European Journal of Operational Research 139 (2002) 84–95

www.elsevier.com/locate/dsw

Stochastics and Statistics

An efficient and flexible mechanism for constructing membership

functions
Andr
es 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 B
ezier curve-based mechanism for constructing membership functions of convex normal
fuzzy sets. The mechanism can fit 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. Ó 2002 Elsevier Science B.V. All rights reserved.

Keywords: Fuzzy sets; Membership functions; B
ezier curves; Measures of information; Fuzzy numbers

1. Introduction cates the degree of belonging to A~ for each element

of X. One of the most important concepts of fuzzy
Fuzzy set theory was first introduced in the sets is the concept of an a-cut. Given a fuzzy set A~
1960s by Zadeh [15] as a way to capture uncer- defined on X and a 2 ð0; 1, the a-cut is defined as
a ~
tainty and vagueness often overlooked in complex A ¼ fx 2 X : lA~ðxÞ P ag. For continuity pur-
systems. It was pointed out that fuzzy set theory is poses, we take 0 A~ ¼ lima!0 a A~. A fuzzy set A~ is
a generalization of the classical set theory. convex if and only if each of its a-cuts is a convex
A fuzzy set A~ is characterized by its membership set. A fuzzy set A~ is normal if 1 A~ 6¼ ;.
function lA~, which maps each element of the uni- Even though there is no universal agreement on
verse X to the interval ½0; 1. This function indi- 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
Corresponding author. Tel.: +1-919-515-2192; fax: +1-919-
515-5281.
demand for membership functions with the fol-
E-mail address: [email protected] (S.-C. Fang). lowing properties: they should be piecewise
URL: www.ie.ncsu.edu/fangroup. monotone nonincreasing or nondecreasing; they

0377-2217/02/$ - see front matter Ó 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 0 1 ) 0 0 1 5 7 - 6
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 85

should achieve null and full membership for at test problems found in the literature are used to
least two different elements in the universal set; illustrate the proposed mechanism and compare its
and they should be able to represent fuzzy convex performance with that of two methods which ap-
sets. Commonly seen examples are the simple tri- pear in the literature. Finally, conclusions and
angular, trapezoidal, and bell-shaped membership current research directions are given in Section 5.
functions.
Problem formulations based on fuzzy sets can
have greater expressive power than their counter-
2. Membership function generation
parts based on crisp sets, but the applicability of
fuzzy technology depends on the ability to con-
2.1. Overview
struct membership functions that appropriately
represent various concepts in different contexts [8].
Membership functions can be constructed from
To fully exploit the benefits provided by fuzzy
data when it is available. This data can be elicited
technology, we need an efficient membership
by interacting with experts using a direct approach
function generating mechanism with the following
(or direct rating) [8,11,13]. The direct approach
desirable characteristics:
requires the degree of membership of a collection
1. Accurate. In the presence of data, the resulting
of points in the universal set. A membership
membership functions should reflect the knowl-
function that describes the underlying concept is
edge contained in the data in the most accurate
fitted to the collected data points. This is known as
way possible. Data in the form of membership
data-driven membership function estimation.
values for points in the universe is usually ob-
Sometimes this approach can be overly precise in
tained from experts.
capturing subjective judgment. By formulating
2. Flexible. The methodology should provide a
easier and simpler questions, knowledge can also
broad family of membership functions.
be acquired through an indirect approach. We will
3. Computationally affordable. The method should
not deal with the indirect approach in this paper,
be computationally tractable in order to be of
but the reader is referred to the paper by Chameau
any practical use. Medasani [10] has highlighted
and Santamarina [1] and the book by Klir and
the importance of having membership functions
Yuan [8].
that can be easily tuned and adjusted. Other
When data are not available in the form of
authors have expressed the need for methods
value-membership pairs, a membership function
in which computer graphics can facilitate the
has to be constructed subjectively. In this case, the
process of constructing membership functions
conventional approach is to first pick the shape of
by allowing the user an easy and direct manip-
the membership function from a list of families,
ulation of different shapes [1].
and then to fine-tune the values of the parameters
4. Easy to use. Once a membership function has
of that function. It is always desirable to have a
been generated, it should be easy to find lA~ðxÞ
parsimonious, meaningful parameterization of
for a given x; and it should be easy to find a A~
membership functions [4].
for a given a.
In this paper we propose a mechanism that
exploits the properties of B
ezier curves to address
these issues and to provide the user with a flexible 2.2. Current methods
and efficient way of generating membership func-
tions. In the literature, fuzzy sets are commonly
The paper is organized as follows. In Section 2, modeled by triangular, trapezoidal, and bell-
we review the basic techniques used for generating shaped membership functions. However, other
membership functions. Section 3 describes the parameterized functional shapes are useful in
proposed mechanism and some fundamental defi- particular situations. More details can be found in
nitions and properties of B
ezier curves. In Section 4, Dombi [4] and Medasani et al. [10].
86 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95

An effort to create a broad class of functions a measure of distance. Even though these models
was made by Zysno [17] and Zimmermann and provide flexibility for estimating S-shaped func-
Zysno [16]. In their model, the membership func- tions, they fail to provide more general monotonic
tion for a fuzzy set A~ is given by curves.


