Constructing Membership Functions Using

T
Interpolation and Measurement Theory
Joseph E. Cheny
Kevin N. Ottoz
Department of Mechanical Engineering
Massachusetts Institute of Technology

AF April 4, 1994

Abstract
Imprecision is well suited for representing uncertainty in choice during an
engineering design process, and in particular for computer aided engineering
design and analysis tools. To propagate imprecise understanding through en-
gineering tools, however, rst the membership functions must be constructed
based on the understandings of the design engineer. To help construct suit-
able membership functions, measurement theory oers an axiomatically based,
easy to use method. For any given variable, the best and worst values are
determined, and the remaining values are assigned a degree of membership by
comparison with the best and the worst. On real variables, however, this would
require an innite number of questions. Instead, a continuity assumption can be
made, and the remaining membership values determined through interpolation.
DR
Traditional interpolation schemes, however, fail to satisfy the restrictions of a
membership function. The [0 1] boundedness condition and the fuzzy-convex
;

property in particular present diculty. A simple and ecient constrained in-

terpolation scheme is developed for tting a membership function to a nite
number of known membership values. Thus, a simple method enabling design
engineers to construct membership functions is presented.

1 Introduction
Imprecision in specication is a natural phenomena which arises with any human
conceptualization activity. In our particular interest, imprecision is an intrinsic
aspect of product development , the process of understanding a customer's need, and
based upon this generating and embodying a product which can satisfy this need.

Manuscript prepared for submission to Fuzzy Sets and Systems .
yGraduate Research Assistant
zAssistant Professor of Mechanical Engineering, Room 3-449, M.I.T., 77 Massachusetts Ave.,
Cambridge MA, 02139

1
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 2

Of particular interest to us is imprecision as it occurs in engineering design , which

involves developing a product using engineering models and analysis.
Imprecision and vagueness are intrinsic aspects of engineering design. If (at the
start of a design process) a proposed solution were neither imprecise nor vague, its

T
description would be precise and it would therefore be a completed design. While
stochastic uncertainty typically remains in a completed design description (e.g.,
dimensional tolerances), the nominal desired dimensions are precise. However, much
of the early description of a design concept (physical dimensions, material properties,
etc.) is vague and imprecise. At the early stages, a design team simply is not sure
what values to use. Engineering calculations become dicult, because values for
variables are simply unknown.
In this frame, imprecise mathematics becomes useful to represent the possibili-

AF
ties. Zadeh's extension principle, for example, can be used to propagate imprecise
understanding through relevant calculations and models. A comprehensive review of
modeling imprecision and uncertainty in engineering design is given in [8]. Further
work can be found in [6, 7, 9, 10, 11, 12].
For imprecision to be used during a design process, it must be represented. This
means that a designer's understanding of the usefulness of values in a model must
be represented, say with fuzzy numbers. A well dened procedure for specifying
the fuzzy numbers must be given, however, for design engineers to manipulate their
models in an imprecise manner. Our approach is to ask the designers and customers
their preferences for various aspects of the design. They indicate their rank of
membership (on a scale from 0 to 1) for every value of each variable describing the
design and its performance and constraints.
The membership information is usually (but not always) a convex fuzzy number.
We use the mathematics of fuzzy sets to operate on these imprecise descriptions of
the design. For example, we can use membership data in conjunction with the usual
engineering computations encountered in design. We expand the process to map
DR
not just single values, but the entire imprecise set of variables specifying a design,
such as geometric lengths. We map these memberships through calculations to de-
pendent variables such as material stresses, dynamic responses, or costs. Further,
dependent variable memberships can also be back-mapped onto the variables spec-
ifying a design. This provides the designer an ability to manipulate the imprecise
aspects of a design in an understandable and rapid way. The reader is referred
to [8, 9, 10, 11, 12].
A two fold problem arises. The rst is to provide a well dened method for a
design engineer to specify the membership values. The second problem is to provide
a method for variables with an uncountable number of points, such as a common
real-valued variable. A design engineer can only specify the membership for a nite
subset of the points, and the remaining must have their membership determined
through some form of interpolation. This paper will present a method we have found
useful for specifying a complete membership function across real valued variables.
We present methods to ask a design engineer for the membership values on a nite
subset, and present methods to interpolate the remaining.
The interpolation problem is not solvable by a simple least-squares or spline
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 3

method. Membership functions have constraints which the simple methods do not
satisfy. In particular, a membership function is bounded in [0; 1]. A constrained in-
terpolation is required. We present a simple, ecient method to calculate a properly
constrained membership function to a nite set of specied membership values.

