Stochastics and Statistics
An ecient and exible mechanism for constructing membership
functions
Andrees L. Medaglia, Shu-Cherng Fang
*
, Henry L.W. Nuttle, James R. Wilson
Department of Industrial Engineering & Graduate Program in Operations Research,
College of Engineering, Operations Research and Industrial Engineering,
North Carolina State University, Raleigh, NC 27695-7906, USA
Received 13 September 2000; accepted 27 March 2001
Abstract
This paper introduces a Beezier curve-based mechanism for constructing membership functions of convex normal
fuzzy sets. The mechanism can t any given data set with a minimum level of discrepancy. In the absence of data, the
mechanism can be intuitively manipulated by the user to construct membership functions with the desired shape. Some
numerical experiments are included to compare the performance of the proposed mechanism with conventional
methods. 2002 Elsevier Science B.V. All rights reserved.
Keywords: Fuzzy sets; Membership functions; Beezier curves; Measures of information; Fuzzy numbers
1. Introduction
Fuzzy set theory was rst introduced in the
1960s by Zadeh [15] as a way to capture uncer-
tainty and vagueness often overlooked in complex
systems. It was pointed out that fuzzy set theory is
a generalization of the classical set theory.
A fuzzy set
~
AA is characterized by its membership
function l
~
AA
, which maps each element of the uni-
verse X to the interval 0; 1. This function indi-
cates the degree of belonging to
~
AA for each element
of X. One of the most important concepts of fuzzy
sets is the concept of an a-cut. Given a fuzzy set
~
AA
dened on X and a 2 0; 1, the a-cut is dened as
a
~
AA fx 2 X : l ~
AA
x Pag. For continuity pur-
poses, we take
0
~
AA lim
a!0
a
~
AA. A fuzzy set
~
AA is
convex if and only if each of its a-cuts is a convex
set. A fuzzy set
~
AA is normal if
1
~
AA 6 ;.
Even though there is no universal agreement on
the question, Dombi [4] reported that there are
some characteristics shared by the majority of
continuous membership functions found in the
literature. Among others, there is an apparent
demand for membership functions with the fol-
lowing properties: they should be piecewise
monotone nonincreasing or nondecreasing; they
European Journal of Operational Research 139 (2002) 8495
www.elsevier.com/locate/dsw
*
Corresponding author. Tel.: +1-919-515-2192; fax: +1-919-
515-5281.
E-mail address:
[email protected] (S.-C. Fang).
URL: www.ie.ncsu.edu/fangroup.
0377-2217/02/$ - see front matter 2002 Elsevier Science B.V. All rights reserved.
PII: S0377- 2217( 01) 00157- 6
should achieve null and full membership for at
least two dierent elements in the universal set;
and they should be able to represent fuzzy convex
sets. Commonly seen examples are the simple tri-
angular, trapezoidal, and bell-shaped membership
functions.
Problem formulations based on fuzzy sets can
have greater expressive power than their counter-
parts based on crisp sets, but the applicability of
fuzzy technology depends on the ability to con-
struct membership functions that appropriately
represent various concepts in dierent contexts [8].
To fully exploit the benets provided by fuzzy
technology, we need an ecient membership
function generating mechanism with the following
desirable characteristics:
1. Accurate. In the presence of data, the resulting
membership functions should reect the knowl-
edge contained in the data in the most accurate
way possible. Data in the form of membership
values for points in the universe is usually ob-
tained from experts.
2. Flexible. The methodology should provide a
broad family of membership functions.
3. Computationally aordable. The method should
be computationally tractable in order to be of
any practical use. Medasani [10] has highlighted
the importance of having membership functions
that can be easily tuned and adjusted. Other
authors have expressed the need for methods
in which computer graphics can facilitate the
process of constructing membership functions
by allowing the user an easy and direct manip-
ulation of dierent shapes [1].
4. Easy to use. Once a membership function has
been generated, it should be easy to nd l
~
AA
x
for a given x; and it should be easy to nd
a
~
AA
for a given a.
In this paper we propose a mechanism that
exploits the properties of Beezier curves to address
these issues and to provide the user with a exible
and ecient way of generating membership func-
tions.
The paper is organized as follows. In Section 2,
we review the basic techniques used for generating
membership functions. Section 3 describes the
proposed mechanism and some fundamental de-
nitions and properties of Beezier curves. In Section 4,
test problems found in the literature are used to
illustrate the proposed mechanism and compare its
performance with that of two methods which ap-
pear in the literature. Finally, conclusions and
current research directions are given in Section 5.
