You are on page 1of 17

MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC

DISTANCE APPROACH
A. M. JUBRIL AND L. O. KEHINDE
Abstract. This paper considers the design of static state feedback controller
that stabilizes a linear system while minimizing a multiple performance objec-
tives.
The vector optimization problem is reduced to an equivalent quadratic
distance problem by using the method of Pontryagin. The resulting problem
is then solved as minimizing a square norm subject to a linear matrix inequality
(LMI).
The method is applied to linear quadratic regulator problem
1. Introduction
Most natural decision and control problems are multi-objective. Finding the con-
trol or decision variable that simultaneously satisfy a set of objectives is a classical
and fundamental problem, especially when the objectives are conicting.
The basic goal of multi-criterion optimization is to identify the specication that
are on the boundary between achievable and unachievable specications or design.
This generally leads to Pareto-Optimal design.
Many techniques have been proposed and developed for handling such problems.
One of these is the extension of the well developed techniques of optimization of a
single objective functional.
Previous work relating to this subject was considered by Nelson(1964), where
several inequality functionals
i
are considered,

i

i
, i = 0, . . . , N
where
i
are constants. The method sets to minimize one of the functional, say
k
,
subject to the constraint

i

i
; i = 0, . . . , N, i ,= k
The approach proposed by Waltz(1967), called hierarchical optimization crite-
ria, also ranked the cost functionals in order of importance as
1
,
2
,. . . ,
n
. The
optimization is then carried out sequentially with respect to these functionals ac-
cording to their ranking. Another important approach is the use of vectored-valued
cost functional suggested by Zadeh(1963) which was extended by Chang (1966) and
Salama and Gourishanker(1971). This approach reduces the multiple cost function-
als to a single objective functional which is a linear combination of the components
of cost functionals. Various methods of combining cost functionals have been pro-
posed(Boyd and Barrat, 1991).
Key words and phrases. multi-objective optimal control, quadratic distance problem, linear
matrix inequality.
1
2 A. M. JUBRIL AND L. O. KEHINDE
A considerable work has been done in scalarization of vector objectives. The
common scalarization method is the linear combination of the component of the
vector objectives in the form (Zadeh (1963);Makila(1989))
(1.1) (x, ) =
q

i=1

i
(x)
Another similar method is called the min-max. This forms a single objective
functional
obj
as the maximum of the weighted cost functionals (Boyd and Barrat,
1991),
(1.2)
obj
= max
1

1
, . . . ,
n

where
i
are non-negative weight factors.
In this paper, the modication of the weighted linear combination of the com-
ponent functional will be considered. It has been shown in (Khargoneka & Rotea,
1991) that the squaring each of the component objectives of the vector objectives
does not change the Pareto optimality, therefore an auxiliary cost function,
aux
,
which is a linear combination of the square of the component objectives will be
investigated.
(1.3)
aux
(x, ) =
q

i=1

2
i
(x)
2. Single Objective Optimal Control
Since the scalarization reduces the problem to the general single objective opti-
mization problem, a brief review of the basic optimal control problem. The basic
optimization problem is that of nding a minimizer of the function denes on a
curved polyhedron described by the intersection of the equality and inequality con-
straint sets

i
and
j
respectively. This problem is a special case of the more
general abstract problem of mathematical programming which is posed as
Problem 2.1. Find the minimum and the minimizers of the function f(x), x R
n
in the domain dened by the intersection of
1
, . . . ,
k
,
i
R
n
, (i = 1, . . . , k)
If it is assumed that there exist at least a solution x

to the problem (more than


one solution may exist). The solution space may then be divided into two: the set
(2.1)
o
= x[f(x) < f(x

) x

and the set


(2.2)
c
o
= x[f(x) > f(x

)
The solution to this problem is given by the following theorem.
Theorem 2.1. Let x

be a point in the set


o
. The point x

is a minimizer of the
real-valued function f(x) dened on if and only if the intersection
o
only
consists of the point x

(2.3)
o
=

1
. . .

q
= x

Proof. (See Boltyanski et al,1999)[Theorem 2.2 pp16]


MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 3
2.1. Separation Principle. The following denition is considered necessary.
Denition 2.1. Two sets X
1
, X
2
R
n
are said to be separated if there is a
hyperplane = x : a, x) = b in R
n
such that
X
1
x : a, x) b
X
2
x : a, x) b
where the hyperplane is called the separating hyperplane and vector a is the
normal vector to the hyperplane
The graph and epigraph, or supergraph of the function f(x), are written respec-
tively, as
grf = (x, f(x)) : x K
and
epif = (x, r) / R : r f(x)
consisting of all points x, r R
n+1
with r f(x), x R
n
.
Consider the graph x = (x, f(x)) of f(x); using the rst-order approximation,
(2.4) f(x) = f(x