T
2 Measurement Theory
Measurement theory [2] provides a mathematically axiomatic method to construct
membership values as used in this work. A comprehensive presentation of using
measurement theory to construct evaluations in design is given in [5]. To construct
a membership function for a variable, a set of values of the variable must be known,

AF
and a reason for making the preference specications. If a design engineer supplies
a membership value of 1:0 to a variable value x1 but a membership value of 0:0 to a
dierent variable value x0 , there must be a reason for the dierence between x1 and
x0 . A design engineer must have a reason for making this distinction. This informal
reason behind the dierence in formal membership we will denote by f , which is
not a function, merely a label attached to the informal reason.
Once a design engineer has constructed the set of values X and has determined
a reason f for specifying membership values, a membership function  can be con-
structed over X using measurement theory [2, 5]. If the design engineer wishes to
construct the membership function relative to the points in X , an interval scale can
be constructed.
An interval scale construction will allow a designer to construct a real valued
scale  which re ects the informal objective f . Then the designer must rst identify
which points in X have the least and most amount of the objective f , denoted
xworst and xbest respectively. These are designated with amounts zero and one on
the membership scale being constructed. Then for each other element xi 2 X , the
designer must answer:
DR
\On a scale of zero to one, what is your belief  that you are indierent
between:
1) receiving the objective performance provided by xi ,
or
2) receiving the objective performance provided by xbest with certainty 
and receiving the objective performance provided by xworst with certainty
(1 ? )?"
This constructs a real valued (measurable) membership function  : X ! [0; 1] which
directly preserves the partial ordering of the elements of X , and also the relative
separation. A dierence of 0:5 compared to 0:3 in  means a larger dierence in the
informal reason f for constructing the membership function.
There are extensions to this method for sets which have additional structure,
such as a well dened zero, or a well dened unit of measure [5]. Constructing such
ratio and extensively measurable scales is beyond the scope of this paper.
Historically this method has been associated to subjective probability. We main-
tain that for engineering design purposes, this method is suitable for constructing
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 4

Table 1: Elicited membership values for maximum stress.

Stress 
200 MPa 1.00

T
210 MPa 0.95
225 MPa 0.50
230 MPa 0.10
250 MPa 0.00

membership functions for propagating imprecision in choice through engineering

AF
systems. In any case, for the interpolation algorithms presented below, how the
nite subset of membership values is determined is irrelevant. A resulting subset of
pairs f(x1 ; 1 ); : : : ; (xn ; n )g is all that is required.
This basic measurement approach obviously assumes nite sets. On uncountable
sets, further assumptions are required. The designer can only provide answers on a
nite subset, which must then be interpolated between. The next section will present
the problem formulation and an ecient solution to the interpolation problem.

3 Interpolating Over Real Valued Variables

Membership functions have constraints which conventional least-squares and spline
methods do not satisfy. The rst constraint is that membership functions are usu-
ally monotonic and convex. A second constraint is that membership functions are
bounded in [0; 1]. These constraints will be detailed.
A membership function is usually monotonic and convex, dened as a fuzzy
number using the fuzzy-convex property:
DR
(xi + (1 ? )xj )  minf(xi ); (xj )g
where  2 [0; 1] and xi ; xj 2 R. Least squares and simple splines may not preserve
the monotonicity and/or convexity. For example, consider a set of known mem-
bership values as shown in Table 1 for a set of material stress values. The data is
fuzzy-convex. However, when using least squares or cubic splines to interpolate, the
result is not fuzzy-convex. Figure 1, 2 show the interpolations using least squares
and cubic splines. Both interpolations violate the fuzzy-convex property.
Another constraint that any proposed interpolation scheme must address is to
keep the membership function bounded within [0; 1]. Again least squares and sim-
ple splines do not guarantee this boundary condition. For example, if two known
membership values are closely spaced in the domain but widely separated in , then
overshoots tend to occur, which may force the interpolation below 0 or above 1.
Figure 1, 2 show the interpolations using least squares and cubic splines. Again,
both interpolations exhibit membership overshoot pathologies.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 5

T
AF
Figure 1: Graph of interpolated membership using least-squares.
DR

Figure 2: Graph of interpolated membership using a common cubic spline.

