Some Special Cases of Optimizing over
the Efficient Set of a Generalized Convex
Multiobjective Programming Problem
Tran Ngoc Thang
Tran Thi Hue
School of Applied Mathematics and Informatics
Faculty of Management Information System
Hanoi University of Science and Technology
The Banking Academy of Vietnam
Email:
Email:
Abstract—Optimizing over the efficient set is a very hard
fi , i = 1, . . . , m are linear (resp. convex), called a
and interesting task in global optimization, in which local
linear multiobjective programming problem (resp. a con-
optima are in general different from global optima. At the
vex multiobjective programming problem), have received
same time, this problem has some important applications
in finance, economics, engineering, and other fields. In this
special attention in the literature (see the survey in [21]
article, we investigate some special cases of optimization
and references therein). However, to the best of our
problems over the efficient set of a generalized convex
knowledge, there is little result about numerical methods
multiobjective programming problem. Preliminary com-
in the nonconvex case where fi , i = 1, . . . , m, are
putational experiments are reported and show that the
nonconvex (see [4], [16],...).
proposed algorithms can work well.
The main problem in this paper is formulated as
AMS Subject Classification: 90 C29; 90 C26
min Φ(x) s.t. x ∈ XE ,
(PX )
Keywords: Global optimization, Efficient set, Generalized convexity, Multiobjective programming problem
where Φ : X → R is a continuous function and XE is
the efficient solution set for problem (GM OP ), i.e.
I. I NTRODUCTION
XE = {x0 ∈ X | ∃x ∈ X : f (x0 ) ≥ f (x), f (x0 ) = f (x)}.
The generalized convex multiobjective programming
problem (GM OP ) is given as follows
Minf (x) = (f1 (x), . . . , fm (x))T s.t. x ∈ X ,
As usual, the notation y 1 ≥ y 2 , where y 1 , y 2 ∈ Rm , is
used to indicate yi1 ≥ yi2 for all i = 1, . . . , m.
It is well-known that, in general, the set XE is non-
where X ⊂ Rn is a nonempty convex compact set
convex and given implicitly as the form of a standard
and fi , i = 1, . . . , m, are generalized convex functions
mathematical programming problem, even in the case
on X . In the case m = 2, problem (GM OP ) is
m = 2, the objective functions f1 , f2 are linear and
called a generalized convex biobjective programming
the feasible set X is polyhedral. Hence, problem (PX )
problem. The special cases of problem (GM OP ), where
is a global optimization problem and belongs to NP-
hard problem class. This problem has many applications
in economics, finance, engineering, and other fields.
Recently this problem has a great deal of attention
In the case h is differentiable, if h is quasiconvex on
X , we have
h(x1 ) − h(x2 ) ≤ 0 ⇒ ∇h(x2 ), x1 − x2 ≤ 0
(1)
from researchers (for instance, see [1], [5], [6], [7],
[8], [12], [15], [16] [17], [19] and references therein).
Like problem (GM OP ), there is only few numerical
for all x1 , x2 ∈ X , where ∇h(x2 ) is the gradient vector
of h at x2 (see Theorem 9.1.4 in [10]).
algorithms to solve problem (PX ) in the nonconvex
A vector function f (x) = (f1 (x), . . . , fm (x))T is
case (see [1], [16],...). In this article, simple convex
called convex (resp., pseudoconvex, quasiconvex) on X if
programming procedures are proposed for solving three
its component functions fi (x), i = 1, . . . , m, are convex
special cases of problem (PX ) where XE is the efficient
solution set for problem (GM OP ) in the nonconvex
(resp., pseudoconvex, quasiconvex) functions on X .
Recall that a vector function f is called scalarly
m
i=1
case. These special-case procedures require quite little
pseudoconvex on X if
computational effort in comparison to ones required by
X for every λ = (λ1 , . . . , λm ) ≥ 0 (see [16]).
algorithms for the general problem (PX ).
In Section 2, the theoretical preliminaries are presented
to analyze three special cases of optimization over the
efficient solution set of problem (GM OP ). Section 3
λi fi is pseudoconvex on
By definition, if f is convex then it is scalarly pseudoconvex, and f is scalarly pseudoconvex then it is pseudoconvex. Hence, the convex multiobjective programming
problem is a special case of problem (GM OP ).
proposes the algorithms to solve these cases, including
Example II.1. Consider the vector function f (x) over
some computational experiments to illustrate the algo-
the set X = {x ∈ R2 | Ax ≥ b, x ≥ 0}, where
rithms. Some conclusions are given in the last section.
