You are on page 1of 20

Fuzzy Sets and Systems 146 (2004) 167186

www.elsevier.com/locate/fss
Fuzzy nonlinear programming for mixed-discrete design
optimization through hybrid genetic algorithm
Ying Xiong

, Singiresu S. Rao
Department of Mechanical Engineering, University of Miami, Coral Gables, FL 33124-0624, USA
Received 28 November 2002; received in revised form 6 August 2003; accepted 3 September 2003
Abstract
Many practical engineering optimization problems involve discrete or integer design variables, and often
the design decisions are to be made in a fuzzy environment in which the statements might be vague or
imprecise. A mixed-discrete fuzzy nonlinear programming approach that combines the fuzzy z-formulation
with a hybrid genetic algorithm is proposed in this paper. This method can nd a globally compromise
solution for a mixed-discrete fuzzy optimization problem, even when the objective function is nonconvex and
nondierentiable. In the construction of the objective membership function, an error from the early research
work is corrected and the right conclusion has been made. The illustrative examples demonstrate that more
reliable and satisfactory results can be obtained through the present method.
c 2003 Elsevier B.V. All rights reserved.
Keywords: Hybrid genetic algorithm; Fuzzy programming; Mixed-discrete optimization; Membership function;
Engineering design
1. Introduction
The traditional approaches to nonlinear programming problems treat the design variables as con-
tinuous quantities. However, in real life, many engineering optimization problems frequently involve
discrete or integer design variables. This is due to limitations such as availability of components in
standard sizes and constraints imposed by the construction and manufacturing practices. Examples
of discrete/integer variables are the number of gear teeth, the value of gear module, the number of
screws or rivets used in a structural assembly, the number of bars in a truss, etc. Optimization prob-
lems involving both continuous and discrete/integer variables are often referred to as mixed-discrete
nonlinear programming problems (MDNLP).

Corresponding author.
E-mail address: yingxiong2@yahoo.com (Y. Xiong).
0165-0114/$ - see front matter c 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.fss.2003.09.001
168 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
As early as late 1960s, people were showing interest in the research of MDNLP. However, many
research works were not reported until the last decade. Some of the well-known approaches can
be classied as: (1) branch-and-bound method [13,14,28,29]; (2) sequential linearization algorithm
[20,21]; (3) sequential linearization coupled with the branch-and-bound method [4]; (4) penalty func-
tion approach [5,11,32]; (5) dual strategy [17]; simulated annealing [18]; and (6) genetic algorithm
[8,25]. More details of the research work done in this eld can be seen in the literature surveys of
Loh and Papalambros [20], and Arora and Huang [2].
Generally, the design problems are always stated in precise mathematical forms. It must be recog-
nized that many practical problems encountered by designers and decision makers would take place
in an environment in which the statements might be vague or imprecise. Usually it is dicult to
describe the goals and constraints of such optimization problems by crisp relations through equations
and}or descriptions. It may be possible that a small violation of a given constraint may lead to a
more ecient and practical solution of the problem. Fuzzy set theory can build a model to repre-
sent a subjective estimation of possible eect of the given values on the problem and permit the
incorporation of vagueness in the conventional set theory that can be used to deal with uncertainty
quantitatively. Hence, it has proven to be an ecient tool for the treatment of fuzzy problems, and
the mathematical developments have provided the theoretical basis necessary for use in practical
applications.
Fuzzy set theory was originally established by Zadeh [38] in 1965. Since then this theory has
been widely developed and applied to various scientic elds. The fuzzy set-based optimization
was introduced by Bellman and Zadeh in their seminal paper on decision making in a fuzzy en-
vironment [3], in which the concepts of fuzzy constraint, fuzzy objective and fuzzy decision were
introduced. These concepts were subsequently profusely used and applied by many investigators.
Fuzzy optimization is a exible approach that permits a more adequate solution of real problems
in the presence of vague information. In the last two decades, the principles of fuzzy optimization
were critically studied, and the technologies and solution procedures have been investigated within
the scope of fuzzy sets. Today, similar to the developments in crisp optimization, dierent kinds of
mathematical models have been proposed and many practical applications have been implemented
to solve fuzzy optimization problems in various engineering elds, such as mechanical design and
manufacturing [6,7,9,16,26,35], power systems [1,12,22,40], water resources research [36,37], control
systems [15,19,23,30,33,34,39], etc.
However, to the authors knowledge, only a limited work has been done in the development of
optimization techniques for fuzzy nonlinear programming problems in mixed-discrete design space,
and the literature discussing the application of such techniques is scarce. Therefore, seeking a reli-
able and eective optimization method seems to be necessary and meaningful to researchers in the
engineering design elds.
In this paper, a mixed-discrete fuzzy nonlinear programming (MDFNLP) approach is presented.
Specically, the fuzzy z-formulation is combined with a hybrid genetic algorithm for solving fuzzy
optimization problems with mixed-discrete design variables. The genetic algorithm starts with a
population of designs instead of one design point and converges to a design space in the neighbor-
hood of the global optimum point. Hence, the genetic algorithm is used as the primary scheme of
optimization in this work. The basic concepts of the genetic algorithm are combined with a tradi-
tional gradient-based optimization strategy to develop a new mixed-discrete hybrid genetic algorithm
(MNHGA), with a detailed description given in the following sections. In the construction of the
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 169
objective membership function, an error from the paper of Shih and Lai is corrected and right conclu-
sions have been reached. Both linear and quadratic functions are used in constructing the constraint
membership functions. Although linear membership functions are commonly used, mainly for sim-
plicity, a quadratic membership function can also be used when the preference values corresponding
to dierent constraint limits follow a nonlinear relationship.
2. Mathematical model
The mathematical model of a general MDFNLP problem can be stated as
P(MDFNLP) : nd X which minimizes [(X)
s.t. q
)
(X) G
)
, ) = 1, 2, . . . , m,
X = [x
1
, . . . , x
nq
, . . . , x
nd
, . . . x
n
]
T
,
x
L
i
6x
i
6x
U
i
; i = 1, 2, . . . , n
where G
)
indicates the allowable interval of the constraint function q
)
. The lower and upper bounds
of the design variables are indicated by x
L
i
and x
U
i
, respectively.
The fuzzy feasible region is dened by considering all the constraints as
S =
m

