FS V: Fuzzy rule-based systems

Triangular norms
Triangular norms were introduced by Schweizer and Sklar to model the
distances in probabilistic metric spaces. [ Associative functions and ab-
stract semigroups, Publ. Math. Debrecen, 10(1963) 69-81].
In fuzzy sets theory triangular norm are extensively used to model the log-
ical connective and.
Denition 1. (Triangular norm.) A mapping
T : [0, 1] [0, 1] [0, 1]
is a triangular norm (t-norm for short) iff it is symmetric, associative, non-
decreasing in each argument and T(a, 1) = a, for all a [0, 1]. In other
words, any t-norm T satises the properties:
Symmetricity:
T(x, y) = T(y, x), x, y [0, 1].
Associativity:
T(x, T(y, z)) = T(T(x, y), z), x, y, z [0, 1].
Monotonicity:
T(x, y) T(x
/
, y
/
) if x x
/
and y y
/
.
One identy:
T(x, 1) = x, x [0, 1].
1
These axioms attempt to capture the basic properties of set intersection.
The basic t-norms are:
minimum: min(a, b) = mina, b,
ukasiewicz: T
L
(a, b) = maxa +b 1, 0
product: T
P
(a, b) = ab
weak:
T
W
(a, b) =
_
mina, b if maxa, b = 1
0 otherwise
Hamacher:
H

(a, b) =
ab
+ (1 )(a +b ab)
, 0
Dubois and Prade:
D

(a, b) =
ab
maxa, b,
, (0, 1)
Yager:
Y
p
(a, b) = 1 min1,
p
_
[(1 a)
p
+ (1 b)
p
], p > 0
Frank:
F

(a, b) =
_

_
mina, b if = 0
T
P
(a, b) if = 1
T
L
(a, b) if =
1 log

_
1 +
(
a
1)(
b
1)
1
_
otherwise
2
All t-norms may be extended, through associativity, to n > 2 arguments.
The minimum t-norm is automatically extended and
T
P
(a
1
, . . . , a
n
) = a
1
a
2
a
n
,
T
L
(a
1
, . . . a
n
) = max
n

i=1
a
i
n + 1, 0
A t-norm T is called strict if T is strictly increasing in each argument.
Triangular co-norms are extensively used to model logical connectives or.
Denition 2. (Triangular conorm.) A mapping
S: [0, 1] [0, 1] [0, 1],
is a triangular co-norm (t-conorm) if it is symmetric, associative, non-
decreasing in each argument and S(a, 0) = a, for all a [0, 1]. In other
words, any t-conorm S satises the properties:
S(x, y) = S(y, x) (symmetricity)
S(x, S(y, z)) = S(S(x, y), z) (associativity)
S(x, y) S(x
/
, y
/
) if x x
/
and y y
/
(monotonicity)
S(x, 0) = x, x [0, 1] (zero identy)
If T is a t-norm then the equality
S(a, b) := 1 T(1 a, 1 b),
denes a t-conorm and we say that S is derived from T. The basic t-
conorms are:
3
maximum: max(a, b) = maxa, b
ukasiewicz: S
L
(a, b) = mina +b, 1
probabilistic: S
P
(a, b) = a +b ab
strong:
STRONG(a, b) =
_
maxa, b if mina, b = 0
1 otherwise
Hamacher:
HOR

