was
rearranged in order to facilitate the visualisation of the trade-offs among the objectives,
such as TV, ||S||∞ and settling time. Figure 3.8 illustrates the “best” closed-loop step
responses obtained by the MOGA.
The MOGA controller (3.23) was chosen and compared with the LMI controller (3.24).
The first comparison is in Figure 3.7, where the MOGA controller is marked (◊) and the
LMI controller is marked (o); both solutions are in bold lines. The selected MOGA
59
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
controller offers the best compromise among the Pareto front. For example, it offers best
compromise between settling time and ||S||∞. The step response produced by the selected
MOGA controller is compared with the “best” step response produced by the LMI
controller (Henrion et al. 2004). The objective measures of those controllers are shown in
Table 3.4.
MOGA Trade-off
Cost
1 2 3 4 5 6 7
Objective no.
1 2 3 4 5 6 7
Objective ||T||∞ Overshoot TV ||S||∞ Settling time K Undershoot
(amplitude) (sec) nth-order (amplitude)
Cost [1] 2 0.008 1.2 1.4 4.3 4 0.1
Cost [0] 0 0 0.99 1.1 2 0 -0.99
Goal 1.5 0.001 1.2 1.5 4 4 0
vector
Table 3.3 Normalization of the design objectives in Figure 3.7 and goal vector.
60
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
Step Response
1
0.9
0.8
0.7 KLMI
KMOGA
0.6
Amplitude
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8
Time (sec)
Figure 3.8. Flexible mode closed-loop step responses of a MOGA controller (‘solid line’)
and the best LMI controller (‘dashed line’).
The closed-loop bandwidth of the MOGA controller is ωb1 ≈0.7079 rad /s (||S||∞ crosses -
3db from below, dashed grey line) whereas the bandwidth of the LMI controller is
ωb 2 ≈0.52481 rad/s. Since ωb1 > ωb 2 , the MOGA controller will have effective control for
61
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
a larger frequency range, and after the frequency ωb1 , the tracking performance (see ||T||∞
) will degrade and reach higher magnitude of resonance than the LMI controller.
0 0
-5 -5
-10 -10
|T(jw)| (dB)
|S(jw)| (dB)
-15 -15
-20 -20
-25 -25
-30 -30
-35 -35
-40 -40
-2 0 2 -2 0 2
10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
Some experiments showed that MOGA with YK parametrization enhances the population
of closed-loop stable controllers providing a reduction in the computational burden more
so than the same formulation without YK parametrization. This experiment was
performed with a fixed-order controller (3.25)
α 0 + α1 s + α 2 s 2
K= (3.25)
β0 + β1 s + s 2
Using the settings of Example 1 and the controller (3.25), MOGA was iterated for 100
generations. Figure 3.10 illustrates a family of preferable closed-loop step responses
obtained by the MOGA-base method and LMI techniques. This figure shows that for the
same number of generations in Example 1, the MOGA results are still inferior to the LMI
results or the MOGA with YK parametrization results. The YK parametrization changes
62
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
the shape of the decision space; the benefit is that it might change the shape of narrow
decision space.
Step Response
1.2
LMI:Approach
1
0.8
Amplitude
0.4
0.2
0 2 4 6 8 10 12 14 16 18 20
Time (Sec)
Figure 3.10 Step responses of controllers resulting from MOGA-base method without YK
parametrization.
The control problem is to design a controller, K, so that the resulting feedback system
exhibits a step response with overshoot no greater than 10% and settling time
approximately 8 sec. The procedure is the same as in Example 1. A new initial controller
was calculated by solving the Diophantine equation (3.5) with arbitrary placement of the
poles at s=-1, then a d x d = ( s + 1) 7 . The controller design problem was then tackled using
the MOGA-based method. All the settings of the MOGA were the same as for Example
1. The goal vector was set up to search for controllers of first, second and third order and
Table 3.5 shows the objective function values (the second and the third row) of the cost
for each objective in Figure 3.11 as well as the initial goal vector (the fourth row). In this
way a Pareto optimal front was found. This information allows the control engineer to
visualise which specification can be achieved and if a reduction in the order of the
63
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
controller will achieve good performance. (see Figure 3.11). For comparison, the LMI
controller, shown in equation (3.28) was included with bold line and marked (o) in Figure
3.11.
Cost MOGA Trade-off
1 2 3 4 5 6 7
Objective no.
1 2 3 4 5 6 7
Objective ||S||∞ ||T||∞ TV Overshoot Undershoot Settling K
(amplitude) (amplitude) time (sec) nth-order
Cost [1] 1.9 1.5 2 0.3 0 20 4
Cost [0] 1.2 0.9 1 0 -0.025 0 1
Goal 2 2 2 0.3 0 20 4
vector
Table 3.5 Normalization of the design objectives and goal vector.
The MOGA controller 3.27 was chosen and compared with the LMI controller 3.27. The
first comparison is in Figure 3.12, where the MOGA controller is marked (◊) and the
LMI controller is marked (o); both solutions are in bold lines and cannot be distinguish
because they overlap.
64
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
MOGA Trade-off
Cost
1 2 3 4 5 6 7
Objective no.
Figure 3.12. New trade-off graph for the controllers KMOGA for example 2.
1 2 3 4 5 6 7
Objective ||S||∞ ||T||∞ TV Overshoot Undershoot Settling K
(amplitude) (amplitude) time (sec) nth-order
Cost [1] 1.5 1.04 1.17 0.065 -0.004 7.5 3
Cost [0] 1.25 1 1 0 -0.011 3.5 1
Goal 1.5 1.5 1.5 0.1 0 8 2
vector
Table 3.6 Normalization of the design objectives and goal vector.
The selected MOGA controller offers the best compromise among all the objectives, for
example, it offers best compromise among overshoot, undershoot and settling time.
Figure 3.13 illustrates the “best” closed-loop step responses obtained by the MOGA and
LMI techniques (Henrion 2003a). Both controllers produced very similar step responses.
Step Response
1.2
0.8
KLMI
Amplitude
0.6
0.4
0.2
KMOGA
0 2 4 6 8 10 12 14
Time (sec)
Figure 3.13. Flexible Beam closed-loop step responses of MOGA controller (‘solid line’)
and the best LMI controller (‘dashed line’).
65
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
The selected MOGA controller and the LMI controller are:
0.56229⋅ 10−2 + 1.3052s − 1.3691⋅10−2 s 2
K MOGA = (3.27)
3.851+ 3.73s + s 2
Although is impossible to have zero undershoot since the plant is non-minimum phase,
and thus the plant has a right-half-plane (rhp) zero that will appear in T(s). In MOGA, the
goal was set up to be zero in order to enforce minimization of the undershoot. The final
MOGA controller is non-minimum phase. However, the rhp zero is far in the rhp, and
could be ignored. Thus a final controller of
0.56229⋅10−2 + 1.3052s
K MOGA =
3.851+ 3.73s + s 2
would remove the additional rhp zero, and the performance remain almost unaffected
In Figure 3.13 the closed-loop frequency domain performance measures of the MOGA
controller in equation. (3.25) are ||T||∞ = 0 dB and ||S||∞ ≈2.07 dB and those the LMI
controller are ||T||∞ = 0 dB and ||S||∞ ≈1.5 dB. It is important to mention that similar
specifications were achieved with H∞ state space techniques. However, the controller was
of order eight (Doyle et al. 1992). The closed-loop bandwidth of the MOGA controller
and LMI controller are ωb =0.3 rad/s. The LMI controller has slightly better bandwidth.
However, the tracking performance (see ||T||∞) will reach slightly higher magnitude of
resonance than the MOGA controller.
66
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
0 0
-3 -3
-10 -10
-20 -20
-30 -30
|T(iw)| (dB)
|S(iw)| (dB)
-40 -40
-50 -50
-60 -60
-70 -70
-80 -80
-90 -90
-5 0 -2 0 2
10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
The algorithms were run on a standard PC with Pentium IV, 1.7GHz processor. The
computational cost of solving the examples with MOGA was CPU time 15min to 25min.
A single run of LMI algorithms was CPU time 15 sec. The MOGA-based method can
deal with MIMO systems, but because of the controller structure, equation (3.18) relies
on the preliminary results of YK paramatrization by Henrion et al. (2004), and only
applies to the SISO case.
67
Chapter 3 Multiobjective Optimisation of Controller Structure via YK parametrization
using GAs and LMIs
3.5. Conclusions
The LMI optimisation for fixed-order H∞ controller design can be computed quickly.
However, the whole procedure for designing a controller requires several attempts.
Multiple trials changing the central polynomial, bound on the frequency domain
specification and the order of the controller have to be carried out until the specification
seems to be satisfied or no progress is noticed. Despite the computational cost of the
MOGA, the Pareto surface between controller order and closed-loop performance can be
investigated in a single run. Thus, the control engineer has a choice from among the
family of Pareto optimal solutions and design time is saved.
68
Chapter 4
Mixed H2/H∞ problem with LMIs Approach and
MOGA-based method.
4.1 Introduction
This chapter deals with two of the most popular design specifications, H2 (LQG case)
performance and H∞ performance, which have been shown to have competing objectives
(Bernstein and Haddad 1989). This chapter highlights the relative flexibility offered by a
multiobjective genetic algorithm (MOGA) and linear matrix inequalities (LMIs) in
addressing complex problems.
Research into LMIs reveals an extensive list of possible design specifications (Boyd et al.
1994b; Ghaoui and Silviu-lulian 1999), which can be expressed in an LMI formulation.
However, the application usually requires transformation of the controller parameters,
more about this later in section 4.3. This chapter illustrates briefly how the H2 norm case
can be cast as an LMI optimisation problem. Alternatively, MOGA can be applied to
multiobjective control problems by encoding the control problem, defining the vector
spaces and the objectives. If a successful design is easily located, further improvements
such as reduced controller order can be incorporated. Alternatively, if acceptable
solutions are not readily obtained, the designer may choose to relax the design goals.
69
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
These techniques are applied to two control problems. Firstly, the LQG design problem,
which has been used to find the trade-off between process and sensor noise for a simple
but realistic controller design problem (see Section 4.8.1 and 4.8.2). Secondly, these
techniques have also been applied to the mixed H2/H∞ problem (see Section 4.8.3 and
4.8.4).
The LQG problem is an example from Barratt and Boyd (1989), they computed an exact
trade-off involving noise sensitivity and robustness measurements. Such trade-off can
serve as the limit of performance bounds against which controllers of any complexity,
obtained by any design method, can be computed. This chapter shows how this problem
can be formulated as a convex program (LMI formulation) using the equivalence between
LQG control problem and the H2 optimal control theory. This formulation is presented in
continuous-time. However, because the Barratt and Boyd (1989) example is given in
discrete-time, the LMI constraints are derived in discrete-time. The LMI technique and
the MOGA-based method are used to find the same trade-off as Barratt and Boyd (1989)
and the advantages and disadvantages of using each method are addressed.
The mixed H2/H∞ problem has been extensively studied and it remains an open problem
(Bernstein and Haddad 1989; Khargonekar and Rotea 1991a; Kaminer et al. 1993; Doyle
et al. 1994; Zhou et al. 1994). The LMI techniques for multiobjective control design have
been proposed to handle the mixed H2/H∞ control problem (Boyd et al. 1994b). The
approaches can be divided in two. The first approach (Scherer 1995b; Scherer 2000)
expresses the mixed H2/H∞ control problem as LMI conditions by using the Youla
parametrization of the controller. This approach truly solves the multi-objective problem
but at the expense of an infinite dimensional problem. In order to alleviate this problem,
the Youla parameter is approximated (see Section 2.9). The accuracy of the
approximation compromises the order of the controller. The second approach (Scherer
1995a; Scherer et al. 1997) leads to finite dimensional controllers but at the expense of
introducing dependence (see Section 4.4) between objectives. It also makes the problem a
single-objective problem. In this chapter, when using the LMIs, the second approach is
studied.
70
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎛ x& ⎞ ⎛ A B1 B2 B ⎞⎛ x ⎞
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ z1 ⎟ ⎜ C1 D11 D12 E1 ⎟⎜ w1 ⎟
⎜ z ⎟ = ⎜C D21 D22 E2 ⎟⎜ w2 ⎟
(4.1)
⎜ 2⎟ ⎜ 2 ⎟⎜ ⎟
⎜ y⎟ ⎜C
⎝ ⎠ ⎝ F1 F2 0 ⎟⎠⎜⎝ u ⎟⎠
where, x is the state vector, u is the control input, y is the measured output available for
the controller, and the channels from wi (disturbance inputs) to zi (controlled outputs)
serve to specify performance objectives. A controller, K, is a LTI system described as:
⎛ x& c ⎞ ⎡ Ac Bc ⎤⎛ xc ⎞
⎜⎜ ⎟⎟ = ⎢ ⎜ ⎟. (4.2)
⎝ u ⎠ ⎣Cc Dc ⎥⎦⎜⎝ y ⎟⎠
w1 z1
w2 z2
P
u y
K
Figure 4.1 Generalized control structure for a two objective control design problem.
The controller is parameterized by the matrices (Ac, Bc, Cc, Dc). The closed loop system
in figure 4.1 is denoted in equation (4.3),
⎡ Acl Bcl , j ⎤
z j = Tj wj = ⎢ w ,
Dcl , j ⎥⎦ j
(4.3)
⎣C cl , j
where the index, j, refers to a given channel. Then Tzw = Dcl + C cl ( sI − Acl ) −1 Bcl with the
notation:
71
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎛ A + BDC C BC C B j + BDC F j ⎞
⎛ Acl Bcl , j ⎞ ⎜ ⎟
⎜ ⎟=
⎜ Ccl , j Dcl , j ⎟ ⎜ BC C AC BC F j ⎟. (4.4)
⎝ ⎠ ⎜
⎝ C j + E j DC C E j CC D j + E j DC F j ⎟⎠
This chapter deals with the computation of a dynamic output-feedback control law
u = Ky .
Consider the design of a controller that achieves stability and quadratic performance in
the channel wi to zi and suppose have the following performance index:
⎛Q j Sj ⎞
P j = ⎜⎜ T ⎟, Rj ≥ 0 .
