Professional Documents
Culture Documents
Abstract: - The purpose of this work is to present a new fuzzy structure, which is computationally efficient. The
main idea is to modify the original Mamdani fuzzy structure, specially the rule aggregation and defuzzification
procedures. The computational effort is similar to that required by Takagi-Sugeno-Kang structures.
C=
2 2 2
X Y
(3) The aggregation of consequents yields
M
r
Define αs and α t as the respective weights that are 3 Mamdani-like fuzzy structure
provided by the antecedent MF when the input value In spite of the qualities of Mamdani fuzzy structures,
is equal x. the TSK has received widespread acceptance in
fuzzy reasoning, especially when training and
α s =µX1 (x) ; α t =µX2 (x) (14) optimisation of the MF are pursued. This section is
The fuzzy output is the centroid of the area shown in devoted to describe a new fuzzy structure, which is
Figure 4, which can be found by using the general computationally efficient and preserves the main
expression: advantages of the Mamdani structure.
3.1 Defuzzify each consequent MF
individually
To find the output of Mamdani fuzzy structures, it is
Figure 3: Fuzzy output computation, example necessary to perform integration or to use an
approximation such as the discretezation shown in
Figure 2. Here it is proposed an alternative procedure
where defuzzification is performed on individual
weighted MFs corresponding to each rule. Basically,
the centroid of each individual weighted MF can be
Figure 4: Rules aggregation to compute fuzzy output
found directly from its parameters using (10)-(12).
1 3α ( A 2 − B 2 ) + α t ( A 2 + AB − BC − C 2 ) Some authors refers to the problem of equivalence of
yˆ = s
(15) prior defuzzification followed by aggregation or
3 2α s ( A − B) + α t ( A − C ) prior aggregation followed by defuzzification (see
Note that the MF in this example is very simple and pp. 386 of [15]) and also pp. 80 of [12]).
a closed formula for the centroid could be found and, Defining T as the vector of the centroid co-
moreover, discretezation was not needed. ordinates of each consequent MF
Ti =Dfz(Yi ) , ∀ i∈ [1,r] (19) X 1 a1 b1 c1 d1
each rule gives
MF_X= = (25)
X 2 a 2 b2 c 2 d 2
Ci =µXi (x).Ti , ∀ i∈ [1,r] (20) 2. Two trapezoidal consequent MF whose
and the C matrix can be written as parameters are defined as MF_Y rows:
C1 = µ X 1 ( x).T1 Y1 A1 B1 C1 D1
C = µ ( x ).T MF_Y= =
D2
(26)
C=
2
2 X
2
(21) Y2 A2 B2 C2
M 3. Two rules:
Cr = µ X r ( x).Tr rx1
If x is Xi then y is YI , i=1,2 (27)
Finally, aggregation leads to Then, finding the T vector:
r T = [T1 ,T2 ] = [Dfz(Y1 ) , Dfz(Y2 )] (28)
yˆ = ∑ Ci (22) where:
i =1
Dfz(Yi ) = 1 Di + Ci − Bi − Ai + Ci Di − Ai Bi
2 2 2 2
3.2 Properties
3 D+C− B− A
3.2.1 Computational effort
The simplest case of the TSK fuzzy structure is one gets by aggregating the rules to express the
obtained by setting its consequent MF as constant fuzzy output:
2
yˆ = ∑ µ X i ( x).Ti = µ X i ( x ).T1 + µ X 2 ( x).T2
values. The proposed scheme based on the T vector
requires comparable computational effort, which is (29)
i =1
significantly smaller than Mamdani case. This is due
to the fact that Yi is fully characterised by its As Ti is constant ∀ i ∈ [1,r], the fuzzy output mirrors
parameters and its centroid Ti =Dfz(Yi ) is a real the µX1 (x) and µX2 (x) characteristics.