II. T HEORETICAL P RELIMINARIES
f (x) =
and
First, recall that a differentiable function h : X → R
A=
is called pseudoconvex on X if
∇h(x2 ), x1 − x2 ≥ 0 ⇒ h(x1 ) − h(x2 ) ≥ 0
for all x1 , x2 ∈ X . For example, by Proposition 5.20 in
[2], the fractional function r(x)/l(x), where r : Rn → R
is convex on X and l : Rn → R is linear such that
x22 + x1
−x21 − 0.6x1 + 0.5x2
,
x1 − x2 − 2
x2 − x1 + 2
−1
1
2
,
b
=
3
2
6
−2 −1
−10
.
We have
λT f (x) =
λ1 (x21 + 0.6x1 − 0.5x2 ) + λ2 (x22 + x1 )
.
x2 − x1 + 2
It is easily seen that
r(x) = λ1 (x21 + 0.6x1 − 0.5x2 ) + λ2 (x22 + x1 )
l(x) > 0 for all x ∈ X , is a pseudoconvex function.
By the definition in [10, tr. 132], a function h is called
quasiconvex on X if
h(x1 ) − h(x2 ) ≤ 0 ⇒ h(λx1 + (1 − λ)x2 ) ≤ h(x2 ),
for all x1 , x2 ∈ X and 0 ≤ λ ≤ 1. If h is quasiconvex
then g := −h is quasiconcave.
is convex because λ1 , λ2 ≥ 0, and l(x) = x2 −x1 +2 > 0
for all x ∈ X . Therefore, by Proposition 5.20 in [2], the
function λT f (x) = r(x)/l(x) is pseudoconvex , i.e. f (x)
is scalarly pseudoconvex.
Let
Y = {y ∈ Rm |y = f (x) for some x ∈ X }.
As usual, the set Y is said to be the outcome set for
to the outcome set Y and the objective function Φ(x) of
problem (GM OP ).
problem (PX ) is pseudoconvex on X .
Let
yiI
clear that
| y ∈ Y}, i = 1, . . . , m. It is
Case 2. The feasible set XE is the efficient solution set
is also the optimal value of the following
of problem (GBOP ) and the objective function Φ(x)
= min{yi
yiI
programming problem
has the form as Φ(x) = ϕ(f (x)) where ϕ : Y → R and
min fi (x) s.t. x ∈ X .
(Pi )
ϕ(x) = λ, y
(2)
For each i ∈ {1, . . . , m}, if fi is pseudoconvex, we can
with λ = (λ1 , λ2 )T ∈ R2 . This case could happen
apply convex programming algorithms to solve problem
in certain common situations, for instance, when the
(Pi ) (Remark 2.3 in [3]).
objective function of problem (PX ) represents the linear
I
The point y I = (y1I , . . . , ym
) is called the ideal point
of the set Y. Notice that the ideal point y I need not
belong to Y.
composition of the criteria fi (x), i ∈ {1, 2} with the
weighted coefficients λi , i ∈ {1, 2}.
Case 3. The feasible set XE is the efficient solution set
of problem (GBOP ) and the objective function Φ(x)
has the form as Φ(x) = ϕ(f (x)) where ϕ : Y → R is
quasiconcave and decreasing monotonic.
Let Q ⊂ Rm be a nonempty set. A point q 0 ∈ Q
is called an efficient point of the set Q if there is no
q ∈ Q such that q 0 ≥ q and q 0 = q. The set of all
efficient points of Q is denoted by MinQ. It is clear that
0
a point q 0 ∈ MinQ if Q ∩ (q 0 − Rm
+ ) = {q }, where
m
Rm
+ = {y ∈ R |yi ≥ 0, i = 1, . . . , m}.
Fig. 1.
Since the functions fi , i = 1, . . . , m, are continuous
The ideal point y I
and X ⊂ Rn is a nonempty compact set, the outcome
Consider a generalized convex biobjective programming problem
set MinY is nonempty [9]. Let
T
Minf (x) = (f1 (x), f2 (x)) s.t. x ∈ X ,
where X := {x ∈ R
set Y is also compact set in Rm . Therefore, the efficient
n
(GBOP )
| gj (x) ≤ 0, j = 1, . . . , k},
the functions gj (x), j = 1, . . . , k, are differentiable
quasiconvex on Rm and the objective function f is
YE = {f (x)|x ∈ XE }.
The set YE is called the efficient outcome set for problem
(GM OP ). By definition, it is easy to see that
YE = MinY.
scalarly pseudoconvex on X .