⎝S j R j ⎟⎠
A characterization of linear dissipative systems have revealed (Scherer et al. 1997) that
the controller (4.2) renders (4.3) internally stable and leads to
∞ T ∞
⎛ w j (t ) ⎞ ⎛ w (t ) ⎞
∫⎜
⎜ z j (t ) ⎟
0⎝ ⎠ ⎝ ⎠ 0
∫
⎟ P j ⎜ j ⎟dt ≤ −ε w j (t ) T w j (t )dt
⎜ z j (t ) ⎟
72
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
T
⎛ I ⎞ ⎛ I ⎞
⎜ ⎟ Pj ⎜ ⎟
⎜ T j (i ω ) ⎟ ⎜ T j (iω) ⎟ < 0 for all ω ∈ ∞
⎝ ⎠ ⎝ ⎠
With these results, it can be stated that the closed loop system (4.3) is asymptotically
stable and has quadratic performance with index Pj if and only if there exists a matrix x
such that,
x>0, (4.5)
⎜ x xA xB
⎛ AclT + cl cl
⎞ ⎛ 0
⎟+⎜
I ⎞
T
⎟
⎛Qj
⎜ T
S j ⎞⎛ 0
⎟⎜
I ⎞
⎟ < 0. (4.6)
⎜
⎝ x
BclT 0 ⎟ ⎜C
⎠ ⎝ cl Dcl ⎟⎠ ⎜Sj
⎝ R j ⎟⎠⎜⎝ Ccl Dcl ⎟⎠
Since Acl , Bcl , Ccl , Dcl are in terms of the state variables Ac , Bc , Cc , Dc of the controller
and the inequality (4.6) turns into nonlinear after substitution, a transformation is
necessary. A partition of x is proposed:
⎛Y V⎞
x = ⎛⎜⎜UX U⎞
⎟ x −1
= ⎜⎜ T ⎟
* ⎟⎠ * ⎟⎠
T
⎝ ⎝V
⎛Y I⎞ ⎛I 0⎞
y = ⎜⎜V T ⎟ ,
0 ⎟⎠
z = ⎜⎜ X ⎟ to get
U ⎟⎠
yT x = z
⎝ ⎝
Using the matrix y the following two transformations can be performed
⎛Y I⎞
y T
xy = ⎜⎜ I ⎟ = X (v) ,
X ⎟⎠ (4.7)
⎝
where ν denotes the new Lyapunov matrices (X,Y), and the second transformation is
T
⎛ y 0 ⎞ ⎛ x 0 ⎞⎛ Acl Bcl ⎞⎛ y 0 ⎞ ⎛ y T xAcl y y T xBcl ⎞⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟ =⎜
⎜ 0 I ⎟ ⎜ 0 I ⎟⎜⎝ C cl Dcl ⎟⎠⎜⎝ 0 I ⎟⎠ ⎜⎝ C cl y Dcl ⎟, (4.8)
⎝ ⎠ ⎝ ⎠ ⎠
substituting (4.4) into transformation (4.8),
73
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎛ AY 0 Bj ⎞
⎜ ⎟
⎛ y T xAcl y y xBcl ⎞⎟ ⎛⎜ zAcl y
T zBcl ⎞⎟
⎜ = =⎜ 0 XA XB j ⎟ +
⎜ C y Dcl ⎟⎠ ⎜⎝ C cl y Dcl ⎟⎠ ⎜⎜ ⎟
⎝ cl
⎝C jY C j D j ⎟⎠
⎛0 B ⎞
⎜ ⎟ ⎡⎛U XB ⎞⎛ AC BC ⎞⎛V T 0 ⎞⎟ ⎛ XAY 0 ⎞⎤⎛ I 0 0⎞
⎜ I 0 ⎟ ⎢⎜⎜ ⎟⎜ ⎟⎜ +⎜ ⎟⎥⎜ ⎟
+
⎜⎜ ⎟⎟ ⎣⎢⎝ 0 I ⎟⎠⎜⎝ C C DC ⎟⎠⎜⎝ CY I ⎟⎠ ⎜⎝ 0 0 ⎟⎠⎦⎥⎜⎝ 0 C F j ⎟⎠ . (4.9)
⎝0 E j ⎠
The term inside the square brackets contains all the non-linear products. By introducing
⎛K L⎞
the parameters ⎜⎜ ⎟ , the expression (4.9) becomes affine in the new set of variables
⎝M N ⎟⎠
if:
⎛K L ⎞ ⎛U XB ⎞⎛ AC BC ⎞⎛ V T 0 ⎞⎟ ⎛ XAY 0⎞
⎜⎜ ⎟=⎜ ⎟⎜ ⎟⎜ +⎜ ⎟, (4.10)
⎝M N ⎟⎠ ⎜⎝ 0 I ⎟⎠⎜⎝ CC DC ⎟⎠⎜⎝ CY I ⎟⎠ ⎜⎝ 0 0 ⎟⎠
⎛ ⎛ AC BC ⎞ ⎞ ⎛ ⎛K L ⎞⎞
⎜ ,⎜ x ⎟⎟ → v = ⎜⎜ X , Y , ⎜⎜ ⎟⎟ .
⎜ ⎜C
⎝ ⎝ C DC ⎟⎠ ⎟⎠
⎝ ⎝M N ⎟⎠ ⎟⎠
where ν now denotes the new Lyapunov matrices (X,Y) and the new controller variables
(K, L, M, N). This nonlinear transformation v relies on a simple mapping of all LMIs,
(e.g. H2 performance) into a set of affine constraints on the new controller variables,
⎛Y I⎞
X (v) = ⎜⎜ ⎟
⎝I X ⎟⎠ ,
⎛ B j + BNF j ⎞
⎛ y T xAcl y y T xBcl ⎞ ⎛⎜ A(v) B j (v) ⎞ ⎜ AY + BM A + BNC
⎟
⎜ ⎟= ⎟=⎜ K XA + LC XB j + LF j ⎟ . (4.11)
⎜ C y Dcl ⎟⎠ ⎜⎝ C j (v) D j (v) ⎟⎠ ⎜
⎝ cl
⎝ C jY + E j M C j + E j NC D j + E j NF j ⎟⎠
as a result the LMIs (4.5) and (4.6) are transformed into a set of affine variables v .
X (v ) > 0 , (4.12)
74
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
T
⎛ A(v) T + A(v) B j (v) ⎞ ⎛ 0 I ⎞ ⎛Q j S j ⎞⎛ 0 I ⎞
⎜ ⎟+⎜ ⎟ ⎜ T ⎟⎜ ⎟ < 0. (4.13)
⎜ B T
( v ) 0 ⎟ ⎜ C j (v ) D j (v ) ⎟ ⎜S j R j ⎟⎠⎜⎝ C j (v) D j (v) ⎟⎠
⎝ j ⎠ ⎝ ⎠ ⎝
After having solved the synthesis inequalities for v , one factorizes I − XY into square
and non-singular blocks U and V T using SVD and the controller by the following
equations:
−1 −1
⎛ AC BC ⎞ ⎛U XB ⎞ ⎛ K − XAY L ⎞ ⎛V T 0⎞
⎜⎜ ⎟=⎜ ⎟ ⎜⎜ ⎟⎜ ⎟ , (4.14)
⎝ CC DC ⎟⎠ ⎜⎝ 0 I ⎟⎠ ⎝ M N ⎟⎠ ⎜⎝ CY I ⎟⎠
Proof. The necessity part has been proven in the construction of the theorem. The
sufficiency part of the proof leads to an Ac that has the same size as A.
4.3.2 H2 Performance
This section shows the formulation of H2 performance specifications as an LMI
constraint on the design variables. The Frobenius norm is used to obtain the H2 norm of
G(s) and is defined by integrating over frequency.
1 ∞
G( s) 2
= ∫
2π −∞
trace(G(iω) H G(iω))dω (4.16)
75
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
∫ trace(g )
T
G ( s ) 2 = g (t ) 2 = ( τ ) g ( τ ) dτ ,
0
⎧ ∞ T ⎫
= trace ⎨ B T ∫ e A t C T Ce At dt B ⎬ = trace ( B T QB ) ,
2
G (s) 2
⎩ 0 ⎭
⎧ ∞ ⎫
G ( s) 2 = trace⎨C ∫ e At BBT e A t dtC T ⎬ = trace(CPC T ) ,
2 T
⎩ 0 ⎭
where Q and P are the observability and controllability Gramians respectively. The
system (A, B) is state controllable if and only if the controllable Gramian P is positive
defined (P>0) and thus has full rank n. P may also be obtained as the solution to the
Lyapunov equation:
AP2 + P2 AT = − BB T , (4.17)
where P2 denotes the controllability Gramian for the H2 norm. The controllability
Gramian is the unique positive solution to the Lyapunov equation. This is equivalent to
2
A is stable and G ( s ) 2
<γ, (4.18)
This work requires the discrete-time Lyapunov equation. The Lyapunov equations for the
76
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
These Lyapunov equations (4.20) and (4.21) are affine in the variable P2 . To convert
(4.20) and (4.21) into LMI constraints, a Schur complement method is applied,
⎛ P2 −1 C clT ⎞
⎜ ⎟>0, trace( z ) < γ . (4.23)
⎜C T ⎟
Z − Dcl Dcl ⎠
⎝ cl
Observe that the matrix inequalities in equations (4.22) and (4.23), are non-linear, as they
involve multiplication of variables, and the variables ( Acl , Bcl , C cl , Dcl ) are expressed in
multiplying dig ( P2−1 , I ) on the left and right hand side, the change of variable x=P 2
−1
must be applied as well as the Schur complement for the terms x − xAx−1A x T
and
Z − Dcl DclT . The results are LMI (4.24) and LMI (4.25).
⎛ x xA xB ⎞
⎜ T cl
⎟
cl
The second more sophisticated transformation (see Section 4.3.1) has been proposed by
(Scherer et al. 1997). The idea is to rewrite the H2 norm inequalities involving y , which
leads to inequalities in the blocks y T xy , y T xAcl y , y T xBcl , C cl y , Dcl . Multiplying
77
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
the LMI (4.24) by diag ( y, y, I ) T on its left hand side and diag ( y , y , I ) on its right hand
side gives the LMI (4.26). Multiplying the LMI (4.25) by diag ( y , I , I ) T on its left hand
side and with diag ( y , I , I ) on the right hand side gives the LMI (4.27).
⎛ X (v ) A(v) B j (v) ⎞
⎜ T ⎟
⎜ A (v ) X (v ) 0 ⎟ >0, (4.26)
⎜ B T (v ) 0 I ⎟⎠
⎝ j
⎛ X (v ) 0 C Tj (v) ⎞
⎜ ⎟
⎜ 0 I D Tj (v) ⎟ > 0 . (4.27)
⎜C v D v Z ⎟⎠
⎝ j( ) j( )
where X(v), A(v), Bj(v), Cj(v), Dj(v) are linear and affine on the controller variables K, L,
M and N. Using equation (4.11) the LMIs (4.26 and 4.27) can be transformed into LMIs
(4.28 and 4.29). Then the LMI optimisation is tractable numerically and the controller
can be recovered by reversing the transformation using Theorem 4.1.
⎛ Y I AY + BM A + BNC B j + BNF j ⎞
⎜ ⎟
⎜ I X K XA + LC XB j + LF j ⎟
⎜ ( AY + BM ) T KT Y I 0 ⎟ > 0,
⎜ ⎟
⎜ ( A + BNC ) T ( XA + LC ) T
I X 0 ⎟
⎜ B + BNF T ⎟
⎝( j j) ( XB j + LF j )T 0 0 I ⎠
(4.28)
⎛ Y I 0 (C j Y + E j M )T ⎞
⎜ ⎟
⎜ I X 0 (C j + E j NC )T ⎟
⎜ >0. (4.29)
⎜ 0 0 I ( D j + E j NF j )T ⎟⎟
⎜C Y + E M C j + E j NC D j + E j NF j Z ⎟
⎝ j j ⎠
4.3.3 H∞ Performance
The H ∞ performance has been formulated as an LMI constraint (Scherer and Weiland
1999). Using the Kalman-Yacubovich-Popov lemma together with a specific quadratic
supply function reduces frequency domain inequalities to equivalent systems of LMIs
78
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
and this is known as the Bounded Real Lemma (Scherer 1990). Using this lemma, let A
be asymptotically stable and T j < γ ∞ if and only if there exists a solution P∞=(P∞)T
∞
where P∞ denotes the positive definite matrix for the H∞ norm. Using the LMI approach
general procedure from section 4.3.1, the LMI in equation (4.30) can be turned into
equation (4.31), which is affine in v , (see equation (4.11)).
The smallest possible upper bound of the L2-induced gain of the system can be computed
by minimizing γ ∞ > 0 over all variables in v satisfying equation (4.31).
The multiobjective H2/H∞ control problem is to keep bounds on the H2 norm of, say, the
first channel w1 → z1 and on the H∞ norm of, say, the second channel w2 → z 2 of the
controlled system:
Despite the recent advances in robust control theory, the robust performance problem
formulated in the multiobjective H2/H∞ framework largely remains an open problem.
First of all, this problem is an infinite-dimensional optimisation problem. Approximations
to the finite dimensional problem using Youla parameter Q have been developed (Scherer
1995b; Hindi et al. 1998). System impulse responses were truncated to a finite horizon
but LMIs have shown that when a state-space description is available, then many of the
79
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
infinite horizon cost and constraints can be represented as LMIs and minimized
efficiently as semidefinite programs (SDPs). In this way, errors by truncations are
eliminated. Another alternative is to introduce dependence among the different
objectives, which leads to finite dimensional controllers of the same order of the
generalized plant (Scherer 1995a; Scherer et al. 1997). This last approach is known as the
mixed H2/H∞ problem and it is used in this work.
Mixed H2/H∞ problems can be motivated in different ways. As a matter of fact, there are
many different mixed H2/H∞ problems. The mixed H2/H∞ problem studied in this work is
to find an internal stabilizing controller K(s) which:
A controller that solves this mixed H2/H∞ problem will ensure that the closed-loop system
is robustly stable to all finite-gain stable perturbations Δ, and at the same time TZ1W
2
represents the steady-state variance of the output z1 when w2 = 0 and w1 is white noise
with unit intensity.
Investigating the trade-off between H∞ norm and the H2 norm constraints consist of plot a
[ ]
curve of optimal values, by varying γ ∞ in some interval γ ∞l ,..., γ ∞u , where the lower
bound γ ∞l could be taken close to the smallest achievable H∞ norm of transfer function
TZ 2W .
∞
In order to solve the problem (4.32), the mixed H2/H∞ problem is reformulated in terms
of the solvability of LMIs. In fact, Acl is stable and (4.32) holds if only if there exist
symmetric P1, P2, Z such that trace(Z) < γ2 and
⎛ AclT P∞ + P∞ Acl
⎜ P∞ Bcl ,1 C clT ,1 ⎞⎟
⎜ BclT ,1 P∞ − γ∞ I DclT ,1 ⎟ < 0 , P∞ > 0 , (4.33)
⎜ ⎟
⎜ C cl ,1 Dcl ,1 − γ∞ I ⎟
⎝ ⎠
80
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
This reformulation includes two different LMI constraints (4.33, 4.34), making the
problem non-convex (bilinear matrix inequality). In order to recover convexity, a
common Lyapunov function is used by equating the Lyapunov matrices P2 from H2
performance to the P∞ from H∞ performance:
P = P2 = P∞ . (4.35)
This leads to a conservative result, which will be illustrated in section 4.8.3. The
controller can be recovered by using the general procedure (see Section 4.3.1)
Typically the problems of H2 optimal control and H∞ optimal control have been studied
independently when using EAs. For example the H∞ design has been assisted with the
MOGA in order to satisfy a number of conflicting design criteria (Dakev et al. 1997;
Griffin et al. 2000; Schroder et al. 2001). Other researchers (Hunt 1992; Patton and Liu
1994; Chen and Cheng 1998) have been directly optimized over the mixed sensitivity
functions (S, T, KS). Their GA has been used to search for a controller which minimized
the following criteria:
⎧⎡ S ⎤ ⎫
⎪ ⎪
min sup σ ⎨⎢⎢ T ⎥⎥ ⎬ ,
K ω∈Ω ⎪⎢ KS ⎥ ⎪
⎩⎣ ⎦ ⎭
where Ω denotes the frequency range and σ denotes the maximum singular value. The
demand of controllers with multiple criteria like robustness and performance has
motivated the research into mixed H2/H∞ problems. This problem is originally a
multiobjective control problem that has been found difficult to solve because of the non-
convexity in the controller parameters. Recently, this problem has been tackled with GAs
(Chen et al. 1995; Herreros et al. 2002; Takahashi et al. 2004). The idea of optimising the
norm of transfer functions and the previous work of Fonseca and Fleming (1994) have
motivated, in this thesis, the study of the multiobjective H2/H∞ problem. Details are
shown in Section 4.8.4.