  Chen and Otto [2] present a novel method for
1 1 1 constructing membership functions using interpo-
lA~ðxÞ ¼ mid 0; c þ ;1
1 þ e aðxþbÞ d 2 lation and measurement theory. Following a sys-
8x 2 X  R; ð1Þ tematic approach, their method is able to
construct general monotonic functions from data.
where However, their methodology does not provide a
mechanism for adjusting or building a membership
a; b 2 R; 0 6 c 6 1; and 0 6 d 6 2 minð1 c; cÞ:
function in the absence of data.
The function midð0; f ðxÞ; 1Þ is defined such that In the area of fuzzy system identification, so-
phisticated methods based on neural networks and
midð0; f ðxÞ; 1Þ ¼ f ðxÞ; if 0 6 f ðxÞ 6 1; midð0; f ðxÞ; 1Þ ¼ 0; evolutionary algorithms have been proposed to
if f ðxÞ < 0; and midð0; f ðxÞ; 1Þ ¼ 1; generate and tune both fuzzy rules and member-
if f ðxÞ > 1: ship functions. However, they are basically case by
case approaches [7,9].
Even though the model provides the user with a
In the following section we shall introduce an
commonly used family of S-shapes, the determi-
interactive and efficient approach for both data-
nation of the parameters from empirical data po-
driven and subjective estimation of membership
ses some problems and there is no direct numerical
functions. Based on B
ezier curves, the method is
method for optimal parameter estimation [16,17].
able to generate a broad family of functions.
The model may be used for estimating membership
functions subjectively, with the parameters a, b, c,
and d being fixed by the expert.
Dombi [4] proposed a model with properties 3. Proposed mechanism
similar to the one presented by Zysno and Zim-
mermann. In his model a membership function for 3.1. B

ezier curves
fuzzy set A~ is constructed using the S-shaped
monotonically increasing function One of the major breakthroughs in computer
aided design (CAD) is the theory of B
ezier curves
k 1 k
ð1 mÞ ðx aÞ and surfaces, independently developed by P. de
lA~ðxÞ ¼ ð2Þ Casteljau and P. B
ezier while working for the
ð1 mÞk 1 ðx aÞk þ mk 1 ðb xÞk
French automakers Citro€en and Renault, respec-
and/or the S-shaped monotonically decreasing tively [6].
function The theory of B
ezier curves provides a mathe-
k 1 k
matical foundation for representing a smooth
ð1 mÞ ðb xÞ curve that passes through the vicinity of a set of
lA~ðxÞ ¼ k 1 k k
; ð3Þ
ð1 mÞ ðb xÞ þ mk 1 ðx aÞ control points. Definition 1 gives a formal expres-
sion of a B
ezier curve in terms of Bernstein poly-
where x 2 ½a; b; a; b 2 R; the steepness is given by nomials.
k P 1; and the inflection point is determined by
0 < m < 1. When data are available, Dombi pro-
Definition 1. A B
ezier curve with n þ 1 control
posed a method for estimating the parameters
points p , ðp0 ; . . . ; pn Þ is given by
based on linearized forms of (2) and (3).
Both of these models provide similar member- X
n
ship functions because they use the same under- f ðt; n; pÞ , Bn;k ðtÞpk ;
lying form, i.e., lA~ðxÞ ¼ 1=ð1 þ dðxÞÞ, where dðxÞ is k¼0
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 87

T 
where t 2 ½0; 1, pk , ðxk ; yk Þ , and Bn;k ðtÞ ¼ nk tk Condition 1. The membership function lA~ is a
n k
ð1 tÞ are the Bernstein polynomials. Since mapping from the universal set X to ½0; 1, i.e.,
f ðt; n; pÞ 2 R2 , we usually denote f ðt; n; pÞ ¼ ½fx lA~ : X ! ½0; 1.
T T
ðt; n; xÞ; fy ðt; n; yÞ , where x , ðx0 ; . . . ; xn Þ , y ,
T
ðy0 ; . . . ; yn Þ . Condition 2. There exist x1 ; x2 2 X such that
lA~ðx1 Þ ¼ 1 and lA~ðx2 Þ ¼ 0. In other words, we say
B
ezier curves have several properties that are that x1 2 X fully belongs to the set A~, while x2 2 X
particularly useful in the context of this paper [6]. does not belong to A~.