) +f(x

), x x

)
then there exist an outward vector v such that v, x x

) 0. If x

= (x

, f( x

))
and x = (x, f(x)), then
(2.5) x x

= (x x

, f(x) f(x

))
and combining the last equation with 2.4, we have
(2.6) (f(x

), 1), (x x

, f(x) f(x

))) 0
which shows that the vector v = (f(x

), 1) is the outward normal to the graph


at x

Lemma 2.2 (Farkas Lemma). If a


1
, . . . , a
k
and b are nite set of vectors in /,
then
b, x) 0 for all x / satisfying
a
i
, x) 0 i = 1, . . . , k
if and only if
b =
k

i=1

i
a
i
with
i
0 for i = 1, . . . , k
Since to every convex set , an approximating convex cone K can be formed
in which the original set is contained in the neighbourhood of the point of
consideration, the separating property of (convex) set is therefore given in terms of
the convex cones.
Denition 2.2. A system of (convex) cones with a common apex x

is separable,
if there exist a hyperplane R
n
through x

that separates one of the cones from


the intersection of the others
The above denition leads to the next theorem.
4 A. M. JUBRIL AND L. O. KEHINDE
Theorem 2.3. Let K
o
, K
1
, . . . , K
t
be a system of convex cones in R
n
with a com-
mon vertex at x

. For separability of the cones, it is necessary and sucient that


there exist vectors a
i
K

, i = 0, 1, . . . , t at least one of which is distinct from zero


such that
a
o
+a
1
+. . . +a
t
= 0
Proof. By Farkas lemma, assume t = 2, then there exists vectors a
1
, a
2
and b; and

1
,
2
0 such that
b, x) 0 a
1
, x) 0; a
2
, x) 0
then b =
1
a
1
+
2
a
2
; setting b = a
o
, and
1
= 1,
2
= 1 then
a
o
+a
1
+a
2
= 0

2.2. Pontryagin Maximum Principle. Consider the basic optimal control prob-
lem. Given a dynamic system
(2.7) x = f(x, u) , x R
n
, u | R
m
Assume the initial state x(t
o
) = x
o
R
n
and the terminal set M R
n
are given.
For an admissible control u(t) |, t
o
t t
f
, the corresponding trajectory is
x(t). The optimal control problem is posed as
Problem 2.2 (Optimal Control Problem). Find an admissible control u that takes
the state x
o
to a point on the terminal set x(t
f
) M such that a given real-valued
function (x, u) is minimized.
The control u(t) which yields this solution is called an optimal control corre-
sponding to a transition from x
o
to x(t
f
) and the corresponding trajectory x(t) is
called an optimal trajectory.
Denition 2.3. For a piece-wise continuous u(t) : [t
0
, t
f
] R
m
, for any t
0
t
1

t
2
t
f
(2.8) x(t
2
) = (t
2
, t
1
, x(t
1
), u)
denotes the state of the solution of the dynamic equation x = f(x, u), at time t
2
,
if it is in state x(t
1
) at time t
1
when the control u is applied
Denition 2.4. For an initial point x
0
S
0
, the set of states reachable at time
t
1
, starting at t
0
in state x
0
= x(t
0
) and using admissible control is called the
attainable or reachable set, /(t
1
, t
0
, x(t
0
))
(2.9) /(t
1
, t
0
, x(t
0
)) = x(t
1
, t
0
, x(t
0
), u)[u |
Denition 2.5. Let (t, ) be transition matrix function, i.e. the solution of the
homogeneous part of the dynamic equation of the optimization problem
(2.10)

t
(t, ) = A(t)(t, )
with the boundary condition (t, ) = I
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 5
Denition 2.6. The Hamiltonian is expressed as the inner product of the co-state
or the momentum vector, p, and the coordinate vector, v
(2.11) H(x, p, u) = p, f(x, u)) =
n

i=0
p
i
f
i
(x, u)
, ) is the standard inner product.
The optimal control problem was reformulated by Pontryagin et al(1963) by
introducing an additional state variable, x
n+1
,
(2.12) x
n+1
=
_
t
f
to
f(x, u)dt ; x R
n
, u R
m
or,
x
n+1
= f(x, u)
The augmented state vector x is given by
x = (x, x
n+1
) R
n+1
and the augmented system dynamics becomes
(2.13)

x = f( x, u)
and initial state x(t
o
) = (x
o
, 0).
It is obvious that 2.13 admits the solution x = (x(t), x
n+1
(t)) where
x
n+1
(t) =
_
t
to
f(x(), u())d
and at the nal time t = t
f
,x(t
f
) = x
1
and
x
n+1
(t
f
) =
_
t
f
to
f(x(), u())d =
The solution of the augmented system passes through the initial point (x
o
, 0)and
the terminal point x = (x
1
, ) at t
f
.
In the reformulation, a line , parallel to the x
n+1
axis and passes through the
point (x
1
, 0) was dened, and the solution of the augmented dierential equation
equation can be seen as an intersection of the trajectory of x(t) and the line .
With the above explanation of the problem, the optimal control problem can be
reformulated as:
Problem 2.3. Find the intersection of the cone of the set
o
= x
n+1
[x
n+1
<
x