81
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
The LQG control refers to an optimal control problem where the plant model is linear, the
cost function is quadratic, and the test conditions consist of random initial condition w , a
white noise disturbance input wd , and a white measurement noise wn . The plant is
described by the following linear state equation:
x& = Ax + Bu + wd ,
y = Cx + wn .
The cost function
1 ⎡ T
∞
⎤
{
J ( x (t ), u (t )) = E ⎢ ∫ x (t )Qx (t ) + u T (t ) Ru (t ) dt ⎥ .
2 ⎣0
} (4.36)
⎦
This cost function is equal to the cost functions specified for the Stochastic LQR version.
A reasonable controller design for the LQG control problem can be obtained by using the
LQR feedback gain matrix K operating on the state estimate generated by the Kalman
filter u (t ) = − Kxˆ (t ) ,
1 ⎡ x (t ) ⎤ ⎤
]⎡⎢Q
∞
[
1/ 2
J= E ⎢ ∫ (Q1 / 2 x(t ))T ( R1 / 2 u (t ))T 1/ 2 ⎥dt ⎥
2 ⎢⎣ 0 ⎣ R u (t ) ⎦ ⎥⎦
82
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎡Q1 / 2 0 ⎤ ⎡ x⎤
Defining an error signal z as z=⎢ ⎥⎢ ⎥ ,
⎣ 0 R1 / 2 ⎦ ⎣u ⎦
1 ⎡ T
∞
⎤
J = E ⎢ ∫ z (t ) z (t )dt ⎥ ,
2 ⎣0 ⎦
∞
⎡ T
⎤
J = E ⎢lim
1
T →∞ 2T ∫
{ }
z T (t ) z(t )dt⎥ = trE z(t ) z T (t ) =
1
2π ∫ tr (G ( jω )
H
G ( jω ))dω . (4.37)
⎣ −T ⎦ −∞
2
In 4.37 Parseval’s theorem was applied and it is equal to G (s ) 2 , which is H2 norm. In
order to finish with the stochastic interpretation of H2 norm and LQG, the input w has to
be the white noise of the unit intensity. That is
{ }
E w(t ) w(τ ) T = S wδ (t − τ ) .
⎧ ⎡ w (t ) ⎤ ⎫ ⎡W
[
E ⎨⎢ d ⎥ wd (τ ) T ]
wn (τ ) T ⎬ = ⎢
0⎤
V ⎥⎦
δ (t − τ ) .
⎩⎣ wn (t ) ⎦ ⎭ ⎣0
In summary, by minimizing the H2 norm, the output power of the system, due to a unit
intensity white noise input, is minimized; we are minimizing the root-mean-square
(RMS) value of z. Thus, the LQG can be cast as an H2 optimisation in the following
general framework:
Stochastic inputs wd , wn as Output signal regulation
⎡ wd ⎤ ⎡W 1 / 2 0 ⎤ ⎡Q1 / 2 0 ⎤ ⎡ x⎤
⎢w ⎥ = ⎢ ⎥w , z=⎢ ⎥⎢ ⎥ .
⎣ n⎦ ⎣ 0 V 1/ 2 ⎦ ⎣ 0 R1 / 2 ⎦ ⎣u ⎦
Looking at these two stochastic conditions, the input and the output signal of the LQG
problem leads to a control configuration of the H2 optimal problem. See Figure 4.2 and
equation (4.38).
83
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎧ W1/2
process noise
R1/2 ⎫
w ⎨ ⎬ z
⎩ ⎭
+ x
B (sI-A)-1 Q1/2
+
+ y
u V 1/2
measurement noise
Figure 4.2 The LQG problem formulated in the H2 optimal control problem.
In Figure 4.2, w is a white noise process of unit intensity. Finally the state-space
description of the LQG problem is
x& = Ax + (W 1 / 2 0)w + Bu ,
⎛ Q1 / 2 ⎞ ⎛ 0 0 ⎞ ⎛ 0 ⎞
z = ⎜⎜ ⎟ x + ⎜⎜ ⎟⎟ w + ⎜⎜ 1 / 2 ⎟⎟u , (4.38)
⎟
⎝ 0 ⎠ ⎝0 0⎠ ⎝R ⎠
y = Cx + (0 V 1 / 2 )w + (0)u .
It was shown that LQG problem can be cast as an H2 optimal control problem. Thus,
under this framework, the LQG problem can be cast using linear matrix inequalities. This
is shown in Section 4.8.1.
84
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
[ ]
J = E y T (∞ )Qy (∞ ) + u T (∞ ) Ru (∞ ) , (4.39)
where the infinity symbol is used as an argument to indicate that the given quantity has
reached steady state, y is the output regulation, and u is the control effort. The polynomial
LQG problem (Hunt 1992) is another approach where the optimisation problem aims to
select the regulator in such a way that the cost function J is minimal. In the genetic
algorithm, J serves as the evaluation function in the search. Hunt used the cost function J
in the frequency domain and introduces another parameter to check stability.
Fonseca and Fleming (1994) tackled the LQG problem from a multiobjective point of
view. LQG controller design was used to find a trade-off between steady-state output and
actuator variances due to the presence of process and measurement noise. The aim was to
find the minimum of the linear combination of the two variances, in equation (4.39).
Fonseca and Fleming used the MOGA-based method to minimize the cost function J by
encoding the controller parameters (poles zeros and gain). The closed-loop system
evaluation of RMS(u) and RMS(y) served as multi-evaluation functions. The MOGA was
used to search for a Pareto optimal set or non-dominated solutions, these results are
recalculated in section 4.8.2 and compared with the Pareto optimal resulting from the
LMI approach.
85
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
(4 − s)
P(s) = (4.40)
s 2 (4 + s)
The objective is to investigate the trade-offs in the design of a regulator, K(s), for this
plant, where the closed loop system shown in figure 4.3 is disturbed by an input process
noise, wP. The only signal available to the controller is yp=y+vs , where vs is sensor noise
and y is the output,
u wP
+ +
r =0 K P y
- + +
vS
+
Figure 4.3 Closed-loop regulator system with noise injection wP and vs.
The process and sensor noise variances are zero mean, white noise, with root-mean-
square (RMS) values of wP = 10 for and vs = 1. The trade-off under study is between the
steady-state output variance lim k →∞ E y k2 { } and the steady state actuator
{ }
variance lim k →∞ E u k2 . This trade-off can be computed finding the LQG optimal
regulator for different penalty values on the control effort, ρ , which minimises (over all
stabilizing regulators) the linear combination, JLQG, of actuator and output variance.
{ }
J LQG = lim E y k2 + ρ u k2 = lim E x kT Qx k + u kT Ru
k→∞ k→∞ k
{ } (4.50)
Alternatively, this trade-off can be found using H2 performance measures by the fact that
LQG is a special case of H2 optimal control and it can be solved via LMI optimisation.
The smallest possible upper bound of the H2 norm of the transfer function can be
calculated by minimizing the criterion trace (z) over the variables X, Y, K, L, M and N
that satisfy LMIs (4.28 and 4.29).
The trade-off was computed as follows. The NMP plant (4.49) was discretized using a
zero-order hold at 10 Hz and its state space realization, equation (4.51), was used.
86
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
⎛ 0.6703 0 0 0.0824 ⎞
⎜ ⎟
⎛ A B ⎞ ⎜ 0.0824 1 0 0.0044 ⎟ . (4.51)
⎜⎜ ⎟⎟ =
⎝ C D ⎠ ⎜ 0.0044 0.1 1 0.0002 ⎟
⎜ ⎟
⎝ 0 −1 4 0 ⎠
The H 2 optimal problem under consideration can be stated using the state-space
description of the plant for the LQG problem (4.38). One could obtain the LTI plant
(4.52) with some H2 performance γ>0 to find an LTI control law u = K ( z ) y such that:
Twz < γ where Twz (z ) denotes the closed-loop transfer function from w to z. The LTI
controller K (z ) can be represented with the state space realization (4.2) and the
generalized state-space description of the plant in equation (4.51) for the LQG problem is
derived from equation (4.38) with Q = CT C , R = ρ and the stochastic inputs are wp=[W1/2
0]w and vs=[0 V1/2]w. The resulting realization is state-space realization (4.52).
⎛ 0.6703 0 0 ⎞ ⎛ 1 0 ⎞ ⎛ 0.0824 ⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
x& = ⎜ 0.0824 1 0 ⎟ x + ⎜ 0 0 ⎟ w + ⎜ 0.0044 ⎟u ,
⎜ 0.0044 0.1 1 ⎟ ⎜ 0 0 ⎟ ⎜ 0.0002 ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
⎛0 0 0 ⎞ ⎛0 0⎞ ⎛0⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ 0 0 .2425 0. 9791 ⎟ ⎜0 0⎟ ⎜0⎟ , (4.52)
z =⎜ x+ w + ⎜ ⎟u
0 0.9701 3.8806 ⎟ ⎜ 0 0 ⎟ 0
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜0 0 ⎟⎠ ⎜⎝ 0 ⎟ ⎜ ⎟
⎝ 0 0⎠ ⎝ρ⎠
The optimal LQG cost, JLQG, was calculated through the specification G (s) 2 ≤ γ , which
is achievable if and only if γ2≥ JLQG (Boyd and Barratt 1991). In order to cast the H2
optimal problem with LMI techniques, the H2 performance can be formulated as the
linear objective minimization problem.
min γ (4.53)
87
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
2
where P2 = P2T > 0 and also, G ( s ) 2
= trace(C T PC ) < γ 2 .
The optimisation problem was solved implementing the LMIs (4.28 and 4.29) in the
MATLAB® LMI Control Toolbox (Gahinet et al. 1995). Single runs changing ρ from
0.0001 to 10 (marked ♦) were computed to search for the Pareto optimal surface, which
represents the trade-off between noise sensitivities, as shown in figure 4.4. The result
obtained was the same as that in Barratt and Boyd (1989). Note that this LMI technique
found third order controllers. In general, the output-feedback technique applied will
always find a controller with the same order as the plant.
1
Trade-off between noise sensitivites
10
rho=10
rho=1
(RMS) value of y
rho=0.1
rho=0.01
rho=0.001
rho=0.0001
0
10
0 1 2
10 10 10
(RMS) value of u
Figure 4.4 Trade-off between noise sensitivities achieved with LMI techniques.
88
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
were used, the average, α , and the deviation, β , of each pair of roots (Kristinsson and
Dumont 1992).
K (z + a1 )( z + a 2 )
C= (4.54)
( z + a3 )( z + a 4 )
where a1, 2 = α1 ± β1 and a3, 4 = α 2 ± β 2 . A positive deviation indicates real roots and negative
deviation indicates complex conjugate roots. A pair of zeros and a pair of poles were set
up to search in the interval [-1, 1], and the remainder were set up to search the interval [0,
1]. The gain was set up to search the interval [0, 100].
The objectives under consideration were the RMS values of noise sensitivity for the
output y and actuator u. These variances RMS(y) and RMS(u) for a given discrete-time
controller plant were computed as two separated objectives. In order to avoid unstable
closed-loop systems, another objective was added. The maximum of the magnitudes of
the poles of the corresponding close loop systems were evaluated minimizing the degree
of instability of the closed-loop system. Since closed-loop stability is an absolute design
requirement, it was set as a high priority objective, or constraint, with an associated
priority level of 1. The region of the trade-off curve to be evolved was delimited with the
goal vector u rms = 100 and y rms = 10 . A population of 100 individuals were generated as real
parameters. The recombination operator was simulated binary crossover (SBX) for
continuous search space and the mutation operator was a polynomial mutation (Deb
2001).
In figure 4.5 the evolution of the individuals (family of controllers) is shown. Notice how
the solution evolves at generation 50, 100 and 250 and how they have found a
preliminary trade-off. The Pareto surface at generation 250 (marked o) is close to the
desired trade-off (solid line). Fonseca and Fleming’s (1994) results used 100 generations,
but in order to improve the diversity, SBX was used, which requires a population size
equal to 100 and 250 generations as a standard setting. Note from Table 4.1 that MOGA
was fixed to find second order controllers; see equation (4.54). In Table 4.1, six
controllers, resulting from the LMI optimisation, have been chosen. They are third order
controllers. For each of these controllers, a controller resulting from MOGA optimisation
89
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
has been chosen by finding the closest MOGA controller (marked ♦ in figure 3) to the
LMI controller in the Pareto front.
Generation 100
rho = 1 Generation 50
(RMS) value of y
rho = 0.01
rho = 0.001
rho = 0.0001
“LMI Approach”
0
10 0 1 2
10 10 10
(R M S) value of u
Figure 4.5 Trade-off between input and output RMS noise sensitivities achieved with
MOGA and LMIs.
LMI . MOGA .
ρ (rho) 3rd order controllers RMS (u) RMS (y) 2nd order controllers
10 0.85715( z + 0.004227 )( z − 0.6511)( z − 0.9715) 3.3125 8.4555 1.8398( z − 0.9547 )( z − 0.1961)
2
( z − 0.5464)( z − 1.57 z + 0.6831) ( z − 0.7181)( z − 0.4166 )
0.1 4.0011( z + 0.01609)( z − 0.6393)( z − 0.9449 ) 8.7479 3.6898 1.728( z − 0.9772)( z − 0.2097 )
2
( z − 0.415)( z − 1.502 z + 0.7137 ) ( z + 0.1479)( z − 1.204 )
0.01 9.2423( z − 0.9318)( z − 0.657 )( z + 0.001701) 16.027 2.8678 7.8333( z − 0.9371)( z − 0.0482)
2
( z − 0.3651)( z − 1.374 z + 1.7106) ( z + 2.462)( z − 3.69 )
0.001 21.3759( z − 0.9209)( z − 0.6617 )( z + 0.01372) 26.962 2.4089 1.728( z − 0.9773)( z − 0.2096)
2
( z − 0.2566)( z − 1.11z + 0.6779) ( z + 0.1349)( z − 1.195)
0.0001 45.8816( z − 0.914 )( z − 0.663)( z + 0.05016) 72.593 2.0798 11.6629( z − 0.8942)( z − 0.542)
2
( z − 0.1311)( z − 0.6331z + 1.6107 ) ( z + 2.442)( z − 3.844)
Table 4.1 LMI and MOGA Controllers from the Pareto front for the Trade-off between
input and output RMS noise sensitivities.
90
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
The second problem is the mixed H2/H∞ output-feedback problem and the plant
considered is the following LTI system, which was studied by Herreros et al. (2002),
⎛ − 21 − 120 − 100 1⎞
⎜ ⎟
⎛ A B⎞ ⎜ 1 0 0 0⎟ . (4.55)
⎜⎜ ⎟⎟ = ⎜
⎝ C D ⎠ ⎜ 0 1 0 0⎟
⎟
⎜ 0 0 150 0 ⎟⎠
⎝
The control system under consideration is the output-feedback problem for the
generalized plant in its state-space realization form shown in equation (4.56),
⎛ x& ⎞ ⎛ A I B⎞
⎜ ⎟ ⎜ ⎟⎛ x ⎞
⎜ z1 ⎟ ⎜ C 0 D ⎟⎜ ⎟ , (4.56)
⎜z ⎟ = ⎜A ⎜ w⎟
0 0 ⎟⎜ ⎟
⎜ 2⎟ ⎜ ⎟ u
⎜ y ⎟ ⎜C
⎝ ⎠ ⎝ 0 D ⎟⎠⎝ ⎠
where x is the system state vector, u is the control input vector and w is the exogenous
disturbance vector, z1 is the regulated output that satisfies H2 performance and z2 is the
regulated output that satisfies H∞ performance. An assumption is: w1 = w2 , which is a
necessary condition to obtain a convex approximation of the problem (Khargonekar and
Rotea 1991a).