2. Membership function generation
2.1. Overview
Membership functions can be constructed from
data when it is available. This data can be elicited
by interacting with experts using a direct approach
(or direct rating) [8,11,13]. The direct approach
requires the degree of membership of a collection
of points in the universal set. A membership
function that describes the underlying concept is
tted to the collected data points. This is known as
data-driven membership function estimation.
Sometimes this approach can be overly precise in
capturing subjective judgment. By formulating
easier and simpler questions, knowledge can also
be acquired through an indirect approach. We will
not deal with the indirect approach in this paper,
but the reader is referred to the paper by Chameau
and Santamarina [1] and the book by Klir and
Yuan [8].
When data are not available in the form of
value-membership pairs, a membership function
has to be constructed subjectively. In this case, the
conventional approach is to rst pick the shape of
the membership function from a list of families,
and then to ne-tune the values of the parameters
of that function. It is always desirable to have a
parsimonious, meaningful parameterization of
membership functions [4].
2.2. Current methods
In the literature, fuzzy sets are commonly
modeled by triangular, trapezoidal, and bell-
shaped membership functions. However, other
parameterized functional shapes are useful in
particular situations. More details can be found in
Dombi [4] and Medasani et al. [10].
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 85
An eort to create a broad class of functions
was made by Zysno [17] and Zimmermann and
Zysno [16]. In their model, the membership func-
tion for a fuzzy set
~
AA is given by
l
~
AA
x mid 0;
1
1 e
axb
_ _
c
_
1
d
1
2
; 1
_
8x 2 X R; 1
where
a; b 2 R; 0 6c 61; and 0 6d 62 min1 c; c:
The function mid0; f x; 1 is dened such that
mid0; f x; 1 f x; if 0 6f x 61; mid0; f x; 1 0;
if f x < 0; and mid0; f x; 1 1;
if f x > 1:
Even though the model provides the user with a
commonly used family of S-shapes, the determi-
nation of the parameters from empirical data po-
ses some problems and there is no direct numerical
method for optimal parameter estimation [16,17].
The model may be used for estimating membership
functions subjectively, with the parameters a, b, c,
and d being xed by the expert.
Dombi [4] proposed a model with properties
similar to the one presented by Zysno and Zim-
mermann. In his model a membership function for
fuzzy set
~
AA is constructed using the S-shaped
monotonically increasing function
l
~
AA
x
1 m
k1
x a
k
1 m
k1
x a
k
m
k1
b x
k
2
and/or the S-shaped monotonically decreasing
function
l
~
AA
x
1 m
k1
b x
k
1 m
k1
b x
k
m
k1
x a
k
; 3
where x 2 a; b; a; b 2 R; the steepness is given by
k P1; and the inection point is determined by
0 < m < 1. When data are available, Dombi pro-
posed a method for estimating the parameters
based on linearized forms of (2) and (3).
Both of these models provide similar member-
ship functions because they use the same under-
lying form, i.e., l ~
AA
x 1=1 dx, where dx is
a measure of distance. Even though these models
provide exibility for estimating S-shaped func-
tions, they fail to provide more general monotonic
curves.
Chen and Otto [2] present a novel method for
constructing membership functions using interpo-
lation and measurement theory. Following a sys-
tematic approach, their method is able to
construct general monotonic functions from data.
However, their methodology does not provide a
mechanism for adjusting or building a membership
function in the absence of data.
In the area of fuzzy system identication, so-
phisticated methods based on neural networks and
evolutionary algorithms have been proposed to
generate and tune both fuzzy rules and member-
ship functions. However, they are basically case by
case approaches [7,9].
In the following section we shall introduce an
interactive and ecient approach for both data-
driven and subjective estimation of membership
functions. Based on Beezier curves, the method is
able to generate a broad family of functions.
3. Proposed mechanism
3.1. B eezier curves
One of the major breakthroughs in computer
aided design (CAD) is the theory of Beezier curves
and surfaces, independently developed by P. de
Casteljau and P. Beezier while working for the
French automakers Citroeen and Renault, respec-
tively [6].
The theory of Beezier curves provides a mathe-
matical foundation for representing a smooth
curve that passes through the vicinity of a set of
control points. Denition 1 gives a formal expres-
sion of a Beezier curve in terms of Bernstein poly-
nomials.
Denition 1. A Beezier curve with n 1 control
points p,p
0
; . . . ; p
n
is given by
f t; n; p ,
n
k0
B
n;k
tp
k
;
86 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495
where t 2 0; 1, p
k
,x
k
; y
k
T
, and B
n;k
t
n
k
_ _
t
k
1 t
nk
are the Bernstein polynomials. Since
f t; n; p 2 R
2
, we usually denote f t; n; p f
x
t; n; x; f
y
t; n; y
T
, where x,x
0
; . . . ; x
n
T
, y,
y
0
; . . . ; y
n
T
.