Now we will describe the three sets of conditions
associated with three special cases of problem (PX )
under consideration.
Case 1. The feasible set XE is the efficient solution
set of problem (GM OP ), the ideal point y I belongs
(3)
(4)
The relationship between the efficient solution set XE
and the efficient set MinY is described as follows.
Proposition II.1.
i) For any y 0 ∈ MinY, if x0 ∈ X satisfies f (x0 ) ≤ y 0 ,
then x0 ∈ XE .
ii) For any x0 ∈ XE , if y 0 = f (x0 ), then y 0 ∈ MinY.
Proof: i) Since y 0 ∈ MinY, by definition, we have
0
0
0
m
(y 0 − Rm
+ ) ∩ Y = y . Moreover, f (x ) ∈ (y − R+ ) and
f (x0 ) ∈ Y because x0 ∈ X and f (x0 ) ≤ y 0 . Therefore,
0
0
f (x ) = y ∈ MinY. Combined this fact with (3) and
(4), we imply x0 ∈ XE .
The following interesting property of Z (see Theorem
3.2 in [20]) will be used in the sequel .
Proposition II.2. MinZ = MinY.
Now we will consider the first special case where the
ideal point y I belongs to the outcome set Y.
Proposition II.3. If y I ∈ Y then MinY = {y I }.
ii) This fact follows immediately from (3) and (4).
Proof: Since y I ∈ Y, there exists xI ∈ X such that
y I = f (xI ), i.e. yiI = fi (xI ) for all i = 1, 2, . . . , m. For
each i ∈ {1, 2, . . . , m}, since yiI is the optimal value
of problem (Pi ), xI is an optimal solution of problem
(Pi ). Hence, xI ∈ Argmin{fi (x) | x ∈ X } for all
i = 1, . . . , m. By definition, xI is an efficient solution
of problem (GM OP ), i.e. xI ∈ XE . From Proposition
II.1(ii), it implies y I ∈ MinY. Since yiI is the optimal
value of problem (Pi ) for i = 1, 2, . . . , m, by definition
of efficient points, y I is the only efficient point of Y.
Let
Fig. 2.
The efficient set MinY
X id = {x ∈ X | fi (x) ≤ yiI , i = 1, 2, . . . , m}.
By [10, Theorem 9.3.5], for each i = 1, 2, . . . , m,
Let
if fi is the pseudoconvex function, fi is quasiconvex.
Z=Y+
Rm
+
= {z ∈ R
m
| z ≥ y for some y ∈ Y}.
Therefore, X id is a convex set because every lower level
set of continuous quasiconvex functions is convex. The
It is clear that Z is a nonempty, full-dimension closed
set but it is nonconvex in general.
following assertion provides a property to detect whether
y I belongs to Y.
Proposition II.4. If X id is not empty then y I ∈ Y and
XE = X id . Otherwise, y I does not belong to Y.
Proof: By definition, if X id = ∅, y I ∈ Z. Since
Y ⊆ Z, y I ∈ Y. Otherwise, X id is not empty, y I ∈ Z.
Therefore, y I ∈ Y because yiI is the optimal value of
problem (Pi ) for i = 1, 2, . . . , m. By Proposition II.3
and Proposition II.1(i), we get MinY = {y I } and XE =
{x ∈ X | f (x) ≤ y I } = X id .
In the next two cases, we consider problem (PX )
Fig. 3.
The set Z
where XE is the efficient solution set of problem
(GBOP ) and Φ(x) has the form as Φ(x) = ϕ(f (x))
¯ y − y¯ ≥ 0 for all y ∈ Y. For i = 1, 2, set
i.e. λ,
where ϕ : Y → R. Then the outcome-space reformula-
ˆi = λ
¯ i if y¯i = w
ˆ i = 0 if y¯i = w
λ
¯i and λ
¯i . It is easy to
tion of problem (PX ) can be given by
check that
min ϕ(y) s.t. y ∈ YE .
ˆ w
ˆ ≥ 0, λ
ˆ = 0.
λ,
¯ − y¯ = 0 and λ
(PY )
Combining (4) and Proposition II.2, problem (PY ) can
be rewritten as follows
{y ∈ R
min ϕ(y) s.t. y ∈ MinZ.
(PZ )
Therefore, instead of solving problem (PY ), we solve
problem (PZ ).