Mixed H2/H∞ output-feedback has a trade-off that is interesting to investigate because the
H∞ norm and the H2 norm are objectives that mutually compete. In order to find such a
trade-off, the mixed H2/H∞ problem seeks an internal controller, K(s), as shown in
equation (4.57).
min TZ1W subject to TZ 2W < γ2 . (4.58)
2 ∞
The trade-off between the H∞ norm and the H2 norm constraints consists of plotting the
TZ1W obtained from equation (4.57). For this problem γ 2 = [10…600], where γ 2l was
2
chosen to be close to the smallest achievable TZ W . As was discussed in section 4.4, this
2 ∞
91
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
problem can be cast as LMI constraints. The LMI Algorithm was set up with the
generalized plant (4.56) and single optimisation for fifty values within the range of γ 2
was computed and the resulting trade-off is plotted in figure 4.6. The synthesis values
obtained by the LMI approach (marked ◊) are just an upper bound. The trade-off curve
looks non-convex. This might be a result of controller computation, the existence of the
controller depends on the non-singularity of the matrix I-XY as I-XY=UVT with non-
singular U and V. A simple test shows that the minimal eigenvalues of matrix XY of the
fifty LMI controllers are close to one, which means that the controllers are ill-
conditioned; this issue is studied in Chapter 5, Section 5.4.1.
500
300
"Synthesis values: LMI Approach"
200
100
0
20 30 40 50 60 70 80 90 100 110 120 130
||Tz1w (s) ||2
In order to illustrate the conservatism of this approach, a second curve is plotted and
labelled “Analysis values: LMI Approach” (marked ∆). For example, considering the
following problem, see equation (26), for γ 2 =600, and solving this problem with LMI
optimisation, gives 26.5376 as the best constrained H2 performance, which was achieved
with the following LMI controller:
92
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
The loop with the plant (see equation (4.56)) is then closed, and the H2 norm computed.
The result will be 22.2081 (16% lower than the LMI optimal value). When computing the
H∞ norm, the result will be 492.6874 (18% lower than the LMI constraint 600). In this
way, the second curve, labelled “Analysis values: LMI Approach”, was computed. This
conservatism of the results is due to the introduction of the common Lyapunov matrix
(see Section 4.4). It is clear that LMI optimal values are only an upper bound on the
closed-loop H2 performance.
The lower and upper boundaries on the decision variables were -5e+15 and -5e+4
respectively and the binary resolution was 32 bit Gray coding with exponential scaling in
each of the six decision variables space. Objectives were taken into account: The first
objective was Closed-Loop Stability. It was evaluated to minimise the degree of
instability of the closed-loop system and it was set up as a high priority objective, with an
associated goal of 0. The second objective was min ||Tz1w||2 norm with an associated goal
of 120.
∞
1
TZ1W = ∫ tr (T ( jω ) H TZ1W ( jω ))dω , (4.60)
2π
2 Z1W
−∞
⎡ A + BDC C BC C B j + BDC F j ⎤
⎢ ⎥
where TZ 1W = ⎢ BC C AC BC F j ⎥. (4.61)
⎢C1 + E1 DC C E1C C D1 + E1 DC F j ⎥⎦
⎣
93
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
The third objective was min ||Tz2w||∞ norm with an associated goal of 600, where
TZ 2W ( s) = max TZ 2W ( jω ) , (4.61)
∞ ω
⎡ A + BDC C BC C B j + BDC F j ⎤
⎢ ⎥
where TZ 2W = ⎢ BC C AC BC F j ⎥. (4.62)
⎢C 2 + E 2 DC C E 2 CC D2 + E 2 DC F j ⎥⎦
⎣
MOGA used mating restriction and crossover shuffle with reduced surrogate with
probability 0.7. A mutation rate of 0.09 was applied to all individuals after crossover. The
population size was 200 candidate solutions and the result after 250 generations is
presented in Figure 4.7 (marked o) to compare with the corresponding LMI results.
400
|| Tz2w ( s ) || i n f
200
100
“MOGA Approach”
0
10 20 40 60 80 100 120 130
|| Tz1w ( s ) || 2
Figure 4.7 Trade-off Mixed H2/H∞ achieved with MOGA and LMIs.
Results obtained from 5 independent runs of the algorithm provide a clear estimate of
how likely the various regions of the objective space were to be attained in a run of GA.
The Pareto fronts are presented in Figure 4.8. The Pareto front obtained using MOGA is
better than the one obtained with the LMI approach. Another achievement was that this
94
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
Pareto front is better than the one obtained by Herreros et al. (2002). Note the way in
which both problems were solved. When the LMI approach is used, the multi-objective
problem is reduced to a problem that optimises a single objective function. In the case of
the MOGA approach, there is no need for such a reduction and the original objectives are
optimised retaining the multiobjective formulation.
For the LMI approach, the difference between the synthesis values and analysis values
only shows the conservatism of the approach. A good estimation can be achieved only
through lower bound computation (Scherer 1999) and the use of Youla techniques to
approximate the optimal value (Scherer 2000).
500 500
400 400
||Tz2w (s) ||inf
||Tz2w (s) ||inf
300 300
200
200
100
100
0
0 20 40 60 80 100 120 140 0
0 20 40 60 80 100 120 140
||Tz1w (s) ||2
||Tz1w (s) ||2
500 500
400 400
||Tz2w (s) ||inf
300 300
200
200
100
100
0
0 20 40 60 80 100 120 140 0
0 20 40 60 80 100 120 140
||Tz1w (s) ||2
||Tz1w (s) ||2
Figure 4.8 Evolution of the Multiobjective H2/H∞ Trade-off achieved with MOGA (5
runs).
95
Chapter 4 Mixed H2/H∞ problem with LMIs Approach and MOGA-based method
4.9. Conclusions
In conclusion, solving control design problems with LMI techniques lacks full flexibility.
There is not a general method that converts a control specification into LMI constraints
and even if this is achieved it might be non-convex as in the case of the reduced order
controller.
A significant issue with using MOGA is that the search space is not always known. For
complex problems where there is no indication where the solutions are, MOGA will need
more prior knowledge until a set of Pareto optimal solutions can be found. However,
once the results are found the design requirements can be refined, by introducing more
constraints into the design, such as minimising the order of the controller.
The order of the controllers is of great relevance in control design. The flexibility of
MOGA permits the structure of the controller to be changed and a reduction in the order
of the controller to be found. Using the LMI framework, the problem of controller order
reduction is still a hard constraint on the optimisation and is often not tractable
numerically. This comparison was illustrated by the problem trade-off of noise
sensitivities.
96
Chapter 5
Multiobjective Control Design for the Turbofan
Engine
5.1 Introduction
Both the multiobjective H2/H∞ control design: evolutionary approach and mixed H2/H∞
control design: LMI approaches are applied to the optimisation of the Turbofan engine
model MIMO controller. In this chapter, a numerical analysis on the Pareto optimal result
is carried out. Numerical issues are assessed and a solution based on the evolutionary
approach is proposed. The gain scheduling controller design is considered and practical
issues are addressed. A multiobjective H2/H∞ gain scheduling controller design is
proposed, for two operating points.
Three basic types of jet engines are in current use. The turbojet is the best for high
subsonic and supersonic flight speeds. The turbofan is chosen for subsonic commercial
97
Chapter 5 Multiobjective Control Design for the Turbofan Engine
airplanes. The turboprop or turboshaft is the best for power plants, helicopters and for
small lower speed aircraft.
This chapter is concerned with the turbofan engine whose core consists of a compressor,
combustor and turbine which drive the compressor. It has a fan in front of the core
compressor and a second power turbine behind the core turbine to drive the fan as shown
in Figure 5.1a.
B Combustor C Exhaust
E
Fuel flow
Inlet, A
(a)
Combustion
(heat energy added)
B
C
Expansion
(output work to
run compressor)
Compression
Pressure
(pressure energy
added)
D
Output work for
Pressure is increased Thrust
through compressor
as volume is reduced
Inlet Ambient Air
Exhaust Heat
A E
Volume
(b)
Figure 5.1 (a) Gas engine schematic (b) Brayton cycle pressure-volume diagram.
98
Chapter 5 Multiobjective Control Design for the Turbofan Engine
The turbofan operates on the working cycle as depicted in Figure 5.1b, which is the
thermodynamic Brayton cycle. This cycle is based on the relation between pressure and
volume. The cycle can be described as a sequence of five stages:
In the compression system, one of the most critical mechanical problems occurs. Two
modes of mechanical instability can occur: “surge” which is a longitudinal flow
oscillation over the length of the compressor and turbine and “stall” which is the lack of
pressure rise between the compressor and blades. Stall occurs at low rotor speeds and
surge at high rotor speeds. Both surge and stall generate violent axial oscillations of the
internal air column which can cause substantial damage to both the compressor and the
engine. Operating characteristics of a typical compressor are shown in terms of corrected
parameters in the compressor map in Figure 5.2.
Surge line
Compressor
Characteristics
Zone of poor
compression
efficiency
Mass flow
99
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Also, the control system must be designed to avoid the compressor becoming too close to
the surge line and to maintain a sufficient design margin, for example a surge avoidance
line, see Figure 5.2. A design parameter to avoid surge is known as surge margin and is
based on the total pressure ratio between outlet pressure P0 and inlet pressure Pi at the
compressor. The surge margin (SM) is defined as follows (Tavakoli et al. 2004):
⎛ Po ⎞ ⎛P ⎞
⎜⎜ ⎟⎟ − ⎜⎜ o ⎟⎟
⎝ Pi ⎠ surge ⎝ Pi ⎠ surge avoidance
SM = .
⎛ Po ⎞
⎜⎜ ⎟⎟
⎝ Pi ⎠ surge avoidance
The overall function of the engine controller is to provide thrust in response to throttle
position. It must achieve the requested thrust with the lowest specific fuel consumption. It
must also ensure that limits are not exceeded, such as maximum compressor speed,
maximum turbine temperature, fan stall and compressor stall.
The overall control must control each limiting parameter so that none of the limits are
exceeded. Part power engine operation should occur below all limits at the lowest
specific fuel consumption for the thrust requested.
Because of the nature of thrust, commercial jet engines use an indirect measurement
called integrated engine pressure ratio (IEPR) to control it. However, the only closed loop
control in commercial engines is the main fuel control. On modern turbo engines, thrust
is more accurately controlled by setting the fan speed or the engine pressure ratio, which
is defined as the ratio of low pressure turbine to inlet pressure (Spang and Brown 1999).
100
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Commercial engines are generally controlled by a single closed loop main fuel control.
There is minimal interaction between each open loop control and the main fuel control.
Consequently, single input single output design techniques are usually adequate (Spang
and Brown 1999). Multivariable control has been developed for military engines or
turbojets and much of the literature on multivariable engine control is focused on the
regulator problem, which maintains the engine near the desired operating trajectory.
Polley et al. (1989) found a nonlinear controller by using appropriate engine corrected
parameters to schedule multivariable linear compensator gains designed at selected
operating points of the flight envelope. A KQ (K matrix compensator, Q-desired
response) multivariable control design technique was used to design the compensator
using linear state-space models. The KQ compensator design methodology is a frequency
domain technique, which uses closed-loop Nyquist and Bode plots. Parameters of the
matrix compensator K are estimated by minimizing the mean square error between the
desired closed-loop response Q and the actual closed-loop response. This method is
known as Edmunds’ model matching design technique, and is described in Edmunds
(1979).
Frederick et al. (2000) points out that the use of Edmunds’ design technique is easy to
use. However, it does not guarantee stability and does not take into account actuator
rates. They proposed using H∞ controller design to include weights on actuator rates and
enhance robustness as well as a simplified scheduling scheme to cover a range of power
levels from idle to maximum. They addressed the following areas of future research:
design of a low-order robust H∞ controller, reduction of gain scheduling effort and
complexity. Balas et al. (1998) have proposed applying linear parameter varying (LPV)
techniques for the design of gas turbine controllers. They demonstrated that LPV
101
Chapter 5 Multiobjective Control Design for the Turbofan Engine
techniques provide a methodology for designing H∞ gain scheduling controllers with low
design effort and complexity. These new techniques have been developed as a result of
recent improvements in the area of LMI in control systems.
LMI was first introduced in another related area, aircraft design, by Niewoehner and
Kaminer (1996). They developed a new methodology that provided a numerical
framework for the integrated aircraft-controller design. They considered a problem of
combining the design of some aircraft parameters with the control system development
using linear matrix inequalities. The remainder of this chapter investigates multivariable
control strategies using LMI techniques and evolutionary algorithms. The following
section will describe the GTE undertaken for this study.
Bypass duct
Variable inlet High Pressure Low Pressure
Fan Guide vanes Compressor (HPC) High Pressure Turbine (LPT)
(XNL) (ZIGV) Speed (ZXNH) Turbine (HPT)
102
Chapter 5 Multiobjective Control Design for the Turbofan Engine
The engine comprises of 3 spools; the fan being driven by the low pressure turbine
(LPT), the intermediate pressure compressor (IPC) which is driven by the intermediate
pressure turbine (IPT) and the high pressure compressor (HPC) which is driven by the
high pressure compressor (HPC) (see Figure 5.3). The engine has separated ducts; the
outermost part of the flow enters from the fan to the bypass duct which exhausts through
the cold nozzle. The combustor is selected as staged or conventional type. Variable inlet
guide vanes (ZIGV) are provided at IPC entry. The flight envelope was constrained to sea
level, zero Mach number and an outside environment of 15° C and 14.696 lb/in2.
The model is a nonlinear, thermodynamic type. Linear models were obtained for nine
steady-state operating points. The linearisation of the nonlinear plant was carried out by
perturbation of the states or Jacobian linearisation using the software MATRIXx. The
operating point was changed by varying the specified HPC speed (ZXNH), see Appendix
A.
Based on the literature of control for gas turbine engines (Postlethwaite et al. 1995;
Spang and Brown 1999; Frederick et al. 2000), the selected inputs and outputs are
proposed for control; Inputs: u1 is the specified engine fuel flow (ZWFE) and u2 is the
variable guide vane angle in the compressor (ZIGV). Outputs: y1 is the fan speed (XNL),
y2* is the surge margin IP (SMI) and y3* is the surge margin HP (SMH).
•
⎛ x1 ⎞ ⎛ x1 ⎞
⎜ ⎟ ⎜ ⎟ ⎛ u1 ⎞
⎜ x 2 ⎟ = Ai ⎜ x 2 ⎟ + Bi ⎜⎜ ⎟⎟ ,
⎜x ⎟ ⎜x ⎟ ⎝ u2 ⎠
⎝ 3⎠ ⎝ 3⎠
⎛ x1 ⎞
⎛ y1 ⎞ ⎜ ⎟ ⎛u ⎞
⎜⎜ ⎟⎟ = C i ⎜ x 2 ⎟ + Di ⎜⎜ 1 ⎟⎟ , (5.1)
⎝ y2 ⎠ ⎜x ⎟ ⎝u2 ⎠
⎝ 3⎠
103
Chapter 5 Multiobjective Control Design for the Turbofan Engine
i = 1 L 9.
where the states: x1 is the specified HP speed (ZXNH) , x2 is the specified IP speed (ZXNI)
and x3 is the specified LP speed (ZXNL).