number. In other words, the computational effort is Using the antecedent MF as proposed in equation
basically the same since instead of crisp values, the (25), each one of them is characterised by four linear
centroid co-ordinates are used in the final stage of parts:
defuzzification process. 1. µXi (x) = 0 3. µXi (x) = 1
2. µXi (x) = x − ai 4. µXi (x) = x − di
3.2.2 Shape of the input-output function bi − ai ci − d i
The fuzzy output is computed as (22), or component The next tables summarise the possible results for
wise the possible combinations (superpositions):
yˆ = µ X 1 ( x ).Dfz(Y1) + ... + µ X r ( x ).Dfz(Yr ) i) For µX1 (x) = 0
(23)
where Ti =Dfz(Yi )=constant ∀ i ∈ [1,r] µX2 (x) µX1 (x)
By inspection of (23), it is easily noticed that the 0 yˆ = 0
fuzzy output follows the weighted sum of antecedent x − a2 T2 a 2 .T2
yˆ = x −
MF. For instance, if µXi (x) are piecewise linear ∀ i ∈ b2 − a2 b2 − a 2 b2 − a 2
[1,r], i.e. µXi (x)=a i .x+bi (triangular or trapezoidal 1 yˆ = T2
antecedent MF) then the fuzzy output becomes a sum x − d2 T2 d2 .T2
yˆ = x −
of straight lines (see example in the next section). c2 − d 2 c2 − d2 c2 − d 2
r
yˆ = ∑ ( a i .x + bi ).Ti (24) ii) For µX1 (x) = x − a1
i =1
b1 − a1
µX2 (x) µX1 (x)
3.2.3 Universal approximation
As seen in the last section, the linearity between the T a .T
yˆ = 1 x − 1 1
height and the centroid co-ordinates of consequent 0 b1 − a1 b1 − a1
MFs allows the shaping of the output as a connection x − a2 T1 T2 a1 .T1 a .T
yˆ = + x − + 2 2
of portions of the antecedent MFs. This property can b2 − a2 b1 − a1 b2 − a 2 b1 − a1 b2 − a 2
be explored for universal approximation of T a .T
yˆ = 1 x + T2 − 1 1
functions, as illustrated by the following example. 1 −
1 1
b a b1 − a1
1 yˆ = T1 + T2
x − d2 T2 d .T
yˆ = x + T1 − 2 2
c2 − d 2 2
c − d 2 c2 − d2
iv) For µX1 (x) = x − d1
b1 − d1
µX2 (x) µX1 (x)
T d1.T1
yˆ = 1 x −
Figure 6: TSK results with weighted average
0 c1 − d1 c1 − d1
x − a2 T T2 d1 .T1 T2 Figure 6 shows the results corresponding to the
yˆ = 1 + x − +
b2 − a2 c1 − d1 b2 − a2 c1 − d1 b2 − a2 weighted average formula. Non-linear as expected.
T d .T Now, by using the weighted sum to obtain the fuzzy
yˆ = 1 x + T2 − 1 1
1 c1 − d1 c1 − d1 output for the same examples, one can see in Figure
x − d2 T1 T2 d1 .T1 .T2 7 that example 1 is still non-linear. In example 2
yˆ = + x − +
c2 − d 2 1
c − d 1 c 2 − d 2 1
c − d 1 c 2 − d2
however the weighted sum procedure combined with
crisp set outputs is a combination of linear segments.
Note that for whole antecedent MF conditions, the
fuzzy output result will always be a linear
composition. Recall that a large class of functions
can be approximated to an arbitrary precision by a
sufficient number of line-segments, and so the
proposed fuzzy structure can be used to provide a
versatile tool in many applications where such ability
is required.
4 Numerical examples
For the next examples define the input UD as [0,10]
and consider two antecedent MFs
Figure 7: TSK results with weighted sum
X 1 − 1 2 5 7
MF_X= = (30) 4.2 Mamdani fuzzy output
X 2 3 4 8 11 Define the consequent MF as
that are presented in Figure 5. Y1 0 3 5 6
MF_Y= = , (33)
Y2 4 7 8 9
which is graphically represented in Figure 8, and
consider the simple rule-base:
If x is X1 then y = Y1
(34)
If x is X2 then y = Y2
Figure 5: Input MF