Beezier curves have several properties that are
particularly useful in the context of this paper [6].
Property 1. The Beezier curve f t; n; p defined over
t 2 0; 1, lies in the convex hull of the polygon de-
fined by the control points p,p
0
; . . . ; p
n
.
Property 2. The Bernstein polynomial B
n;k
t
achieves its unique maximum at t k=n. If the
control point p
k
is moved, then the curve is mostly
affected in the region around the parameter t k=n.
Property 3. The Beezier curve interpolates its first
(p
0
) and last (p
n
) control points. In other words,
f 0; n; p p
0
and f 1; n; p p
n
.
These properties have practical eects in the
curve design process. Property 1 guarantees that
the curve will not fall outside the control poly-
gon. By using this property along with Property
2, a Beezier curve can be designed by exaggerating
the target shape using the control polygon. Even
though a single control point displacement will
change the whole curve, this pseudo-local con-
trol property gives us the sense that the control
points work locally as magnets on the curve.
Property 3 is very useful for breaking the con-
struction of a complex curve into simpler parts.
A complete discussion on Beezier curves and its
properties can be found in the book by Farin [6].
3.2. Mathematical framework
In this section we give the mathematical
framework of a broad family of membership
shapes based on Beezier curves.
Let
~
AA be a fuzzy set on the universal set X. The
following conditions are commonly required for its
membership function, l ~
AA
.
Condition 1. The membership function l ~
AA
is a
mapping from the universal set X to 0; 1, i.e.,
l ~
AA
: X ! 0; 1.
Condition 2. There exist x
1
; x
2
2 X such that
l
~
AA
x
1
1 and l
~
AA
x
2
0. In other words, we say
that x
1
2 X fully belongs to the set
~
AA, while x
2
2 X
does not belong to
~
AA.
Condition 3. For x
1
; x
2
2 X and k 2 0; 1, we have
l ~
AA
kx
1
1 kx
2
P minfl ~
AA
x
1
; l ~
AA
x
2
g.
Condition 1 is conventional in the fuzzy liter-
ature. The normality requirement implicit in Con-
dition 2 (i.e., existence of x 2 X such that
l
~
AA
x 1) can be easily relaxed, but we preserve it
for the sake of clarity in our presentation. Con-
dition 3 guarantees that the fuzzy set
~
AA is convex.
A convenient, parametric form for expressing
our membership function model is
l ~
AA
xt
0 if xt < m
L
c;
l
~
AA
L
xt if m
L
c 6xt 6m
L
;
1 if m
L
< xt < m
R
;
l ~
AA
R
xt if m
R
6xt 6m
R
b;
0 if xt > m
R
b;
_
_
4
where c and b are the left and right spreads, re-
spectively; m
L
; m
R
2 X are the lowest and highest
values with full membership, respectively; and
l ~
AA
L
xt and l ~
AA
R
xt are the left and right
membership values. Assume that p
L
p
L;0
; . . . ;
p
L;n
L
T
and p
R
p
R;0
; . . . ; p
R;n
R
T
are n
L
1 and
n
R
1 control points for generating the left and
right membership functions, respectively. The left
and right membership functions are part of the
following parametric expressions:
xt; l
~
AA
L
xt
T
~ll
~
AA
L
t; n
L
; p
L
5
,
n
L
k0
B
n
L
;k
tp
L;k
;
xt; l
~
AA
R
xt
T
~ll
~
AA
R
t; n
R
; p
R
6
,
n
R
k0
B
n
R
;k
tp
R;k
;
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 87
where ~ll ~
AA
L
and ~ll ~
AA
R
are the Beezier curves for
the left and right membership functions, respec-
tively; t 2 0; 1; p
L;k
,x
L;k
; y
L;k
T
is the kth Beezier
control point for the left membership function (for
k 0; . . . ; n
L
); p
R;k
,x
R;k
; y
R;k
T
is the kth Beezier
control point for the right membership function
(for k 0; . . . ; n
R
); and B
n
L
;k
t and B
n
R
;k
t are
Bernstein polynomials. As before, in two-dimen-
sional space, we denote ~ll ~
AA
L
t; n
L
; p
L
f
x
t; n
L
;
x
L
; f
y
t; n
L
; y
L
T
and ~ll ~
AA
R
t; n
R
; p
R
f
x
t; n
R
;
x
R
; f
y
t; n
R
; y
R
T
, where x
L
,x
L;0
; . . . ; x
L;n
L
T
,
y
L
,y
L;0
; . . . ; y
L;n
L
T
, x
R
,x
R;0
; . . . ; x
R;n
R
T
and
y
R
,y
R;0
; . . . ; y
R;n
R
T
.