)=1
G
)
(1)
and the membership degree of any design vector X to the fuzzy feasible region S is given by
j
s
(X) = min
)
{j
q
)
(X)}, (2)
which is the minimum degree of satisfaction of the design vector X to all of the constraints. The
design vector X can be considered feasible if 0j
s
(X)1. The objective function denes a fuzzy
domain O in S that is expressed as
O = {j
[
(X)}
_
_
_
m

)=1
j
q
)
(X)
_
_
_
, (3)
where j
[
(X) and j
q)
(X) denote the membership functions of the objective and )th constraint func-
tion, respectively. Thus, the optimum solution X

can be selected by maximizing the smallest mem-


bership function in the fuzzy domain O such that
j
O
(X

) = max j
O
(X), (4)
where
j
O
(X) = min{j
[
(X), j
q1
(X), . . . , j
qm
(X)}.
170 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
This maxmin problem can be solved by using the fuzzy z-formulation technique that can be
stated mathematically as
maximize z (5)
s.t. z 6j
[
(X),
z 6j
q)
(X), ) = 1, . . . , m,
(6)
where
j
[
(X) =
_

_
0 if [(X) [
max
,
[
max
[(X)
[
max
[
min
if [
min
[(X) 6[
max
,
1 if [(X) 6[
min
(7)
and
j
q)
(X) =
_

_
0 if q
)
(X) b
)
+ d
)
,
1
_
q
)
(X) b
)
d
)
_
if b
)
q
)
(X) 6b
)
+ d
)
,
1 if q
)
(X) 6b
)
.
(8)
The values of b
)
and d
)
denote the constant and the allowable fuzzy transition intervals, respectively,
for the )th inequality constraint. The values of [
max
and [
min
are obtained by optimization in crisp
and fuzzy feasible domains, respectively.
In cases where two or more solutions exist with the same value of z
max
with one solution having
systematically better or equal ratings for all constraint membership values compared to other solu-
tions, Dubois and Fortemps [10] gave a detailed procedure for choosing the right solution. Also a
theoretical framework is developed in their paper to include a broad class of maxmin optimization
problems. For simplicity, the following iterative method has been used in this work:
Consider X

and X

as two solutions (global optima) of the problem in the sense of the criterion
of Eq. (4), let j
O
(X

) =j
O
(X

), j
qi
(X

)j
qi
(X

) for all i =0, 1, . . . , m (by letting q


0
=[), and
j
q)
(X

)j
q)
(X

) for at least one ). Then to nd the right solution X

instead of X

, rst group all


the constraints into two setsone for X

and the other for X

. In each set, arrange the constraints


in the increasing order of their membership function values. Then compare the smallest, second
smallest, third smallest, . . . , values of the membership functions from one set with the values of
corresponding membership functions from the other set. During this process, at kth stage, the situation
j
(k)
(X

)j
(k)
(X

) is encountered, then X

is chosen as the right solution to the problem. This


iterative process is easy to implement to ensure that the best solution among the maxmin optimal
solutions is chosen for the mixed-discrete fuzzy programming problem, though it involves some
additional computational work.
2.1. Construction of the objective membership functions
In continuous design domain, it is known that the membership function of j
[
(X) can be con-
structed by the optimum values of [
min
and [
max
that correspond to crisp and fuzzy optimization.
In discrete design domain, there are four linear membership function combinations composed of
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 171
f
1

f
c
min
f
c
max
f
d
min
f
d
max
Fig. 1. Four possible combinations of j
[
(X).
([
min
c
, [
max
c
), ([
min
c
, [
max
d
), ([
min
d
, [
max
c
), and ([
min
d
, [
max
d
). Here, [
min
d
and [
max
d
denote the opti-
mum objective values in the mixed-discrete design space with and without fuzziness, and [
min
c
and
[
max
c
are the optimum values of [(X) in the continuous design domain with and without fuzziness,
respectively. But the problem is which combination is the best selection for the construction of the
mixed-discrete membership function of the objective?
Shih and Lai [31] observed that in equation
z 6j
[
(X) =
[
max
[(X)
[
max
[
min
, (9)
the maximum value of z is obtained when the equality is strictly satised. Thus
z =
[
max
[(X)
[
max
[
min
, (10)
which can be rewritten as
[(X) = [
max
z([
max
[
min
). (11)
Thus, obtaining the minimum value of [(X) needs a minimum [
max
and a maximum of z([
max

[
min
). Since z is always maximized as indicated in Eq. (5), it requires both [
max
and [
min
to
be as small as possible. Shih and Lai, therefore, declared that the membership function composed
of ([
min
c
, [
max
c
) will yield the best optimum value and the membership function constructed by
([
min
d
, [
max
d
) will yield the worst nal solution.
This declaration, however, is not always correct, as shown below. Consider a mixed-discrete
optimization problem, and assume that its four optimum values in the continuous and mixed-discrete
design spaces with and without fuzziness have been obtained with the following relationship:
[
min
c
[
max
c
[
min
d
[
max
d
. (12)
This implies that the intervals [[
min
c
, [
max
c
] and [[
min
d
, [
max
d
], do not overlap, as illustrated in Fig. 1.
If as Shih and Lai stated, we select the combination ([
min
c
, [
max
c
) in the construction of the
objective membership function, the optimum solution will be obtained in [[
min
c
, [
max
c
]; however, the
true optimum solution should lie in the range [[
min
d
, [
max
d
]. It seems that Shih and Lais declaration
can be valid only when [[
min
c
, [
max
c
] has an overlap with [[
min
d
, [
max
d
], as shown in Fig. 2. Even
172 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
1
f
c
min
f
d
min
f
c
max
f
d
max
f