5.3.1 Scaling
For controller design purposes, the plant was scaled as suggested in Skogestad and
Postlethwaite (1996). A condition number was obtained from the ratio between the
maximum and the minimum singular values of the plant; this condition number was
reduced by a factor of 34.261 (from 3866.7 to 112.86) by pre- and post-multiplying the
plant with diagonal scaling matrices. The inputs were scaled in a way that the maximum
allowed fuel flow (ZWFE) and the maximum allowed angle of the variable guide vine in
the compressor (ZIGV) correspond to the unit at the operating point. The outputs were
scaled in such a way that 2.2% of the fan speed corresponds to the unit and 5% of the
surge margin corresponds to the unit.
104
Chapter 5 Multiobjective Control Design for the Turbofan Engine
z1
Paug
W1
w=r
z2
W2
+
1
G(s) - s
u ym
K(s)
The performance specification ||W1S||∞<1 puts a lower bound on the bandwidth which
provides rejection to disturbance. To achieve robustness or restrict the magnitude of the
input signals, u=KS⋅r, one may place an upper bound 1/|W2|, on the magnitude KS. The
objective in the S/KS mixed-sensitivity is to minimize a suitable norm of the transfer
function from the input w to outputs Z1 and Z2. In order to enforce these two
specifications (performance and robustness), H∞ minimisation is utilised in equation
(5.2). If the minimal value of the left-hand side in (5.2) is greater than 1, then no
controller exists that stabilizes the system for all perturbations. In this case, stability
robustness is only obtained for perturbations satisfying (5.2) with the right-hand side
replaced with γ (Kwakernaak 1993).
W1 S
<1 (5.2)
W2 KS ∞
105
Chapter 5 Multiobjective Control Design for the Turbofan Engine
s / M + ωB
W P (s) = , (5.3)
s + ωB A
Step Response
From: ZWFE From: ZIGV
1.5
1
To: XNL
0.5
0
Amplitude
1.5
1
To: SMmix
0.5
-0.5
-1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (sec)
Figure 5.5 Step responses resulting from the controller obtained form the LMI synthesis.
The results presented in Figure 5.5 and Figure 5.6 have a bound γ= 1.9007. Although the
step response shows a stable close-loop system, this has a numerical problem. A
numerical analysis of the controller reveals that it is rank deficient. This numerical issue
will be studied in the next section.
106
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Singular Values
20
0
1/|wp|
-20
Singular Values (dB)
-40
-60
1/S
-80
-100
-2 -1 0 1 2 3 4 5
10 10 10 10 10 10 10 10
Frequency (rad/sec)
Figure 5.6 Analysis in frequency response of the controller resulting from the controller
obtained form the LMI synthesis.
In order to treat this problem with different specifications in independent channels, the
mixed H2/H∞ approach is proposed. The use of a single-objective can be a limitation
same frequency (Dettori 2001). Thus, channel 1 in Figure 5.4 can be optimized with H2
norm to enhance performance. The proposed optimisation problem can be stated as:
min W1 S 2
subject to W 2 KS ∞
≤ γ∞ (5.4)
The problem was solved using LMI techniques (Section 4.4). Figure 5.7 shows the Pareto
optimal front.
107
Chapter 5 Multiobjective Control Design for the Turbofan Engine
1.5
||W 1 S||2
"Analysis values"
0
0 1 2 3 4 5 6 7 8 9 10
||W 2 KS||inf
Figure 5.7 Trade-off curves for the mixed H2/H∞. problem. {Synthesis values ( ) and
Analysis values (◊)}.
An analysis of the controllers resulting from solving the mixed H2/H∞ problem (equation
5.4) with LMI techniques shows that they are rank deficient, and this leads to very poor
closed-loop stability. The reasons for this are listed as follows:
1) It is known that the H∞ techniques have problems with jω-axis zeros (see
Appendix B2. In this application, there are two invariant zeros at s=-0.001 (close
to the imaginary axis). This problem leads to incorrect estimations of γ∞. This
problem has been investigated before (Scherer 1992; Gahinet and Apkarian 1994;
Gahinet 1996). The LMI formulation by Gahinet and Apkarian (1994) provides a
parametrization of all H∞ suboptimal controllers where some of the free
parameters are pair (X, Y) of positive definite matrices that solve the LMIs and
satisfy some coupling constraints. The pair (X, Y) moves to infinity when γ∞ is
108
Chapter 5 Multiobjective Control Design for the Turbofan Engine
decreased below the upper continuity value and will place some closed-loop poles
very close to the imaginary axis.
2) The existence of the controller depends on the non-singularity of the matrix I-XY
as I-XY= UVT with non-singular U and V (see Section 4.3.1). A simple test shows
that the minimal eigenvalue of matrix XY of each controller is close to one, which
means that the controllers are ill-conditioned, see Table 5.1.
In order to improve the conditioning for reconstruction of the controllers, one should
avoid I-XY becoming near singular. This can be managed by fixing γ2 to be “larger” than
γ2,opt and including the LMI constraint
⎛Y βI ⎞
⎜⎜ ⎟ > 0, (5.5)
⎝ βI X ⎟⎠
with an additional variable β, and maximize β. This procedure maximizes the minimal
eigenvalue of X·Y. The extra inequality is expected to push the smallest singular value of
I-XY away from zero. This improves the conditioning U and V, and renders the
calculation of the controller more reliable (Scherer et al. 1997).
In this case, after a series of trials, it was found that: for each controller between number
1 to 4 (from table 5.2), the optimum fixed value of γ2 should be four times larger than
γ2,opt. For each controller between numbers 5 to 14 (from table 5.1), γ2,opt was fixed to
different factors, and the heuristic rule was to gradually decrease the fixed value from the
109
Chapter 5 Multiobjective Control Design for the Turbofan Engine
fixed value 4.8891, while the minimal eigenvalues of matrix (X·Y) remain away from one
and the closed-loop feedback system remains stable. The results are presented in Table
5.2, plotted in Figure 5.8 and magnified in Figure 5.9, where the synthesis values are
marked with (∇) and the analysis values are marked with (*).
6
"Synthesis values"
||W 1 S||2
2
"ill-conditioned controllers"
1
0
0 1 2 3 4 5 6 7 8 9 10
||W 2 KS||inf
Figure 5.8 Trade-off curves for problem in equation 5.4 with β constraint.
110
Chapter 5 Multiobjective Control Design for the Turbofan Engine
8
"Synthesis values with beta constraint"
7
6
||W 1 S||2
4
"Analysis values with beta constraint"
3
"ill-conditioned controllers"
2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
||W 2 KS||inf
Figure 5.9 Magnification of the trade-off curves for problem in equations 5.4 with β
constraint.
MOGA used mating restriction and crossover shuffle with reduced surrogate with
probability 0.7. A mutation rate of 0.09 was applied to all individuals after crossover. The
111
Chapter 5 Multiobjective Control Design for the Turbofan Engine
population size was 200 candidate solutions and the Pareto front after 200 generations is
presented in Figure 5.10 and 5.11 (marked o) to compare it with the corresponding LMI
results.
Trade off Mixed H2/Hinf
10
7
"Synthesis values with beta constraint"
6
||W 1 S||2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
||W 2 KS||inf
Figure 5.10 Trade-off curves. Comparison between MOGA-based method and LMI
Approach.
A comparison of step responses resulting from the LMI synthesis and the MOGA-based
method is shown in Figure 5.12. The LMI controller (“dashed line”) is the controller 4 in
Tables 5.1 and 5.2 and it has analysis values W2 KS ∞ = 0.33794 and W1 S 2 =1.8759.
The controller resulting from the multiobjective H2/H∞: MOGA-base method (“solid
line”) has synthesis values γ∞= 0.41946 and γ2 = 1.0072.
It was demonstrated that the H∞ controller design technique and the mixed H2/H∞
controller design technique find ill-conditioned controllers for this gas turbine case study.
Although a remedy has been proposed by Scherer et al. (1997), the procedure to obtain
the Pareto-front set requires a lot of trial and errors to fix an appropriate optimal value or
bound γ. This procedure is impractical and time-consuming. On the other hand, the
multiobjective H2/H∞ controller design technique tackled by the MOGA-based method
overcomes these problems.
112
Chapter 5 Multiobjective Control Design for the Turbofan Engine
7
"Synthesis values with beta constraint"
6
||W 1 S||2
4
"Analysis values with beta constraint"
3
1
MOGA
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
||W 2 KS||inf
Figure 5.11 Trade-off curves, comparison between MOGA-based method and LMI
Approach with beta constraint.
Step Response
From: ZWFE From: ZIGV
1.5
1
To: XNL
0.5
0
Amplitude
-0.5
1.5
1
To: SMmix
0.5
-0.5
-1
0 1 2 3 4 5 0 1 2 3 4 5
Time (sec)
Figure 5.12 Step responses resulting form the Mixed H2/H∞: LMI Synthesis (“Dashed
line”) and MOGA-base method (“Solid line”).
113
Chapter 5 Multiobjective Control Design for the Turbofan Engine
In the previous section, the linear controller design was applied to a single operating
point. Now the question will be whether or not this control can be extended to a different
operating point while keeping stability and performance. Since the gas turbine engine is a
nonlinear system, gain scheduling control seems to be an attractive approach, and will
now be studied. A three-step gain scheduling control design procedure has been widely
applied (Shamma and Athans 1990; Rugh 1991):
where x is the state, u is the input. It is known that the dynamical behaviour of a
nonlinear system (5.7) changes with the operating regions, which are physical
variables, such as speed, temperature, altitude, etc. The designer selects several
operating points which cover the range of the dynamics of the plant, then the model is
linearised about one or more operating points (also called set points, equilibrium
points, or operating conditions). In the literature for nonlinear systems, the
equilibrium point is more frequently used. An equilibrium point is a point such
that f ( xo , u o ) = 0 , Associated with this equilibrium point is the output operating
point y o = h( x o , u o ) . For the purpose of this investigation, an equilibrium point will be
referred to as an operating point.
For the system (5.7), suppose an operating point x o , u o , and y o is known. The
difference between the equilibrium vector function and some slightly perturbed
function x(t ) , u (t ) , and y (t ) can be defined by
δx = x(t ) − xo
δu = u (t ) − u o (5.8)
δy = y (t ) − y o
and the equation (5.7) is written as
114
Chapter 5 Multiobjective Control Design for the Turbofan Engine
⎡ ∂f ⎤ ⎡ ∂f ⎤
x& o + δx& = f ( x o + δx, u o + δu ) = f ( x o , u o ) + ⎢ ⎥ δx + ⎢ ⎥ δu + higher − order terms
⎣ ∂x ⎦ o ⎣ ∂u ⎦ o
⎡ ∂h ⎤ ⎡ ∂h ⎤
y o + δy = h( x o + δx, u o + δu ) = h( x o , u o ) + ⎢ ⎥ δx + ⎢ ⎥ δu + higher − order terms
⎣ ∂x ⎦ o ⎣ ∂u ⎦ o
where [ ]o means the derivates are evaluated at the operating point. Since the
operating point solution satisfies equation (5.7), the first terms in the preceding
Taylor series expansions cancel. For sufficiently small δx , δu and δy perturbations,
the higher-order terms can be neglected, leaving the linear equations
⎡ ∂f ⎤ ⎡ ∂f ⎤
δx& = ⎢ ⎥ δx + ⎢ ⎥ δu , (5.9)
⎣ ∂x ⎦ o ⎣ ∂u ⎦ o
⎡ ∂h ⎤ ⎡ ∂h ⎤
δy = ⎢ ⎥ δx + ⎢ ⎥ δu , (5.10)
⎣ ∂x ⎦ o ⎣ ∂u ⎦ o
⎡ ∂f ⎤ ⎡ ∂f ⎤ ⎡ ∂h ⎤ ⎡ ∂h ⎤
where A = ⎢ ⎥ , B = ⎢ ⎥ , C = ⎢ ⎥ , D = ⎢ ⎥ .
⎣ ∂x ⎦ o ⎣ ∂u ⎦ o ⎣ ∂x ⎦ o ⎣ ∂u ⎦ o
When the system has more than one operating point, the linearization family will
require a family of operating points and they may need to be parametrized by the
scheduling variable θ , and then the operating points will be defined as x o (θ ) , u o (θ )
and y o (θ ) as well as the model,
where δy = y (t ) − y o (θ ) .
2) After linearisation, one arrives at a set of linear models at each operating point. In
order to arrive at a set of linear feedback control laws, linear design methods are
applied performing satisfactorily when the closed-loop systems is operated near
the respective operating points. Consider the linear controller family:
δx& c = Ac (θ )[ x c (t ) − x o ,c (θ )] + Bc (θ )[ y (t ) − y o (θ )] ,
δu = C c (θ )[ x c (t ) − x o ,c (θ )] + Dc (θ )[ y (t ) − y o (θ )] . (5.12)
115
Chapter 5 Multiobjective Control Design for the Turbofan Engine
The objective of the controller is that the operating point of the plant matches the
same operating point of the corresponding closed-loop system.
3) The final step is the gain scheduling. This is achieved either by interpolating the
controller parameters between operating points or by using non-smooth switching
strategy between the linear controllers (Astrom and Wittenmark 1989). The result
is a global compensator.
Qualities such as stability and robust performance are then evaluated through exhaustive
simulations, under the assumption that the scheduling variable should very slowly and
that the scheduling variable should capture the nonlinear plant behaviour (Shamma and
Athans 1990). Traditionally, gain scheduling techniques formulated relatively simple
structure control laws and gains were selected for each operating condition by using
classic SISO control methods. As process systems and aerospace applications such as
GTE became more complex, and performance capabilities more demanding, modern
control techniques have to be applied. The resulting controllers became more complex
with many more gains. Consequently, evaluating gain scheduling controllers using
exhaustive simulations results is impractical.
116
Chapter 5 Multiobjective Control Design for the Turbofan Engine
1999), robots (Kajiwara et al. 1998), inductor motors (Prempain et al. 2002) and power
systems (Qiu et al. 2004).
r + u y
K(s,θ) G(s,θ)
-
θ(t)
The LPV structure incorporates the parameter measurement θ in real time. This technique
“automatically” produces gain scheduling controllers which simultaneously ensure
global stability over the operating domain, and linear time invariant robustness properties
(Biannic et al. 1997). It has been shown that the illustrated LPV control problem
formulated in Figure 5.13 can be formulated as a convex optimisation problem (Gahinet
et al. 1995). This formulation is explained as follows:
117
Chapter 5 Multiobjective Control Design for the Turbofan Engine
V ( x cl , θ ) = x clT P (θ ) x cl , (5.14)
118
Chapter 5 Multiobjective Control Design for the Turbofan Engine
The aim of developing an LPV model is to capture the actual non-linear behaviour of the
plant in a global model, although a family of linearized models of a plant can be
parametrized by some signal, and therefore could be presented by an LPV system.
Motivated by the hybrid LPV modelling method (Sun and Postlethwaite 1998), and the
LPV modelling method for the longitudinal control of a high-performance aircraft by
(Biannic et al. 1997), the LPV modelling and control with polytopic systems was selected
to tackle the gas turbine engine control problem.
1) The state-space matrices A(·), B(·), C(·), D(·) depend affinely on the time-varying
parameter θ , which is a vector of physical variables (velocity, angle, temperature,
etc) in the following structure:
119
Chapter 5 Multiobjective Control Design for the Turbofan Engine
i= j
A(θ ) = A0 + ∑ θ i Ai ,
i =1
i= j
B (θ ) = B0 + ∑ θ i Bi ,
i =1
i= j
C (θ ) = C 0 + ∑ θ i C i ,
i =1
i= j
D(θ ) = D0 + ∑ θ i Di . (5.15)
i =1
{( ) ( )} ( )
the domain Ρ ⊆ θ 1 , θ 1 ,..., θ j , θ j , where θ j , θ j are the lower/upper bounds of
θ j , can be put into an rth order polytopic form, where r = 2 j . The measurements
r
where ρi ≥ 0 , ∑ ρi = 1 .
i =1
These vertices represent the extreme values of the parameters. Though not fully general,
this description encompasses many practical situations. The Polytopic system is defined
as follows:
From this characterization, it is clear that the state space matrices evolve in a polytope of
matrices whose vertices are the images of the vertices Θ 1 , Θ 2 ,..., Θ r . For an illustration
of this mapping, see Figure 5.14.