The type of shapes that can be obtained using
the family of membership functions described by
(4) are presented in Fig. 1.
In order to satisfy Conditions 13 we need to
impose some restrictions on the parametric form
expressed by (5) and (6).
For Conditions 1 and 2,
Proposition 1. The first and last control points of
~ll ~
AA
L
are p
L;0
m
L
c; 0
T
and p
L;n
L
m
L
; 1
T
.
Proof. It follows from Property 3 of the Beezier
curves.
Proposition 2. The first and last control points of
~ll ~
AA
R
are p
R;0
m
R
; 1
T
and p
R;n
R
m
R
b; 0
T
.
Proof. It follows from Property 3 of the Beezier
curves.
For Condition 3,
Proposition 3. If the control points p
L
of ~ll
~
AA
L
are
chosen such that x
L;0
6 6x
L;n
L
and y
L;0
6
6y
L;n
L
, then l
~
AA
L
xt is monotonically nonde-
creasing for m
L
c 6xt 6m
L
and xt is mono-
tonically nondecreasing for 0 6t 61.
Proof.
f
0
y
t; n
L
; p
L
l
0
~
AA
L
xt
lim
d!0
l ~
AA
L
xt d l ~
AA
L
xt
d
_ _ _
xt d xt
d
_ _ _ _
f
0
y
t; n
L
; y
L
f
0
x
t; n
L
; x
L
n
L
k0
n
L
B
n
L
1;k1
t B
n
L
1;k
ty
k
n
L
k0
n
L
B
n
L
1;k1
t B
n
L
1;k
tx
k
n
L
1
k0
B
n
L
1;k
tDy
k
n
L
1
k0
B
n
L
1;k
tDx
k
;
where t 2 0; 1, Dy
k
,y
k1
y
k
, and Dx
k
,
x
k1
x
k
, for k 0; . . . ; n
L
1. From the result in
Eq. (7), if Dy
k
P0 and Dx
k
P0, then l
0
~
AA
L
xt P0.
Thus we conclude that l ~
AA
L
xt is monotonically
nondecreasing.
The basic results used in the proof of Proposi-
tion 3 can be found in Farin [6] and Wagner and
Wilson [14].
Similar to Proposition 3, the following result
applies for the monotonically nonincreasing
membership function, ~ll ~
AA
R
.
1.0
)) ( ( ~ t x
A
x(t)
m
R
+ m
R
m
L
m
L
-
)) ( ( ~ t x
L
A
)) ( ( ~ t x
R
A
1.0
)) ( ( ~ t x
L
A
x(t)
m
L
m
L
-
1.0
x(t)
)) ( ( ~ t x
R
A
m
R
+ m
R
)) ( ( ~ t x
A
(a)
(b)
(c)
Fig. 1. Types of membership functions. (a) Monotonic non-
decreasing and nonincreasing. (b) Monotonic nondecreasing
(left). (c) Monotonic nonincreasing (right).
88 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495
Proposition 4. If the control points p
R
of ~ll ~
AA
R
are
chosen such that x
R;0
6 6x
R;n
R
and y
R;0
P
Py
R;n
R
, then l ~
AA
R
xt is monotonically nonin-
creasing for m
R
6xt 6m
R
b and xt is mono-
tonically nondecreasing for 0 6t 61.
The next result follows from Propositions 3
and 4.
Proposition 5. If the control points p
L
of ~ll ~
AA
L
are
chosen such that x
L;0
6 6x
L;n
L
and y
L;0
6
6y
L;n
L
; and the control points p
R
of ~ll
~
AA
R
are
chosen such that x
R;0
6 6x
R;n
R
and
y
R;0
P Py
R;n
R
, then the fuzzy set
~
AA is convex
and satisfies Condition 3.
3.3. Methodology
3.3.1. Basic operations
In the previous section we imposed conditions
on the placement of the control points to guar-
antee the generation of membership functions that
satisfy Conditions 13. It remains to discuss how
to calculate l ~
AA
x given x and
a
~
AA given a. Assum-
ing the location of the control points p
L
and p
R
are
known, the following algorithms may be used.