Fig. 2. The only possible situation according to Shih and Lais statement.
though, there is still a necessary condition for Shih and Lais approach to be valid. That is, the
optimum value obtained must be located in the interval [[
min
d
, [
max
c
]; otherwise, the solution will be
a false one. Since [[
min
d
, [
max
c
] is a subset of [[
min
d
, [
max
d
], the optimum solution can still be found
if the membership function is constructed using the combination ([
min
d
, [
max
d
).
Thus, it can be concluded that the correct objective membership function to be used in the fuzzy
mixed-discrete nonlinear programming should be composed of ([
min
d
, [
max
d
) that are obtained in the
discrete design space. Therefore, the best optimum solution can be determined using the proposed
optimization algorithm, but not by constructing the membership function with ([
min
c
, [
max
c
).
2.2. Construction of the constraint membership functions
Recently, some methods have been proposed for using statistical data in the construction of mem-
bership functions. In the absence of additional information, the linear variation has been used com-
monly. In this work, nonlinear membership functions, based on a second-order polynomial, are used
in addition to the linear membership functions for the constraints. The quadratic membership function
corresponding to q
)
(X) is given by
j
q)
(X) = a
0
q
)
(X)
2
+ a
1
q
)
(X) + a
2
, (13)
where a
0
, a
1
and a
2
are the coecients to be determined. The fuzzy constraint can be stated as
q
)
(X) 6b
)
+ d
)
, ) = 1, 2, . . . , m (14)
with j
q)
(X) =1 when q
)
(X) =b
)
. Similarly, j
q)
(X) =0 when q
)
(X) =b
)
+d
)
. In addition to these two
conditions, in order to determine the coecients a
0
, a
1
, and a
2
, arbitrary values of q
)
(X) (b
)
, b
)
+d
)
)
and j
q)
(X) (0, 1) can be chosen as the third condition. Dierent j
q)
(X) and q
)
(X) indicate dierent
shapes of the polynomial. In this work, j
q)
(X) is assumed 0.5 when q
)
(X) =(b
)
+ d
)
)}2. Thus, the
following equations are solved to nd the coecients a
0
, a
1
, and a
2
in Eq. (13):
j
q)
(X) = 0 at q
)
(X) = b
)
+ d
)
or a
0
(b
)
+ d
)
)
2
+ a
1
(b
)
+ d
)
) + a
2
= 0,
j
q)
(X) = 0.5 at q
)
(X) = (b
)
+ d
)
)}2 or a
0
_
b
)
+ d
)
2
_
2
+ a
1
_
b
)
+ d
)
2
_
+ a
2
= 0.5,
j
q)
(X) = 1 at q
)
(X) = b
)
or a
0
b
2
)
+ a
1
b
)
+ a
2
= 1.
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 173
The resulting quadratic constraint membership function can be represented as
j
q)
(X) =
_

_
0 if q
)
(X) b
)
+ d
)
,
(3b
2
)
+2b
)
d
)
+d
2
)
)q
)
(X)2b
)
q
)
(X)
2
(b
2
)
+d
2
)
)(b
)
+d
)
)
(b
2
)
d
2
)
)d
)
if b
)
q
)
(X) 6b
)
+ d
)
,
1 if q
)
(X) 6b
)
.
(15)
It is observed that for any specic example the optimum results obtained with linear and quadratic
constraint membership functions diered very little. This shows that the optimization solution is
relatively insensitive to the actual shape of the membership function. A possible reason is that the
assumed nonlinear membership function diered very little from the linear relationship.
3. Mixed discrete hybrid genetic algorithm
Genetic algorithms (GAs) are a family of algorithms that have the same basic structure, and dier
from one another with respect to several strategies and parameters that control the search. Often the
choice of parameters (e.g. population size, crossover probability and mutation probability) can have
a signicant impact on the eectiveness of GAs, i.e. each combination of the parameters may result
in a dierent optimum solution. It is a troublesome business to tune these adjustable parameters if
we want to obtain acceptable solutions. Another disadvantage of GAs is their high computational
cost, i.e. the number of function evaluations required for a global optimum result is very large,
especially with nite element type of analysis for large-scale problems.
In order to overcome these shortcomings of traditional GAs, a mixed-discrete hybrid genetic
algorithm (MDHGA) is proposed in this paper for solving nonlinear optimization problems with
mixed-discrete design variables. This algorithm has some distinguishable advantages compared with
other GAs, in terms of crossover, mutation and regeneration operations. The operators in a genetic
algorithm involve string copying, substring copying and occasional alteration of bits. A number of
studies have been made to establish not only an intuitive convergence of the method, but also a
rigorous mathematical convergence proof [27]. This characteristic of the GAs is used in this work
in the initial phase of the search. The MDHGA is mainly used to nd a small feasible region
surrounding the global optimum point. Subsequently, a discrete iterative method such as the branch-
and-bound method, penalty function approach and sequential linearization method can be employed to
nd the nal optimum result, as these deterministic iterative methods usually have faster convergence
speed and higher convergence eciency compared with the random search methods when searching
within a xed discrete space. Secondly, although the GAs converge to the global optimum point,
they require a large number of function evaluations to reach the actual optimum point. However, the
convergence of the method to the neighborhood of the global optimum point can be accomplished
in a reasonable number of function evaluations. Therefore, a gradual reduction of the feasible region
along a favorable search direction not only ensures the location of the optimal point always within
the shrink region but also reduces the computational expense such that the eectiveness and eciency
of the algorithm is improved.
The three kinds of design variables are handled using the following mapping functions:
(a) Discrete variables with equal spacing:
x
i
= x
L
i
+ (N
i
1)dsp
i
, i = 1, . . . , nq. (16)
174 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
(b) Discrete variables with unequal spacing:
x
i
= q
N
i
,inq
, nq i 6nd. (17)
(c) Continuous variables:
x
i
= x
L
i
+ (N
i
1)c
i
, nd i 6nn, (18)
where x
i
denotes the ith design variable, N
i
is the natural number corresponding to x
i
, q represents
the matrix of the values of discrete variables with unequal spacing.
Corresponding to each variable x
i
, only one value of N
i
exists. Thus, the problem of nding the
optimal design variables can be transformed into that of nding the optimal values of N
i
. Thereupon,
all operations of the iterative procedure are to determine suitable values of N
i
, which in turn, can
be used to obtain the physical values of the design variables.
The scheme of MDHGA can be summarized as follows:
1. Binary encoding of the problem: Each design variable N
i
is expressed as a nite length binary
digit string. These strings represent articial chromosomes. Every digit in the string is an articial
gene.
2. Initialization of population: The size of the population M is predened and xed throughout.
The initial population is created randomly within the feasible design space.
3. Evaluation of population: The tness is a quality value that is a measure of the reproductive
eciency of living creatures according to the principle of survival of the ttest. In the genetic
algorithm, the tness function is chosen as a measure of goodness to be maximized. The tness
function used to evaluate individual chromosomes is chosen as
[
)
[X] = [
)
[X] + R
m