n+1

n+1
and the cone of the reachable set = x(t
1
, t
0
, x(t
0
), u)[u |
with x

n+1
as the vertex
The theorem below gives necessary condition of optimality, which is also called
the Maximum Principle
Theorem 2.4 (Pontryagin Maximum Principle(PMP)). Let u(t), t
0
t t
1
be
an admissible control such that the corresponding trajectory x(t) which begins at
the point x
0
, at t
0
passes at t
1
, through a point on the line parallel to the new
coordinate. In order that u(t) and x(t) be optimal, it is necessary that there exists a
non-zero continuous vector p(t) = (p
1
, . . . , p
n
) corresponding to u(t) and x(t) such
that
(1) for every t, t
0
t < t
1
, the function H(x, p, u) of the variable u | attains
maximum /(x, p) at the point u(t
1
), and
6 A. M. JUBRIL AND L. O. KEHINDE
(2) at the terminal time t
1
, the relation
(2.14) p
0
(t
1
) 0, /(p(t
1
), x(t
1
)) = 0
are satised
3. Vector-valued Optimal Control
In the basic (scalar) optimization problem, we have assumed a single objective
functional and also that the solution space is completely ordered. In vector optimal
control problem, a class of system, H
1
, assessed based on a performance objective
T
1
may be superior to another class of systems H
2
, and inferior to it based on
another performance T
2
. Therefore, the following denitions will be necessary.
Denition 3.1. A set X is completely ordered by a preference ordering (with
denoting better than or equal to) if for every pair x, y X, either x y or x < y
is true, with the relation being
(1) reective (x x)
(2) transitive (x y and y z implies x z)
(3) (x y and y x implies x = y)
A set X is partially ordered if it has the above properties except that not all pairs
of elements in X are comparable i.e. there may exist x and y such that neither
x y nor y x
We shall state the multi-objective optimal control problem as:
Problem 3.1. Find the control, u(t) or set of control that minimizes the set of
functionals
(3.1)
i
=
_
t
f
to
l
i
(x(), u())d; i = 1, . . . , p
of a xed dynamic system
(3.2) x(t) = f(x, u), x(t
o
) = x
o
x R
n
, u R
m
and the inequality constraint stated by the boundedness on the
control variable in the form
i
(u)
i
which describe the admissible controls |, or
| = u [
i

i
, i = 1, . . . , m
Handling multiple objectives optimization problem posed a greater challenge
because of the partial ordering of its solution space. This implies that, unlike the
scalar optimization problem where the solution space is divided into two: class
of systems superior to x

, and the class of systems inferior or equal to x

, the
vector-valued optimal control problem is divided into three classes (Zadeh, 1963)
Assuming a partial ordering on the solution space (feasible set) /, then / can
be divided into three disjoint subclasses with respect to a system x
(1) the subclass
o
or /

(x), of all systems inferior to or equal to x


(2) the subclass
1
or /
>
(x), of all systems better than or superior to H
(3) the subclass
2
or /
=
(x), of all systems which are not comparable to H
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 7
A system x

is said to be noninferior in / if the intersection of / and /


>
(x

) is
empty. In the same vein, the system x

is said to be optimal in / if / is contained


in /

(x

), which implies that every system in / is inferior to or equal to x

.
It is obvious that the determination of the set /
=
(x

) makes the vector-valued


optimization problem very dicult. However, if there exist a scalar-valued objective
function which completely orders the set /, then the subclass /
=
(x

) disappears,
and the vector optimization problem reduces to a scalar optimization. The above
observation has been well researched, which has led to the development of various
forms of scalarization techniques (Barrat and Boyd, 1994).
Consider a vector-valued objective function
(3.3) (x) = (
1
(x), . . . ,
p
(x))
where
i
(x), i = 1, . . . , p are the component scalar objective functionals. If x

is
optimal, then there exist, for every x

on the boundary, the intersection of /


>
(x

)
with the feasible set /. If x

is a member of the optimal solution set, then for any


feasible point x
(3.4)
i
(x) =
i
(x

) +
i
(x

)(x x

) , i = 1, . . . , p
and the class /
>
(x

) is the dual polar cone of the cone spanned by


i
(x

)
Also, since the feasible solution set is convex, then the set of all noninferior
points on the boundary of / is the Pareto optimal set x
i
of points through
which hyperplanes separating / and /
>
(x

) pass. Also let be the normal to the


separating hyperplane at each of the optimal solution on the boundary of /, which
belong to the dual polar cone spanned by
i
(x

)
(3.5) =
p

i=1

i
(x

) ,
i
0 , i = 1, . . . , p
which implies that the vector-valued function (x) can be reduced to a scalar equiv-
alent,
(3.6) (x, ) =
1