Algorithm 1 (Finding l
~
AA
x given x).
if x 6m
L
c or x Pm
R
b
then l
~
AA
x 0.
if m
L
6x 6m
R
then l ~
AA
x 1.
if m
L
c < x < m
L
then
Find t 2 0; 1 such that
n
L
k0
n
L
k
_ _
t
k
1 t
n
L
k
x
L;k
x
and compute
l ~
AA
x
n
L
k0
n
L
k
_ _
t
k
1 t
n
L
k
y
L;k
:
if m
R
< x < m
R
b
then
Find t such that
n
R
k0
n
R
k
_ _
t
k
1 t
n
R
k
x
R;k
x
and compute
l ~
AA
x
n
R
k0
n
R
k
_ _
t
k
1 t
n
R
k
y
R;k
:
return l
~
AA
x.
The computational burden of Algorithm 1 is
the solution of a root nding problem on a
polynomial of degree n
L
or n
R
. This problem can
be solved eciently using the bisection method [3]
or the methods proposed by M uuller or Laguerre
[12].
Algorithm 2 (Finding
a
~
AA given a).
if a 0
then l m
L
c, u m
R
b.
else
if a 1
then l m
L
, u m
R
.
else
if c 6 0
then
Find t 2 0; 1 such that
n
L
k0
n
L
k
_ _
t
k
1 t
n
L
k
y
L;k
a
and compute
x
n
L
k0
n
L
k
_ _
t
k
1 t
n
L
k
x
L;k
:
set l x.
else l m
L
if b 6 0
then
Find t such that
n
R
k0
n
R
k
_ _
t
k
1 t
n
R
k
y
R;k
a
and compute
x
n
R
k0
n
R
k
_ _
t
k
1 t
n
R
k
x
R;k
:
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 89
set u x.
else u m
R
set
a
~
AA l; u.
return
a
~
AA.
Again the computational bottleneck of Algo-
rithm 2 is a root nding problem on a polynomial
of degree n
L
or n
R
.
3.3.2. Data-driven estimation
In a direct approach to knowledge acquisition,
experts are required to provide the degree of
membership for each of a collection of points in
the universal set [8]. The resulting set of value-
membership pairs is used to construct the mem-
bership function of the underlying concept. This
section provides a mechanism for constructing
membership functions from data by determining
the number of control points and their locations in
the x; lx space.
The left side of the membership function can
be estimated independently from the right side.
We formulate a mathematical model and propose
an algorithm for estimating the monotonically
nonincreasing portion (right side) of a member-
ship function. A similar approach can be used
for estimating the nondecreasing (left side)
portion.
Let the given data points be d
R;i
xx
R;i
; yy
R;i
T
for i 1; . . . ; M
R
, where M
R
is the total number of
data points and yy
R;i
is the membership given by the
expert through the direct approach to the ith value
xx
R;i
2 X. Without loss of generality, assume there
are at least three data points (i.e., M
R
P3) which
are sorted in ascending order by their rst com-
ponent. Also let the n
R
1 control points be
p
R
x
R;0
; y
R;0
x
R;n
R
; y
R;n
R
T
.
Let the decision variables be x
R;k
and y
R;k
, the
rst and second coordinates of the kth control
point (k 0; . . . ; n
R
); t
i
, the parameter value of the
Beezier curve for the ith data point (i 1; . . . ; M
R
);
and n
R
, the maximum value of the index associated
with the control points to be placed. By Proposi-
tion 2, the rst and last control points are xed in
p
R;0
xx
1
; 1
T
and p
R;n
R
xx
M
R
; 0
T
. Thus the nal
value of some variables is known before perform-
ing any optimization, namely, x
R;0
xx
R;1
, y
R;0
1,
x
R;n
R
xx
R;M
R
, y
R;n
R
0, t
1
0, and t
M
R
1.
The following mathematical program mini-
mizes the sum of the squared errors (SSE) between
the tted membership function and the empirical
data.
min
M
R
1
i2
yy
R;i
_
n
R
k0
n
R
k
_ _
t
k
i
1 t
i
n
R
k
y
R;k
_
2
8
subject to:
n
R
k0
n
R
k
_ _
t
k
i
1 t
i
n
R
k
x
R;k
xx
R;i
for i 2; . . . ; M
R
1;
t
i
6t
i1
for i 1; . . . ; M
R
1;
x
R;k
6x
R;k1
for k 0; . . . ; n
R
1;
y
R;k
Py
R;k1
for k 0; . . . ; n
R
1;
x
R;k
Pxx
R;1
for k 1; . . . ; n
R
1; 9
x
R;k
6xx
R;M
R
for k 1; . . . ; n
R
1;
y
R;k
P0 for k 1; . . . ; n
R
1;
y
R;k
61 for k 1; . . . ; n
R
1;
t
i
P0 for i 2; . . . ; M
R
1;
t
i
61 for i 2; . . . ; M
R
1;
n
R
2 f2; 3; . . .g: 10
The fact that the number of control points is
unknown and integer increases dramatically the
complexity of the problem described by (8)(10).