120
Chapter 5 Multiobjective Control Design for the Turbofan Engine
( )
Θ 2 = (θ 1 , θ 2 ) , Θ 3 = (θ 1 , θ 2 ) and Θ 4 = θ 1 , θ 2 . Mapping the polytopic parameters into a
polytopic of matrices creates a set of state-space systems which evolve into a polytope of
matrices with four vertices.
⎛ A(Θ 4 ) B (Θ 4 ) ⎞
⎜⎜ ⎟⎟
⎛ A(Θ 3 ) B(Θ 3 ) ⎞ ⎝ C (Θ 4 ) D (Θ 4 ) ⎠
θ2 Θ4 Θ3 S(θ2) ⎜⎜ ⎟⎟
⎝ C (Θ 3 ) D (Θ 3 ) ⎠
⎛ A(Θ1 ) B (Θ1 ) ⎞
Ξ ⎜⎜ ⎟⎟
⎝ C (Θ 1 ) D (Θ 1 ) ⎠
Θ1
Θ2 ⎛ A(Θ 2 ) B (Θ 2 ) ⎞
⎜⎜ ⎟⎟
⎝ C (Θ 2 ) D (Θ 2 ) ⎠
θ1 S(θ1)
Polytopic of parameters Polytopic of matrices
The LPV system is extended to augmented plants. Consider an augmented LPV plant,
mapping exogenous inputs w and control inputs u to controlled outputs z and measured
outputs y,
x& = A(θ ) x + B1 (θ ) w + B 2 u ,
z = C1 (θ ) x + D11 (θ ) w + D12 u ,
y = C 2 x + D 21 w + D 22 u . (5.18)
A3) The pairs ( A(θ ) ,B2) and ( A(θ ) ,C2) are quadratically stabilizable and quadratically
detectable over Ξ respectively.
121
Chapter 5 Multiobjective Control Design for the Turbofan Engine
If the assumption A2 is not satisfied, the computation of a LPV controller (next Section)
requires the solution of a problem with an infinite number of constraints. However, this
difficulty can be alleviated by pre- and/or post-filtering of the control inputs u and/or the
measured outputs y. In real world problems, the plant will include actuator and sensor
dynamics, and then the control and measurement matrices are parameter independent.
x&c = Ac (θ ) xc + Bc (θ ) y
(5.19)
u = Cc (θ ) xc + Dc (θ ) y
Following the vertex property, the controller system can be mapped into an LPV system
with polytopic form:
(5.20)
The controller state-space matrices at any operating point θ(t) are obtained by convex
interpolation of the LTI vertex controllers, in the following way,
⎛ A (Θ ) B c (Θ 1 ) ⎞ ⎛ A (Θ ) Bc (Θ 2 ) ⎞ ⎛ A (Θ ) Bc (Θ r ) ⎞
K (θ ) := ρ1 ⎜⎜ c 1 ⎟⎟ + ρ 2 ⎜⎜ c 2 ⎟⎟ + L + ρ r ⎜⎜ c r ⎟⎟, (5.21)
⎝ C c (Θ 1 ) D c (Θ 1 ) ⎠ ⎝ C c (Θ 2 ) Dc (Θ 2 ) ⎠ ⎝ C c (Θ r ) Dc (Θ r ) ⎠
where ρ1 ≥ 0 , ρ 2 ≥ 0 , …, ρ r ≥ 0 and ρ1 + ρ 2 + L + ρ r = 1 .
122
Chapter 5 Multiobjective Control Design for the Turbofan Engine
The existence of an LPV controller depends on the assumptions (A1)-(A3) and the
following two statements.
1) Given some positive scalar γ, there exists a kth-order LPV controller solving the
quadratic H∞ problem where the Lyapunov function V ( x cl ) = xclT X cl xcl ensures
asymptotic stability, and the L2 gain of the input/output map is bounded by γ. This
is
y 2
<γ u 2.
2) There exists some (n+k)x(n+k) positive definite matrix X cl and the LTI
⎛ Ac ,i B c ,i ⎞
controllers K i = ⎜⎜ ⎟ such that the LMI (5.22) derived from the bounded
⎝ C c ,i Dc ,i ⎟⎠
After computing an adequate Xcl, the problem is reduced to solving a system of LMIs for
all the vertices. The LMI-based solvability conditions are given in the following theorem.
T ⎛ Ai S + SAi C1Ti ⎞
T
SB1i
⎛ NS 0⎞ ⎜ ⎟
T ⎛ NS 0⎞
⎜⎜ ⎟ ⎜ B1Ti S − γI D11 i ⎟⎜
⎜ ⎟<0, I = 1, 2, …r (5.24)
⎝ 0 I ⎟⎠ ⎜ ⎟ 0 I ⎟⎠
⎜
⎝ C1i D11i − γI ⎟⎠⎝
123
Chapter 5 Multiobjective Control Design for the Turbofan Engine
⎛R I ⎞
⎜⎜ ⎟⎟ ≥ 0 , (5.25)
⎝ I S⎠
where N R and N S denote bases of the null spaces of ( B2T , D12T ) and (C 2 , D21 )
respectively. In order to retain convexity, the order of the controller must be the same as
the order of the plant. Proof of this theorem is described in Apkarian and Gahinet (1995).
Once adequate matrices R and S have been computed, the computation of the controller is
as follows: The Lyapunov Xcl, common to all inequalities (5.22), is derived from SVD
computation of R and S. The vertex controllers Ki, can be computed using explicit
formulae of Gahinet and Apkarian (1994), which involves standard linear algebra.
In summary, using the notion of quadratic H∞ performance will enforce the L2 gain
constraint, and observing that the closed-loop system is a polytopic system, the design of
a gain-schedule controller K (θ ) satisfying the vertex property will exhibit the following
characteristics:
In the following section, this application of the polytopic approach will be investigated.
124
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Recalling that the linearisation of the nonlinear plant was carried out by changing the
speed of the fan ZXNH, nine LTI plants were obtained from this process. Figure 5.15
plots the pole map of the corresponding nine LTI plants. The nine LTI plants have three
eigenvalues and Figure 5.15 shows, with numbers, one of the poles of each plant.
Pole Map
1.5
8
7
9 5
6
0.5 4
2
1
Imaginary Axis
3
-0.5
-1
-1.5
-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
Real Axis
Figure 5.15 Open-loop system poles for the nine LTI plants.
There is not a general method to derive an LPV model. One can start by analyzing the
behaviour of the nine state space matrices A, B, C, D versus the variable ZXNH that was
used to linearize the nonlinear model. Plots of these variations are shown in Figures 5.16
to 5.18, where the values of the entries are marked as (°). In this case, the model is strictly
proper, thus the matrix D = 0. The idea is to capture the nonlinear behaviour of the GTE,
which is represented in Figures 5.16 to 5.18, into an LPV model. The first attempt to
develop an affine model is to consider each of the entries, e.g. aij , of the matrices A, B, C
of the model (5.1) as independent variables. For example, following the notation in (5.15)
for the matrix A(θ ).
125
Chapter 5 Multiobjective Control Design for the Turbofan Engine
⎛0 0 0⎞ ⎛1 0 0⎞ ⎛ 0 1 0⎞ ⎛ 0 0 0⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
A(θ ) = ⎜ 0 0 0 ⎟ + a11 ⎜ 0 0 0 ⎟ + a12 ⎜ 0 0 0 ⎟ + L + a 33 ⎜ 0 0 0 ⎟ (5.26)
⎜ 0 0 0⎟ ⎜ 0 0 0⎟ ⎜ 0 0 0⎟ ⎜0 0 1⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠
This affine model will have 21 independent variables, and the polytopic model will have
r = 2,097,152 vertices and it means solving a system of 2r+1 LMIs. It has been reported
(Dettori and Scherer 2002) that the LMI solver is not able to solve a problem above 1000
LMIs. Therefore this approach is not tractable.
10 10 10
5 5 5
a11
a12
a13
0 0 0
-5 -5 -5
5 5 5
a21
a22
a23
0 0 0
-5 -5 -5
5 5 5
a31
a32
a33
0 0 0
-5 -5 -5
126
Chapter 5 Multiobjective Control Design for the Turbofan Engine
10 10 10
5 5 5
c11
c12
c13
0 0 0
-5 -5 -5
0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3
ZXNH 4 ZXNH 4 ZXNH 4
x 10 x 10 x 10
10 10 10
5 5 5
c21
c22
c23
0 0 0
-5 -5 -5
0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3
ZXNH 4 ZXNH 4 ZXNH 4
x 10 x 10 x 10
10 10
5 5
b11
b12
0 0
-5 -5
-10 -10
0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3
ZXNH 4 ZXNH 4
x 10 x 10
10 10
5 5
b21
b22
0 0
-5 -5
-10 -10
0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3
ZXNH 4 ZXNH 4
x 10 x 10
10 10
5 5
b31
b32
0 0
-5 -5
-10 -10
0.9 1 1.1 1.2 1.3 0.9 1 1.1 1.2 1.3
ZXNH 4 ZXNH 4
x 10 x 10
A second attempt of affine modelling was investigated, the purpose being to reduce the
number of vertices. For example, Figure 5.18 shows that b22 and b32 do not vary more
than 12% compared with the maximum variation in matrix B. Also, variables a11, a12, and
a21 can be approximated to a linear function dependent on ZXNH (see the dotted lines in
Figure 5.16):
127
Chapter 5 Multiobjective Control Design for the Turbofan Engine
0 1
a11 ≈ a11 + a11 ZXNH . (5.28)
This approach has been applied to missile autopilots (Wu et al. 1995; Biannic and
Apkarian 1999) and a high-performance aircraft (Biannic et al. 1997). In order to
combine the first attempt and the linear approximation idea, an affine LPV model was
developed, based on the hybrid method of affine LPV models by Sun and Postlethwaite
(1998), and trials on the reduction of vertices were carried out. More entries on the state-
space matrices were identified to be approximated to a linear function of ZXNH via least
square optimisation. For example, the matrix A(θ ).
⎛ a11
0 0
a12 0⎞ ⎛ a11
1 1
a12 0⎞ ⎛0 0 0⎞ ⎛ 0 0 0⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
A(θ ) ≈ ⎜ a21
0
0 0 ⎟ + ZXNH ⎜ a121 0 0 ⎟ + a22 ⎜ 0 1 0 ⎟ + L + a33 ⎜ 0 0 0 ⎟ (5.27)
⎜⎜ ⎟ ⎜⎜ ⎟ ⎜0 0 0⎟ ⎜0 0 1⎟
⎝ 0 0 0 ⎟⎠ ⎝ 0 0 0 ⎟⎠ ⎝ ⎠ ⎝ ⎠
An affine model of approximately half of the vertices was developed. The experiments on
A(ZXNH) were carried out, attempting to match the pole map and reducing the
interpolation errors or residuals. However, these simplifications on the affine LPV model
did not result in a tractable model. Although these attempts result in very computationally
expensive models, it was decided to study further these ideas in local operating regions.
The motivation was that the polytopic form is an attractive controller structure to
investigate its application in controller design methods with Evolutionary strategies,
whereas the application of LFT and parametric-dependent Lyapunov functions
approaches are not clear. The problem under study was reduced to only two operating
points, and then was simplified to two vertices. The gain scheduling design for the GTE
was carried out between 80% and 85% of power. A linear affine model was derived as
well as its polytopic form, as follows,
x& = ( ρ1 A1 + ρ2 A2 )x + ( ρ1 B1 + ρ2 B2 )u ,
y = ( ρ1C1 + ρ 2 C 2 ) x . (5.28)
where (A1, B1, C1) and (A2, B2, C2) are the state-space matrices of the operating points at
80% and 85% respectively, and the polytopic coordinates are,
128
Chapter 5 Multiobjective Control Design for the Turbofan Engine
ZXNH − ZXNH (t )
ρ1 = ,
ZXNH − ZXNH
ZXNH (t ) − ZXNH
ρ2 = ,
ZXNH − ZXNH
ρ1 + ρ 2 = 1 ,
[
ZXNH (t ) ∈ ZXNH ]
ZXNH = [12426rpm 13157 rpm] . (5.29)
Once the LPV model has been derived, an attempt to design a self-scheduled controller
can be investigated. For the purpose of the controller design method, the variation on the
matrix B1 and B2 was neglected, the reason being to satisfy the assumptions of section
5.9. The controller structure is derived from Section 5.2. The augmented plant with
weighting functions and the scheduling parameter ZXNH is depicted in Figure 5.19.
Paug z1
W1
w=r
z2
W2
+
1
G(s, ZXNH) - s
u ym
ZXNH
K(s, ZXNH)
W1 S
<γ (5.30)
W2 K ( s, ZXNH ) S ∞
129
Chapter 5 Multiobjective Control Design for the Turbofan Engine
In order to solve this optimisation problem for the LPV augmented plant Paug (see Figure
5.19), the quadratic H∞ performance approach was used. The weighting functions W1 and
W2 were the same selected in Section 5.3 and were fixed for the two operating points. The
LPV synthesis of the gain-scheduled structure (see Figure 5.19) was solved using the
LMIs (equations 5.23-5.25). The software is available in the MATLAB LMI Toolbox
(Gahinet et al. 1995). A performance level of γ=0.76 was achieved. Then, the LPV
polytopic controller can be constructed from the convex decomposition 5.21, as follows,
where ρ1 and ρ 2 are the same as the plant, see polytopic coordinates equation (5.29).
The state-space matrices AC1 , BC1 , C C1 , DC1 and AC2 , BC2 , C C2 , DC2 are obtained by fixing
the polytopic coordinates on the vertices (operating points 80% and 85% respectively).
Therefore, any plant (5.28) between 80% and 85% can be controlled by convex
interpolation of the LTI vertex controllers. An assessment of the LPV controller was
carried out by step responses of closed-loop LTI systems. These closed-loop systems
were obtained by fixing the scheduling parameter ZXNH, both for the plant and the
controller at 20 operating points between 80% and 85%. The step responses are in Figure
5.21
u y
r + 1 K(s, ZXNH) G(s, ZXNH)
s
-
ZXNH(t)
130
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Step Response
From: ZWFE From: ZIGV
1.5
1
To: XNL
0.5
0
Amplitude
1.5
0.5
To: SMmix
-0.5
-1
-1.5
0 1 2 3 0 1 2 3
Time (sec)
Figure 5.21 Step responses of 20 operating points in between 80% and 85%.
A numerical study of the LPV controllers revealed that they are ill-conditioned for the
same reasons explained in Section 5.4.1. One attempt to get more reliable controllers was
to include the LMI constraint (5.5) proposed by Scherer et al. (1997). This constraint
slightly improved the rank of the controllers but not enough to get full order on the rank
of the controller. The motivation of this section was to study the polytopic form of
controller structure for its application in controller design methods with evolutionary
strategies. Therefore, no further investigation on the numerical issue of the LMI
techniques was carried out.