i=1
max{0, q
i,)
}, ) = 1, . . . , M,
Fit
)
[X] =
1
|[
)
[X] C|
, ) = 1, . . . , M,
(19)
where R is a large enough positive constant, and C is a small enough negative constant. The bigger
the tness value is, the better the )th chromosome.
4. Selection: Some pairs of strings are randomly selected as parents to reproduce osprings according
to the selection rule related to their tness values. For this, the candidates are sorted according
to their tness. The crossover rate is predened such that a xed proportion of the population is
randomly generated. For each random number, it is transformed into an ordinal number of the ranked
candidates, and the corresponding string is selected as crossover parents of the next generation.
5. Crossover: Three dierent operators are often considered for crossover strategy. They are single-
point crossover, two-point crossover and multi-point crossover. In practical applications, it was found
that when the population size is large enough (n100), there is no notable dierences in the
eciencies of the solutions among these three strategies. If the population size is comparatively
small, the multi-point crossover has the highest eciency, followed by the two-point crossover, with
the one-point crossover having the lowest eciency. Therefore, the multi-point crossover strategy is
selected; it can be stated as follows: Multi points are randomly chosen, which cut the binary strings
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 175
of parents into several segments. Some segments of father string can be exchanged with those of
mother string. An example is as follows:
Parents Ospring
01010010 01100001
10101001 10011010
6. Mutation: Mutation is the occasional random alteration on a bit-by-bit basis. Similar to the
crossover, multi-point random mutation proved to be the best choice in this step. An example is:
11001001 11111010.
7. Population regeneration: Combine parent population and their osprings into a whole population,
sort them in the order from the highest tness value to the lowest, and select the rst M chromosomes
with better values of tness to obtain a new population. The new population is then used to repeat
steps 47 until the population generated tends to have a stable tness value; and then go to step 8.
8. Determination of search direction: Once a stable population is obtained, it is time to nd an
iterative search direction along which the optimum point can be further approached. Considering the
population as vertices of a composite polygon, the centroid (X
c
) of all chromosomes except the one
with the lowest tness value, is dened as
X
c
=
1
k 1
k

i=1
X
i
, X
i
= X
w
, (20)
where, X
w
is the chromosome with the worst (lowest) tness value. A favorable search direction
can then be determined as

d =
X
c
X
w
X
c
X
w

. (21)
This equation takes all the individual chromosomes into account, and better reects the ecological
environment of the whole population. Hence, it can achieve a better result.
9. Shrinking of feasible search region: The best chromosome is taken as the center of the search
region, and the feasible region is reduced along the favorable search direction

d, ensuring that the
optimum point always lies within the reduced region.
Set r
i1
as the range of the ith variable before the region is shrunk, r
i2
as the range after the region
is shrunk; r
+
i
as the range along the +

d direction, and r

i
as the range along the

d direction, so
that
r
i2
= r
+
i
+ r

i
. (22)
Dening

i
= r
i2
}r
i1
, (23)
[
i
= r
+
i
}r

i
, (24)
176 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
where r
+
i
and r

i
can be expressed as
r
+
i
= (
i
[
i
r
i1
)}([
i
+ 1), (25)
r

i
= (
i
r
i1
)}([
i
+ 1). (26)
Eqs. (22)(26) indicate that if
i
and [
i
are known, r
+
i
and r

i
can be derived, which means that
the range of r
i2
will also be known. In practice, the value of [
i
may be taken as [
i
=2, and the
value of
i
is limited to 01. In this work
i
is chosen according to the following equations:
d
min
= min{|d
i
|, i = 1n}, (27)

i
= min{1.0, RA|d
i
|}d
min
}, i = 1n, (28)
where d
i
is the ith component of the vector

d, and RA is the value of
i
corresponding to d
min
. RA
is usually chosen in the range of 6090%, and it is chosen as RA=85% in this work.
After the new shrunk region is determined, the search goes back to step 2 to repeat the iterative
procedure.
10. Finding the optimal solution: When the search region shrinks to a space within an acceptable
precision, the population tness value will become stable. The best chromosome of the population
is chosen as the starting point for the deterministic discrete iterative methods, which then are used
to replace GA to nd the nal optimum result. The hybrid negative sub-gradient or PRP conjugate
gradient method combined with discrete one-dimensional search, is used for this search procedure, as
they usually have faster convergence speed and higher convergence eciency compared to random
search methods when searching within a deterministic discrete space.
4. Computational procedure
1. Starting from any trial design vector X
0
, minimize the objective function [(X) subject to the con-
straints q
)
(X)6b
)
and subject to the constraints q
)
(X)6b
)
+d
)
, ) =1, 2, . . . , m, respectively, using
mixed-discrete hybrid genetic algorithm (MDHGA) proposed in this paper. Let the corresponding
optimum values be [
max
d
and [
min
d
.
2. Construct the fuzzy objective membership function as
j
[
(X) =
_