1
(x) +. . . +
p

p
(x)
which is a weighted combination of the component functional objectives
i
(x).
From equation 3.6, Zadeh(1963) has showed the equivalence of the weighted sum
objective to the original vector-valued objective function.
He also showed that the normal vector is spanned by the inward normals,
h
i
(x

), of the active constraint set, at the point x

on the boundary of the feasible


set / i.e.
(3.7) =
1
h
1
(x

) +. . . +
r
h
r
(x

) ,
i
0 , i = 1, . . . , r
Combining eqns. 3.5 and 3.7, gives
(3.8) (
1

1
(x

) +. . . +
p

p
(x

) +
1
h
1
(x

) +. . . +
r
h
r
(x

)) = 0
The equivalent relation in equation 3.8 leads to the Pontryagin Maximum Principle
for vector-valued optimal control problem(Torres, 2005)
Theorem 3.1 (Maximum Principle for MOP). If the pair (x

, u

) is a Pareto-
optimal solution of the probem 3.1, then there exist a continuos co-state function,
8 A. M. JUBRIL AND L. O. KEHINDE
R
n
, having piece-wise continuous derivatives, and constant multipliers R
p
,
where (, ) ,= 0, satisfying the pseudo-Hamiltonian system
(3.9)
x(t) =
H

(t, x(t), u(t), (t), )


(t) =
H
x
(t, x(t), u(t), (t), )
and the maximal condition
(3.10) H(t, x(t), u(t), (t), ) = max
uU
H(t, x(t), u, (t), )
and the Hamiltonian is dened by
(3.11) H(t, x(t), u(t), (t), ) = l(t, x(t), u(t)) +
Moreover,
i
0, i = 1, . . . , p; and H(t, x(t), u(t), (t), ) is a continuous function
of t and on each interval of continuity of u, is dierentiable and satises the equality
(3.12)
dH
dt
(t, x(t), u(t), (t), ) =
H
t
(t, x(t), u(t), (t), )
Sussman(1997) also gave a necessary condition of optimality in terms of the
separation principle as
Theorem 3.2. Given that (x

, u

) are noninferior pair, then


(3.13) (x, u) (x

, u

)
does not belong to the closed negative orthant from the Pareto set except (x, u) =
(x

, u

)
The formulation in Zadeh(1963) thus showed that optimization of the vector
objective is equivalent to that of an auxiliary scalar objective function,
aux
, which
is a weighted linear combination of the component objective functions i.e.
aux
=

p
i=1

i
4. Main Contribution
As observed earlier, the subclass /
=
(H) makes the vector-valued optimal control
tasking, and as suggested, a very good solution is that of nding a scalar-value
performance objective which completely orders the set. In view of this, the following
proposition is made.
Proposition 4.1. Consider a pvector objective function (x) = (
1
(x), . . . ,
p
(x)).
If each of the component objective function
i
(x) by
i
, then the vector objective
optimization problem is equivalent to the optimizing a scalar auxiliary objective
function
aux
=

p
i=1

2
i
Proof. For a single objective functional, p = 1, () =
2
1
(). Optimizing
2
1
() is
equivalent to
1
(), the proof is therefore trivial.
If p 2,
aux
=

p
i=1

2
i
. The canonical quadratic form representation of

aux
describes an ellipsoid.
For any given , the ellipsoid can, by nonsingular transformation = M, be
reduced to a ball,
aux
=

p
i=1

2
i
, where M = diag(

i
, . . . ,
_

p
)
The target set or the trade-o surface, S
1
, is dened by is described by the p
equations of
i
,
(4.1)
i
=
i
=
_
T
t0
l
i
(x(), u())d i = 1, . . . , p
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 9
Re-writing the equations as
(4.2) r
i
=
i

_
T
t0
l
i
(x(), u())d = 0 i = 1, . . . , p
and S
1
becomes
(4.3) S
1
= (x(T), )[r
i
() = 0, i = 1, . . . , p
and R() as
(4.4) R() = (r
1
(), . . . , r
p
())
Assumption 4.1. It is assumed that S
1
is a smooth dierentiable manifold, and
r
i
() are smooth functions
The tangent hyperplane at any point on S
1
is given by the components of
R()=(grad r
1
,. . ., grad r
p
). The tangent plane T
f
is given by the transversality
condition as
(4.5) T
f
=
_
[
r
i

) = 0, i = 1, . . . , p
_
multiplying each by a constant variable
i
, we have

r
i

) = 0, i = 1, . . . , p

i
r
i

) = 0, i = 1, . . . , p
(4.8)
p

k=1

i
r
i

) = 0
(4.9)
p

i=1

i
r
i

) = 0
and by transversality condition,
(4.10) ,

) =
p

i=1

i
r
i

) = 0
Then, we write as
(4.11) =
p

i=1

i
r
i

R
which is a normal vector to the surface S
1
.
Remark 4.1. By Farkas Lemma, is normal to the trade-o surface S
1
, and it
is a linear combination of the normal vectors to tangent plane of support of the
manifold S
1
dened by R() = 0
10 A. M. JUBRIL AND L. O. KEHINDE
Let K

be a convex cone approximating the ellipsoid (or the ball) at the point
of intersection of the ball and the trade-o surface S
1
(4.12) K