131
Chapter 5 Multiobjective Control Design for the Turbofan Engine
where the state-space matrices AC1 , BC1 , C C1 , DC1 and AC2 , BC2 , C C2 , DC2 are the decision
variables and they have the MIMO canonical structure form (Kailath 1980),
⎛ 0 1 0 0 0 0 0 0 0 0 ⎞
⎜ ⎟
⎜ 0 0 1 0 0 0 0 0 0 0 ⎟
⎜ an a 2n a3n a 4n a5n a 6n a7n 1 n
b22 n ⎟
b23
⎜ 1 ⎟
⎜ 0 0 0 0 1 0 0 0 0 0 ⎟
⎛ ACn BCn ⎞ ⎜ n n ⎟
⎜ ⎟ = a8 a9n n
a10 n
a11 n
a12 n
a13 n
a14 0 1 b24
⎜C n
⎝ C DCn ⎟⎠ ⎜ ⎟, (5.33)
⎜ 0 0 0 0 0 0 1 0 0 0 ⎟
⎜ n n n n n n n ⎟
⎜ a15 a16 a17 a18 a19 a 21 a 22 0 0 1 ⎟
⎜cn n
c 26 n
c 27 n
c 28 n
c 29 n
c30 n
c31 n
d 39 0 0 ⎟
⎜⎜ 25 ⎟
n
⎝ c32
n
c33 n
c34 n
c35 n
c36 n
c37 n
c38 0 0 0 ⎟⎠
where n = 1, 2.
In the LMI approach, the vertex property alleviates the problem of having an infinite
dimensional problem. In other words, the LMI approach solves the problem at the
vertices and guarantees stability and robustness for the entire operating domain. The
same approach was taken here, different attempts being carried out. For example, the
controller at 80% (obtained from Section 5.5) was fixed and the controllers at two
operating points 82.5% and 85% were optimised. However, the resulting controllers were
unstable between the operating points of 80% and 85%.
An appropriate strategy for optimizing the gain scheduling controllers was found to be
one where optimisation is allowed at the two vertices and at a third operating point at the
mid-point. The evaluation function was implemented to evaluate the stability as a
constraint and to evaluate H2 performance and H∞ performance at 80% and 85%. In an
attempt to ensure the same properties over the whole operating domain, the third
operating point was evaluated at the mid-point, 82.5%. In summary, the evaluation
function contains one constraint and two objectives for each operating point, see Table
5.3.
132
Chapter 5 Multiobjective Control Design for the Turbofan Engine
K 82.5% =
Controller ⎛ A1 BC1 ⎞ ⎛ A2 BC2 ⎞
K ( ZXNH (t )) K 80% = ⎜⎜ C1 ⎟ ⎛ A1 BC1 ⎞ ⎛ A2 BC2 ⎞ K 85% = ⎜⎜ C2 ⎟
⎝ CC DC1 ⎟⎠ 0.5⎜⎜ C1 ⎟ + 0.5⎜ C
⎟ ⎜C 2
⎟
⎟ ⎝ CC DC2 ⎟⎠
⎝ CC DC1 ⎠ ⎝ C DC2 ⎠
Constraint (1 max(λ( Acl −80% ) ) < 0 (2) max(λ( Acl −82.5% ) ) < 0 (3) max(λ( Acl −85% ) ) < 0
(5) W1 S 80% 2
(9) W1 S 82.5% 2
(7) W1 S 85% 2
Table 5.3 Objectives, constraints and controllers for the evaluation function.
As can be seen in Table 5.3, the evaluation function used three controllers. Two of them
are the controllers at the vertices/operating points 80% and 85% and the third controller,
at the middle operating point 82.5%, is calculated by convex interpolation.
Note, the decision variables are the controller matrices AC1 , BC1 , C C1 , DC1 , AC2 , BC2 , C C2 ,
and DC2 . The third controller uses a convex interpolation of the controllers at the vertices;
thus it does not use extra decision variables.
MOGA used mating restriction and “crossover shuffle with reduced surrogate” with
probability 0.7. A mutation rate of 0.09 was applied to all individuals after crossover, and
the population size was 250 candidate solutions. The Pareto-front was obtained after 100
generations, and is presented in Figure 5.22 (parallel coordinates), the corresponding
values of the normalized coordinates (y-axis, Figure 5.22) are in Table 5.4 as well as the
goal vector. As in the LMI approach, an assessment of the LPV controller was carried out
using step responses of closed-loop LTI systems. These closed-loop systems were
obtained by fixing the scheduling parameter ZXNH both for the plant and the controller at
20 operating points between 80% and 85%. The step responses are plotted in Figure 5.23.
133
Chapter 5 Multiobjective Control Design for the Turbofan Engine
MOGA Trade-off
Cost
1 2 3 4 5 6 7 8 9
Objective no.
Figure 5.22 Trade-off of the Gain-scheduling design for the GTE using MOGA-based
method.
These results demonstrate that the MOGA-based method can be used to deal with
uncertain systems or gain-scheduling controllers. These results are preliminary studies
and further investigations are required to establish conclusions about the application of
LPV techniques with evolutionary algorithms. In addition, the MOGA-based method can
deal with nonlinear problems in a different way such as the hybrid Fuzzy-MOGA
approach by Chipperfield et al. (2002).
4 W1 S 2
@ 80% 5 20 20
5 W2 ( KS ) ∞
@ 80% 0.1 0.8 3
W1 S @ 85%
Objectives
6 2 4 15 20
7 W2 ( KS ) ∞
@ 85% 0.2 0.8 3
8 W1 S 2
@ 82.5% 3 6 20
9 W2 ( KS ) ∞
@ 82.5% 0.2 0.45 3
Table 5.4 Normalization of the design objectives in Figure 5.22 and goal vector.
134
Chapter 5 Multiobjective Control Design for the Turbofan Engine
Step Response
From: ZWFE From: ZIGV
1.5
To: XNL
0.5
0
Amplitude
1.5
1
To: SMmix
0.5
-0.5
-1
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (sec)
Figure 5.23 Step responses of 20 operating points in between 80% and 85%.
The step responses in Figure 5.23 can be improved by optimising the weighting function
defined in Section 5.3.2. When using H∞ techniques, the weighting function has to be
designed with trial and error. Thus, MOGA can be used to optimise weighting functions
as it was proposed by Dakev et al. (1997); Griffin et al. (2000); Schroder et al. (2001).
135
Chapter 5 Multiobjective Control Design for the Turbofan Engine
5.12 Conclusions
The gas turbine engine study case is one of the more challenging applications in the
aerospace industry. The results in this chapter, when controller design is applied, reveal
an important numerical limitation when using LMI techniques. Alternatively, the
multiobjective H2/H∞ control design: evolutionary approach offers a more reliable
optimisation. The well-conditioned controllers were achieved by using the MOGA-based
method. Another achievement by this evolutionary approach was that results from
Chapter 4 were extended to the MIMO case.
When further studies in gain-scheduling control design were considered, it became clear
from section 5.10 that the LMI Control Toolbox solver faces another limitation with the
number of variables that this solver can support. From these preliminary results, it is
concluded that the gain scheduling controller design for GTE is still a theoretical issue
that does not have a clear methodology to parameterize a nonlinear system into an LPV
model. This is a salient issue that might open areas of research in system identification.
The controller design result shown in Section 5.11 demonstrated that the MOGA-based
method has potential in addressing the LPV gain-scheduling controller design.
136
Chapter 6
Conclusions and Future Work
This comparative study of two emerging optimisation techniques looks at the difference
between a convex optimisation technique over LMI constraints, which is single objective
and the multiobjective genetic algorithm (MOGA). At first, it seems unfair to compare
two techniques which are different in nature. However, far from deciding which one is
the best one; this investigation demonstrated the way the second optimisation technique
can assist by improving results or offering a solution to a control problem.
It has been shown, that a control problem, which is formulated into a linear matrix
inequality constraint, can be efficiently solved as a convex optimisation problem by the
interior-point methods. However, to translate a control engineering specification into a
LMI constraint is not a trivial task.
The convex optimisation over LMI constraints or semidefinite programming (SDP) is the
area of research which develops the mathematical theory for translating control
engineering specifications into LMI constraints. Many of these control engineering
specifications are not convex, and SDP theory seeks for a way to overcome non-
convexity. This requires a very deep understanding of theories such as dual theory and
LMI relaxations.
137
Chapter 6 Conclusions and Future Work
When the multiobjective control problem is tackled with LMI techniques, the
optimisation problem is a necessarily single objective problem, due to the nature of
convex optimisation. Recall that this technique is a general form of linear programming
(LP) (Parrillo and Lall 2003).
Although current research has shown advances in nonconvex optimisation (Kocvara and
Stingl 2003; Henrion and Lasserre 2004) and that it might handle problems such as the
mixed H2/H∞ in a more proper way, it is found that LMI techniques are not entirely
prepared to handle truly multiobjective control problems.
6.2 Suitability
On the other hand, MOGA can be applied to multiobjective control problems by
encoding the control problem, defining the vector spaces and the objectives. If a
successful design is easily located, further improvements such as reduced controller order
can be incorporated. Alternatively, if acceptable solutions are not readily obtained, the
designer may choose to relax the design goals. However, all this flexibility has to be
treated with caution; MOGA was not designed only to solve control system problems. In
the case of MOEAs making the wrong choice of EA parameters, exceedingly poor results
can be produced, as was pointed out by Fleming and Purshouse (2002).
Thus, the suitability of MOGA has to be defined in comparison with LMI techniques.
First, in control systems theory, any method has to be accompanied by analysis and
design. MOGA is not suitable for analysis; it is an optimisation tool that must be used for
138
Chapter 6 Conclusions and Future Work
139
Chapter 6 Conclusions and Future Work
• The case study of the gas turbine engine showed practical/numerical issues of the
state of the art LMI solvers. It was demonstrated that controllers resulting from
the LMI approached are ill-conditioned, whereas MOGA offers a robust
optimisation and provides well-conditioned controllers. In an attempt to develop
gain scheduling controllers and in the literature of applications, the LMI solvers
have shown that it is not possible to handle more than 1000 decision variables.
This is due to the LMI formulation. In the case of MOGA, it is not well
understood how to gain schedule controllers in a global way. It has been shown
assist in tuning controllers at several operating points (locally) when the structure
is designed by a Fuzzy approach (Chipperfield et al. 2002).
140
Chapter 6 Conclusions and Future Work
The MOGA-based method can be extended to deal with nonlinear problems, using gain-
scheduling techniques. Research can start by using the results from this thesis in
conjunction with the hybrid Fuzzy-MOGA approach by Stirrup et al. (2001),
Chipperfield et al. (2002) and Stirrup and Chipperfield (2002)
Research into uncertain systems will also provide the background to MOGA to be able to
handle gain-scheduling controllers. One possibility is to study linear parametric varying
(LPV) modelling and LPV control to understand what kind of structures can be used by
MOEAs such as the MOGA. In the study case of the gas turbine engine, this result has
not been tested with the nonlinear plant; this research requires a more extensive analysis
of the results with nonlinear theory.
Finally, in the case of the mixed H2/H∞ problem, in order to understand how accurate the
Pareto optimal is, it will be very interesting to use the lower bound computation by
Scherer (1999) and the use of Youla techniques (Scherer 2000) to approximate the
optimal value.
141
Appendix A
Linear models of the ANTLE gas turbine engine were obtained for nine steady-state
operating points. The linearisation of the nonlinear plant was carried out by perturbation
of the states or Jacobian linearisation using the software MATRIXx.
Table A1. Initial conditions of the outputs (yo) for the nine LTI plants
Num Power ZXNH (rpm) ZXNI (rpm) ZXNL (rpm) ZWFE (lb/hr) ZIGV
state x1 state x2 state x3 input u1 (degrees)
input u2
1 50% 8040.4 3257.9 479.29 591.975 37.7
2 55% 8771.4 3891.2 610.85 772.778 37.7
3 60% 9502.4 4732.4 761.51 1051.31 36.4
4 65% 10233 5772.8 953.39 1499.22 14.718
5 70% 10964 6490.8 1440.1 2622.25 -7.3744
6 75% 11695 7048.5 1875.3 4370.62 -15.5
7 80% 12426 7390.3 2292 6916.6 -15.5
8 85% 13157 7797.3 2686.3 10763.2 -15.5
9 90% 13888 8188.1 2992.3 15525.1 -15.5
Table A2. Initial conditions of the states (xo) and inputs (uo) for the nine LTI plants.
142
Appendix B
B1 Partition of the Plant
⎧⎪⎛ x& ⎞ ⎛ a b ⎞⎛ x ⎞
G = ⎨⎜⎜ ⎟⎟ = ⎜⎜ ⎟⎜ ⎟
⎪⎩⎝ y ⎠ ⎝ c d ⎟⎠⎜⎝ u ⎟⎠
Augmented plant
⎛ Z 2 ( s) ⎞ ⎛W ( s ) ⎞ ⎛ P11( 2) ( s ) P12( 2) ( s ) ⎞
⎜⎜ ⎟⎟ = Paug ⎜⎜ ⎟⎟ , where Paug ( 2) = ⎜⎜ ⎟.
⎟
⎝ Y (s) ⎠ ⎝ U (s) ⎠ ⎝ P21( 2) ( s ) P22( 2) ( s ) ⎠
where
⎧⎪⎛ x& ⎞ ⎛ a p b1 p ⎞⎛ x ⎞ ⎧⎪⎛ x& ⎞ ⎛ a b2 p ⎞⎛ x ⎞
P21( ∞ ) ( s ) = ⎨⎜⎜ ⎟⎟ =⎜⎜ ⎟⎜ ⎟ ; P12( ∞ ) ( s ) = ⎨⎜⎜ ⎟⎟ =⎜⎜ p ⎟⎜ ⎟
⎪⎩⎝ y ⎠ ⎝ c y d y1 ⎟⎠⎜⎝ u ⎟⎠ ⎪⎩⎝ z ⎠ ⎝ c i d i 2 ⎟⎠⎜⎝ u ⎟⎠
B2 Invariant zeros
Definition (Zhou et al. 1996): A complex number zo ∈C is called an invariant zero of the
system realization (A,B,C,D) if it satisfies:
⎡ A − zo I B⎤ ⎡ A − sI B⎤
rank ⎢ ⎥ < normalrank ⎢ .
⎣ C D⎦ ⎣ C D ⎥⎦
143
B3 Numerical analysis of the LTI plant corresponding to 80% of power
Plant
⎛ − 2.6229 − 1.594 − 0.91842 − 3.3568 1.3904 ⎞
⎜ ⎟
⎜ − 0.43951 − 2.5555 − 0.62366 10.728 0.79366 ⎟
⎛a b ⎞ ⎜
G = ⎜⎜ ⎟⎟ = − 2.1691 6.0409 − 2.7649 − 9.8496 0.36931 ⎟
⎝ c d ⎜
⎠ ⎜ − 1.4496 ⎟
1.9607 0.39378 0 0 ⎟
⎜ ⎟
⎝ 2.3984 − 2.4476 1.9979 0 0 ⎠
Poles: Zeros:
-2.4039 -11.626
-2.7697 + 1.0526i
-2.7697 - 1.0526i
Poles: Zeros:
-3e-006 -2.4039
-3e-006 -2.7697 + 1.0526i
-0.0001 -2.7697 - 1.0526i
-0.0001
-2.7697 + 1.0526i
-2.7697 - 1.0526i
-2.4039
⎡a − z o I b2 p ⎤
rank ⎢ p =7
⎣ ci d i 2 ⎥⎦
⎡a − sI d2p ⎤
normalrank ⎢ p =9
⎣ ci d i 2 ⎥⎦
144
Poles of the controller number 1 (tables 5.1)
-2.2371e+005
-32693
-8016.6
-9513.6
-7.7902
-3.5378e-006
-3.0161e-006
145
Poles of the controller 1 with beta constraint (table 5.2):
-3.4563 + 9.3871i
-3.4563 - 9.3871i
-3.1819 + 4.2961i
-3.1819 - 4.2961i
-5.1852
-2.4958e-006
-1.8223e-006
146
References:
Apkarian, P., Noll, D. and Tuan, H. D. (2003). Fixed-order H-Infinity Control Design via
A Partially Aumented Lagrangian Method. IFAC Control Applications of
Optimisation, Hungary: 69-74.