Condition 3. For x1 ; x2 2 X and k 2 ½0; 1, we have

Property 1. The B
ezier curve f ðt; n; pÞ defined over lA~ðkx1 þ ð1 kÞx2 Þ P minflA~ðx1 Þ; lA~ðx2 Þg.
t 2 ½0; 1, lies in the convex hull of the polygon de-
fined by the control points p , ðp0 ; . . . ; pn Þ. Condition 1 is conventional in the fuzzy liter-
ature. The normality requirement implicit in Con-
Property 2. The Bernstein polynomial Bn;k ðtÞ dition 2 (i.e., existence of x 2 X such that
achieves its unique maximum at t ¼ k=n. If the lA~ðxÞ ¼ 1) can be easily relaxed, but we preserve it
control point pk is moved, then the curve is mostly for the sake of clarity in our presentation. Con-
affected in the region around the parameter t ¼ k=n. dition 3 guarantees that the fuzzy set A~ is convex.
A convenient, parametric form for expressing
our membership function model is
Property 3. The B
ezier curve interpolates its first
(p0 ) and last (pn ) control points. In other words, 8
>
> 0 if xðtÞ < mL c;
f ð0; n; pÞ ¼ p0 and f ð1; n; pÞ ¼ pn . >
>
< A~L ðxðtÞÞ
l if mL c 6 xðtÞ 6 mL ;
lA~ðxðtÞÞ ¼ 1 if mL < xðtÞ < mR ;
These properties have practical effects in the >
>
>
> l ~ ðxðtÞÞ if mR 6 xðtÞ 6 mR þ b;
curve design process. Property 1 guarantees that : AR
0 if xðtÞ > mR þ b;
the curve will not fall outside the ‘‘control poly-
gon’’. By using this property along with Property ð4Þ
2, a B
ezier curve can be designed by exaggerating
the target shape using the control polygon. Even where c and b are the left and right spreads, re-
though a single control point displacement will spectively; mL ; mR 2 X are the lowest and highest
change the whole curve, this ‘‘pseudo-local con- values with full membership, respectively; and
trol’’ property gives us the sense that the control lA~L ðxðtÞÞ and lA~R ðxðtÞÞ are the left and right
points work locally as magnets on the curve. membership values. Assume that pL ¼ ðpL;0 ; . . . ;
T T
Property 3 is very useful for breaking the con- pL;nL Þ and pR ¼ ðpR;0 ; . . . ; pR;nR Þ are nL þ 1 and
struction of a complex curve into simpler parts. nR þ 1 control points for generating the left and
A complete discussion on B
ezier curves and its right membership functions, respectively. The left
properties can be found in the book by Farin [6]. and right membership functions are part of the
following parametric expressions:
T
½xðtÞ; lA~L ðxðtÞÞ ¼ ~
lA~L ðt; nL ; pL Þ ð5Þ
3.2. Mathematical framework
X
nL
, BnL ;k ðtÞpL;k ;
In this section we give the mathematical k¼0
framework of a broad family of membership
shapes based on B
ezier curves. ½xðtÞ; lA~R ðxðtÞÞT ¼ ~
lA~R ðt; nR ; pR Þ ð6Þ
Let A~ be a fuzzy set on the universal set X. The XnR
following conditions are commonly required for its , BnR ;k ðtÞpR;k ;
membership function, lA~ðÞ. k¼0
88 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95

