Professional Documents
Culture Documents
Relationships Between Input and Output Membership Functions Shapes in Fuzzy Systems
Relationships Between Input and Output Membership Functions Shapes in Fuzzy Systems
Relationships Between Input and Output Membership Functions Shapes in Fuzzy Systems
In this work we will consider a simple fuzzy system, with one input and one output, and we will
focus on the relationships between different choices for the membership functions (MF) for both
the input and the output variables, taking into account also some different defuzzification
operators. We will do this using the FIS (Fuzzy Inference System) provided by Matlab.
The system we consider is a system which models proportionality between the input and the
output variables. In order to build the MF with “simple” numbers, the input variable will range from
0 to 9 and the output variable from 0 to 8, so that the proportionality constant is not 1, but 8/9.
Both the input and the output variables will have three MF: “Low”, “Medium”, “High”, and the rules
2 Triangle:
Low = [0 0 4.5]; Medium = [0 4.5 9]; High = [4.5 9 9];
3 Gauss’ Bell
Low = [1.25 0]; Medium = [1.25 4.5]; High = [1.25 9];
4 Trapeze
Low = [0 0 2 4]; Medium = [2 4 5 7]; High = [5 7 9 9];
For each of these shapes we will use four different shapes for the output variable:
1 Singleton:
Low = [0 0 0] ; Medium = [4 4 4] ; High = [8 8 8] ;
2 Triangle:
Low = [0 0 4] ; Medium = [0 4 8] ; High = [4 8 8] ;
3 Gauss’ Bell
Low = [1.2 0] ; Medium = [1.2 4] ; High = [1.2 8] ;
4 Trapeze
Low = [0 0 1 3.5] ; Medium = [1 3.5 4.5 7] ; High = [4.5 7 8 8];
and for each input-output shapes combination we will plot the output value of the system as a
function the input value, varying over all its value range. The output value will be obtained by
1. Centroid;
centroid
MoM
LoM
We will compare all these plots, looking for the core properties of these combinations, trying also
to get some “thumb rules” on the behaviour of the system, which would be useful when building
real systems.
After this, we will do the same for a system modelling inverse proportionality, in order to see what
All the plots from this first system are shown at the end of this section.
From the input shape “Rectangle” we get a first result: using defuzzificator “centroid” means that
the output variable will never reach its border values (0 and 8), if its MFs have shapes different
from “Singleton”. This straight follows from the definition of “centroid” as an average: as soon as
some output values different from 0 (resp. 8) gets truth value different from 0, the final outcome
Moving to defuzzificator “LoM” and “MoM”, the only choice for the output’s shape that does not
allow the output to get its minimum and maximum value is “Trapeze”. This is again quite easy to
explain, thinking about the shape of “Trapeze”: as concerning defuzzificator “LoM”, since all the
values of the output variable between 0 and 1 belong to “Low” with truth degree 1, when the first
rule has truth degree 1 the Largest of Maximum is 1 (as shown in the next picture).
For the same reason, the output with “MoM” is 0.48 for input < 3 and 7.52 for input > 6 (see
picture below)
Looking at input shape “Triangle”, we see that output shape “Singleton” with defuzzificator
“Centroid” gets the right proportionality shape (more, it is the ONLY combination that gets it). The
other output shapes keep the output a continue and monotonically increasing function but make
the plot concave for small values of the input and convex for high values (the sequence concave-
convex is because of symmetry reasons). Again, when the output shape is not “Singleton”, the
“MoM” still keeps the output function non decreasing but loses the continuity, showing a step
function behaviour. This is because of the definition of “MoM” as an average just of the outputs
with the maximum truth values; since all MFs “Medium” are symmetrical with respect to 4, the
“LoM” gets a piecewise behaviour as well (for the same definition’s reasons of MoM), but even
loses the non decreasing property, and shows a particular trend (we shall say “LoM” shape), which
We can also remark that both “LoM” and “MoM” plots are polygonal chains, except for when the
output shape is “Bell”: in this case, the polygonal chain becomes a piecewise curve, still keeping
Moving to the input shape “Trapeze”, we observe that the resulting plots can be obtained by those
from input shape “Triangle”, by adding constant intervals to the “Triangle” input/output plots,
approximation of “Trapeze”; the only difference will be in the regularity of the plots: actually, the
shape “Bell” is not only continuous but also differentiable, and therefore the output plots will be
The last remark is that using “LoM” or “MoM”, when the input and the output shapes are the
Speaking about the inverse proportionality system, we can easily see that the plots for the output
of this system can be easily obtained by the previous’ one, with the expression:
centroid
Largest of Max.
Medium of Max
MF IN: RECTANGLE
Defuzzy\MF out Singletons Triangles Gauss bells Trapezes
centroid
Largest of Max.
Medium of
Max.
MF IN : TRIANGLE
Defuzzy\MF out Singletons Triangles Gauss bells Trapezes
centroid
Largest of Max.
Medium of
max.
MF IN : TRAP
Defuzzy\MF out Singletons Triangles Gauss bells Trapezes
centroid
Largest of Max.
Medium of
Max.
MF IN : BELL
INVERSE PROPORTIONALITY SYSTEM
Centroid
Largest of Max.
Medium of
Max.
MF IN : TRIANGLE
Defuzzy\mf out Singletons Triangles Gauss bells Trapezes
centroid
Largest of Max.
Medium of
Max.
MF IN : TRAP
3) CONCLUSIONS
In conclusion, we can state some facts about the relationships between input and output
MF shapes and the defuzzification operators, that would also be useful when building real
systems:
1. there are some basic shapes for the final outcome of the system, according to the
defuzzification operator used: “centroid” gets a kind of ramp (even if the output
2. only some combinations of input-output shape and defuzzificator let the output
“Centroid” is the only combination able to get the correct plot, both in the
true);
3. Changing the input shape can improve the regularity of the outcoming function; in
general we observe that the changes in the input shape reflect on the properties of