Fortunately, in most practical applications the
number of control points required is small. By
treating this number as a parameter, we can solve a
series of nonlinear programs, instead of dealing
directly with a more dicult mixed integer non-
linear program. For a given n
R
, the nonlinear
program has 2n
R
M
R
4 continuous variables,
M
R
2 nonlinear constraints, M
R
2n
R
1 linear
constraints, and 2M
R
2n
R
4 lower and upper
bounds.
Given n
R
, let en
R
be the sum of the square
errors between the tted membership function and
the empirical data when n
R
1 control points are
used. Let NLPd
R
; n
R
be a function that solves the
nonlinear program described by (8) and (9).
NLPd
R
; n
R
takes the empirical data d
R
and a
90 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495
specied value of n
R
as its arguments and returns
the optimal value of the objective function de-
scribed in (8), en
R
, and the optimal locations of
the control points, p
R
n
R
. Then Algorithm 3 can
be used to solve the data-driven estimation for the
right membership functions. The algorithm stops
when the improvement in SSE is less than a given
small quantity
0
(say,
0
0:0010) or when the
maximum number of control points to be placed is
reached.
Algorithm 3 (Data-driven estimation of the right
membership function).
set
0
, n
R
1, e1 1.
do
n
R
n
R
1
en
R
; p
R
n
R
NLPd
R
; n
R
if en
R
1 en
R
6 or n
R
M
R
1
if en
R
1 en
R
< 0
then return p
R
n
R
1, en
R
1,
n
R
1.
else
return p
R
n
R
, en
R
, n
R
.
end
Note that the algorithm may terminate with an
increase in the SSE. In this case a local minimum
has been obtained.
4. Performance
4.1. Flexibility
In current practice, users choose the shape of
the membership functions from a pool of com-
monly used parameterized families. After the
shape is selected, the parameters are manipulated
to tune the shape. As discussed in Section 2.1 the
pool of parameterized families of membership
functions include triangular, trapezoidal, Gauss-
ian, generalized bell curve, sigmoid, and S-shaped.
In contrast, our approach can be used to produce
the membership function of almost any imprecise
concept. Basically, our approach can be viewed as
a generalized free form generator of membership
functions that satisfy the basic requirements pre-
sented in Section 3.2.
The example in Table 1 and Fig. 2 illustrates
the ease with which a membership function can be
Table 1
Control points (before change)
k p
L;k
p
R;k
x
L;k
y
L;k
x
R;k
y
R;k
0 10 0.0 50 1.0
1 25 0.1 60 0.3
2 30 0.8 70 0.2
3 50 1.0 75 0.0
0
0.2
0.4
0.6
0.8
1
0 20 40 60 80
x
D
e
g
r
e
e
0
0.2
0.4
0.6
0.8
1
0 20 40 60 80
x
D
e
g
r
e
e
(a)
(b)
Fig. 2. Eect on the change of a single control point. (a) before;
(b) after.
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 91
constructed and tuned interactively using our ap-
proach. By placing the control points in the loca-
tions shown in Table 1, the membership function
depicted in Fig. 2(a), with the control points being
represented by black dots, can be obtained. By
changing the location of the second control point
on the left side (k 1) from 25; 0:1 to 15; 0:5,
the curve bends toward the new point as if there
were some magnetic attraction between the control
point and the membership function (left portion).
This is shown in Fig. 2(b). Moreover, due to the
Property 2 of the Beezier curves presented in Sec-
tion 3.1, we observe that, even though this change
aects the whole left membership function, the
change is more noticeable in the vicinity of the
control point.
This new exible and interactive way of building
and tuning a membership function can be lever-
aged by using a graphical user interface (GUI).
Currently, we are developing a GUI that helps the
user add, move, and delete control points to obtain
the desired free-form membership function.
4.2. Numerical examples
For data-driven estimation, we tested our ap-
proach using data originally published by Zysno
[17] and compared its performance to that of the
methods reported in Zysno [17] and Dombi [4].
Sixty-four persons from 21 to 25 years of age were
asked to rate 52 dierent statements related to age
concepts. The group was divided into four sub-
groups of 16. The individuals within a subgroup
were asked to rate one of the 4 concepts: very
young man, young man, old man, and very old man.