where ~ lA~L ðÞ and ~ lA~R ðÞ are the B
ezier curves for Proposition 1. The first and last control points of
T T
the left and right membership functions, respec- lA~L ðÞ are pL;0 ¼ ðmL c; 0Þ and pL;nL ¼ ðmL ; 1Þ .
~
T
tively; t 2 ½0; 1; pL;k , ðxL;k ; yL;k Þ is the kth B
ezier
control point for the left membership function (for Proof. It follows from Property 3 of the B
ezier
k ¼ 0; . . . ; nL ); pR;k , ðxR;k ; yR;k ÞT is the kth B
ezier curves. 
control point for the right membership function
(for k ¼ 0; . . . ; nR ); and BnL ;k ðtÞ and BnR ;k ðtÞ are Proposition 2. The first and last control points of
T T
Bernstein polynomials. As before, in two-dimen- lA~R ðÞ are pR;0 ¼ ðmR ; 1Þ and pR;nR ¼ ðmR þ b; 0Þ .
~
sional space, we denote ~ lA~L ðt; nL ; pL Þ ¼ ½fx ðt; nL ;
T
xL Þ; fy ðt; nL ; yL Þ and ~ lA~R ðt; nR ; pR Þ ¼ ½fx ðt; nR ; Proof. It follows from Property 3 of the B
ezier
T T
xR Þ; fy ðt; nR ; yR Þ , where xL , ðxL;0 ; . . . ; xL;nL Þ , curves. 
T T
yL , ðyL;0 ; . . . ; yL;nL Þ , xR , ðxR;0 ; . . . ; xR;nR Þ and
yR , ðyR;0 ; . . . ; yR;nR ÞT . For Condition 3,
The type of shapes that can be obtained using
the family of membership functions described by lA~L ðÞ are
Proposition 3. If the control points pL of ~
(4) are presented in Fig. 1. chosen such that xL;0 6    6 xL;nL and yL;0 6   
In order to satisfy Conditions 1–3 we need to 6 yL;nL , then lA~L ðxðtÞÞ is monotonically nonde-
impose some restrictions on the parametric form creasing for mL c 6 xðtÞ 6 mL and xðtÞ is mono-
expressed by (5) and (6). tonically nondecreasing for 0 6 t 6 1.
For Conditions 1 and 2,
Proof.

fy0 ðt; nL ; pL Þ ¼ l0A~L ðxðtÞÞ

µ A~ ( x(t )) lA~L ðxðt þ dÞÞ lA~L ðxðtÞÞ
1.0
¼ lim
d!0 d

µ A~L ( x(t )) µ A~R ( x(t )) xðt þ dÞ xðtÞ
d
fy0 ðt; nL ; yL Þ
x(t) ¼ 0 ð7Þ
mL-γ mL mR mR+β
(a) fx ðt; nL ; xL Þ
PnL
µ A~L ( x (t )) nL ½BnL 1;k 1 ðtÞ BnL 1;k ðtÞyk
¼ Pnk¼0
k¼0 L ½BnL 1;k 1 ðtÞ BnL 1;k ðtÞxk
L
1.0 n
PnL 1
BnL 1;k ðtÞDyk
¼ Pnk¼0 L 1
;
k¼0 BnL 1;k ðtÞDxk

x(t) where t 2 ½0; 1, Dyk , ykþ1 yk , and Dxk ,

(b) mL-γ mL
xkþ1 xk , for k ¼ 0; . . . ; nL 1. From the result in
µ A~ ( x(t ))
Eq. (7), if Dyk P 0 and Dxk P 0, then l0A~ ðxðtÞÞ P 0.
L
1.0
Thus we conclude that lA~L ðxðtÞÞ is monotonically
µ A~R ( x(t )) nondecreasing. 

The basic results used in the proof of Proposi-

x(t) tion 3 can be found in Farin [6] and Wagner and
(c) mR mR+β
Wilson [14].
Fig. 1. Types of membership functions. (a) Monotonic non- Similar to Proposition 3, the following result
decreasing and nonincreasing. (b) Monotonic nondecreasing applies for the monotonically nonincreasing
(left). (c) Monotonic nonincreasing (right). membership function, ~ lA~R ðÞ.
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 89

nR 
lA~R ðÞ are
Proposition 4. If the control points pR of ~ X nR  nR k
chosen such that xR;0 6    6 xR;nR and yR;0 P    tk ð1 tÞ xR;k ¼ x
k¼0
k
P yR;nR , then lA~R ðxðtÞÞ is monotonically nonin-
creasing for mR 6 xðtÞ 6 mR þ b and xðtÞ is mono- and compute
tonically nondecreasing for 0 6 t 6 1. nR 
X nR  nR k
lA~ðxÞ ¼ tk ð1 tÞ yR;k :
The next result follows from Propositions 3 k¼0
k
and 4.
return lA~ðxÞ.
Proposition 5. If the control points pL of ~ lA~L ðÞ are
The computational burden of Algorithm 1 is
chosen such that xL;0 6    6 xL;nL and yL;0 6   
the solution of a root finding problem on a
6 yL;nL ; and the control points pR of ~ lA~R ðÞ are
polynomial of degree nL or nR . This problem can
chosen such that xR;0 6    6 xR;nR and
be solved efficiently using the bisection method [3]
yR;0 P    P yR;nR , then the fuzzy set A~ is convex
or the methods proposed by M€ uller or Laguerre
and satisfies Condition 3.
[12].