ˆ y−w
λ,
¯
Hence,
m
(7)
≥ 0 for all y ∈ Z. Set H(w)
¯ =
ˆ y−w
¯
| λ,
≥ 0}. Then Z ⊂ H(w)
¯ for all
w
¯ ∈ ∂Z. By [14, Theorem 6.20], we imply that Z is a
convex set.
Since Z is a nonempty convex subset in R2 by
Proposition II.5, it is well known [13] that the efficient set
Proposition II.5. If f is scalarly pseudoconvex, the set
MinZ is homeomorphic to a nonempty closed interval
Z = f (X) + R2+ is a convex set in R2 .
of R. By geometry, it is easily seen that the problem
Proof: Let w
¯ be an arbitrary point in the boundary
min{y2 : y ∈ Z, y1 = y1I }
(PS )
∂Z of the set Z. By geometry, there exists y¯ ∈ MinZ
such that y¯ ≤ w.
¯ From Proposition II.2, we imply that
y¯ ∈ MinY. Hence, by Proposition II.1(ii), there exists
has a unique optimal solution y S and the problem
min{y1 : y ∈ Z, y2 = y2I }
(PE )
x
¯ ∈ X such that y¯ = f (¯
x) and x
¯ is an efficient solution
has a unique optimal solution y E . Since Z is convex,
of problem (GM OP ).
Let J = {j ∈ {1, . . . , k} | gj (¯
x) = 0} and s = |J| .
problems (PS ) and (PE ) are convex programming prob-
For any vector a ∈ Rs , we denote aJ := {aj , j ∈ J}.
lems. If y I ∈ Y then, by Propositions II.3 and II.4, y I
Since x
¯ ∈ XE , by [11, Corollary 3.1.6], there exist a
is the only optimal solution to problem (PZ ) and X id is
¯ ∈ R2 \ {0} and a vector µ
¯J ∈ Rs+ , such that
vector λ
+
the optimal solution set of problem (PX ).
¯ T ∇f (¯
λ
x) + µ
¯TJ ∇gJ (¯
x) = 0 which means
¯ T ∇f (¯
λ
x) = −¯
µTJ ∇gJ (¯
x).
(5)
Since µ
¯J ≥ 0, gJ (x) ≤ 0 for all x ∈ X and
gJ (¯
x) = 0, we have µ
¯TJ gJ (x) − µ
¯TJ gJ (¯
x) ≤ 0 for
all x ∈ X . Combined this fact with the condition
gj , j = 1, . . . , k, are differentiable quasiconvex and (1),
we get µ
¯TJ ∇gJ (¯
x), x − x
¯ ≤ 0 for all x ∈ X . Thus, by
¯ T ∇f (¯
(5), one has λ
x), x − x¯ ≥ 0 or
¯ T f (¯
x) , x − x
¯ ≥0
∇ λ
∀x ∈ X .
(6)
¯T f is pseudoconvex on X because f is
Moreover, λ
scalarly pseudoconvex on X . Therefore, (6) implies that
¯ T f (x) − λ
¯ T f (¯
λ
x) ≥ 0 ∀x ∈ X ,
Fig. 4.
The efficient curve MinZ
If y I ∈ Y then y S = y E and the efficient set MinZ
is a curve on the boundary of Z with starting point y S
and the end point y E such that
Consider the following convex problem
y1E > y1S and y2S > y2E .
˜
min λ, y s.t. y ∈ Z.
(8)
Note that we also get the efficient solutions xS , xE ∈ XE
(CP 0 )
that has the explicit reformulation as follows
such that y S = f (xS ) and y E = f (xE ) while solving
min
problems (PS ) and (PE ). For the convenience, xS , xE
s.t.
is called to be the efficient solutions respect to y S , y E ,
λ, y
f (x) − y ≤ 0
x ∈ X,
respectively.
(CP 1 )
c, y ≤ α,
In the second case, we consider problem (PZ ) where
ϕ(y) = λ, y . Direct computation shows that the equation of the line through y S and y E is c, y = α, where
c = E 1 S , S 1 E ,
y1 −y1 y2 −y2
(9)
α = Ey1E S + Sy2E E .
y −y
y −y
1
1
2
2
where the vector c ∈ R2 and the real number α is
determined by (9).
Proposition II.6. Suppose that (x∗ , y ∗ ) is an optimal
solution of problem (CP 1 ). Then x∗ is an optimal
solution of problem (PX ).