= [(

),

) 0
and K
S1
be the convex approximating cone for the trade-o surface at

(4.13) K
S1
= [R(

),

) 0
Consider minimizing an auxiliary functional = ||
2
subject to the equality
constraint R() = 0 in the space u |, then this is solved by minimizing the
Langrangian L(, )
(4.14) L(, ) = () +
T
R()
or,
(4.15) +
T
R = 0 and R() = 0
which can be written as
(4.16) =
T
R
Remark 4.2. Equation 4.16 implies that the normal vector to the ellipsoid at

i.e. (

), is spanned by the inward normal vectors of the trade-o surface


Also, according to theorem 2.3, the two set are separable and there exist a
separating hyperplane that passes through the point of intersection

, and the
separating hyperplane is given by the tangent plane of support to both set
(4.17) (

),

) =

T
R(

),

_
= 0
Given that
aux
=
T
S, S = diag() then
(4.18) (

) = 2
T

Thus showing that the normal vector, , is equivalent to the twice the linear
combination of the weight weighted by the value of the component functional at


4.1. Quadratic Distance Problem. It is clear form the proposition above that
minimizing the weighted sum of square of the component objectives is equivalent to
the minimizing the weighted sum of the component objectives, it is then convenient
to proceed to re-pose the problem as,
Problem 4.1. minimize
aux
= ||
2
S
, x R
n
such that
x = Ax +Bu x(0) = x
0

i
=
_
T
0
l
i
(x, u)dt i = 1, . . . , p
and
u |
The problem can now be viewed geometrically as the minimization of Euclidean
distance of the trade-o surface dened by
i
from the point (x(T),0). This problem
can be generalized by considering the canonical quadratic distance problem in which
each of the component functional is weighted equally, or by considering a linear
transformation = S.
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 11
Denition 4.1 (Chesi, 2000). The canonical quadratic distance problem(CQDP)
is of the form
minimize
aux
= ||
2
R
p
such that
w() =
m

i=0
w
2i
() = 0
where w() is an even polynomial of degree 2m and w
2i
() are homogeneous forms
of degree 2i of polynomial w()
The above problem is generalized by considering the problem of minimizing the
weighted Euclidean distance from a given point to the generic polynomial surface.
The generalized form of the problem is of the form,
(4.19) c = min | x
0
|
2
S
R
n
such that p() = 0.
It is obvious that this problem can be written in the form of a CQDP by a
suitable transformation
x = P
f
( x
0
) or, = P
1
f
x +x
0
where x R
n
, P
f
R
nn
is a non-singular matrix, such that P = P
f

P
f
.
The transformation above helps to handle cases of dierent weighting factors

i
of the component objective functionals
i
or
i
. For such cases, the objective
functional , can be written as
(4.20)
=

p
i=1

2
i
=
1

2
1
+
2

2
2
+. . . +
p

2
p
=
T
diag(
1
, . . . ,
p
)
=
T
_
_
_

1
0
.
.
.
.
.
.
.
.
.
0
p
_
_
_
=
T
S
Writing Q = M
T
M, we have = z
T
M
T
Mz. Therefore, the transformation =
M gives
(4.21) =
T
=
p

i=1

2
i
= ||
2
= ||
2
S
The solution now involves searching for the point of tangency of the ball B
c
=
[|| = c, x(T) S
1
or the ellipsoidal surface B
c
=
T
S = c, x(T) S
1

and the boundary of the trade-o surface S


1
. The ellipsoidal surface has diameter

i
along the
i
-axis
Looking at the objective functional in problem Eq(4.21), it is obvious that the
objective functional is not a norm of the state-space but a semi-norm. However, it
can be modied to a form of a CQDP, in the form
(4.22)
minimize

p
i=1
x
2
i
x R
n
1 p n
such that w(x) = 0 (w(x) is even)
12 A. M. JUBRIL AND L. O. KEHINDE
5. Multi-Objective LQR Problem
In this section, problem 4.1 will be investigated with the functionals
i
substi-
tuted with quadratic performance functions.
(5.1)
i
=
_

0
(x
T
Q
i
x +u
T
R
i
u)dt
The problem can then be posed as
Problem 5.1.
(5.2)
min = ||
2
x = Ax +Bu