_
0 i (X) [
max
d
,
[
max
d
[(X)
[
max
d
[
min
d
if [
min
d
[(X) 6[
max
d
,
1 if [(X) 6[
min
d
.
(29)
3. Represent the fuzzy contraint membership functions using linear relationships (of course, other
types of relationships are also applicable) as
j
q)
(X) =
_

_
0 if q
)
(X) b
)
+ d
)
,
b
)
+ d
)
q
)
(X)
d
)
if b
)
q
)
(X) 6b
)
+ d
)
,
1 if q
)
(X) 6b
)
.
(30)
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 177
4. Formulate the mixed-discrete fuzzy optimization problems using the z-formulation as
minimize z (31)
s.t. z 6j
[
(X),
z 6j
q)
(X) ) = 1, . . . , m,
X = [x
1
, . . . , x
nq
, . . . , x
nd
, . . . x
n
]
T
,
x
L
i
6x
i
6x
U
i
; i = 1, 2, . . . , n.
(32)
5. Solve this problem using the MDHGA program again to obtain the maximum value of z. The
values of X and [(X), corresponding to the maximum value of z, denote the optimum solution
of the problem.
4.1. Program ow diagram
On the basis of MDHGA program structure, a MDFNLP computer program has been developed to
solve fuzzy nonlinear programming problems with mixed-discrete design variables. The ow diagram
of the program is given in Fig. 3.
This gure shows that when solving a mixed-discrete optimization problem with fuzzy information,
the MDHGA program is used twice during the solution procedure. In the rst time, it is used is
to obtain [
max
d
and [
min
d
, the optimum objective values of [(X) in the mixed-discrete design space
with and without fuzziness. In the second time, it is used to obtain the true optimum values of X
and [(X) of the problem. Similarly, the mathematical modeling le and the input data le are also
established twice; once for the crisp information, and the second time for the fuzzy information.
5. Illustrative examples
5.1. Design of a welded beam
This classic example, originally from Ref. [24], has been used in a study of the mixed-discrete
programming algorithms in [13,14]. The problem is to nd the minimum cost design of a structural
welded beam shown in Fig. 4, with constraints on weld stresses, buckling loads, bar deections, and
bending stresses. There are four design variables in the problem, and they are the weld thickness h,
the weld length !, the bar thickness t and the bar breadth b.
5.1.1. Objective function
The minimization of the cost is considered as the objective of the problem. The major cost
components of such a welded beam include: (a) set-up labor cost, (b) welding labor cost, and
(c) material cost. Using the design vector
X = [x
1
, x
2
, x
3
, x
4
]
T
= [t, b, h, !]
T
the objective function can be obtained as
[
1
( x) = (1 + c
1
)x
2
3
x
4
+ c
2
x
1
x
2
(L + x
4
). (33)
178 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
Fuzzy mathematical
modeling module
Crisp mathematical
modeling module
Crisp data
input file
Contain f
d
max
and f
d
min
Users Interface
Program
MDFNLP Main Program
Contain fuzzy
information
Yes
MDHGA module
No
Yes
No
Fuzzy data
input file
Optimum
solution
without fuzzy
information
MDHGA module
Optimum solution
with fuzzy
information
Start
Stop
Fig. 3. Flow diagram of the MDFNLP program.
Fig. 4. A welded beam.
5.1.2. Functional constraints
(a) Maximum shear stress in weld: t(x)
t( x) = [(t

)
2
+ 2t

cos 0 + (t

)
2
]
1}2
(34)
with
t

= F}
_

2x
3
x
4
_
,
t

= MR}J,
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 179
M = F(L + x
4
}2),
R = {x
2
4
}4 + [(x
1
+ x
3
)}2]
2
}
1}2
,
J = 2 0.707x
3
x
4
{x
2
4
}12 + [(x
1
+ x
3
)}2]
2
},
cos 0 = x
4
}(2R).
(b) Maximum normal stress in beam: o(x)
o( x) = 6FL}(x
2
x
2
1
). (35)
(c) Buckling load of the beam: P
c
(x)
P
c
( x) =
4.013

EI:
L
2
_
1
x
1
2L
_
EI
:
_
,
I =
1
12
x
1
x
3
2
(36)
with
: =
1
3
Gx
1
x
3
2
.
(d) End deection of the beam: DEL(x)
DEL( x) = 4FL
3
}(Ex
3
1
x
2
). (37)
The optimization problem can be formulated as follows:
Min [
1
( x) = (1 + c
1
)x
2
3
x
4
+ c
2
x
1
x
2
(L + x
4
) (38)
s.t. q
1
( x) = t( x) t
d
60,
q
2
( x) = o( x) o
d
60,
q
3
( x) = x
3
x
2
60,
q
4
( x) = F P
c
( x) 60,
q
5
( x) = 0.125 x
3
60,
q
6
( x) = DEL( x) 0.25 60,
(39)
where c
1
= 0.37 0.283, c
2
= 0.17 0.283, L = 14, F = 6000, t
d
= 13600, o
d
= 30000, E =
3e7, G = 12e6.
5.1.3. Results and analysis
Case A: The variables t and b are considered as discrete variablesinteger multiples of 0.5; and
h and ! are treated as integers. The optimum solution obtained is given in Table 1.
Case B: For the optimal design of the problem in a fuzzy environment, an allowable deviation
of 10% for each design constraint is considered a fuzzy transition zone, which is represented by a
linear membership function. Table 2 shows the nal designs obtained using the MDFNLP program.
The optimum design vector and the optimum objective function value can be seen to be
X = (4.0, 1.0, 1, 2) and [(X) = 5.28846.
180 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
Table 1
Crisp optimization results of the welded beam
Quantity Starting point Continuous solution Optimal solution Remarks
x
1
10.0 8.29 4.5 Discrete
x
2
5.0 0.24 1.0 Discrete
x
3
5 0.24 1 Integer
x
4
10 6.22 2 Integer
[(X) 333.9095 2.38 5.67334
Infeasible point
Table 2
Fuzzy optimization results of the welded beam
Quantity Starting point [
max
d
[
min
d
([
min
d
, [
max
d
) Remarks
MDNLP MDNLP MDFNLP
x
1
10.0 4.5 4.0 4.0 Discrete
x
2
5.0 1.0 1.0 1.0 Discrete
x
3
5 1 1 1 Integer
x
4
10 2 2 2 Integer
[(X) 333.9095 5.67334 5.28846 5.28846
Infeasible point z = 0.5
From Tables 1 and 2, it also can be found that [
max
d
=5.67334, [
min
d
=5.28846, [
max
c
=2.38, and
the four optimum values of the problem are in the order:
[
min
c
[
max
c
[
min
d
[
max
d
and no overlap exists between intervals [[
min
c
, [
max
c
] and [[
min
d
, [
max
d
]. If the ([
min
c
, [
max
c
) combination
is used in the formulation, the optimum solution [