From (8), it is easily seen that the vector c is strictly
Proof: It is well known that a convex programming
positive. Now, let
problem with the linear objective function has an optimal
Z˜ = {y ∈ Z | c, y ≤ α}
solution which belongs to the extreme point set of the
and
feasible solution set [18]. Therefore, problem (CP 0 ) has
Γ = ∂ Z˜ \ (y S , y E ),
an optimal solution y ∗ ∈ Γ. This fact and (10) implies
where (y S , y E ) = {y = ty S + (1 − t)y E | 0 < t < 1}
˜ it implies that y ∗ is
that y ∗ ∈ MinZ. Since MinZ ⊂ Z,
˜
and ∂ Z˜ is the boundary of the set Z.
an optimal solution of problem (PZ ).
Since MinZ = YE = MinY, by definition, we have
λ, y ∗ ≤ λ, y for all y ∈ YE and y ∗ ∈ MinY. Then
λ, y ∗ ≤ λ, f (x) , ∀x ∈ XE .
(11)
Since (x∗ , y ∗ ) is a feasible solution of problem (CP 1 ),
we have f (x∗ ) ≤ y ∗ . By Proposition II.1, x∗ ∈ XE .
Furthermore, we have f (x∗ ) ∈ Y and y ∗ ∈ MinY.
The definition of efficient points infers that y ∗ = f (x∗ ).
Combining this fact and (11), we get Φ(x) ≥ Φ(x∗ ) for
all x ∈ XE which means x∗ is an optimal solution of
Fig. 5.
problem (PX ).
The convex set Z˜
It is clear that Z˜ is a compact convex set because Z
In the last case, we consider problem (PZ ) where the
is convex. By the definition and geometry, we can see
function ϕ : Y → R is quasiconcave and decreasing
that Γ contains the set of all extreme points of Z˜ and
monotonic. The following assertion presents the special
MinZ = Γ.
(10)
property of the optimal solution to problem (PZ ).
By Proposition II.4, to detect whether the ideal point
y I belongs to Y and solve problem (PX ) in this case,
we solve the following problem
min Φ(x) s.t. x ∈ X id .
(CP id )
Since Φ(x) is pseudoconvex on X and X id is convex,
we can apply convex programming algorithms to solve
problem (CP id ) (Remark 2.3 in [3]). The procedure for
this case is described as follows.
Procedure 1.
Fig. 6.
The illustration to Case 3
Step 1. For each i = 1, . . . , m, find the optimal value yiI
Proposition II.7. If the function ϕ is quasiconcave
of problem (Pi ).
and decreasing monotonic then the optimal solution of
Step 2. Solve problem (CP id ).
problem (PZ ) is attained at either y S or y E .
If problem (CP id ) is not feasible Then STOP (Case 1
Proof: Let Z △
=
conv{y S , y I , y E }, where
does not apply)
conv{y S , y I , y E } stands for the convex hull of the points
Else Find an optimal solution x∗ to problem (CP id ).
{y S , y I , y E }. Since MinZ ⊂ Z △ , we have
STOP (x∗ is an optimal solution to problem (PX )).
min{ϕ(y) | y ∈ MinZ} ≥ min{ϕ(y) | y ∈ Z △ }. (12)
Below we present a numerical example to illustrate
Procedure 1.
It is obvious that the optimal solution of the problem
min{ϕ(y) | y ∈ Z △ }, where the objective function ϕ
Example III.1. Consider the problem (PX ), where XE
is quasiconcave, belongs to the extreme point set of Z △
is the efficient solution set to the following problem
[18]. Therefore,
Min
Argmin{ϕ(y) | y ∈ Z △ } ∈ {y S , y I , y E }.
(f1 (x), f2 (x)) = (0.5x21 − x1 + 0.3x2 , x22 + x1 )
s.t. x21 + x22 − 4x1 − 4x2 ≤ −6
Moreover, since ϕ is also decreasing, we have
−x1 + x2 ≥ α
Argmin{ϕ(y) | y ∈ Z △ } ∈ {y S , y E }.
x1 + x2 ≥ 2
x1 ≥ 1
Since y S , y E ∈ MinZ, this fact and (12) imply
min{ϕ(y) | y ∈ MinZ} = min{ϕ(y) | y ∈ Z △ }
and Φ(x) = min{0.5x21 + x2 + 0.2; 2x22 − 4.6x1 + 5.8}.
•
and Argmin{ϕ(y) | y ∈ MinZ} ∈ {y S , y E }.
III. P ROCEDURES
AND
C OMPUTING E XPERIMENTS
In the case α = 0:
Step 1. Solving problems (P1 ) and (P2 ), we obtain the
ideal point y I = (−0.2000, 2.0000).