= x
T
Q
i
x +u
T
R
i
u i = 1, . . . , q
Since = ||
2
, then

= 2
T
. Starting with the formulation of Hamiltonian
of the dynamic system,
(5.3)
H = p
T
(Ax +Bu) +

q
i=1
i
2
(x
T
Q
i
x +u
T
R
i
u) 1

q
i=1
i
2
(x
T
Q
i
x +u
T
R
i
u)
= p
T
(Ax +Bu) +
1
2

q
i=1
(
i

i
)(x
T
Q
i
x +u
T
R
i
u)
setting
i
=
i

i
(5.4) H = p
T
(Ax +Bu) +
q

i=1

i
2
(x
T
Q
i
x +u
T
R
i
u)
and also,
(5.5)
H
x
= p = A
T
p +
q

i=1

i
Q
i
x
(5.6)
H

i
=
i
=
1
2
_
x
T
Q
i
x +u
T
R
i
u
_
(5.7)
H
u
= 0 = B
T
p +
q

i=1

i
R
i
u
It could be observed from (5.6) that
i
=

i
or rather
i

i
= 0, which implies
(5.8)
_

i=1
(
i

i
)dt =
i

i
=
i
= constant
with the general assumption that
i
(t
f
) = 0,
i
=
i
.
Also, (5.7) gives
(5.9) u = (
q

i=1

i
R
i
)
1
B
T
p
Substituting this expression for u in (5.4)-(5.6) gives the Hamiltonian
(5.10)
x = Ax B
_
q

i=1

i
R
i
_
1
B
T
p
p =
_
q

i=1

i
Q
i
_
x A
T
p
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 13
or,
(5.11)
_
x
p
_
=
_

_
A B
_
q

i=1

i
R
i
_
1
B
T

i=1

i
Q
i
A
T
_

_
_
x
p
_
which leads to the next theorem,
Theorem 5.1. The Hamiltonian 5.11 admits a solution u = (
_
q

i=1

i
R
i
_
1
B
T
P)
where P is the solution of the Riccati Equation
(5.12)

P +PA+A
T
P +
q

i=1

i
Q
i
PB(
q

i=1

i
R
i
)
1
B
T
P = 0
Proof. Setting p = Px, then p =

Px + P x. Substituting these in equations 5.10,
gives the Riccati equation 5.12. Substituting p = Px in equation 5.10 gives the
required control law
5.1. LMI Constraint. In order to utilize the LMI tools, the dierential constraints
in equations 5.11 is converted to inequality constraints by considering the following
proposition.
Proposition 5.2. If x = Ax + Bu is closed-loop stabilizable, there exists a Lya-
punov function V () =
T

P where = (x, x
n+1
, . . . , x
n+q
)
T
and

P =
_
P 0
0 I
_
Proof. Given that x
n+i
(t
o
) = 0, is obvious that V () = 0 when = 0 and V () > 0
for ,= 0. It remains to show that

V () 0.
For simplicity of representation, let = (x
n+1
, . . . , x
n+q
), and write = (x, )
T
and V () = V (x, ) or,
V () =
_
x

_
T
_
P 0
0 I
__
x

_
Then,
(5.13)

V () =
_
x

_
T
_
P 0
0 I
__
x

_
+
_
x

_
T
_
P 0
0 I
__
x

_
By expansion, it is clear that

V () can be written as

V =

V
1
+

V
2
where
(5.14)

V
1
= x
T
Px +x
T
P x and

V
2
=

T
+
T

= 2

(5.15)

V = (Ax +Bu)
T
Px +x
T
P(Ax +Bu) + 2

= x
T
A
T
Px +u
T
B
T
Px +x
T
PAx +x
T
PBu + 2
q

i=1

i
(x
T
Q
i
x +u
T
R
i
u)
= x
T
(PA+A
T
P + 2
q

i=1

i
Q
i
)x +u
T
B
T
Px +x
T
PBu +u
T
(2
q

i=1

i
R
i
)u
14 A. M. JUBRIL AND L. O. KEHINDE
and in matrix form
(5.16)

V =
_
x
u
_
T
_

_
PA+A
T
P + 2
q

i=1

i
Q
i
PB
B
T
P 2
q

i=1

i
R
i
_

_
_
x
u
_
If x = Ax+Bu is closed-loop stabilizable, then there exist a controller u = Kx
and a Lyapunov matrix P such that
(5.17)

V
1
= x
T
(P(ABK) + (ABK)
T
P)x < 0
From the representation

i
=
_
t
f
to
(x
T
Q
i
x +u
T
R
i
u)dt =
_
t
f
to
x
T
(Q
i
+K
T
R
i
K)xdt 0

i
can be written as

i
= x
T
(Q
i
+K
T
R
i
K)x 0, i = 1, . . . , q
Since

and are positive denite, which implies that

T
0. It can therefore
be shown that there exists region of the achievable set for which

V
1

V
2
, with
the limit on the achievable set on the trade-o surface dened by

V
1
=

V
2
.
The problem reduces to that of minimizing
aux
=
2
1
+
2
2
given that

V () 0.
Writing V () = x
T
Px +x
T
P x +

T
+
T
and
(5.18)

V () = x
T
(P(ABK) + (ABK)
T
P)x + 2
T

with = (
1
, . . . ,
q
) and

i
= x
T
(Q
i
+K
T
R
i
K)x, i = 1, . . . , q
(5.19)