J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 6

3.1 Desired Properties in Interpolated Membership Functions

We present axioms about human decision making that we choose to adopt for the
purposes of the developed interpolation scheme. These may be excessively restrictive
for all types of human decision-making, which is beyond our purposes. Instead, we

T
seek to explicitly lay out the axioms which must be adopted to use this interpolation
scheme (the strong continuity assumptions, for example). We then demonstrate
resulting properties of the derived membership functions.
Axiom 1 A membership function is a numerical scale bounded in [0; 1].
Axiom 2 Among the set of points considered by a designer, at least one will be
a utopian point with membership 1, and at least one more will be an unacceptable

AF
point of membership 0.
These axioms are actually the foundation for using measurement theory to elicit
membership values. The worst and best cases are assigned membership value 0 and
1. The rest are assigned membership values between 0 and 1.
To t a membership function to a set of elicited membership values, the level
of continuity in the resulting function must be known. We choose to t a smoothly
varying curve to the membership data. This implies that a designer's choice among
values is smoothly varying.
Axiom 3 The rate that a designer changes membership for values in a domain is
continuous.
This axiom implies that a membership function is rst derivative continuous and
thus the function itself is dierentiable. This implies that:
Proposition 1 If (x) is 0 or 1, then 0(x) = 0.
DR
Proof. Either points in the right- or left-hand neighborhood of x have the same value
of , or x is a local maximum or minimum for x. In the rst case, 0 (x) must be
zero since  is dierentiable and 0 is zero in the neighborhood of x. In the second
case, 0 (x) is zero since x is a local maximum or minimum.
This means when we t a membership function to some data points, we will assume
that the slope at the end-points of support of the membership function will ramp
up slowly. We also will assume the membership is smooth about any peak points.
Another important feature of approximate reasoning is in dealing with fuzzy
numbers, when the decision making involves grades of membership in sets. We
choose not to restrict to fuzzy numbers. However, if a designer elicits membership
data which has the fuzzy-convex property, we will insist on maintaining the fuzzy-
convex property across all of the domain.
These restrictions form a precise statement of what we seek from an interpolation
function. We seek to nd
 : R ! [0; 1]
such that
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 7

1. (x) 2 [0; 1] 8 x,
2.  is dierentiable,
3. (xi ) = i for a nite set of known pairs f(x1 ; 1 ); : : : ; (xn; n )g.

T
4. If the set of known pairs f(x1 ; 1 ); : : : ; (xn ; n )g is a fuzzy-convex set, then 
is fuzzy-convex.
This problem statement is related to a calculus of variations problem, which
is a method to determine a function that satises some constraints and simulta-
neously minimizes an \energy" function. For constrained curve tting, such an
energy-minimized curve approach is typically used. The problem would be to min-

AF
imize \energy" represented by the overall curve length of the tted curve, and/or
minimizing various \size" functions of the derivatives along the curve. We choose
not to be so explicit with an energy function, which, as discussed below, introduces
articial variables into the problem formulation. Instead, we adopt the last restric-
tion which is related to the degree of continuity and overall curve length. We also
constrain ourselves to seek simple and ecient methods.
3.2 Constrained Interpolation Methods
The last constraint we imposed on the interpolation scheme is a constraint to pre-
serve the \local shape" of the elicited membership data. Investigations into inter-
polation schemes that preserve the shape of the data center mainly on the notions
of \splines under tension" [1]. Following a calculus of variations approach, these
techniques basically nd splines that minimize tension energy under dierent \ten-
sion parameters," similar to an sti band under a parameterized amount of bending
tension passing through knot points. A diculty in this approach is nding a mean-
ingful interpretation of a \tension parameter" for a decision maker. How should a
DR
value be determined?
A second problem with splines under tension as a means of tting a membership
function is that the resulting curve may still overshoot the required [0; 1] boundary.
Even when the slopes at the boundary are set equal to zero, the spline may still
have an overshoot problem because the overshoot may help minimize the tension
energy (see Figure 3). An interpolation scheme for tting membership functions
must eliminate such overshoots.
Another issue is the slope at any local extremum and the boundary. According to
our axioms and propositions, the slope vanishes at these points. Equally important
is to preserve the \shape," or local monotonicity and convexity of the data. If the
data elicited is fuzzy-convex, we demand that the resulting interpolated membership
function be fuzzy-convex.
3.3 Constrained Interpolation Using Bernstein Polynomials
McAllister and Roulier [3, 4] developed an algorithm which produces a monotonicity
and convexity preserving second degree Bernstein polynomial. Further, it is very
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 8