Case 1. The feasible set XE is the efficient solution
Step 2. Solving problem (CP id ), we can find an optimal
set of problem (GM OP ), the ideal point y I belongs
solution x∗ = (1.0000, 1.0000). Then x∗ is the optimal
to the outcome set Y and the objective function Φ(x) of
solution to problem (PX ) and Φ(x∗ ) = 1.7000 is the
problem (PX ) is pseudoconvex on X .
optimal value of problem (PX ) .
•
X = x ∈ R2 | (x1 − 1)2 + (x2 − 2)2 ≤ 1, 2x1 − x2 ≤ 1 ,
In the case α = −1:
Step 1. Solving problems (P1 ) and (P2 ), we obtain the
ideal point y I = (−0.2299, 1.8917).
Step 2. Solving problem (CP id ), we can find that it is
not feasible. It means that the ideal point y I does not
belong to Y.
and Φ(x) = λ1 f1 (x) + λ2 f2 (x).
It is easily seen that the function f is scalarly pseudoconvex on X because f1 , f2 are convex on X . Therefore,
we can apply Procedure 2 to solve this problem.
Case 2. The feasible set XE is the efficient solution set of
Step 1. The optimal value of the problems (P1 ) and (P2 ),
problem (GBOP ) and the objective function Φ(x) has
respectively, is y1I = 1.0000 and y2I = 1.0000.
the form as (2).
Step 2. Solving problem (CP2id ), we can find that it is
In this case, the procedure for solving problem (PX )
is established by Proposition II.4 and Proposition II.6.
not feasible. Then go to Step 3.
Step 3. Solve problems (PS ) and (PE ) to obtain
Recall that if y I ∈ Y, X id = Argmin(PX ). Therefore,
y S = (1.9412, 1.0000), y E = (1.0000, 1.9326)
we can obtain an optimal solution of problem (PX ) by
solving the following convex programming problem
min e, x s.t. x ∈ X id ,
(CP2id )
and
xS = (0.9612, 2.9991), xE = (0.0011, 2.0435)
where e = (1, . . . , 1) ∈ Rn .
respect to y S , y E , respectively.
Procedure 2.
Step 4. For each λ = (λ1 , λ2 ) ∈ R2 , solve problem
Step 1. For each i = 1, 2, find the optimal value yiI of
(CP 1 ) to find the optimal solution (x∗ , y ∗ ). Then x∗ is
problem (Pi ).
an optimal solution and Φ(x∗ ) is the optimal value of
Step 2. Solve problem (CP2id ).
(PX ). The computational results are shown in Table I.
If problem (CP2id ) is not feasible Then Go to Step 3
λ
x∗
y∗
Φ(x∗ )
(y ∈ Y).
(0.0, 1.0)
(0.9612, 2.9991)
(1.9412, 1.0000)
1.0000
Else Find an optimal solution x∗ to the problem (CP2id ).
(0.2, 0.8)
(0.4460, 2.8325)
(1.1989, 1.0281)
1.0622
(0.5, 0.5)
(0.2929, 2.7071)
(1.0858, 1.0858)
1.0858
(0.8, 0.2)
(0.1675, 2.5540)
(1.0208, 1.1989)
1.0622
Step 3. Solve problem (PS ) and problem (PE ) to find the
(1.0, 0.0)
(0.0011, 2.0435)
(1.0000, 1.9326)
1.0000
efficient points y S , y E and the efficient solutions xS , xE
(−0.2, 0.8)
(0.9654, 2.9992)
(1.9412, 1.0000)
0.4118
(0.8, −0.2)
(0.0011, 2.0435)
(1.0000, 1.9326)
0.4146
I
STOP (x∗ is an optimal solution to problem (PX )).
respect to y S , y E , respectively.
TABLE I
1
Step 4. Solve problem (CP ) to find an optimal solution
C OMPUTATIONAL RESULTS OF E XAMPLE III.2
(x∗ , y ∗ ). STOP (x∗ is an optimal solution to (PX )).
Below are some examples to illustrate Procedure 2.
Example III.2. Consider problem (PX ), where XE is
the efficient solution set to problem (GBOP ) with
f1 (x) =
x21
2
+ 1, f2 (x) = (x2 − 3) + 1,
Example III.3. Consider problem (PX ), where Φ(x) =
λ1 f1 (x) + λ2 f2 (x) and XE is the efficient solution set
to problem (GBOP ) with
f (x) =
x22 + x1
−x21 − 0.6x1 + 0.5x2
,
x1 − x2 − 2
x2 − x1 + 2
and X = x ∈ R2 | Ax ≥ b, x ≥ 0 , where
−1
1
2
A= 3
2 , b =
6 .