V () = x
T
_
P(ABK) + (ABK)
T
P + 2

q
i=1

i
(Q
i
+K
T
R
i
K)
_
x
= x
T
_
P(ABK) + (ABK)
T
P + 2

q
i=1

i
Q
i
+ 2

q
i=1

i
K
T
R
i
K
_
x
= x
T
_
P(ABK) + (ABK)
T
P + 2Q+ 2K
T
RK
_
x
where Q =

q
i=1

i
Q
i
and R =

q
i=1

i
R
i
This can be put in LMI framework by considering a semidenite relaxation
(5.20)
T
and reducing the problem to
(5.21)
minimize trace
such that
T
0
and P(ABK) + (ABK)
T
P + 2Q+ 2K
T
RK 0
Considering the last inequality,
P(ABK) + (ABK)
T
P + 2Q+ 2K
T
RK 0
let =

q
i=1

i
, multiply the rst part of the inequality by

q
i=1
i

(5.22)

q
i=1
i

_
P(ABK) + (ABK)
T
P

+ 2

q
i=1

i
(Q
i
+K
T
R
i
K)
=

q
i=1
_
i

(P(ABK) + (ABK)
T
P)
_
+ 2

q
i=1

i
(Q
i
+K
T
R
i
K)
=

q
i=1

i
_
1

(P(ABK) + (ABK)
T
P) + 2(Q
i
+K
T
R
i
K)

MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 15


Also, let = 1/ and c = (1, . . . , 1) = 1
q
, and then write c
T
= 1, > 0,
i
> 0;
or in the matrix form
(5.23)
_
1
1 c
T

_
0
and the inequality (5.22), from the converse of the o-procedure, can be written as
(5.24)
_
(P(ABK) + (ABK)
T
P) + 2Q
i
+ 2K
T
R
i
K
_
0 i = 1, . . . , q
Introducing change of variables P = Y
1
and K = XY
1
(Boyd et al,1994), (5.24)
becomes
(5.25)
_
(Y A+A
T
Y BX X
T
B
T
) + 2Y Q
i
Y + 2X
T
R
i
X
_
0 i = 1, . . . , q
The problem can now be cast in the LMI framework as
(5.26)
minimize trace
such that
_

T
1
_
> 0,
_
1
1 c
T

_
> 0
and
_

_
BX +X
T
B
T
Y A
T
AY Y Q
1/2
i
X
T
R
1/2
i
Q
1/2
i
Y

2
I 0
R
1/2
i
X 0

2
I
_

_ > 0 i = 1, . . . , q
5.2. Parametrization of the Pareto Set. We have solve the problem for equal
weighting factor
i
=
j
, i, j = 1, . . . , q. To generate the Pareto set, /, for the
controllers, we consider weighting the performances indices,
i
, by
i
. Thus param-
eterizing the solution set by = (
1
, . . . ,
q
)
T
,
(5.27)
=

q
i=1

2
i

2
i
=
2
1

2
1
+
2
2

2
2
+. . . +
2
q

2
q
= diag(
2
1
, . . . ,
2
q
)
=
T
G
Let G = S
T
S, so that =
T
S
T
S, and S = diag(
1
, . . . ,
q
). The functionals
expressed in the form

i
=

_
0
_
x
T
Q
i
x +u
T
R
i
u
_
dt (5.28)
(5.29)
Let Q
i
= Q
1/2
i
T
Q
1/2
i
and R
1/2
i
T
R
1/2
i
, then
(5.30)
i
=
_

0
_
x
T
Q
1/2
i
T
Q
1/2
i
x +u
T
R
1/2
i
T
R
1/2
i
u
_
dt
16 A. M. JUBRIL AND L. O. KEHINDE
and

i
=

_
0
_
x
T
Q
1/2
i
T

i
Q
1/2
i
x +u
T
R
1/2
i
T

i
R
1/2
i
u
_
dt (5.31)
=

_
0
_
x
T
(
_

i
Q
1/2
i
)
T
(
_

i
Q
1/2
i
)x +u
T
(
_

i
R
1/2
i
)
T
(
_

i
R
1/2
i
)u
_
dt (5.32)
Setting
Q
1/2
=
_
_
_
_
Q
1/2
1
.
.
.
Q
1/2
q
_
_
_
_
, R
1/2
=
_
_
_
_
R
1/2
1
.
.
.
R
1/2
q
_
_
_
_
M
q
=
_

1
I
nQ1
0 0
0

2
I
nQ2
0
.
.
.
.
.
.
.
.
. 0
0 0
_

q
I
nQq
_

_
M
r
=
_

1
I
nR1
0 0
0

2
I
nR2
0
.
.
.
.
.
.
.
.
. 0
0 0
_

q
I
nRq
_

_
where nRi and nQi are the dimensions of the R
i
and Q
i
matrices. With the above
representations, equation 5.26 becomes
(5.33)
minimize trace
such that
_
r
r
T
1
_
> 0,
_
1
1 c
T
r
_
> 0
and
_
_
BX