3.3. Methodology Algorithm 2 (Finding a A~ given a).

if a ¼ 0
3.3.1. Basic operations then l mL c, u mR þ b.
In the previous section we imposed conditions else
on the placement of the control points to guar- if a ¼ 1
antee the generation of membership functions that then l mL , u mR .
satisfy Conditions 1–3. It remains to discuss how else
to calculate lA~ðxÞ given x and a A~ given a. Assum- if c 6¼ 0
ing the location of the control points pL and pR are then
known, the following algorithms may be used. Find t 2 ½0; 1 such that
nL 
X nL  k
Algorithm 1 (Finding lA~ðxÞ given x). t ð1 tÞnL k yL;k ¼ a
k¼0
k
if x 6 mL c or x P mR þ b
then lA~ðxÞ ¼ 0. and compute
if mL 6 x 6 mR
nL 
then lA~ðxÞ ¼ 1. X nL  k
x¼ t ð1 tÞnL k xL;k :
if mL c < x < mL k¼0
k
then
Find t 2 ½0; 1 such that set l x.
else l mL
nL 
X nL  k nL k if b 6¼ 0
t ð1 tÞ xL;k ¼ x then
k¼0
k
Find t such that
and compute nR 
X nR  k nR k
nL  t ð1 tÞ yR;k ¼ a
X nL  k nL k k¼0
k
lA~ðxÞ ¼ t ð1 tÞ yL;k :
k¼0
k
and compute
if mR < x < mR þ b nR 
then
X nR  k nR k
x¼ t ð1 tÞ xR;k :
Find t such that k¼0
k
90 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95

set u x. The following mathematical program mini-

else u mR mizes the sum of the squared errors (SSE) between
set a A~ ½l; u. the fitted membership function and the empirical
return a A~. data.
!2
Again the computational bottleneck of Algo- X
MR 1 XnR

nR k nR k
rithm 2 is a root finding problem on a polynomial min yR;i ti ð1 ti Þ yR;k
i¼2 k¼0
k
of degree nL or nR .
ð8Þ
3.3.2. Data-driven estimation
In a direct approach to knowledge acquisition, subject to:
experts are required to provide the degree of nR 
X nR  nR k
membership for each of a collection of points in tik ð1 ti Þ xR;k ¼ xR;i
k
the universal set [8]. The resulting set of value- k¼0

membership pairs is used to construct the mem- for i ¼ 2; . . . ; MR 1;

bership function of the underlying concept. This ti 6 tiþ1 for i ¼ 1; . . . ; MR 1;
section provides a mechanism for constructing
xR;k 6 xR;kþ1 for k ¼ 0; . . . ; nR 1;
membership functions from data by determining
the number of control points and their locations in yR;k P yR;kþ1 for k ¼ 0; . . . ; nR 1;
the ðx; lðxÞÞ space. xR;k P xR;1 for k ¼ 1; . . . ; nR 1; ð9Þ
The left side of the membership function can xR;k 6 xR;MR for k ¼ 1; . . . ; nR 1;
be estimated independently from the right side.
yR;k P 0 for k ¼ 1; . . . ; nR 1;
We formulate a mathematical model and propose
an algorithm for estimating the monotonically yR;k 6 1 for k ¼ 1; . . . ; nR 1;
nonincreasing portion (right side) of a member- ti P 0 for i ¼ 2; . . . ; MR 1;
ship function. A similar approach can be used ti 6 1 for i ¼ 2; . . . ; MR 1;
for estimating the nondecreasing (left side)
portion. nR 2 f2; 3; . . .g: ð10Þ
T
Let the given data points be dR;i ¼ ðxR;i ; yR;i Þ
for i ¼ 1; . . . ; MR , where MR is the total number of The fact that the number of control points is
data points and yR;i is the membership given by the unknown and integer increases dramatically the
expert through the direct approach to the ith value complexity of the problem described by (8)–(10).
xR;i 2 X . Without loss of generality, assume there Fortunately, in most practical applications the
are at least three data points (i.e., MR P 3) which number of control points required is small. By
are sorted in ascending order by their first com- treating this number as a parameter, we can solve a
ponent. Also let the nR þ 1 control points be series of nonlinear programs, instead of dealing
T
pR ¼ ððxR;0 ; yR;0 Þ    ðxR;nR ; yR;nR ÞÞ . directly with a more difficult mixed integer non-
Let the decision variables be xR;k and yR;k , the linear program. For a given nR , the nonlinear
first and second coordinates of the kth control program has 2nR þ MR 4 continuous variables,
point (k ¼ 0; . . . ; nR ); ti , the parameter value of the MR 2 nonlinear constraints, MR þ 2nR 1 linear
B
ezier curve for the ith data point (i ¼ 1; . . . ; MR ); constraints, and 2ðMR þ 2nR 4Þ lower and upper
and nR , the maximum value of the index associated bounds.
with the control points to be placed. By Proposi- Given nR , let eðnR Þ be the sum of the square
tion 2, the first and last control points are fixed in errors between the fitted membership function and
T T
pR;0 ¼ ðx1 ; 1Þ and pR;nR ¼ ðxMR ; 0Þ . Thus the final the empirical data when nR þ 1 control points are
value of some variables is known before perform- used. Let NLPðdR ; nR Þ be a function that solves the
ing any optimization, namely, xR;0 ¼ xR;1 , yR;0 ¼ 1, nonlinear program described by (8) and (9).
xR;nR ¼ xR;MR , yR;nR ¼ 0, t1 ¼ 0, and tMR ¼ 1. NLPðdR ; nR Þ takes the empirical data dR and a
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 91