−2 −1
−10
In this case, the procedure for solving problem is
established by Proposition II.4 and Proposition II.7. Let
xopt be an optimal solution of problem (PX ).
Procedure 3.
By Example II.1, it is verified that f is scalarly
Step 1. For each i = 1, 2, find the optimal value yiI of
pseudoconvex on X . Therefore, we can apply Procedure
problem (Pi ).
2 to solve this problem.
Step 2. Solve problem (CP2id ).
Step 1. The optimal value of the problems (P1 ) and (P2 ),
If problem (CP2id ) is not feasible Then Go to Step 3
respectively, is y1I = −0.4167 and y2I = 1.5400.
(y I ∈ Y).
Step 2. Solving problem (CP2id ), we can find that it is
Else Find an optimal solution x∗ to the problem (CP2id ).
not feasible. Then go to Step 3.
STOP (x∗ is an optimal solution to problem (PX )).
Step 3. Solve problems (PS ) and (PE ) to obtain
Step 3. Solve problem (PS ) and problem (PE ) to find the
y S = (−0.2000, 1.5400), y E = (−0.4167, 8.3332)
efficient points y S , y E and the efficient solutions xS , xE
respect to y S , y E , respectively.
and
Step 4. If ϕ(y S ) > ϕ(y E ) Then xopt = xE Else xopt =
xS = (0.4000, 2.4000), xE = (0.0000, 9.9997)
respect to y S , y E , respectively.
xS . STOP (xopt is an optimal solution to problem (PX )).
We give below an example to illustrate Procedure 3.
Step 4. For each λ = (λ1 , λ2 ) ∈ R2 , solve problem
(CP 1 ) to find the optimal solution (x∗ , y ∗ ). Then x∗ is
Example III.4. Consider problem (PX ), where XE is
the efficient solution set to problem (GBOP ) with
an optimal solution and Φ(x∗ ) is the optimal value of
(PX ). The computational results are shown in Table II.
f1 (x) = x21 + 2x1 x2 + 3x22 , f2 (x) = x22 − 0.5x1 + 0.3x2 ,
X = x ∈ R2 | g1 (x) ≤ 0, g2 (x) ≤ 0 ,
λ
x∗
y∗
Φ(x∗ )
(0.0, 1.0)
(0.4000, 2.4000)
(−0.2000, 1.5400)
1.5400
g1 (x) = 9(x1 −2)2 +4(x2 −3)2 −36, g2(x) = x1 −2x2 −3,
(0.2, 0.8)
(0.4000, 2.4000)
(−0.2000, 1.5400)
1.1920
(0.5, 0.5)
(0.4000, 2.4000)
(−0.2000, 1.5400)
0.6700
(0.8, 0.2)
(0.0900, 2.8650)
(−2.2870, 1.7378)
0.1180
(1.0, 0.0)
(0.0000, 9.9997)
(−0.4167, 8.3332)
−0.4167
(−0.2, 0.8)
(0.4000, 2.4000)
(−0.2000, 1.5400)
1.2720
pseudoconvex on X because f1 , f2 are convex on X .
(0.8, −0.2)
(0.0000, 9.9997)
(−0.4167, 8.3332)
−2.0000
Moreover, the function ϕ is quasiconcave and decreasing
TABLE II
C OMPUTATIONAL RESULTS OF E XAMPLE III.3
and Φ(x) = ϕ(f (x)) with ϕ(y) = −y12 − y22 .
It is easily verified that the function vector f is scalarly
monotonic on R2 . Therefore, we can apply Procedure 3
to solve this problem.
Case 3. The feasible set XE is the efficient solution
set of problem (GBOP ) and the objective function Φ(x)
Step 1. The optimal value of the problems (P1 ) and (P2 ),
respectively, is y1I = 2.2192 and y2I = −1.2623.
has the form as Φ(x) = ϕ(f (x)) where ϕ : Y → R is
Step 2. Solving problem (CP2id ), we can find that it is
quasiconcave and decreasing monotonic.
not feasible. Then go to Step 3.
Step 3. Solve problem (PS ) and problem (PE ) to obtain
[6] J. Fulop and L. D. Muu, “Branch-and-bound variant of an
outcome-based algorithm for optimizing over the efficient set of a
y S = (9.2894, −1.2623), y E = (2.2192, −0.4012)
bicriteria linear programming problem”, J. Optim. Theory Appl.,
vol. 105, pp. 37-54, 2000.
and
[7] R. Horst, N. V Thoai, Y. Yamamoto and D. Zenke, “On optimiza-
xS = (2.7848, 0.2406), xE = (1.1432, 0.2892)
respect to y S , y E , respectively.