(a, b) =
a +b (2 )ab
1 (1 )ab
, 0
Yager:
Y OR
p
(a, b) = min1,
p

a
p
+b
p
, p > 0.
Lemma 1. Let T be a t-norm. Then the following statement holds
T
W
(x, y) T(x, y) minx, y, x, y [0, 1].
Proof. From monotonicity, symmetricity and the extremal condition we
get
T(x, y) T(x, 1) x, T(x, y) = T(y, x) T(y, 1) y.
This means that T(x, y) minx, y.
Lemma 2. Let S be a t-conorm. Then the following statement holds
maxa, b S(a, b) STRONG(a, b), a, b [0, 1]
4
Proof. From monotonicity, symmetricity and the extremal condition we
get
S(x, y) S(x, 0) x, S(x, y) = S(y, x) S(y, 0) y
This means that S(x, y) maxx, y.
Lemma 3. T(a, a) = a holds for any a [0, 1] if and only if T is the
minimum norm.
Proof. If T(a, b) = min(a, b) then T(a, a) = a holds obviously. Suppose
T(a, a) = a for any a [0, 1], and a b 1. We can obtain the following
expression using monotonicity of T
a = T(a, a) T(a, b) mina, b.
From commutativity of T it follows that
a = T(a, a) T(b, a) minb, a.
These equations show that T(a, b) = mina, b for any a, b [0, 1].
Lemma 4. The distributive law of t-norm T on the max operator holds for
any a, b, c [0, 1].
T(maxa, b, c) = maxT(a, c), T(b, c).
Denition 3. (t-norm-based intersection) Let T be a t-norm. The T-intersection
of A and B is dened as
(A B)(t) = T(A(t), B(t)),
for all t X.
5
Example 1. Let
T(x, y) = LAND(x, y) = maxx +y 1, 0
be the ukasiewicz t-norm. Then we have
(A B)(t) = maxA(t) +B(t) 1, 0,
for all t X.
Let A and B be fuzzy subsets of
X = x
1
, x
2
, x
3
, x
4
, x
5
, x
6
, x
7

and be dened by
A = 0.0/x
1
+ 0.3/x
2
+ 0.6/x
3
+ 1.0/x
4
+ 0.6/x
6
+ 0.3/x
6
+ 0.0/x
7
B = 0.1/x
1
+ 0.3/x
2
+ 0.9/x
3
+ 1.0/x
4
+ 1.0/x
5
+ 0.3/x
6
+ 0.2/x
7
.
Then A B has the following form
AB = 0.0/x
1
+0.0/x
2
+0.5/x
3
+1.0/x
4
+0.6/x
5
+0.0/x
6
+0.2/x
7
.
The operation union can be dened by the help of triangular conorms.
Denition 4. (t-conorm-based union) Let S be a t-conorm. The S-union
of A and B is dened as
(A B)(t) = S(A(t), B(t)),
for all t X.
Example 2. Let
S(x, y) = LOR(x, y) = minx +y, 1
6
be the ukasiewicz t-conorm. Then we have
(A B)(t) = minA(t) +B(t), 1,
for all t X.
Let A and B be fuzzy subsets of
X = x
1
, x
2
, x
3
, x
4
, x
5
, x
6
, x
7

and be dened by
A = 0.0/x
1
+ 0.3/x
2
+ 0.6/x
3
+ 1.0/x
4
+ 0.6/x
5
+ 0.3/x
6
+ 0.0/x
7
B = 0.1/x
1
+ 0.3/x
2
+ 0.9/x
3
+ 1.0/x
4
+ 1.0/x
5
+ 0.3/x
6
+ 0.2/x
7
Then A B has the following form
AB = 0.1/x
1
+0.6/x
2
+1.0/x
3
+1.0/x
4
+1.0/x
5
+0.6/x
6
+0.2/x
7
.
If we are given an operator C such that
mina, b C(a, b) maxa, b, a, b [0, 1]
then we say that C is a compensatory operator.
A typical compensatory operator is the arithmetical mean dened as
MEAN(a, b) =
a +b
2
Averaging operators
Fuzzy set theory provides a host of attractive aggregation connectives for
integrating membership values representing uncertain information. These
7
connectives can be categorized into the following three classes union, in-
tersection and compensation connectives.
Union produces a high output whenever any one of the input values repre-
senting degrees of satisfaction of different features or criteria is high.
Intersection connectives produce a high output only when all of the in-
puts have high values. Compensative connectives have the property that
a higher degree of satisfaction of one of the criteria can compensate for a
lower degree of satisfaction of another criteria to a certain extent.
In the sense, union connectives provide full compensation and intersec-
tion connectives provide no compensation. In a decision process the idea
of trade-offs corresponds to viewing the global evaluation of an action as
lying between the worst and the best local ratings. This occurs in the pres-
ence of conicting goals, when a compensation between the corresponding
compabilities is allowed. Averaging operators realize trade-offs between
objectives, by allowing a positive compensation between ratings.
Denition 5. An averaging operator M is a function
M: [0, 1] [0, 1] [0, 1],
, satisfying the following properties
Idempotency
M(x, x) = x, x [0, 1],
Commutativity
M(x, y) = M(y, x), x, y [0, 1],
Extremal conditions
M(0, 0) = 0, M(1, 1) = 1
8
Monotonicity
M(x, y) M(x
/
, y
/
) if x x
/
and y y
/
,
M is continuous.
Averaging operators represent a wide class of aggregation operators. We
prove that whatever is the particular denition of an averaging operator,
M, the global evaluation of an action will lie between the worst and the
best local ratings:
Lemma 5. If M is an averaging operator then
minx, y M(x, y) maxx, y, x, y [0, 1]
Proof. From idempotency and monotonicity of M it follows that
minx, y = M(minx, y, minx, y) M(x, y)
and M(x, y) M(maxx, y, maxx, y) = maxx, y. Which ends the
proof.
Averaging operators have the following interesting properties:
Property 1. A strictly increasing averaging operator cannot be associa-
tive.
Property 2. The only associative averaging operators are dened by
M(x, y, ) = med(x, y, ) =
_