specified value of nR as its arguments and returns to tune the shape. As discussed in Section 2.1 the
the optimal value of the objective function de- pool of parameterized families of membership
scribed in (8), eðnR Þ, and the optimal locations of functions include triangular, trapezoidal, Gauss-
the control points, pR ðnR Þ. Then Algorithm 3 can ian, generalized bell curve, sigmoid, and S-shaped.
be used to solve the data-driven estimation for the In contrast, our approach can be used to produce
right membership functions. The algorithm stops the membership function of almost any imprecise
when the improvement in SSE is less than a given concept. Basically, our approach can be viewed as
small quantity 0 (say, 0 ¼ 0:0010) or when the a generalized free form generator of membership
maximum number of control points to be placed is functions that satisfy the basic requirements pre-
reached. sented in Section 3.2.
The example in Table 1 and Fig. 2 illustrates
Algorithm 3 (Data-driven estimation of the right the ease with which a membership function can be
membership function).
set  0 , nR 1, eð1Þ þ1. 1
do
nR nR þ 1 0.8

ðeðnR Þ; pR ðnR ÞÞ NLPðdR ; nR Þ

0.6
Degree

if eðnR 1Þ eðnR Þ 6  or nR ¼ MR 1
if eðnR 1Þ eðnR Þ < 0 0.4
then return pR ðnR 1Þ, eðnR 1Þ,
nR 1. 0.2

else 0
return pR ðnR Þ, eðnR Þ, nR . 0 20 40 60 80
end (a) x

Note that the algorithm may terminate with an 1

increase in the SSE. In this case a local minimum

0.8
has been obtained.
0.6
Degree

0.4
4. Performance
0.2
4.1. Flexibility
0
0 20 40 60 80
In current practice, users choose the shape of (b) x
the membership functions from a pool of com-
monly used parameterized families. After the Fig. 2. Effect on the change of a single control point. (a) before;
shape is selected, the parameters are manipulated (b) after.

Table 1
Control points (before change)
k pL;k pR;k
xL;k yL;k xR;k yR;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
92 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95