147
Balakrishnan, V. and Kashyap, R. L. (1999). Robust stability and performance analysis of
uncertain systems using linear matrix inequalities. Journal of Optimization Theory
and Applications 100(3): 457-478.
Balas, G. J., Ryan, J., Shin, J. and Garrard, W. L. (1998). A new technique for design of
controllers for Turbofan Engines. 34th Joint Propulsion Conference, Cleveland,
OH, American Institute of Aeronautics and Astronautics.
Barratt, C. and Boyd, S. (1989). Example of exact trade-offs in linear controller design.
IEEE Control Systems Magazine 9(1): 46-52.
148
Biannic, J. M. (2004). IQC for Robustness Analysis of Fighter Aircraft Control Laws.
Workshop on Linear Matrix Inequalities in Control,, Toulouse, France, LAAS-
CNRS.
Biannic, J. M. and Apkarian, P. (1999). Missile autopilot design via a modified LPV
synthesis technique. Aerospace Science and Technology 3(3): 153-160.
Boyd, S., Balakrishnan, V., Barratt, C., Khraishi, N.,X. Li, D. M. and Norman, S. (1988).
A new CAD method and associated architectures for linear controllers. IEEE
Transactions on Automatic Control 33(3): 268 - 283.
Boyd, S., Balakrishnan, V., Feron, E. and Ghaoui, L. E. (1993). Control system analysis
and synthesis via linear matrix inequalities. Proceeding of the American Control
Conference, San Francisco, USA: 2147-2154.
Boyd, S., Crusius, C. and Hansson, A. (1997). Control applications of nonlinear convex
programming. Journal of Process Control 8(5-6): 313-324.
Boyd, S., Feron, E., Balakrishnan, V. and Ghaoui, L. E. (1994a). History of linear matrix
inequalities in control theory. Proceedings of the American Control Conference 1:
31-34.
149
Boyd, S., Ghaoui, L. E., Feron, E. and Balakrishnan, V. (1994b). Linear Matrix
Inequalities in System and Control Theory, Studies in Applied Mathematics,
SIAM.
Burl, J. B. (1999). Linear Optimal Control H-2 and H-Infinity Methods, Addison-Wesley.
Caruana, R. A., Eshelman, L. A. and Schaffer, J. D. (1989). Representing and hidden bias
II: Eliminating defining length bias in genetic search via shuffle crossover.
Proceedings of the Eleventh International Joint Conference on Artificial
Intelligence: 750-755 Los Altos, Cal.:William Kaufmann, Inc.
Chen, B. S., Cheng, Y. M. and Lee, C. H. (1995). A Genetic Approach to Mixed H-2/H-
Infinity Optimal PID Control. IEEE Control Systems Magazine 15(5): 51-60.
Chilali, M. and Gahinet, P. (1996). H-Infinity design with pole placement constraints: An
LMI approach. IEEE Transactions on Automatic Control 41(3): 358-367.
Chilali, M., Gahinet, P. and Apkarian, P. (1999). Robust pole placement in LMI regions.
IEEE Transactions on Automatic Control 44(12): 2257-2270.
150
Chipperfield, A. and Fleming, P. J. (1996). Multiobjective gas turbine engine controller
design using genetic algorithms. IEEE Transactions on Industrial Electronics
43(5): 583-587.
Darwin, C. (1859). The origin of species ;Charles Darwin ; edited with an introduction
and notes by Gillian Beer, Oxford University Press,1998.
151
Deb, K. (2001). Multi-Objective Optimization using Evolutionary Algorithms, John
Wiley & Sons, Chichester, UK.
Deb, K. and Agrawal, R. B. (1995). Simulated Binary Crossover for Continuous Search
Space. Complex Systems 9: 115--148.
Deb, K. and Goyal, M. (1996). A combined Genetic Adaptive Search (GeneAS) for
Engineeriong Design. Computer Science and Informatics 25(4): 30-45.
DeHoff, R. L. and W. Eral Hall, J. (1979). Optimal Control of Turbine Engines. Journal
of Dynamic Systems, Measurement, and Control 101: 117-126.
Dettori, M. (2001). LMI techniques for control with application to a Compact Disc player
mechanism. Dutch Institute of Systems and control. Delft, The Netherlands,
Technische Universiteit Delft: 248.
Dettori, M. and Scherer, C. W. (2002). MIMO control design for a compact disc player
with multiple norm specifications. IEEE Transactions on Control Systems
Technology 10(5): 635-645.
Doyle, J., Zhou, K. M., Glover, K. and Bodenheimer, B. (1994). Mixed H-2 and H-
Infinity Performance-Objectives .2. Optimal- Control. IEEE Transactions on
Automatic Control 39(8): 1575-1587.
Doyle, J. C., Francis, B. A. and R., T. A. (1992). Feedback Control Theory. New York,
MacMillan.
152
Dullerud, G. E. and Paganini, F. (2000). A Course in Robust Control Theory: A Convex
Approach, Springer-Verlag New York.
Elia, N. and Dahleh, M. A. (1997). Controller design with multiple objectives. IEEE
Transactions on Automatic Control 42(5): 596-613.
Feron, E., Balakrishnan, V., Boyd, S. and Ghaoui, L. E. (1992). Numerical methods for
H-2 related problems. Proceedings of American Control Conference: 2921--2922.
153
Fonseca, C. M. and Fleming, P. J. (1998a). Multiobjective optimization and multiple
constraint handling with evolutionary algorithms. I. A unified formulation. IEEE
Transactions on Systems, Man and Cybernetics-Part A: systems and humans
28(1): 26-37.
Forsgren, A., Gill, P. E. and Wright, M. H. (2002). Interior Methods for Nonlinear
Optimization. SIAM Review 44(4): 525-597.
Frederick, D. K., Garg, S. and Adibhatla, S. (2000). Turbofan engine control design using
robust multivariable control technologies. IEEE Transactions on Control Systems
Technology 8(6): 961-970.
Fujisaki, Y., Dabbene, F. and Tempo, R. (2003). Probabilistic design of LPV control
systems. Automatica 39(8): 1323-1337.
Gahinet, P. and Nemirovski, A. (1997). The projective method for solving linear matrix
inequalities. Mathematical Programming 77(2): 163-190.
Gahinet, P.,Nemirovski, A.,Laub, A. J. and Chilali, M. (1995). LMI Control Toolbox for
use with MATLAB. Natick, MA, The MathWorks, Inc.
154
Ghaoui, L. E. and Silviu-lulian, N. (1999). Advances in linear matrix inequality methods
in control, Society for Industrial and Applied Mathematics SIAM.
Glover, K. and Doyle, J. C. (1988). State-space formulae for all stabilizing controllers
that satisfy an H-Infinity-norm bound and relations to relations to risk sensitivity.
Systems & Control Letters 11(3): 167-172.
Henrion, D. (2000). LMIs for Linear Systems Control by Polynomial Methods. IFAC
Symposium on Robust Control Design, Prague, Czech Republic.
155
Henrion, D. and Lasserre, J.-B. (2004). Solving nonconvex optimization problems.
Control Systems Magazine, IEEE 24(3): 72-83.
Henrion, D., Sebek, M. and Kučera, V. (2003). Positive polynomials and robust
stabilization with fixed-order controllers. IEEE Transactions on Automatic
Control 48(7): 1178-1186.
Herreros, A., Baeyens, E. and Peran, J. R. (2002). MRCD: a genetic algorithm for
multiobjective robust control design. Engineering Applications of Artificial
Intelligence 15(3-4): 285-301.
Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor, The
University of Michigan Press.
Iwasaki, T. and Skelton, R. E. (1994a). All Controllers for the General H-Infinity Control
Problem - LMI Existence Conditions and State-Space Formulas. Automatica
30(8): 1307-1317.
156
Iwasaki, T. and Skelton, R. E. (1994b). A Unified Approach to Fixed Order Controller
Design Via Linear Matrix Inequalities. Proc. American Control Conference: 35-
39.
Kalman, R. E. (1963). Lyapunov function for the problem of Lur'e in automatic control.
Proceedings of the National Academy of Science of the United States of America
49: 201-205.
157
Khargonekar, P. P. and Rotea, M. A. (1991c). Controller Synthesis for Multiobjective
Optimal Control. Control of Uncertain Dynamical Systems. S. P. Battacharyya
and L. H. Keel. Boca Raton, CRC Press: 261-279.
Kocvara, M. and Stingl, M. (2003). PENNON—a code for convex nonlinear and
semidefinite programming. Optimization Methods and Software 8(3): 317-333.
Kučera, V. (1975). Stability of Discrete Linear Feedback Systems. Proc. IFAC World
Congress, Boston, MA.
158
Leith, D. J. and Leithead, W. E. (1998b). Gain-scheduled and nonlinear systems:
dynamic analysis by velocity-based linearization families. International Journal
of Control 70(2): 289 - 317.
Lu, W. and Doyle, J. (1992). H-Infinity Control of LFT systems: An LMI Approach.
Proceedings of the 31st Conference on Decision and Control, Tucson, Arizona.
Lur'e, A. I. (1961). Anatolii Isakovich Lur'e on his sixtieth birthday. Journal of Applied
Mathematics and Mechanics. 25: 887-897.
Lyapunov, A. M. (1892). The general problem of the stability of motion, London : Taylor
& Francis:Translation by A. T. Fuller from the Russian, 1992.
Mei, T. X. and Goodall, R. M. (2000). LQG and GA solutions for active steering of
railway vehicles. IEE Proceedings - Control Theory and Applications 147(1):
111-117.
Michalewicz, Z. and Fogel, D. B. (2000). How to Solve it: Modern Heuristics. Berlin,
Springer.
159
Middleton, R. H. (1991). Trade-offs in linear control system design. Automatica 27(2):
281-292.
Owens, D. H. and Chotai, A. (1981). Norm Bounds for Linear Systems. Sheffield,
Department of Control Engineering, University of Sheffield: Research Report No.
159.
160
Packard, A. (1994). Gain scheduling via linear fractional transformations. Systems &
Control Letters 22(2): 79-92.
Packard, A. and Doyle, J. (1993). The complex structured singular value. Automatica
29(1): 71-109.
Paganini, F. (1999b). Frequency domain conditions for robust H-2 performance. IEEE
Transactions on Automatic Control 44(1): 38-49.
PolyX, L. (2001). The Polynomial Toolbox for Matlab. Prague, Czech Republic.
161
Popov, V. M. (1962). Absolute stability of nonlinear systems of automatic control.
Automation and Remote Control 22(8): 857-875.
Postlethwaite, I., Samar, R., Choi, B. and Gu, D. W. (1995). A Digital Multi-Mode H-
Infinity Controller for Spey Turbofan Engine. Proceedings of the 3rd European
Control Conference, Rome, Italy.
Qiu, W. Z., Vittal, V. and Khammash, M. (2004). Decentralized power system stabilizer
design using linear parameter varying approach. IEEE Transactions on Power
Systems 19(4): 1951-1960.
Rugh, W. J. (1991). Analytical framework for gain scheduling. IEEE Control Systems
Magazine 11(1): 79-84.
162
Scherer, C. (1990). The Riccati Inequality and State-Space H-Infinity Optimal Control.
Wurzburg, Germany, The University of Wurzburg: 267.
Scherer, C. and Weiland, S. (1999). Lecture Notes DISC Course on Linear Matrix
Inequalities in Control. Delft, The Netherlands, Dutch Institute for Systems and
Control: 205.
Scherer, C. W. (2003). when are multipliers exact? In 4th IFAC Symposium on Robust
Control Design, Milano, Italy.
163
Schroder, P. (1998). Multivariable Control of Active Magnetic Bearings. Department of
Automatic Control and Systems Engineering. PhD thesis, The University of
Sheffield, Sheffield UK.
Shamma, J. S. and Athans, M. (1990). Analysis of Gain Scheduled Control for Nonlinear
Plants. IEEE Transactions on Automatic Control 35(8): 898-907.
Shamma, J. S. and Xiong, D. P. (1999). Set-valued methods for linear parameter varying
systems. Automatica 35(6): 1081-1089.
Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis, Chapman
and Hall.
Skelton, R. E. and Iwasaki, T. (1994). All Controllers for the general H-Infinity control
problem - LMI existence conditions and state-space formulas. Automatica 30(8):
1307-1317.
164
Skelton, R. E. and Iwasaki, T. (1995). Increased Roles of Linear Algebra in Control
Education. IEEE Control Systems Magazine 15(4): 76-90.
Spang, H. A. and Brown, H. (1999). Control of jet engines. Control Engineering Practice
7(9): 1043-1059.
Sturm, J. F. (1999). Using SeDuMi 1.02, a Matlab toolbox for optimization over
symmetric cones. Optimization Methods and Software 11-12: 625-653.
Sun, X.-D. and Postlethwaite, I. (1998). Affine LPV modelling and its use in gain-
scheduled helicopter control. Proceedings of the UKACC International
Conference on Control '98 (Conf. Publ. No. 455 IEE), Swansea, UK: 1504-1509.
165
Takahashi, R. H. C., Palhares, R. M., Dutra, D. l. A. and Gonçalves, L. P. S. (2004).
Estimation of Pareto sets in the mixed control problem. International Journal of
Systems Science 35(1): 55-67.
Vandenberghe, L. (2004). Efficient SDP algorithms for applications in control and signal.
Proceedings of the Symposium on Mathematical Theory of Networks and Systems
(MTNS), Leuven, Belgium.
Whidborne, J. F. and Liu, G. P. (1993). Critical Control Systems: Theory, Design and
Applications. Taunton, U.K., Research Studies Press.
166
Willems, J. C. (1971). Least Squares Stationary Optimal Control and the Algebraic
Riccati Equation. IEEE Transactions on Automatic Control AC-16(6): 621-634.
Wu, F., Packard, A. and Balas, G. J. (1995). LPV Control Design for Pitch-Axis Missile
Autopilots. Proceedings of the 34th IEEE Conference on Decision and Control,
New Orelans, LA: 188-193.
Xin, X. (2004). Reduced-order controllers for the H-Infinity control problem with
unstable invariant zeros. Automatica 40(2): 319-326.
Yedavalli, R. K. and Liu, Y. (1995). H-Infinity Control with regional stability constraints.
Automatica 31(4): 611-615.
167
Youla, D. C., Bongiorno, J. J. and Jabr, H. A. (1976a). Modern Wiener-Hopf design of
optimal controllers Part I: The Single-Input-Output Case. IEEE Transactions on
Automatic Control AC-21: 3-13.
Zakain, V. and Al-Naib, U. (1973). Design of dynamical and control systems by the
method of inequalities. Proceedings of the Institution of Electrical Engineers 120:
1421-1427.
Zhou, K., Doyle, J. C. and Glover, K. (1996). Robust and Optimal Control, Englewood
Cliffs, NJ: Prentice Hall.
Zhou, K. M., Glover, K., Bodenheimer, B. and Doyle, J. (1994). Mixed H-2 and H-
Infinity Performance-Objectives .1. Robust Performance Analysis. IEEE
Transactions on Automatic Control 39(8): 1564-1574.
168