S
E
Step 4. Since ϕ(y ) < ϕ(y ), the optimal solution to
problem (PX ) is xopt = xS = (2.7874, 0.2406) and the
tion over the efficient set in linear multicriteria programming”, J.
Optim. Theory Appl., vol. 134, pp. 433-443, 2007.
[8] N. T. B. Kim and T. N. Thang, “Optimization over the Efficient Set
of a Bicriteria Convex Programming Problem”, Pacific J. Optim.,
vol. 9, pp. 103-115, 2013
[9] D. T. Luc, Theory of Vector Optimization, Springer-Verlag, Berlin,
Germany, 1989.
optimal value of problem (PX ) is Φ(xE ) = −87.8812.
[10] O. L. Mangasarian, Nonlinear Programming, McGraw-Hill, New
IV. C ONCLUSION
[11] K. Miettinen, Nonlinear Multiobjective Optimization, Kluwer
York, 1969.
In this article, we have developed the simple convex
Academic Publishers, Boston, 1999.
[12] L. D. Muu and L. Q. Thuy, “Smooth optimization algorithms
programming procedures for solving three special cases
for optimizing over the Pareto efficient set and their application
of optimization problem over the efficient set (PX ).
to Minmax flow problems”, Vietnam J. Math., vol. 39, no. 1, pp.
These special case procedures require quite little com-
31-48, 2011.
putation effort in comparision to that required to solve
the general case because only some convex programming
problems need solving. Therefore, they can be used as
[13] H. X. Phu, “On efficient sets in R2 ”, Vietnam J. Math, vol. 33,
pp. 463-468, 2005.
[14] R. T. Rockafellar and R. B. Wets, Variational Analysis, SpringerVerlag, Berlin, Germany, 2010.
[15] T. N. Thang and N. T. B. Kim, “Outcome space algorithm
screening devices to detect and solve these special cases.
ACKNOWLEDGMENT
This research is funded by Hanoi University of Science
and Technology under grant number T2016-TC-205.
R EFERENCES
for generalized multiplicative problems and optimization over the
efficient set”, J. Ind. Manag. Optim., vol. 12, no. 4, pp. 1417 1433, 2016.
[16] T. N. Thang, D. T. Luc and N. T. B. Kim, “Solving generalized
convex multiobjective programming problems by a normal direction method”, Optim., vol. 65, no. 12, pp. 2269-2292, 2016.
[17] N. V. Thoai, “Reverse convex programming approach in the space
of extreme criteria for optimization over efficient sets”, J. Optim
[1] L. T. H. An, P. D. Tao, N. C. Nam and L. D. Muu, “Method for
Theory and Appl., vol.147, pp. 263-277, 2010.
optimizing over efficient and weakly efficient the sets of an affine
[18] H. Tuy, Convex Analysis and Global Optimization, Kluwer, 1998.
fractional vector optimization program”, Optim., vol. 59, no. 1, pp.
[19] Y. Yamamoto, “Optimization over the efficient set: overview”, J.
77-93, 2010.
[2] M. Avriel, W. E. Diewert, S. Schaible and I. Zang, Generalized
Concavity, Plenum Press, New York, 1998.
Global Optim., vol. 22, pp. 285-317, 2002.
[20] P. L. Yu, Multiple-Criteria Decision Making. Plenum Press, New
York and London, 1985.
[3] H. P. Benson, “On the Global Optimization of Sums of Linear
[21] M. M. Wiecek, M. Ehrgott and A. Engau, “Continuous multiob-
Fractional Functions over a Convex Set”, J. Optim. Theory Appl.,
jective programming”, Multiple criteria decision analysis: State of
vol. 121, no. 1, pp. 19-39, 2004.
[4] H. P. Benson, “A global optimization approach for generating
efficient points for multiobjective concave fractional programs”,
J. Multi Criteria Decis. Anal., vol. 13, pp. 15–28, 2005.
[5] H. P. Benson, “An outcome space algorithm for optimization
over the weakly efficient set of a multiple objective nonlinear
programming problem”, J. Glob. Optim., vol. 52, pp. 553- 574,
2012.
the art surveys, Oper. Res. Manag. Sci., Springer, New York, vol.
233, pp. 739-815, 2016.