+X
T

B
T
Y

A
T
AY

Y
T

M
q
Q
1/2
X
T

M
r
R
1/2
M
q
Q
1/2
Y

2
I 0
M
r
R
1/2
X

0

2
I
_
_
> 0
where r = diag(
1
, . . . ,
1
) and the controller is given by K

= X

Y
1

6. Example
In this section, the proposed method is now applied to pure multi-objective Linear
Quadratic Regulator problem.
6.1. Pure Multi-Objective Control Problem. A simple but true multi-objective
LQR problem, borrowed from Hu et al (1996) is presented. The system is given by
(6.1) x =
_
0 1
0 0
_
x +
_
0
1
_
u, x(0) = x
0
and
(6.2)
i
(x, u) =
1
2
_

0
(x
T
Q
i
x +u
T
R
i
u)dt, i = 1, 2
where
Q
1
=
_
2 1
1 1
_
, Q
2
=
_
1 1
1 3
_
, R
1
= 2 and R
2
= 1
MULTI-OBJECTIVE CONTROL DESIGN: QUADRATIC DISTANCE APPROACH 17
The optimal controller and the optimal cost for each performance objective for the
system were determined using the standard [K P rts]=lqr(A,B,Q,R) function in
matlab. The multi-objective optimal controller using the method proposed in this
study is also determined (see table 6.1 for the results).
Table 1. Results of the Pure Multi-Objective Control
[hp] height Q
1
, R
1
Q
2
, R
2
MOP
Hu
MOP
QDP
K
_
1.4142 1.9566
_ _
1.000 2.0690
_ _
1.0000 2.2361
_ _
0.9517 1.6538
_
P
_
1.7671 1.4142
1.4142 1.9566
_ _
3.2361 1.0000
1.0000 2.2361
_
Not Avail.
_
2.1087 0.8977
0.8977 1.9255
_
rts
0.9783 +j0.6761
0.9783 j0.6761
0.6180
1.6180
0.6180
1.6181
0.8268 +j0.5176
0.8268 j0.5176

opt
40.1442 54.932 55.0208 39.1866
7. Conclusion
Multi-objective optimal control problem has been recast into a quadratic distance
problem which is reduced to a linear matrix problem for which there are tools which
can be used to solve the resulting problem eciently.
References
Boltyanski V., Martini H. & Soltan V. (1999), Geometric methods and Optimization Prob-
lems,Kluwer Academic Publishers, Netherland.
Boyd, S. & Barratt C. (1991):Linear Controller Design: Limits of Performance. Englewood Clis,
N. J., Prentice-Hall
Boyd S., Elghaoui L., Feron E. & Balakrishnan V. (1994), Linear Matrix Inequalities in Systems
and Control Theory, vol. 15 of SIAM Studies in Applied Mathematics, SIAM, Philadelphia.
Chesi G., Genesio R. & Tesi A.(2000), An LMI approach to a class of distance problem,Proc.
of 14th International Symposium on Mathematical Theory of Networks and Systems, Per-
pignan, France,2000
Chesi G.(2000b),New Convexication techniques and their application to the analysis of dynamic
systems and vision problems, Ph.D Thesis, University of Bologna, Padova, Siena.
Hu Z., Salcudean S. E. & Loewen P. D. (1996) Multiple Objective Control Problems Via Non-
Smooth Analysis,Proceedings of the Triennial World Conference of IFAC, San Francisco,
CA
Makila P. M., On Multiple Criteria Stationary Linear Quadratic Control,IEEE Trans. Autom.
Control, 34(12),pp 1311-1313
Nelson, W. L. (1964) On the use of Optimization Theory for Practical Control System Design
IEEE Autom. Control,9, pp469-477
Pontryagin L. S.,Boltyanski V., Gankrelidze R. & Mishchenko (1962) The Mathematical Theory
of Optimal Processes,Interscience Publishers, Inc., New York.
Salama, A. I. A. & Gourishankar, V. (1971) Optimal Control of Systems with a Single Control
and Several Cost Functionals , Internation Journal of Control, vol. 14, no. 4, pp 705-725
Sussmann H. J.(1998) Transversality conditions and a strong maximum principle for systems of
dierential inclusions,Proc. of the 35th IEEE Conference on Decision and Control,Tampa,
FL., pp1-6
Waltz, F.M. (1967) An Engineering Approach: Hierarchical Optimization Criteria, IEEE Trans.,
Vol. AC-12, pp 179-180, April, 1967.
Zadeh L.A.(1963) Optimality and NonScalar Valued Performance Criteria, IEEE Transactions
on Automatic Control, AC8 pp 59-60
Department of Electronic and Electrical Engineering, Obafemi Awolowo Univer-
sity, Ile-Ife

You might also like