_
y if x y
if x y
x if x y
where (0, 1).
9
An important family of averaging operators is formed by quasi-arithmetic
means
M(a
1
, . . . , a
n
) = f
1
_
1
n
n

i=1
f(a
i
)
_
This family has been characterized by Kolmogorov as being the class of
all decomposable continuous averaging operators. For example, the quasi-
arithmetic mean of a
1
and a
2
is dened by
M(a
1
, a
2
) = f
1
_
f(a
1
) +f(a
2
)
2
_
.
The next table shows the most often used mean operators.
Name M(x, y)
harmonic mean 2xy/(x +y)
geometric mean

xy
arithmetic mean (x +y)/2
dual of geometric mean 1
_
(1 x)(1 y)
dual of harmonic mean (x +y 2xy)/(2 x y)
median med(x, y, ), (0, 1)
generalized p-mean ((x
p
+y
p
)/2)
1/p
, p 1
Mean operators.
The process of information aggregation appears in many applications re-
lated to the development of intelligent systems. One sees aggregation in
10
neural networks, fuzzy logic controllers, vision systems, expert systems
and multi-criteria decision aids. In 1988 Yager introduced a new aggrega-
tion technique based on the ordered weighted averaging (OWA) operators.
Denition 6. An OWA operator of dimension n is a mapping F : IR
n
IR,
that has an associated weighting vector W = (w
1
, w
2
, . . . , w
n
)
T
such as
w
i
[0, 1], 1 i n, and
w
1
+ +w
n
= 1.
Furthermore
F(a
1
, . . . , a
n
) = w
1
b
1
+ +w
n
b
n
=
n

j=1
w
j
b
j
where b
j
is the j-th largest element of the bag a
1
, . . . , a
n
).
Example 3. Assume W = (0.4, 0.3, 0.2, 0.1)
T
then
F(0.7, 1, 0.2, 0.6) = 0.4 1 + 0.3 0.7 + 0.2 0.6 + 0.1 0.2 = 0.75.
A fundamental aspect of this operator is the re-ordering step, in particular
an aggregate a
i
is not associated with a particular weight w
i
but rather a
weight is associated with a particular ordered position of aggregate. When
we view the OWA weights as a column vector we shall nd it convenient
to refer to the weights with the low indices as weights at the top and those
with the higher indices with weights at the bottom.
It is noted that different OWA operators are distinguished by their weight-
ing function. In 1988 Yager pointed out three important special cases of
OWA aggregations:
F

: In this case W = W

= (1, 0 . . . , 0)
T
and
F

(a
1
, . . . , a
n
) = maxa
1
, . . . , a
n
,
11
F

: In this case W = W

= (0, 0 . . . , 1)
T
and
F

(a
1
, . . . , a
n
) = mina
1
, . . . , a
n
,
F
A
: In this case W = W
A
= (1/n, . . . , 1/n)
T
and
F
A
(a
1
, . . . , a
n
) =
a
1
+ +a
n
n
.
A number of important properties can be associated with the OWA opera-
tors. We shall now discuss some of these. For any OWA operator F holds
F

