Sie sind auf Seite 1von 15

4

REASONING
WITH IMPRECISE AND/OR UNCERTAIN RULES

4.1 GENERALIZED

MODUS PONENS RULE

Zadeh (1979) introduced the theory of approximate reasoning in order to work with imprecise and uncertain information. The basic problem of approximate reasoning is to find the membership function of the consequence C from the rule base { R1 , , Rn } and the fact A : R1 : if x is A1 then y is C1
R2 :

if

x is

A2 then y is C 2

. Rn : if x is An then y is C n

x is A Fact ________________________________ y is C Consequence


In fuzzy logic and approximate reasoning, the most important fuzzy implication inference rule is the Generalized Modus Ponens (GMP), based on the compositional rule of inference suggested by Zadeh (1973). Definition 4.1. (compositional rule of inference)

Rule if x is A then y is B x is A' Fact _____________________________________ y is B' Consequence where the consequence B' is determined as a composition of the fact and the fuzzy implication operator
B' = A' ( A B )

that is

B' ( v ) = sup min( A' ( u ) , AB ( u , v ) )


uU

where A and A' are fuzzy subsets of the universe U and B and B' are fuzzy subsets of the universe V . In many practical cases instead of sup min composition one use sup T composition, where T is a t-norm: B' = A' T ( A B ) that is

B' ( v ) = sup T ( A' ( u ) , AB ( u ,v ) ) .


u U

Suppose that A, B and A' are fuzzy numbers. The Generalized Modus Ponens should satisfy some rational properties. Suppose we are given a set of fuzzy rules
R1 :

R2 :

x is if x is
if

A1 then y is B1

A2 then y is B2

. Rn : if x is An then y is Bn

x is A' Fact ________________________________ y is C Consequence


The i -th fuzzy rule 92

Ri :

if

x is

Ai then y is Bi

is implemented by a fuzzy implication I i defined as


I i ( u , v ) = Ai Bi ( u , v ) = Ai ( u ) Bi ( v ) .

There are two main approaches to determine the membership function of consequence C . If the combination operator is denoted by {min, max} ( or, more generally, {T , S } ) we can: combine the rules first:

R ( u ,v ) = A1B1 ( u ,v ) An Bn ( u ,v )

C ( v ) = sup T ( A' ( u ) , R ( u ,v ) )
uU

fire the rules first:


B'k ( v ) = sup T ( A' ( u ) , Ak Bk ( u , v ) ) , k {1, 2, n}
u U

C ( v ) = B '1 ( v ) B '2 ( v ) B 'n ( v ) .


A question appears: the two methods give the same result? The answer is given by the following theorems (Fullr, 1995): Theorem 4.1. If T = min or T ( x , y ) = xy and = max the two method give the same result.

4.2 UNCERTAIN
REASONING

GENERALIZED

MODUS

PONENS

We consider the case when the rules and the facts are uncertain, the uncertainty being expressed by belief degrees or by linguistic variables. The reasoning with belief degrees has the form 93

if

x is A then y is B : x is A' :
_______________________ y is B' :

[0, 1] represent the belief degrees corresponding to the , , where rule, the fact and the conclusion, respectively. The conclusion is obtained
as B' = A' T ( A B ) and its associated belief degree is = T ( , ) , where T and are t-norms operators. For multiple premises, the T inferred schema becomes
An A1 C : A' n

: n : 1
C' :

A'1

______________________ where the conclusion C' is obtained with the formula

C' = A'1 T ( T ( A' n T ( An A1 C ) ) ...)


and the belief degree is

= T ( T ( 1 , , n ) ) .

If the same conclusion is obtained from m different rules, with the belief degrees 1 , 2 , , m , then the global belief degree is computed with the formula = S ( 1 , 2 , , m ) where is a t-conorm. Now, S consider the case of linguistic uncertainty (Leung & Lam, 1989; Leung, Wong & Lam, 1989). Suppose there is a rule and a fact
if X is A then Y is B

: :

FN R FN F

X is A'

____________________________________ Y is B' : FN C 94

FN R , FN F and FN C are fuzzy numbers denoting the uncertainty of

the rule, of the fact and of the conclusion, respectively. If the object X is non-fuzzy, A and A' must be the same atomic symbol in order to apply this rule. Therefore, B' = B and the fuzzy uncertainty FN C is calculated using fuzzy number multiplication of FN R and FN F :
FN C = FN R FN F .