T
AF
Figure 3: Interpolation using a spline in tension. The dierent curves correspond
to dierent \tension levels."

w~ i

g(x)
DR
g(x)
B2(g)(x)

~oi

d~i

xi a = (xi + ti )=2 ti

Figure 4: A second order Bernstein polynomial with middle knot that has abscissa
a = (xi + ti )=2.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 9

fast and ecient to implement, and requires no user judgment to adjust any \tension
parameter" values as do the spline under tension techniques.
The ability of this method to preserve the boundedness and monotonicity of a
membership function depends on the following property of Bernstein polynomials

T
on an interval.
Let d~i = (xi ; i ) and ~oi = (ti ; ~i ) be arbitrary points. Let w~ i = (a; b) be an
arbitrary point with a = (xi + ti )=2. Let g be the piece-wise linear spline passing
through the points d~i , ~oi and w~ i with a single discontinuity at a, as shown in Figure 4.
Let B2 [d~i ; w~ i ;~oi ] be the second-degree Bernstein polynomial of g on [xi ; ti ]. That is,
B2 [d~i ; w~ i;~oi ](x) = B2 (g)(x)
2 ti ? x) + g(ti )(x ? ti )2 :
= g(xi )(ti ? x) + 2b(x(t? ?xi )(

AF
Proposition 2 The following properties hold:

2. B20 (g)(xi ) = g0 (xi ), B20 (g)(ti ) = g0 (ti );

i xi )

1. B2 (g)(xi ) = g(xi ) = i , B2 (g)(ti ) = g(ti ) = ~i ;

2

3. if g is monotone on [xi ; ti ], then B2 (g) is monotone on [xi ; ti ];

4. if g is convex (concave) on [xi ; ti ], then B2 (g) is convex (concave) on [xi ; ti ].

Proof. Proofs for the above properties 1 and 2 are trivial. Properties 3 and 4 are
shown in [4], but are essentially derived from the \convex hull" property of any
quadratic functions. Hence B2 (g) \preserves the shape" of g, and B2 (g) has a
continuous rst derivative on [xi ; ti ].
This proposition forms the basis for why the algorithm presented below preserves
DR
the desired membership function properties. Given proper data points, a quadratic
Bernstein polynomial will preserve the local behaviour. A quadratic curves needs 3
coecients to specify the curve. If one attempts to t a quadratic curve directly to
the points (xi ; i ); (xi+1 ; i+1 ), however, the prescription of the slope at xi would
prescribe the slope at xi+1 , which in turn may force the interpolation to not behave
as a membership function.
The solution proposed is to augment the given set of points f(x1 ; 1 ); : : : ; (xn ; n )g
with additional points (ti ; ~i ) which are positioned to ensure the satisfaction of the
membership function properties. The problem then becomes to nd proper posi-
tioning of these additional points.
The complete proposed algorithm to t a membership function to a set of mem-
bership data f(x1 ; 1 ); : : : ; (xn ; n )g is as follows:
1. First determine the slope mi at each known point (xi ; i ). To do this, rst
dene
si = (i ? i?1 )=(xi ? xi?1 ):
The following properties of mi must be met:
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 10

mi
~b ~c d~i+1

T
si+1
d~i

si

tonicity.
d~i?1 AF
Figure 5: Construction of slope mi complying with the local convexity and mono-