(a
1
, . . . , a
n
) F(a
1
, . . . , a
n
) F

(a
1
, . . . , a
n
).
Thus the upper an lower star OWA operator are its boundaries. From the
above it becomes clear that for any F
mina
1
, . . . , a
n
F(a
1
, . . . , a
n
) maxa
1
, . . . , a
n
.
The OWA operator can be seen to be commutative. Let a
1
, . . . , a
n
) be a
bag of aggregates and let d
1
, . . . , d
n
be any permutation of the a
i
. Then
for any OWA operator
F(a
1
, . . . , a
n
) = F(d
1
, . . . , d
n
).
A third characteristic associated with these operators is monotonicity. As-
sume a
i
and c
i
are a collection of aggregates, i = 1, . . . , n such that for
each i, a
i
c
i
. Then
F(a
1
, . . . , a
n
) F(c
1
, c
2
, . . . , c
n
)
where F is some xed weight OWA operator.
Another characteristic associated with these operators is idempotency. If
a
i
= a for all i then for any OWA operator
F(a
1
, . . . , a
n
) = a.
12
k
k+m-1
1
n
1/m
From the above we can see the OWA operators have the basic properties
associated with an averaging operator.
Example 4. A window type OWA operator takes the average of the m ar-
guments around the center. For this class of operators we have
w
i
=
_

_
0 if i < k
1
m
if k i < k +m
0 if i k +m
(1)
Figure 1: Window type OWA operator.
In order to classify OWA operators in regard to their location between and
and or, a measure of orness, associated with any vector W is introduce by
Yager as follows
orness(W) =
1
n 1
n

i=1
(n i)w
i
.
It is easy to see that for any W the orness(W) is always in the unit interval.
Furthermore, note that the nearer W is to an or, the closer its measure is to
one; while the nearer it is to an and, the closer is to zero.
Lemma 6. Let us consider the the vectors
W

= (1, 0 . . . , 0)
T
, W

= (0, 0 . . . , 1)
T
,
13
W
A
= (1/n, . . . , 1/n)
T
.
Then it can easily be shown that
orness(W

) = 1, orness(W

) = 0
and orness(W
A
) = 0.5.
A measure of andness is dened as
andness(W) = 1 orness(W).
Generally, an OWA opeartor with much of nonzero weights near the top
will be an orlike operator, that is,
orness(W) 0.5
and when much of the weights are nonzero near the bottom, the OWA
operator will be andlike, that is,
andness(W) 0.5.
Example 5. Let W = (0.8, 0.2, 0.0)
T
. Then
orness(W) =
1
3
(2 0.8 + 0.2) = 0.6,
and
andness(W) = 1 orness(W) = 1 0.6 = 0.4.
This means that the OWA operator, dened by
F(a
1
, a
2
, a
3
) = 0.8b
1
+ 0.2b
2
+ 0.0b
3
= 0.8b
1
+ 0.2b
2
,
where b
j
is the j-th largest element of the bag a
1
, a
2
, a
3
), is an orlike
aggregation.
14
The following theorem shows that as we move weight up the vector we
increase the orness, while moving weight down causes us to decrease
orness(W).
Theorem 1. (Yager, 1993) Assume W and W
/
are two n-dimensional OWA
vectors such that
W = (w
1
, . . . , w
n
)
T
,
and
W
/
= (w
1
, . . . , w
j
+, . . . , w
k
, . . . , w
n
)
T
where > 0, j < k. Then orness(W
/
) > orness(W).
Proof. From the denition of the measure of orness we get
orness(W
/
) =
1
n 1

i
(n i)w
/
i
=
1
n 1

i
(n 1)w
i
+ (n j) (n k),
orness(W
/
) = orness(W) +
1
n 1
(k j).
Since k > j, orness(W
/
) > orness(W).
In 1988 Yager dened the measure of dispersion (or entropy) of an OWA
vector by
disp(W) =