obtained will be
[
min
c
6[

6[
max
c
= 2.38.
Actually, the true mixed-discrete optimum solution with fuzziness can only be generated from within
the interval [[
min
d
, [
max
d
], i.e.
5.28846 = [
min
d
6[

6[
max
d
= 5.67334.
Thus, the objective membership function used in the fuzzy mixed-discrete nonlinear programming
should be composed of ([
min
d
, [
max
d
). This example veries that Shih and Lais declaration is not
accurate whereas the proposed suggestion leads to an accurate and practical solution of the problem.
5.2. Design of a pressure vessel
A cylindrical pressure vessel capped at both ends by hemispherical heads, as presented by Sandgren
[29], is shown in Fig. 5 and is used as a compressed air storage tank. The problem of design of the
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 181
Fig. 5. Pressure vessel.
vessel is formulated according to the ASME boiler and pressure vessel code. The design variables
are specied as
X = [x
1
, x
2
, x
3
, x
4
]
T
= [1
s
, 1
h
, R, L]
T
which correspond, respectively, to the shell thickness, spherical heads thickness, shell radius and
shell length. The nonconvex objective function is to minimize the total cost of material, forming
and welding of the pressure vessel. The mathematical model of the problem is stated as
min [
2
(X) = 0.6224x
1
x
3
x
4
+ 1.7781x
2
x
2
3
+ 3.1611x
2
1
x
4
+ 19.8621x
2
1
x
3
(40)
s.t. q
1
(X) = x
1
0.0193x
3
0,
q
2
(X) = x
2
0.00954x
3
0,
q
3
(X) = x
2
3
x
4
+
4
3
x
3
3
750 1728 0,
q
4
(X) = 240 x
4
0,
q
5
(X) = x
1
1.1 0,
q
6
(X) = x
2
0.6 0,
(41)
where 1
s
and 1
h
represent discrete values, as integer multiples of 0.0625 in, while R and L are
continuous variables.
For the fuzzy problem, each of the behavior design constraints is assumed to have a 10% transition
zone. Assuming the preference values for dierent settings of the constraints in the transition zone
to be nonlinear, the quadratic constraint membership functions are used as indicated in Eq. (15).
The optimum results of the fuzzy problem in mixed-discrete design space are shown in Table 3,
together with the values reported in Ref. [31].
The third column in Table 3 shows the optimum values in the mixed-discrete design space without
fuzziness. This result can be seen to be better than the optimum result obtained using the objective
membership function composed of ([
min
c
, [
max
c
). The results reported in [31] can be seen to be
not reliable, as well as infeasible. Ref. [31] also reported another improved result of [