If

X and Y are fuzzy objects, the conclusion B' is obtained using Generalized Modus Ponens and the uncertainty FN C is
FN C = ( FN R FN F ) D

where D is the degree of matching between A and A' . If X is fuzzy and Y is not-fuzzy then B' = B and the uncertainty
FN C is computed as

FN C = ( FN R FN F ) M

where M is the similarity between the fuzzy sets F and F' of A and A' , respectively: if ( F ; F' ) > 0.5 then M = ( F ; F' ) else M = ( N ( F ; F' ) + 0.5) ( F ; F' ) ,

( F ; F ') = max( min ( F ( u ) , F ' ( u ) ) ) , ( F ; F' ) =1 ( F ; F' ) .

Because a rule with multiple consequence can be treated as multiple rules with a single conclusion, only the problem of multiple propositions in the antecedent and a single proposition in the consequence need to be considered. If the object in the consequent proposition is nonfuzzy, no special treatment is needed. However, if the consequent proposition is fuzzy, the fuzzy set of the value B' in the conclusion is calculated using the following two basic algorithms (Mizumoto, 1985): a) rule: fact:
if A1 and A2 then Y is B A'1 , A' 2

95

conclusion:

Y is B'

algorithm: the fuzzy set representing B' in the conclusion is obtained by union of the fuzzy sets F1 and F2 , where the fuzzy set F1 is obtained by GMP from the rule if
A'1 , while the fuzzy set if F2 A1 then Y is B and the fact

is obtained from the rule

A2 then Y is B and the fact A' 2 .

b) rule: fact:

if

A1 or A2 then Y is B A'1 , A' 2

Y is B' conclusion: algorithm: the same as above except fuzzy intersection rather than

union should be applied on the fuzzy sets F1 and F2 . The above two algorithms can be applied repeatedly to handle any combination of antecedent propositions. For instance Rule: IF ( the productivity is high OR the production-cost is low ) AND the sales are high THEN the performance of the company should be good Facts: The productivity is very high The production-cost is low The sales are rather high where low, high, very low and rather high are fuzzy concepts. Let F1 be the fuzzy set obtained by making an inference from the rule IF the productivity is high THEN the performance of the company should be good and the fact The productivity is very high;
F2 be the fuzzy set obtained by making an inference from the rule

IF the production-cost is low THEN the performance of the company should be good and the fact 96

The production-cost is low;


F3 be the fuzzy set obtained by making an inference from the rule

IF the sales are high THEN the performance of the company should be good and the fact The sales are rather high. The fuzzy set F representing the fuzzy value of the object the performance of the company in the conclusion is determined as follows:

F = fuzzy union between F12 and F3 , where F12 is the fuzzy


intersection between F1 and F2 . As a result, F will indicate the fuzzy concept good and the conclusion the performance of the company is good is drown. The fuzzy uncertainty of the conclusion deduced from rules with multiple antecedent propositions is calculated by employing fuzzy number arithmetic operators in the formulae used by MYCINs CF model. For example, for: rule: fact: conclusion: we have
if A1 and A2 then C : FN R
A'1 , A' 2

: FN F1 , FN F 2
C' : FN C'

FN C' = min _fn FN F1 , FN F2 FN R


where FN R , FN F1 , FN F 2 and FN C' represent the uncertainty of the rule, the facts and the conclusion, respectively, is the fuzzy multiplication and the minimum of two fuzzy numbers is given by

min_ fn ( A ,B ) ( z ) = max min( A ( x ) , B ( y ) ) .


z = min ( x,y )

Similarly, the maximum is defined as

max_ fn ( A ,B ) ( z ) = max max( A ( x ) , B ( y ) ) .


z = max ( x,y )