i
w
i
ln w
i
.
We can see when using the OWAoperator as an averaging operator Disp(W)
measures the degree to which we use all the aggregates equally.
15
1
x
0
x
0
_
Figure 2: Fuzzy singleton.
Suppose now that the fact of the GMP is given by a fuzzy singleton. Then
the process of computation of the membership function of the consequence
becomes very simple.
For example, if we use Mamdanis implication operator in the GMP then
rule 1: if x is A
1
then z is C
1
fact: x is x
0
consequence: z is C
where the membership function of the consequence C is computed as
C(w) = sup
u
min x
0
(u), (A
1
C
1
)(u, w) =
sup
u
min x
0
(u), minA
1
(u), C
1
(w),
for all w. Observing that x
0
(u) = 0, u ,= x
0
, the supremum turns into a
simple minimum
C(w) = min x
0
(x
0
) A
1
(x
0
) C
1
(w) =
16
A1 C1
x
0
w
u
A1(x0)
C
A1
u w
C1
C
x
0
Figure 3: Inference with Mamdanis implication operator.
min1 A
1
(x
0
) C
1
(w) = minA
1
(x
0
), C
1
(w)
for all w.
and if we use G odel implication operator in the GMP then
C(w) = sup
u
min x
0
(u), (A
1
C
1
)(u, w) =
A
1
(x
0
) C
1
(w)
for all w.
So,
C(w) =
_
1 if A
1
(x
0
) C
1
(w)
C
1
(w) otherwise
17
Inference with G odel implication operator.
rule 1: if x is A
1
then z is C
1
fact: x is x
0
consequence: z is C
where the membership function of the consequence C is computed as
C(w) = sup
u
min x
0
(u), (A
1
C
1
)(u, w) =
A
1
(x
0
) C
1
(w)
for all w.
18
Consider a block of fuzzy IF-THEN rules
1
1
: if x is A
1
then z is C
1
also
1
2
: if x is A
2
then z is C
2
also
. . . . . . . . . . . .
also
1
n
: if x is A
n
then z is C
n
fact: x is x
0
consequence: z is C
The i-th fuzzy rule from this rule-base
1
i
: if x is A
i
then z is C
i
is implemented by a fuzzy implication R
i
and is dened as
R
i
(u, w) = (A
i
C
i
)(u, w) = A
i
(u) C
i
(w)
for i = 1, . . . , n.
19
Find C from the input x
0
and from the rule base
1 = 1
1
, . . . , 1
n
.
20
Interpretation of
sentence connective also
implication operator then
compositional operator
We rst compose x
0
with each R
i
producing intermediate re-
sult
C
/
i
= x
0
R
i
for i = 1, . . . , n.
C
/
i
is called the output of the i-th rule
C
/
i
(w) = A
i
(x
0
) C
i
(w),
for each w.
Then combine the C
/
i
component wise into C
/
by some aggre-
gation operator:
C =
n
_
i=1
C
/
i
= x
0
R
1
x
0
R
n
C(w) = A
1
(x
0
) C
1
(w)
21
A
n
(x
0
) C
n
(w).
22
So, the inference process is the following
input to the system is x
0
fuzzied input is x
0
ring strength of the i-th rule is A
i
(x
0
)
the i-th individual rule output is
C
/
i
(w) := A
i
(x
0
) C
i
(w)
overall system output (action) is
C = C
/
1
C
/
n
.
overall system output = union of the individual rule outputs
Mamdani (a b = a b)
input to the system is x
0
fuzzied input is x
0
ring strength of the i-th rule is A
i
(x
0
)
the i-th individual rule output is
C
/
i
(w) = A
i
(x
0
) C
i
(w)
23
A1
A2
C1
C2 = C'2
C'1
degree of match
degree of match
individual rule output
individual rule output
x0
x0
overall system output
overall system output (action) is
C(w) =
n

i=1
A
i
(x
0
) C
i
(w)
Larsen (a b = ab)
input to the system is x
0
fuzzied input is x
0
ring strength of the i-th rule is A
i
(x
0
)
24
x0
A1
A2
C1
C2
A1( x0)
A2( x0)
C'1
C'2
C = C'2
the i-th individual rule output is
C
/
i
(w) = A
i
(x
0
)C
i
(w)
overall system output (action) is
C(w) =
n

i=1
A
i
(x
0
)C
i
(w)
The output of the inference process so far is a fuzzy set, spec-
ifying a possibility distribution of the (control) action. In the
on-line control, a nonfuzzy (crisp) control action is usually re-
quired. Consequently, one must defuzzify the fuzzy control
25
action (output) inferred from the fuzzy reasoning algorithm,
namely:
z
0
= defuzzifier(C),
where z
0
is the crisp action and defuzzier is the defuzzica-
tion operator.
Denition 7. (defuzzication) Defuzzication is a process to
select a representative element from the fuzzy output C in-
ferred from the fuzzy control algorithm.
26