constructed and tuned interactively using our ap- Empty circles represent data points. Lines are used
proach. By placing the control points in the loca- to display the estimated membership functions.
tions shown in Table 1, the membership function As is customary in the literature, to compare
depicted in Fig. 2(a), with the control points being our method with those of Zysno and Dombi we
represented by black dots, can be obtained. By used the sum of the squared errors (SSE) as the
changing the location of the second control point measure of goodness of fit between a membership
on the left side (k ¼ 1) from ð25; 0:1Þ to ð15; 0:5Þ, function model and the empirical data. Dombi
the curve bends toward the new point as if there used data for subjects 9, 18, 35, 44 and 58 in Zysno
were some magnetic attraction between the control [17] as his benchmark test cases and measured the
point and the membership function (left portion). corresponding SSEs. Zysno estimated the param-
This is shown in Fig. 2(b). Moreover, due to the eters of his model for all the data sets (64 subjects),
Property 2 of the B
ezier curves presented in Sec- but did not provide SSE as the measure of good-
tion 3.1, we observe that, even though this change ness of fit. In order to make valid comparisons, we
affects the whole left membership function, the calculated the SSE for Zysno’s model for the
change is more noticeable in the vicinity of the benchmark test cases chosen by Dombi. Table 2
control point. gives the SSE for the benchmark test cases for the
This new flexible and interactive way of building three models, namely, Dombi, Zysno, and ours.
and tuning a membership function can be lever- The superior performance of our approach is
aged by using a graphical user interface (GUI). clearly seen.
Currently, we are developing a GUI that helps the Table 3 shows the evolution of the SSE for each
user add, move, and delete control points to obtain of the test benchmark cases when our data-driven
the desired free-form membership function. estimation mechanism is used. The resulting esti-
mated membership functions are shown in Figs. 3
and 4. Note that most of the intermediate solu-
4.2. Numerical examples tions shown in Table 3 are better than the final
solutions provided by Zysno and Dombi. By
For data-driven estimation, we tested our ap- monitoring the progress of the SSE, the algorithm
proach using data originally published by Zysno may be interrupted as soon as the user is satisfied
[17] and compared its performance to that of the with the current SSE. We have used a very small
methods reported in Zysno [17] and Dombi [4]. value for  which could potentially cause overfit-
Sixty-four persons from 21 to 25 years of age were ting. However, our method for membership func-
asked to rate 52 different statements related to age tion generation can get arbitrarily close to the
concepts. The group was divided into four sub- empirical data.
groups of 16. The individuals within a subgroup A final remark should be made. After fitting a
were asked to rate one of the 4 concepts: very membership function to data, the user can still go
young man, young man, old man, and very old man. back and tune the membership by moving the
The subjects were asked to give the degree of control points as described in Section 4.1. This
membership in the designated fuzzy set of a man of high level of interaction and flexibility between the
x years of age on a 0% to 100% scale. model and the user is a desirable feature when
Fig. 3 shows the progress of our algorithm designing imprecise concepts.
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 4.3. Computational efficiency
Zysno [17]. In the figure, a black square represents
a control point. A number beside a control point is In the absence of data, our approach requires
used when more than one control point shares the from the user the number and location of the
same location. The number represents the total control points. We have shown in Section 4.1 how
number of control points in the given location. easy it is to change the shape of the membership
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 93

1
1

0.9
0.9

0.8
0.8

0.7
0.7

0.6
µ (x)

0.6
2
0.5

µ(x)
0.5

0.4
0.4

0.3
0.3

0.2 0.2

0.1 0.1

0 0
25 30 35 40 45 50 55 60 65 70 75 25 30 35 40 45 50 55 60 65 70 75
(a) x (b) x

1
1

0.9
0.9
2
0.8
0.8

0.7
0.7

0.6 2 3
0.6
3 2
µ(x)

µ(x)

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
25 30 35 40 45 50 55 60 65 70 75
(c) x (d) 25 30 35 40 45 50
x
55 60 65 70 75

Fig. 3. Data-driven estimation. Data: subject 35 (Old man). (a) nL ¼ 2; SSE ¼ 0:08320; (b) nL ¼ 5; SSE ¼ 0:03517;
(c) nL ¼ 8; SSE ¼ 0:02846. (d) nL ¼ 11; SSE ¼ 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 ( ¼ 0:0010) 0.07149 0.02390 0.02469 0.03610 0.01941

function by displacing the control points. It is lA~ðxÞ for a given x and find a A~ for a given a reduce
important to note that the computational effort to solving a computationally inexpensive root
needed to redraw a membership function, when- finding problem in a closed interval (t 2 ½0; 1).
ever a control point is moved, is just the simple When our approach is used to fit membership
evaluation of (5) and (6). functions to data, it was seen in Section 3.3.2 that
Once a membership function has been gener- the computational bottleneck is finding a solution
ated (either with or without data), as shown in to a nonlinear program with 2nR þ MR 4 vari-
Section 3.3.1, the calculations required to find ables, MR 2 nonlinear constraints, MR þ 2nR 1
94 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95

Table 3
SSE progress for the test benchmark cases ( ¼ 0:0010)a
Control points Data set (subject)
(nR þ 1) 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 0.01941a
7 0.07538 0.02390a 0.03133 0.03665
8 0.07149a 0.03011 0.03610a
9 0.07428 0.02846
10 0.02715
11 0.02549
12 0.02469a
a
Final solution.