If logical OR is used, the calculation is the same except that the fuzzy maximum is taken rather than the minimum. For the combination of 97

antecedent propositions, the two calculations can be applied repeatedly to handle fuzzy uncertainties corresponding to the matched facts and the rule. In some cases, there is more than one rule with the same consequent proposition. Each of these rules can be treated as giving contributed evidence towards the conclusion. The conclusion C R is obtained from the evidence contributed by these rules and facts. For instance: rules facts
r 1: r2 : if A1 then B

if A2 then B

A'1 , A' 2

conclusion C R obtained from C1 : FN C1 and C 2 : FN C2 , where C1 and C 2 are the conclusions obtained from r1 & A'1 and
r2 & A' 2 , respectively, and

FN C1 , FN C2 represent the uncertainties

of the conclusions. If the object involved in the consequent proposition is fuzzy then the fuzzy set corresponding to the combined conclusion C R is obtained by taking the fuzzy intersection between the fuzzy sets corresponding to C1 and
C 2 . The

two uncertainties FN C1 and overall uncertainty, using a

FN C2 can be combined, to obtain the

formula similar to the evidence combination from MYCINs CF model:


FN C = FN C1 FN C2 FN C1 FN C2 .

4.4 FUZZY

LOGIC CONTROL

In fuzzy logic controller (FLC) the dynamic behavior of a fuzzy system is characterized by a set of linguistic description rules based on expert knowledge. The expert knowledge is usually of the form 98

IF (a set of conditions are satisfied ) THEN (a set of consequences can be inferred). Fuzzy logic control systems usually consist of four major parts: Fuzzification interface, Fuzzy rule-based, Fuzzy inference machine and Defuzzification interface.

Figure 4.1: Fuzzy logic controller A fuzzification operator has the effect of transforming crisp data into fuzzy set. In most of cases one use fuzzy singletons as fuzzifiers:
fuzzifier ( x0 ) = x0 where x0 is a crisp input value from a process.

Figure 4.2: Fuzzy singleton as fuzzifier

99

For a system with two inputs and a single output, the fuzzy rule-base has the form
R1: R2 :
Rn :

if x is A1 and y is B1 then z is C1 if x is A2 and y is B2 then z is C 2


if x is An and y is Bn then z is C n

Ai U , Bi V , Ci W i {1, 2, , n}

where

and the rules are

aggregated by union or intersection. If the crisp inputs x0 and y 0 are presented then the control action is obtained as follows: The firing level of the

i = min( Ai ( x0 ) , Bi ( y 0 ) ) ; evidently, any t-norm can be used


output of the

i -th

rule

is

determined

by

instead of min The

i -th

rule

is

computed

by

C'i ( w) = i Ci ( w) w W
The overall system output, C , is obtained from the individual rule outputs C' i by an aggregation operation

C ( w) = Agg C '1 ( w) , , C 'n ( w)


Defuzzification methods

w W .

The output of the inference process is a fuzzy set that specifies the possibility distribution of control action. In the on-line control, a crisp control action is usually required. Consequently, one must defuzzify the fuzzy control action inferred, namely z 0 = defuzzifier ( C ) . The most used defuzzification operators are Center of Area. The defuzzified value of a fuzzy set C is defined by

100

z0 =

z ( z )dz
C W

( z ) dz
C

or z 0 =

z (z ) (z )
j C j C j

depending on the form of membership function C : continuous or discrete First of maxima. The defuzzified value of a fuzzy set C is the smallest maximizing element, i. e.

z0 = min z / C ( z ) = max C ( w)
W

Middle of Maxima. The defuzzified value of a discrete fuzzy set C is defined as a mean of all values { z1 , , z n } of the universe of discourse having maximal membership grades: z0 = If C is not discrete then
z0 =
G

1 n zi . n i =1

zdz
G

dz

where G denotes the set of maximizing element of C . Max Criterion. This method chooses an arbitrary value, from the set of maximizing elements of C :

z0 z / C ( z ) = max C ( w) . W

101

Inference mechanisms We present the most important inference mechanisms in fuzzy logic control systems and, for simplicity, we consider two fuzzy control rules of the form
R1: R2 : if x is A1 and y is B1 then z is C1 if x is A2 and y is B2 then z is C 2
x is x 0 and y is y 0