(a) mi must be consistent with the monotonicity and convexity of the piece-
wise linear function determined by the data points (xi?1 ; i?1 ), (xi ; i ),
(xi+1 ; i+1 ).
(b) mi must vary continuously with respect to changes in si and si+1 when
the signs of si and si+1 agree (i.e., when (xi ; i ) is not a extremum with
xed slope of zero).
DR
(c) Only one knot should be required between two data points. This is to
minimize the complexity of the algorithm.
The construction of mi given below satises the above properties.
 If sisi+1  0, we set mi = 0. This guarantees that local extrema of the
data points are assigned slopes 0. This also segments the entire data into
monotonically non-increasing or non-decreasing subsets.
 If jsij > jsi+1j > 0, we extend a line through d~i = (xi ; i) with slope si
until it intersects the horizontal line through d~i+1 = (xi+1 ; i+1 ) at the
point ~b = (bx ; i+1 ). Refer to Figure 5. We then dene
cx = bx +2xi+1 ; (1)
which is the abscissa of point ~c shown in Figure 5. Slope mi at (xi ; i ) is
dened as
mi = ci+1??xi : (2)
x i
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 11

F d~i+1

T
AF
d~i E

Figure 6: Rectangle Ri and its midpoint segment EF .

Note that
cx > xi +2xi+1 : (3)
 If on the other hand, 0 < jsi+1j < jsi+1j, we reverse the the above pro-
cedure by extending the line through (xi ; i ) with slope mi+1 until it
intersects the horizontal line through (xi?1 ; i?1 ) at the point (bx ; i?1 ).
Then we set cx = (bx + xi?1 )=2 and mi = (i ? i?1 )=(xi ? cx ). The end
DR
point slopes m0 and mn are set to zero explicitly since this is required by
Proposition 1.
2. We now insert a knot point between each xi and xi+1 and t a Bernstein
polynomial to the 2n ? 1 data points.
Let Ri be the rectangle determined by the points (xi ; i ) and (xi+1 ; i+1 ) and
let the midpoint segment of it be the line segment that bisects Ri vertically
and is bounded within each Ri . Refer to Figure 6. Let Li be the line that
passes through (xi ; i ) with slope mi .
Without loss of generality we assume nondecreasing data points; for non-
increasing data sets, the same algorithm can be applied without change, we
use nondecreasing data points for demonstration purposes. There are two
distinct cases regarding the intersection of the neighboring slope lines Li and
Li+1 , depending on whether the knots change the local convexity of the spline
or not.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 12

 Case 1. Li and Li+1 intersect at a point ~z = (ti; zy ) in Ri . Refer to

Figure 7. Let
~vi = (xi + xi +
2
ti ;  + L xi + ti )
i i
2 (4)

T
w~ i = (xi+1 ? xi+12+ ti ; i+1 ? Li+1 xi+12+ ti ) (5)
as shown in Figure 8. Let L be the line joining ~vi and w~ i and dene
~i = L(ti ): (6)
Now we will let ~oi = (ti ; ~i ) be a knot for the spline. Refer to Figure 8.
Dene  on [xi ; xi+1 ] as follows:

AF ( ~
(x) = B [B~o2;[dw~i ;~;vd~i ;~oi](](xx)) on
2 i i i+1 on
[xi ; ti ];
[ti ; xi+1 ]:
From Proposition 2 we immediately see that (x) 2 C 1 [xi ; xi+1 ] and
satises
(xi ) = i ;
(xi+1 ) = i+1 ;
0 (xi ) = mi ;
0 (xi+1 ) = mi+1 :
(7)

Moreover, if the rst-degree spline dened by the points d~i , ~vi , w~ i , and
d~i+1 is convex (concave) and/or monotone, then  is convex (concave)
and/or monotone.
 Case 2. Li and Li+1 do not intersect within Ri . As a consequence of
Proposition 3 (shown below), there are only two possible situations as
depicted in Figure 9. The knot ~oi is determined similarly as in Case
DR
1, but with ti = (xi + xi+1 )=2. The denitions of ~vi , w~ i , ~i , ~oi , and 
remain as in (4), (5), (6), and (7), respectively. Refer to Figure 10. Then
 2 C 1 [xi ; xi+1 ] and  preserves the shape of the data. Note here that
the knot ~oi becomes the point at which the convexity changes.
This algorithm discusses nondecreasing data points for demonstration with the g-
ures; however, it remains equally valid for non-increasing data sets. Also, by slicing
the whole data set at the local maximas and minimas, the whole data set consists
of many independent segments, where each is non-increasing or non-decreasing.
We now prove the proposition that the slope assigned at d~i complies to the local
convexity and monotonicity.
Proposition 3 Consider data points which are nondecreasing. The line Li of slope
mi passing through d~i = (xi ; i ) intersects the midpoint segments of both of the
adjoining rectangles Ri?1 and Ri+1 .
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 13

F d~i+1
Li

T
Li+1 ~z

d~i
AF E

Figure 7: Li and Li+1 intersect within Ri .