The subjects were asked to give the degree of
membership in the designated fuzzy set of a man of
x years of age on a 0% to 100% scale.
Fig. 3 shows the progress of our algorithm
when applied to automatically estimate the mem-
bership function for the fuzzy set old man based on
the data collected by interviewing subject 35 in
Zysno [17]. In the gure, a black square represents
a control point. A number beside a control point is
used when more than one control point shares the
same location. The number represents the total
number of control points in the given location.
Empty circles represent data points. Lines are used
to display the estimated membership functions.
As is customary in the literature, to compare
our method with those of Zysno and Dombi we
used the sum of the squared errors (SSE) as the
measure of goodness of t between a membership
function model and the empirical data. Dombi
used data for subjects 9, 18, 35, 44 and 58 in Zysno
[17] as his benchmark test cases and measured the
corresponding SSEs. Zysno estimated the param-
eters of his model for all the data sets (64 subjects),
but did not provide SSE as the measure of good-
ness of t. In order to make valid comparisons, we
calculated the SSE for Zysnos model for the
benchmark test cases chosen by Dombi. Table 2
gives the SSE for the benchmark test cases for the
three models, namely, Dombi, Zysno, and ours.
The superior performance of our approach is
clearly seen.
Table 3 shows the evolution of the SSE for each
of the test benchmark cases when our data-driven
estimation mechanism is used. The resulting esti-
mated membership functions are shown in Figs. 3
and 4. Note that most of the intermediate solu-
tions shown in Table 3 are better than the nal
solutions provided by Zysno and Dombi. By
monitoring the progress of the SSE, the algorithm
may be interrupted as soon as the user is satised
with the current SSE. We have used a very small
value for which could potentially cause overt-
ting. However, our method for membership func-
tion generation can get arbitrarily close to the
empirical data.
A nal remark should be made. After tting a
membership function to data, the user can still go
back and tune the membership by moving the
control points as described in Section 4.1. This
high level of interaction and exibility between the
model and the user is a desirable feature when
designing imprecise concepts.
4.3. Computational eciency
In the absence of data, our approach requires
from the user the number and location of the
control points. We have shown in Section 4.1 how
easy it is to change the shape of the membership
92 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495
function by displacing the control points. It is
important to note that the computational eort
needed to redraw a membership function, when-
ever a control point is moved, is just the simple
evaluation of (5) and (6).
Once a membership function has been gener-
ated (either with or without data), as shown in
Section 3.3.1, the calculations required to nd
l
~
AA
x for a given x and nd
a
~
AA for a given a reduce
to solving a computationally inexpensive root
nding problem in a closed interval (t 2 0; 1).
When our approach is used to t membership
functions to data, it was seen in Section 3.3.2 that
the computational bottleneck is nding a solution
to a nonlinear program with 2n
R
M
R
4 vari-
ables, M
R
2 nonlinear constraints, M
R
2n
R
1
25 30 35 40 45 50 55 60 65 70 75
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
25 30 35 40 45 50 55 60 65 70 75
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
2
25 30 35 40 45 50 55 60 65 70 75
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
2 3
25 30 35 40 45 50 55 60 65 70 75
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
3 2
2
(a) (b)
(c) (d)
Fig. 3. Data-driven estimation. Data: subject 35 (Old man). (a) n
L
2; SSE 0:08320; (b) n
L
5; SSE 0:03517;
(c) n
L
8; SSE 0:02846. (d) n
L
11; SSE 0:02469.
Table 2
Sum of square errors (SSE)
Model Data set (subject)
9 18 35 44 58
Zysno 0.10074 0.05054 0.17808 0.07572 0.02641
Dombi 0.13204 0.05103 0.14841 0.05284 0.03027
Proposed approach ( 0:0010) 0.07149 0.02390 0.02469 0.03610 0.01941
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 93
0 5 10 15 20 25 30 35 40
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
3
2
30 35 40 45 50 55 60 65
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
2
35 40 45 50 55 60 65 70 75
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
5
35 40 45 50 55 60 65 70 75 80
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
(
x
)
2
(a) (b)
(c)
(d)
Fig. 4. Data-driven estimation ( 0:0010). (a) Subject 9: n
R
7; SSE 0:07149. (b) Subject 18: n
R
6; SSE 0:02390. (c) Subject
44: n
R
7; SSE 0:03610. (d) Subject 58: n
R
5; SSE 0:01941.