Fact:

_____________________________________ z is C Consequence Mamdanis model. The fuzzy implication is modeled by Mamdanis minimum operator, for the aggregation of rules is used the max operator, the conjunction operator is min and the t-norm from Generalized Modus Ponens rule is min . The firing levels of the rules are

1 = min A1 ( x0 ) , B1 ( y 0 ) , 2 = min A2 ( x0 ) , B2 ( y 0 )
Using the GMP rule, the conclusion given by first rule is

)
) ))

C '1 ( w) =
Because

( u ,v )U V

sup

min min x0 ( u ) , y0 ( v ) , min min A1 ( u ) , B1 ( v ) , C1 ( w)

x0 ( u ) = 0 u x0 and y0 ( v ) = 0 y y 0

the supremum becomes minimum and therefore it results

C '1 ( w) = min 1 , C1 ( w) , C '2 ( w) = min 2 , C2 ( w) .


The overall system output is computed as

C ( w) = max C '1 ( w) , C '2 ( w) .

102

Finally, to obtain a deterministic control action, any defuzzification strategy can be employed. Tsukamotos model. All linguistic terms are supposed to have monotonic membership functions. The firing levels of the rules are computed such as in the Mamdanis model. The individual crisp control actions z1 and z 2 are computed from the equations

1 = C1 ( z1 ) , 2 = C2 ( z 2 )

And the overall crisp control action is expressed using the discrete center of area z + 2 z2 z0 = 1 1 . 1 + 2 Sugenos model. Sugeno and Takagi used the following architecture (Takagi & Sugeno, 1985).
R1: R2 : if x is A1 and y is B1 then z 1 = a1 x + b1 y if x is A2 and y is B2 then z 2 = a 2 x + b2 y
x is x 0 and y is y 0
z0

Fact:

_____________________________________________ Consequence The firing levels of the rules are computed as in the previous models, the individual rule outputs are
z1 = a1 x0 + b1 y 0 , z 2 = a 2 x0 + b2 y 0

and the crisp control action is expressed as


z0 =
* 1 z1 + 2 z* 2 . 1 + 2

Larsens model. The fuzzy implication is modeled by Larsens product operator, rules aggregation is made by union and the conjunctive operator is min . The firing levels are computed as 103

1 = min( A1 ( x 0 ) , B1 ( y 0 ) ) , 2 = min( A2 ( x 0 ) , B2 ( y 0 ) )
and the conclusion is given by

C ( w) = max( 1 C1 ( w) , 2 C2 ( w) ) .
In order to obtain a deterministic control action, is used a defuzzification strategy. Example 4.1. As an example, we illustrate Sugenos reasoning method.
R1: R2 : if x is BIG and y is SMALL then z 1 = x + 2 y if x is MEDIUM and y is BIG then z 2 = 3 x 2 y

x is 3 and y is 2 Fact: _________________________________________________

Consequence

z0

Figure 4.3: 104

Sugenos inference mechanism

According to the previous figure, we have

BIG ( x 0 ) = BIG ( 3) = 0.9 , SMALL ( y 0 ) = SMALL ( 2) = 0.3


and

MEDIUM ( x0 ) = MEDIUM ( 3) = 0.5 , BIG ( x0 ) = BIG ( 2) = 0.8 .


The firing levels of the rules are

1 = min{ BIG ( x0 ) , SMALL ( y 0 )} = min{ 0.9, 0.3} = 0.3


2 = min{ MEDIUM ( x0 ) , BIG ( y 0 )} = min{ 0.5, 0.8} = 0.5
The individual rule outputs are computed as
z1 = x0 + 2 y 0 = 3 + 4 = 7 , z 2 = 3 x0 2 y 0 = 9 4 = 5

so the crisp control action is


z0 = 7 0.3 + 5 0.5 = 5.75 0.3 + 0.5

105

Das könnte Ihnen auch gefallen