Proof. For mi = 0, the assertion is trivial since the line of slope 0 through (xi ; i )
always forms a horizontal edge to each of the adjacent rectangles and thus inter-
sects the midpoint segment of each at the end point. Therefore we only need to
consider the cases where mi 6= 0. Since the slope assignments in the above step is
symmetrical, we only need to consider the case of an interior point (xi ; i ) with
mi > 0
DR
and
0 < si+1 < si ;
as depicted in Figure 5. From (1), (2), (3), We see that
xi+1 > cx > xi +2xi+1 : (8)
With this and (2) we have
si+1 < mi < 2si+1 (9)
and
0 < m i < si : (10)
Inequality (9) means that the line Li through (xi ; i ) of slope mi intersects the
midpoint segment of Ri . The second inequality (10) means that Li also intersects
the midpoint segment of Ri?1 . Refer to Figure 11.
A feature of the inserted knot points f(ti ; ~i )g which will be required subse-
quently is their monotonicity with their neighbors.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 14

d~i+1

T
w~ i

d~i
AF ~vi
~oi

~z
Li+1

Figure 8: Determination of ~oi for Case 1.

Li

d~i+1 d~i+1
Li
Li+1
DR

Li

Li+1
d~i d~i

Figure 9: Two possible cases when Li and Li+1 do not intersect within Ri .
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 15

d~i+1 d~i+1
w~ i

T
Li
Li+1

w~ i
~oi ~oi

d~i
AF
~vi

Li+1
d~i
~vi

Figure 10: Determination of ~oi for Case 2.

Proposition 4 The points (ti; ~i ) determined in the algorithm are monotonic with
Li

(xi ; i ) and (xi+1 ; i+1 ).

Proof. Without loss of generality, we consider a set of monotonically increasing data.
According to Proposition 3 the slopes assigned must all be non-negative.
In case 1, this proof is trivial since the intersection ~z of Li and Li+1 is inside Ri .
DR
Thus ~vi 2 di z 2 Ri and w~ i 2 zdi+1 2 Ri . The knot ~oi thus satises
~oi 2 viwi 2 Ri :
For case 2, as shown in Figure 12, Li has to intersect midpoint segment EF and
hence it also has to intersect the line segment GH . We have ~vi 2 Ri . By the same
argument Li+1 has to intersect line IJ . Thus
~vi 2 GH 2 Ri;
w~ i 2 IJ 2 Ri ;
and so
~oi 2 viwi 2 Ri :

We now demonstrate the resulting interpolated membership function for some

typical elicited membership data sets, and prove that the method will always result
in a fuzzy-convex membership function when the elicited membership data is fuzzy-
convex.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 16

d~i+1

T
Ri
Li

d~i?1
AF Ri?1
d~i

Figure 11: Line Li intersects the midpoint segments of Ri and Ri+1 .

H F J d~i+1
DR
Li

w~ i
~oi
~vi

Li+1

d~i G E I

Figure 12: Li must intersect with GH and Li+1 must intersect with IJ .
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 17

T
AF
Figure 13: Graph of interpolated membership for stress.

4 Relevance for Imprecise Engineering Calculations

We apply the above interpolation scheme to construct membership functions. As
can be seen, it clearly satises the boundary conditions and preserves the local
DR
monotonicity and/or convexity. We then demonstrate the algorithm's fuzzy-convex
preserving properties, and its convergence properties.
4.1 Engineering Examples
Figure 13 shows the membership function of maximum stress constructed from the
elicited membership values of Table 1. It preserves the properties of membership
functions we present above. Table 2 shows another set of membership values based
on preference of geometric length for a mechanical component and Figure 14 is the
resulting interpolated membership function.
4.2 Fuzzy-Convex Membership Functions
Of particular interest is when the elicited data satises the fuzzy-convex property.
Proposition 5 Given a fuzzy-convex set of elicited membership data, the interpo-
lated membership function determined using the algorithm of Section 3.3 will be
fuzzy-convex.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 18

T
Table 2: Elicited membership values for a geometric length.
length 
0.95 m 0.00
0.96 m 0.25
0.97 m 0.50
0.98 m 0.75
1.00 m 1.00
1.02 m 0.75

AF 1.03 m 0.50
1.04 m 0.25
1.05 m 0.00
DR

Figure 14: Graph of interpolated membership values for a geometric length.