Table 3
SSE progress for the test benchmark cases ( 0:0010)
a
Control points
(n
R
1)
Data set (subject)
9 18 35 44 58
3 0.09231 0.07353 0.08320 0.12333 0.06127
4 0.09044 0.04557 0.05242 0.04838 0.02100
5 0.08822 0.03420 0.04498 0.04071 0.01981
6 0.08167 0.02406 0.03517 0.03800 0.01941
a
7 0.07538 0.02390
a
0.03133 0.03665
8 0.07149
a
0.03011 0.03610
a
9 0.07428 0.02846
10 0.02715
11 0.02549
12 0.02469
a
a
Final solution.
94 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495
linear constraints, and 2M
R
2n
R
4 lower and
upper bounds. For the numerical examples pre-
sented in Section 4.2 we used AMPL as the alge-
braic modeling language and MINOS 5.4 as the
nonlinear optimizer. Using a computer with a
266 MHz Pentium II processor, all the nonlinear
programs took less than 4 seconds to run.
5. Conclusion
We have proposed a new mechanism based on
Beezier curves for generating membership functions
well suited for a broad spectrum of fuzzy model-
ing. By placing control points in dierent loca-
tions, the shape of the membership functions can
be altered in a very natural and intuitive way.
Mechanisms for dealing with subjective and data-
driven estimation of membership functions were
discussed. Some advantages of this approach are
its exibility, ease of use, computational eciency,
and suitability for a graphical interactive imple-
mentation. The major advantage is its immense
power of tting data as close as possible without a
priori assumption of the shape of the function.
Several aspects of this work are in progress.
First, a tailored interior point algorithm that can
exploit the structure of the nonlinear program pre-
sented in (8) and (9) is currently under investigation
[5]. We are currently exploring applications of this
methodology to the elds of fuzzy engineering de-
sign, fuzzy decision making, and fuzzy control.
References
[1] J.L. Chameau, J.C. Santamarina, Membership functions I:
Comparing methods of measurement, International Jour-
nal of Approximate Reasoning 1 (1987) 287301.
[2] J.E. Chen, K.N. Otto, Constructing membership functions
using interpolation and measurement theory, Fuzzy Sets
and Systems 73 (1995) 313327.
[3] W. Cheney, D. Kincaid, Numerical Mathematics and
Computing, Brooks-Cole Publishing, Monterey, CA, 1980.
[4] J. Dombi, Membership function as an evaluation, Fuzzy
Sets and Systems 35 (1990) 121.
[5] S.-C. Fang, S. Puthenpura, Linear Programming and
Extensions, Prentice-Hall, Englewood Clis, NJ, 1993.
[6] G.E. Farin, Curves and Surfaces for Computer Aided
Geometric Design: A Practical Guide, fourth ed., Aca-
demic Press, New York, 1997.
[7] J.-S. Roger Jang, C.-T. Sun, E. Mizutani, Neuro-Fuzzy
and Soft Computing: A Computational Approach to
Learning and Machine Intelligence, Prentice-Hall, Upper
Saddle River, NJ, 1997.
[8] G.J. Klir, B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory
and Applications, Prentice-Hall, Upper Saddle River, NJ,
1995.
[9] C.T. Lin, C.S. George Lee, Neural Fuzzy Systems: A
Neuro-Fuzzy Synergism to Intelligent Systems, Prentice-
Hall, Upper Saddle River, NJ, 1996.
[10] S. Medasani, J. Kim, R. Krishnapuram, An overview of
membership function generation techniques for pattern
recognition, International Journal of Approximate Rea-
soning 19 (1998) 391417.
[11] A.M. Norwich, I.B. Turksen, A model for the measure-
ment of membership and the consequences of its empirical
implementation, Fuzzy Sets and Systems 12 (1984) 125.
[12] W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flan-
nery, Numerical Recipes in C The Art of Scientic
Computing, second ed., Cambridge University Press,
Cambridge, 1992.
[13] I.B. Turksen, Measurement of membership functions and
their acquisition, Fuzzy Sets and Systems 40 (1991) 538.
[14] M.A. Wagner, J.R. Wilson, Using univariate Beezier distri-
butions to model simulation input, IIE Transactions 28
(1996) 699711.
[15] L.A. Zadeh, Fuzzy sets, Information and Control 8 (3)
(1965) 338353.
[16] H.J. Zimmermann, P. Zysno, Quantifying vagueness in
decision models, European Journal of Operational Re-
search 22 (1985) 148158.
[17] P. Zysno, Modelling membership functions, in: B. Rieger
(Ed.), Empirical Semantics, Brockmeyer, Bochum, 1981,
pp. 350375.
A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 8495 95