1 1

0.9 0.9

0.8 0.8
3

0.7 0.7

0.6 0.6
µ(x)

µ(x)

0.5 0.5
2

0.4 0.4

0.3 0.3

2
0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 30 35 40 45 50 55 60 65
(a) x (b) x

1 1
2
0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
µ(x)
µ(x)

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 5 0.1

0 0
35 40 45 50 55 60 65 70 75 35 40 45 50 55 60 65 70 75 80
(c) x (d) x

Fig. 4. Data-driven estimation ( ¼ 0:0010). (a) Subject 9: nR ¼ 7; SSE ¼ 0:07149. (b) Subject 18: nR ¼ 6; SSE ¼ 0:02390. (c) Subject
44: nR ¼ 7; SSE ¼ 0:03610. (d) Subject 58: nR ¼ 5; SSE ¼ 0:01941.
 A.L. Medaglia et al. / European Journal of Operational Research 139 (2002) 84–95 95

linear constraints, and 2ðMR þ 2nR 4Þ lower and [2] J.E. Chen, K.N. Otto, Constructing membership functions
upper bounds. For the numerical examples pre- using interpolation and measurement theory, Fuzzy Sets
and Systems 73 (1995) 313–327.
sented in Section 4.2 we used AM P L as the alge- [3] W. Cheney, D. Kincaid, Numerical Mathematics and
braic modeling language and MI N O S 5.4 as the Computing, Brooks-Cole Publishing, Monterey, CA, 1980.
nonlinear optimizer. Using a computer with a [4] J. Dombi, Membership function as an evaluation, Fuzzy
266 MHz Pentium II processor, all the nonlinear Sets and Systems 35 (1990) 1–21.
programs took less than 4 seconds to run. [5] S.-C. Fang, S. Puthenpura, Linear Programming and
Extensions, Prentice-Hall, Englewood Cliffs, NJ, 1993.
[6] G.E. Farin, Curves and Surfaces for Computer Aided
Geometric Design: A Practical Guide, fourth ed., Aca-
5. Conclusion demic Press, New York, 1997.
[7] J.-S. Roger Jang, C.-T. Sun, E. Mizutani, Neuro-Fuzzy
We have proposed a new mechanism based on and Soft Computing: A Computational Approach to
Learning and Machine Intelligence, Prentice-Hall, Upper
B
ezier curves for generating membership functions
Saddle River, NJ, 1997.
well suited for a broad spectrum of fuzzy model- [8] G.J. Klir, B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory
ing. By placing control points in different loca- and Applications, Prentice-Hall, Upper Saddle River, NJ,
tions, the shape of the membership functions can 1995.
be altered in a very natural and intuitive way. [9] C.T. Lin, C.S. George Lee, Neural Fuzzy Systems: A
Mechanisms for dealing with subjective and data- Neuro-Fuzzy Synergism to Intelligent Systems, Prentice-
Hall, Upper Saddle River, NJ, 1996.
driven estimation of membership functions were [10] S. Medasani, J. Kim, R. Krishnapuram, An overview of
discussed. Some advantages of this approach are membership function generation techniques for pattern
its flexibility, ease of use, computational efficiency, recognition, International Journal of Approximate Rea-
and suitability for a graphical interactive imple- soning 19 (1998) 391–417.
[11] A.M. Norwich, I.B. Turksen, A model for the measure-
mentation. The major advantage is its immense
ment of membership and the consequences of its empirical
power of fitting data as close as possible without a implementation, Fuzzy Sets and Systems 12 (1984) 1–25.
priori assumption of the shape of the function. [12] W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flan-
Several aspects of this work are in progress. nery, Numerical Recipes in C – The Art of Scientific
First, a tailored interior point algorithm that can Computing, second ed., Cambridge University Press,
Cambridge, 1992.
exploit the structure of the nonlinear program pre-
[13] I.B. Turksen, Measurement of membership functions and
sented in (8) and (9) is currently under investigation their acquisition, Fuzzy Sets and Systems 40 (1991) 5–38.
[5]. We are currently exploring applications of this [14] M.A. Wagner, J.R. Wilson, Using univariate B
ezier distri-
methodology to the fields of fuzzy engineering de- butions to model simulation input, IIE Transactions 28
sign, fuzzy decision making, and fuzzy control. (1996) 699–711.
[15] L.A. Zadeh, Fuzzy sets, Information and Control 8 (3)
(1965) 338–353.
[16] H.J. Zimmermann, P. Zysno, Quantifying vagueness in
References decision models, European Journal of Operational Re-
search 22 (1985) 148–158.
[1] J.L. Chameau, J.C. Santamarina, Membership functions I: [17] P. Zysno, Modelling membership functions, in: B. Rieger
Comparing methods of measurement, International Jour- (Ed.), Empirical Semantics, Brockmeyer, Bochum, 1981,
nal of Approximate Reasoning 1 (1987) 287–301. pp. 350–375.