J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 19

Proof. By assumption, the elicited membership data is monotonically increasing on

fx0 ; : : : ; xk g and monotonically decreasing on fxk ; : : : ; xng. Using the construction
of the algorithm in Section 3.3, we insert a point o~i = (ti ; ~i ) monotonic with
(xi ; i ); (xi+1 ; i+1 ), by Proposition 4. Then using the algorithm in Section 3.3, a

T
monotonic g (as in Proposition 2) is dened on [xi ; ti ] and on [ti ; xi+1 ], consisting
of Li and Li+1 respectively. Given this, then by Denition (7) and Proposition 2, 
is monotonic between [xi ; ti ] and [ti ; xi+1 ], which with the monotonic ti imply  is
monotonic on [xi ; xi+1 ]. But together with the monotonic data points this implies 
is monotonically increasing on [x1 ; xk ] and monotonically decreasing on [xk ; xn ].
Thus, the interpolation scheme is very suitable for imprecise calculations and deci-
sion making in a fuzzy environment.

AF
4.3 Convergence Properties of the Algorithm
Another concern which arises is the behaviour of the interpolation when additional
elicited points fxi ; i g are added to interpolate with. The membership function
should become more accurate as any arbitrary point fxi ; i g is added. This is in
contrast to some simple interpolation schemes, which can become worse, for example,
if the points then become not evenly spaced. The algorithm of Section 3.3 exhibits
no such properties, and will converge to the actual membership function.
Denote by n (x) an interpolated membership function derived using the algo-
rithm of Section 3.3, with n elicited data points. Denote by a (x) the theoretical
elicited membership function if the interval scale question were asked on all points
x 2 R.
Proposition 6
lim  (x) = a (x)
n!1 n
DR
Proof. Let B [a; b] be the space of bounded continous functions on [a; b] and J
be the interval [x0 ; xn?1 ] and Ji be the interval [xi?1 ; xi ]. By Axiom 1 and 3,
a (x) 2 B [x0 ; xn?1 ]. By Proposition 4, all the points fti ; ~ig are contained inside
the rectangles fRi g and thus a (x) 2 B [x0 ; xn?1 ]. The metric between n and a is
d(n; a ) = max jn(x) ? a (x)j
= max (max jin (x) ? a (x)j)
i x2J i

< max i
Hi
= H :
Here Hi is the height of Ri and H  is the largest of Hi . When n ! 1, intervals
Ji ! 0 and
lim H  = 0
n!1
since a (x) is continous. Thus
lim d(n ; a ) = 0
n!1
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 20

T
and
AF
Figure 15: Convergence to a dierentiable membership function.

lim  (x) = a (x):

n!1 n
DR
Thus, one can start with a few sample points of elicited membership, do engineering
calculations, and interpolate to t the results. One can then add points as necessary
to increase the accuracy of the calculations. But most importantly, the points can
be added at any additional x 2 R, and the accuracy will not decrease. The points
need not be evenly spaced, for example. This is important for engineering calcula-
tions, since this allows one to focus on critical regions by eliciting membership from
points there, with the condence that the interpolated membership will not become
skewed. Again, the interpolation scheme is very suitable for imprecise calculations
and decision making in a fuzzy environment.
As an example, consider a fuzzy number whose membership function is given
exactly by (1
2 [?; ]
a (x) = 02 (cos(x) + 1) xotherwise
but this expression remains unknown to the decision maker. The interval scale
construction can be used to nd a nite number of points to approximate a with.
This is shown in Figure 15 for dierent numbers of points.
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 21

T
AF
Figure 16: Convergence to a non-dierentiable membership function.

As a non-dierentiable example, consider a membership function given exactly

by 8
>
<1 x<0
a (x) = > ? 31 x3 + 32 x2 ? 116 x + 1 x 2 [0; 3]
:0
DR
x>3
but again this expression remains unknown to the decision maker. The interval scale
construction can be used to nd a nite number of points to approximate a with.
This is shown in Figure 16 for dierent numbers of points.
Finally, notice the method satises our requirement for a simple and ecient
implementation. Quadratic interpolation is minimal to satisfy the dierentiability
requirement across the entire curve. Thus, the method satises all of the stated
needs of fuzzy mathematics, but also does so in an ecient manner.