=7107.4,
obtained with a dierent starting point. These two results are found to be worse than the solution
obtained using MDFNLP with the choice of ([
min
d
, [
max
d
) (shown in column 5 of Table 3).
5.3. Design of a reinforced concrete beam
This problem is to optimize the total cost of concrete and reinforcing steel of the beam shown in
Fig. 6. It was previously solved by Amir and Hasegawa as a mixed-discrete variable problem, and by
182 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
Table 3
Optimal results of the pressure vessel
Quantity Starting point [
max
d
[
min
d
[
min
d
[
max
d
[
min
c
[
max
c
Remarks
MDNLP MDNLP MDFNLP [31]
x
1
1.25 1.1875 1.125 1.125 1.125 Discrete
x
2
0.625 0.625 0.625 0.625 0.625 Discrete
x
3
50 61.44833 60.5984 57.34544 56.075 Continuous
x
4
120 27.40367 20.30786 42.3119 59.127 Continuous
[(X) 9590.7 7284.0208 6547.1744 6964.3398 7462.1
z 0.434
Fig. 6. Reinforced concrete beam.
Shih and Lai [31] as a mixed-discrete fuzzy programming problem using modied branch-and-bound
approach. The design variables are dened as follows: The depth of the concrete beam, represented
by x
1
, is assumed to be an integer variable. The cross-sectional area of the reinforcement, x
2
, is
considered as a discrete variable, which can only take certain predetermined values from Table 4.
The variable x
3
, representing the width of beam, is treated as a continuous variable.
The beam is assumed to be simply supported with a span of 30 ft and subjected to a live load of
2.0 klb
f
and a dead load of 1.0 klb
f
, which includes the weight of the beam. The beam should be
proportioned to satisfy the strength requirements specied by the ACI Building Code 318-11 (1977).
The optimization problem can be stated in a simplied mathematical form as
min [
3
(X) = 29.4x
2
+ 0.6x
1
x
3
(42)
s.t. q
1
(X) = (x
1
}x
3
) 4.0 60,
q
2
(X) = 180.0 + 7.735(x
2
2
}x
3
) x
1
x
2
60,
X = [x
1
, x
2
, x
3
]
T
.
(43)
The optimal design in fuzzy environment is considered to be consistent with Shih and Lais report
using an allowance of 10% of each design constraint for the fuzzy transition zone. This transition
zone is represented by a linear membership function. The optimum solutions obtained using the
MDFNLP program are shown in Table 5, and the results reported by Shih and Lai [31] are shown
in Table 6.
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 183
Table 4
The discrete design space for reinforced concrete beam
No. Bar type x
2
, area (in
2
) No. Bar type x
2
, area (in
2
)
1 1#4 0.20 40 5#8 3.95
2 1#5 0.31 41 9#6 3.96
3 2#4 0.40 42 4#9 4.00
4 1#6 0.44 43 13#5 4.03
5 3#4,1#7 0.60 44 7#7 4.20
6 2#5 0.62 45 1435 4.34
7 1#8 0.79 46 10#6 4.40
8 4#4 0.80 47 15#5 4.65
9 2#6 0.88 48 6#8 4.74
10 3#5 0.93 49 8#7 4.80
11 5#4,1#9 1.00 50 11#6 4.84
12 6#4,2#7 1.20 51 5#9 5.00
13 4#5 1.24 52 12#6 5.28
14 3#6 1.32 53 9#7 5.40
15 7#4 1.40 54 7#8 5.53
16 5#5 1.55 55 1338 5.72
17 2#8 1.58 56 10#7,6#9 6.00
18 8#4 1.60 57 14#6 6.16
19 4#6 1.76 58 8#8 6.32
20 9#4,3#7 1.80 59 15#6,11#7 6.60
21 6#5 1.86 60 7#9 7.00
22 10#4,2#9 2.00 61 9#8 7.11
23 7#5 2.17 62 12#7 7.20
24 11#4,5#6 2.20 63 13#7 7.80
25 3#8 2.37 64 10#8 7.90
26 12#4,4#7 2.40 65 8#9 8.00
27 8#5 2.48 66 14#7 8.40
28 13#4 2.60 67 11#8 8.69
29 6#6 2.64 68 15#7 9.00
30 9#5 2.79 69 12#8 9.48
31 14#4 2.80 70 13#8 10.27
32 15#4,5#7,3#9 3.00 71 11#9 11.00
33 7#6 3.08 72 14#8 11.06
34 10#5 3.10 73 15#8 11.85
35 4#8 3.16 74 12#9 12.00
36 11#5 3.41 75 13#9 13.00
37 8#6 3.52 76 14#9 14.00
38 6#7 3.60 77 15#9 15.00
39 12#5 3.72
Because the four optimum values of the problem are ordered as
[
min
c
[
min
d
[
max
c
[
max
d
and the optimum objective value is found to be within the interval [[
min
d
, [
max
c
], Shih and Lais
statement might be valid for the problem. But only after nding the nal results, one can tell
184 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
Table 5
Optimal results of the reinforced beam through MDFNLP
Quantity Starting point [
max
d
[
min
d
[
min
d
[
max
d
Remarks
MDNLP MDNLP MDFNLP
x
1
30 34 34 35 Integer
x
2
5.00 6.32 5.72 5.72 Discrete
x
3
6.0 8.8643 7.80 8.47548 Continuous
[(X) 255.0 366.640 327.275 346.153
Infeasible point z =0.463
Table 6
Comparison with optimal results by Shih and Lai
Quantity [
max
c
[
min
c
[
max
d
[
min
c
[
min
c
[
max
c
[
min
d
[
max
d
Remarks
x
1
34 34 Integer
x
2
6.00 6.16 Discrete
x
3
8.258 8.091 Continuous
[(X) 361.09 326.01 364.86 327.115 344.865 346.165
whether their method is correct or not. Also it has been found that the results obtained by the author
using ([
min
d
, [
max
d
) are very close to those found using ([
min
c
, [
max
c
). The latter results were obtained
by using several dierent starting points. Therefore, it can again be concluded that the best choice
for the objective membership function is the combination of ([
min
d
, [
max
d
).
6. Conclusion
A synthetic optimization method (MDFNLP) that combines the fuzzy z-formulation with a hybrid
genetic algorithm is presented in the paper. This method can nd a globally compromise solution
for fuzzy optimization problems containing mixed-discrete design variables, even when the objective
functions are nonconvex and}or nondierentiable. As expected, the illustrative examples indicate that
more reliable and satisfactory results can be obtained through the MDFNLP. An error from the early
research work is corrected and the right conclusion has been made that the correct objective mem-
bership function to be used in the fuzzy mixed-discrete nonlinear programming should be composed
of ([
min
d
, [
max
d
) that are obtained in the discrete design space.
References
[1] A.R. Abdelaziz, S.A.M. Gouda, Geno-fuzzy optimization of power system reliability, Proc. Mediterranean
Electrotechnical Conf.MELECON, 10th Mediterranean Electrotechnical Conf., 2000, pp. 991994.
[2] J.S. Arora, M.W. Huang, Methods for optimization of nonlinear problems with discrete variables: a review, Struct.
Optim. 8 (1994) 6985.
[3] R.E. Bellman, L.A. Zadeh, Decision making in a fuzzy environment, Management Sci. 17 (1970) 141164.
Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186 185
[4] M. Bremicker, P.Y. Papalambros, H.T. Loh, Solution of mixed-discrete structural optimization problems with a new
sequential linearization algorithm, Comput. Struct. 37 (1990) 451461.
[5] J. Cai, G. Thierauf, Discrete optimization of structures using an improved penalty function method, Eng. Optim. 17
(1993) 293306.
[6] Y.-D. Chen, A. Hui, R. Du, Fuzzy expert system for the design of machining operations, Internat. J. Mach. Tools
Manuf. 35 (12) (1995) 16051621.
[7] L. Chen, S.S. Rao, A fuzzy approach for the manipulation of uncertainties in determination of optimum machining
conditions under multiple criteria, J. Manuf. Sci. Eng. 119 (2) (1997) 186192.
[8] B.K.-S. Cheung, A. Langevin, H. Delmaire, Coupling genetic algorithm with s grid search method to solve mixed
integer nonlinear programming problems, Comput. Math. Appl. 34 (12) (1997) 1323.
[9] A.K. Dhingra, S.S. Rao, Integrated kinematic-kinetostatic approach to optimal design of planar mechanisms using
fuzzy theories, J. Mech. Transmissions, Automat. Des. 113 (3) (1991) 306311.
[10] D. Dubois, P. Fortemps, Computing improved optimal solutions to maxmin exible constraint satisfaction problems,
European J. Oper. Res. 118 (1999) 95126.
[11] J.F. Fu, R.G. Fenton, W.L. Cleghorn, A mixed integer-discrete-continuous programming method and its application
to engineering design optimization, Eng. Optim. 17 (1991) 263280.
[12] X.-H. Guan, W.-H.E. Liu, A.D. Papalexopoulos, Application of a fuzzy set method in an optimal power ow, Electric
Power Systems Res. 34 (1995) 1118.
[13] O.K. Gupta, A. Ravindran, Nonlinear integer programming and discrete optimization, ASME J. Mech. Transmissions
Automat. Des. 105 (1983) 160164.
[14] P. Hajela, C.J. Shih, Optimal design of laminated composites using a modied mixed integer and discrete
programming algorithm, Comput. Struct. 32 (1) (1989) 213221.
[15] U.D. Hanebeck, G.K. Schmidt, Genetic optimization of fuzzy networks, Fuzzy Sets and Systems 79 (1) (1996)
5968.
[16] S.H. Huang, H.C. Zhang, S. Sun, H.H. Li, Function approximation and neural-fuzzy approach to machining process
selection, IEEE Trans. Components Packaging Manuf. Technol. 19 (1) (1996) 918.
[17] O. Jonsson, T. Larsson, Lagrangean relaxation and subgradient optimization applied to optimal design with discrete
sizing, Eng. Optim. 16 (1990) 221233.
[18] R.K. Kincaid, S.L. Padula, Minimizing distortion and internal forces in truss structures by simulated annealing,
AIAA-90-1095-CP (1990) 327333.
[19] H.-B. Kuntze, M. Sajidman, A. Jacubasch, Fuzzy-logic concept for highly fast and accurate position control of
industrial robots, Proc. 1995 IEEE Internat. Conf. on Robotics and Automation Part 1 (of 3), 1995, pp. 11841190.
[20] H.T. Loh, P.Y. Papalambros, A sequential linearization approach for solving mixed-discrete nonlinear design
optimization problems, ASME J. Mech. Des. 113 (1991) 325334.
[21] H.T. Loh, P.Y. Papalambros, Computational implementation and tests of a sequential linearization algorithm for
mixed-discrete nonlinear design optimization, ASME J. Mech. Des. 113 (1991) 335344.
[22] A.M. Luciano, D. Lauria, E. Napoli, Algorithm for automatic generation of fuzzy rules applied to power system
controllers, IEE Proc. Generat. Transmission Distrib. 145 (2) (1998) 161167.
[23] J.A. Momoh, Y. Makarov, M. Elfayoumy, Articial neural network, fuzzy logic and expert systems approaches to
hybrid electric vehicle control system, NASA Conf. Publication, Proc. 1998 HBCUs Research Conf., 1998.
[24] K.M. Ragsdell, D.T. Phillips, Optimal design of a class of welded structures using geometric programming, ASME
J. Eng. Ind. 98 (3) (1976) 10211025.
[25] S. Rajeev, C.S. Krishnamoorthy, Discrete optimization of structures using genetic algorithms, ASCE J. Struct. Eng.
118 (1992) 12331250.
[26] S.S. Rao, L. Chen, Determination of optimal machining conditions: a coupled uncertainty model, J. Manuf. Sci. Eng.
122 (1) (2000) 206214.
[27] G. Rudolph, Convergence analysis of canonical genetic algorithms, IEEE Trans. Neural Network 5 (1) (1994)
96101.
[28] E. Salajegheh, G.N. Vanderplaats, Optimum design of trusses with sizing and shape variables, Struct. Optim. 6
(1993) 7985.
[29] E. Sandgren, Nonlinear integer and discrete programming in mechanical design optimization, ASME J. Mech. Des.
112 (1990) 223229.
186 Y. Xiong, S.S. Rao / Fuzzy Sets and Systems 146 (2004) 167186
[30] W. Sheng, D. Miao, R. Dai, Two-level optimization strategy for fuzzy control design, Proc. 1997 IEEE Internat.
Conf. on Intelligent Processing Systems, ICIPS97 Part 1 (of 2), 1998, pp. 276280.
[31] C.J. Shih, T.K. Lai, Mixed-discrete fuzzy programming for nonlinear engineering optimization, Eng. Optim. 23
(1995) 187199.
[32] D.K. Shin, Z. Gurdal, O.H. Grin Jr., A penalty approach for nonlinear optimization with discrete design variables,
Eng. Optim. 16 (1990) 2942.
[33] J.M. Sousa, R. Babuska, H.B. Verbruggen, Fuzzy predictive control applied to an air-conditioning system, Control
Eng. Pract. 5 (10) (1997) 13951406.
[34] R. Sutton, Optimization of fuzzy autopilots, Proc. 1997 11th Ship Control Systems Symp. Part 1 (of 2), 1997, pp.
6376.
[35] J. Tang, D. Wang, Modeling and optimization for a type of fuzzy nonlinear programming problems in manufacturing
systems, Proc. 35th IEEE Conf. on Decision and Control, 1996, pp. 44014405.
[36] R.S.V. Teegavarapu, S.P. Simonovic, Modeling uncertainty in reservoir loss functions using fuzzy sets, Water Resour.
Res. 35 (9) (1999) 28152823.
[37] H. Xiang, B.P. Verma, G. Hoogenboom, Fuzzy irrigation decision support system, Proc. 12th National Conf. on
Articial Intelligence Part 2 (of 2), Seattle, WA, 1994.
[38] L. Zadeh, Fuzzy sets, Inform. Control 8 (1965) 338353.
[39] D. Zhao, J. Zhu, Fuzzy variable-structure-adaptive control design for robot manipulators, CIRP Ann. Manuf. Technol.
44 (1) (1995) 363368.
[40] J.-X. Zhu, M.-Y. Chow, Review of emerging techniques on generation expansion planning, IEEE Trans. Power
Systems 12 (4) (1997) 17221728.

You might also like