5 Conclusion
This paper has presented methods construct membership functions using measure-
ment theory and constrained interpolation, for use in imprecise computer-aided
engineering and decision making tools. Measurement theory oers a suitable frame-
work for constructing a membership function in cases where the membership is
based on subjective preferences. This is quite common in engineering design, where
J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 22

formal modeling is necessary to develop the product being designed. Given a -

nite set of membership values determined using a measurement scale construction,
the remaining membership values must be determined using interpolation. Further,
constrained interpolation must be used to ensure the interpolated results remain a

T
membership function. The method presented does so, and is simple and ecient.
The clearly over-restrictive assumptions needed to use the algorithm were rst
presented as axioms. We sought a smoothly varying interpolation, which may not
always be appropriate. For example, there might be a point which has a slope dis-
continuity in membership, such as a \turning-point" value in the decision. Above a
particular value, one action must be taken (e.g., use steel), and below that value an-
other decision must be taken (e.g., use aluminum). A membership function re ecting
this may not be dierentiable at that value. The interpolation algorithm presented

AF
can be easily extended to accommodate any such discontinuities in slope or posi-
tion, simply by segmenting the entire membership function into pieces, and using
appropriate end-conditions, rather than the continuous zero slope end-conditions
presented.
Acknowledgments
This work was made possible in part by an initialization grant awarded from the
Sloan Research Fund at the Massachusetts Institute of Technology, sponsored in
part by the DuPont Company, and from an MIT Leaders for Manufacturing award.
Any opinions, ndings, conclusions, or recommendations are those of the author and
do not necessarily re ect the views of the sponsors.

References
[1] A. K. Cline. Scalar- and planar-valued curve tting using splines under tension.
DR
Communications of the ACM, 17(4), April 1974.
[2] D. Krantz, R. Luce, P. Suppes, and A. Tversky. Foundations of Measurement,
volume I. Academic Press, New York, 1971.
[3] David F. McAllister and John A. Roulier. Algorithm 574, shape-preserving
osculatory quadratic splines. ACM Transactions on Mathematical Software,
7(3), September 1981.
[4] David F. McAllister and John A. Roulier. An algorithm for computing a shape-
preserving osculatory quadratic spline. ACM Transactions on Mathematical
Software, 7(3), September 1981.
[5] K. Otto. Measurement methods for product evaluation. Research in Engineer-
ing Design, 1994. Submitted for publication. Also in Proceedings of the 1993
ASME Design Theory and Methodology Conference, Albuquerque, NM, 1993,
pp. 157{166.
 J. Chen & K. Otto, Constructing Membership Functions, April 4, 1994 23

[6] K. N. Otto and E. K. Antonsson. Trade-O Strategies in Engineering Design.

Research in Engineering Design, 3(2):87{104, 1991.
[7] K. N. Otto and E. K. Antonsson. Design Parameter Selection in the Presence

T
of Noise. Research in Engineering Design, 1994. Accepted for publication.
[8] K. N. Otto and E. K. Antonsson. Modeling imprecision for product design. In
Proceedings of the 1994 FUZZ-IEEE Conference, 1994. Submitted for publica-
tion. Engineering Design Reseach Laboratory TR 93-3, Massachusetts Institute
of Technology.
[9] K. N. Otto, A. D. Lewis, and E. K. Antonsson. Fuzzy Preference Induced on
Manifolds by Vector Fields and Flows. Engineering Design Research Laboratory

AF
Report EDRL-TR 91f, California Institute of Technology, 1991.
[10] K. N. Otto, A. D. Lewis, and E. K. Antonsson. Determining optimal preference
points with dependent variables. Fuzzy Sets and Systems, 60(1), November
1993.
[11] K. L. Wood and E. K. Antonsson. Computations with Imprecise Parameters in
Engineering Design: Background and Theory. ASME Journal of Mechanisms,
Transmissions, and Automation in Design, 111(4):616{625, December 1989.
[12] K. L. Wood, K. N. Otto, and E. K. Antonsson. Engineering Design Calculations
with Fuzzy Parameters. Fuzzy Sets and Systems, 52(1):1{20, November 1992.
DR

All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately.