You are on page 1of 23

INFORMATION SCIENCES 45,129-151(1988)

129

Fuzzy fhntrok UnderVarious Fuzzy Reasonhq Methods


MASAHARU

MLZUMOTO

Department of Management Engineering,


Osaka Electra-Communication
University, Neyagawa,

Osaka 572, Japan

ABSTRACT
A fuzzy logic controller consists of iinguistic control rules tied together by means of two
concepts: fuzzy implications and a compositional rule of inference. Most of the existing fuzzy
logic controllers are based on the approximate reasoning method due to Mamdani. This paper
introduces other fuzzy implications, such as the arithmetic rule and maximin rule, for linguistic
control rules and compares control results for a plant model with first order delay under
various approximate reasoning methods. Moreover, control results are compared when the
widths of fuzzy sets of linguistic control rules are changed.

1.

~~ODUC~ON

A number of studies on fuzzy logic controllers have been reported since


Mamdani [l} implemented a fuzzy logic controller on a boiler steam engine. A
fuzzy logic controller consists of linguistic control rules tied together by means
of two concepts: fuzzy implications and a compositional rule of inference. Most
of the existing fuzzy logic controllers are based on the approximate reasoning
method due to Mamdani, in which the implication for a control rule If x is A
then y is B is expressed as the direct product A X B of the fuzzy sets A
and B.
In this paper we introduce other fuzzy implications, such as the arithmetic
rule and the maximin rule, for linguistic control rules due to Yamazaki and
Sugeno [4], and compare control results for a plant model with first order delay
under the various fuzzy reasoning methods. Furthermore, we investigate how
control results are influenced when the widths of fuzzy sets of linguistic control
rules are changed.
@F,lsevier Science Publishing Co., Inc. 1988
52 Vanderbilt Ave., New York, NY 10017

OOZO-0255,88/$03.50

~S~RU

130

2.

MIZUMOTO

FUZZY REASONING METHODS

We shall consider the following form of inference in which a fuzzy implication is contained, where A, A are fuzzy sets in U, and B, B are fuzzy sets in V,
Anti:
Ant2:

IfxisAthenyisB
x is A

cons: y is B
The consequence B is deduced from Ant1 and Ant2 by taking the max-min
composition 0 of the fuzzy set A and the fuzzy relation A -+ B obtained from
the fuzzy implication if A then B . Namely, we have
B-Ao(A+B),

When the fuzzy set A is a singleton uo, that is, pLAP(


ua) = 1 and pA,(u) = 0
for u # u,,, the consequence B is simplified as

When the fuzzy implication A + B is represented by the direct product


A x B of fuzzy sets A and B as in the case of Mamdanis method [l], B is
given as

We list several fuzzy implications A + B in Table 1[2] which will be used in


the discussion of fuzzy logic controls.
EXAMPLE 1, Let A and 3 be fuzzy sets in U and V, respectively, as in Figure
1. Then the consequence 3 of (1) at A = u. under the fuzzy implications
A + B in Table 1 are depicted in Figure 2, where pA(us) = o with u = 0.3
(dotted line) and (I = 0.7 (solid line).

VARIOUS

FUZZY

REASONING

131

METHODS
TABLE 1

PI

Fuzzy Implications pA _ & uo, 0) = BA(UO)-+ Ps(U)


Rc:

[by Mamdani]
[by Larsen]
[bounded product]

Rp:
Rbp:

PAUO)APa(U)
a(uo)~Pa(u)
o[P,(~o)+cs(u)-ll

Rdp:

PA(UO). Pa(U)=l
P&?(U)>PA(UO)
=l
PA(UOhUB(U)<l
i 0,

[drastic product]

Rg:

[arithmetic rule by Zadeh]


1~~1-cL(~o)+PB(~)l
[maximin rule by Zadeh]
~PA~~O~~PB~~~I~~-~~~O~I
[Boolean implication]
[1-PAUo)IPe(U)
[by Bandler]
l- cLAuo)+P(UOh(U)
~~-~,~~O~ILB~~~I~~B~~~~~~-CL~~O~I
[by Bandler]
A[PB(u)~-PB(u)~
1. PA(UO)Sk?(U)
[standard sequence]
i 0. IL(UO)>Be(U)
1,
P(UO)4Pa(U)

RA:

[by Gougen]

Ra:
Rm:
Rb:
R*:
Rft:
Rs:

i PB(U>Y PA(UO)>BB(U)
1,
P(UO)satJ(u>
CB(UVP(UO)l BA(UO)BB(U)

[Giidelian logic]

Fig. 1. Fuzzy sets A and B.

We shall next consider the following form of inference in which the hypothesis of a fuzzy conditional proposition If.. . then.. . contains two fuzzy propositions x is A and y is B combined using the connective and.
Ant1 :
Ant2:

If x is A and y is B then z is C
x is A and y is B

Cons:

z is C

where A, A are fuzzy sets in U, and B, B in V, and C, C in W.

(2)

132

MASAHARU MIZUMOTO

G
@i
.r
_-

.-._.

-5

VARIOUS

FUZZY

REASONING

METHODS

133

The consequence
C can be deduced from Ant1 and Ant2 by taking the
max-min composition
0 of a fuzzy set (A and B) in U x Y and a fuzzy
relation (A and B) + C in U X V X W. Namely, we have
C=(Aand

B)o[(A

and B) -C]

In the case of Mamdanis method Rc in Table 1, the fuzzy implication


[(A and B) + C] is translated into pA(
pB(u)A
p=(w) in virtue of a + b =
a A b. Thus, the consequence C is given as

PC(W)= v {[PA u) N3441 A[C1A(U)ACLg(U)A~UC(W)l}.


(4)
U.

Let Rc(A, B; C) = (A and B) + C, Rc(A; C) = A + C, and Rc( B; C) = B + C


be fuzzy implications by Mamdanis method Rc. Then the consequence C is
reduced from (4) as follows:

r&)=V(

PA<

( u) A~~(u)Arc(w)Av[Ir,,(u)A~,(u)APc(W)I

Therefore, the consequence C = (A and B) 0 Rc( A, B; C) can be obtained as


the intersection of A 0 Rc( A ; C) and B 0 Rc( B; C) for Mamdanis implication
Rc. Namely,
C=(Aand

B)oRc(A,B;C)

=[AoRc(A;C)]n[BoRc(B;C)].
Similarly,

we can have

(Aand

B)o[(Aand

for the fuzzy implications

B) -C]

=[A+4+C)]n[B~(B+C)]

Rp, Rbp, and Rdp in Table 1.

(5)

MASAHARU

134

MIZUMOTO

It is noted that Ra, Rb, R*, Rs, Rg, and RA in Table 1, for which the
equality (u A b) + c = (u --) c) v (b + c) holds, satisfy the following (see [3]):
(Aand

B)#Aand

B) +C]

=[A+4+C)]u[B+-+C)].

(6)

When the fuzzy sets A and B are singletons in (2), i.e., A = u. and B = u,,
the consequence C of (3) is abbreviated as

{l[h~%h~b~))

-P&>]>

= M~o)~PcLg(~cl)I
-+Pc(W).

(7)

For example, in the case of Rc and Ra we have the consequences C at A = ZQ,


and B = u, as follows (the same can be obtained from other fuzzy implications
in Table 1):

RC:

~A(UO)PB(UO)I
kb)~

(8)

In the above discussion, the operation A ( = min) is used as the meaning of


and in the approximate reasoning of (2). It is possible to introduce other
operations, say, algebraic produce . and, more generally, t-norms as and. For
example, the consequence C at A = u,, and B = u, will be

when the algebraic product . is used as and. If the fuzzy implication


Table 1 is used in 4, the consequence C becomes

PlfO =cL(1(o)~cI-~(~~)~c1~(w)~

Rp of

(11)

EXAMPLE 2. Figure 3(a) and (b) show the consequences C by Rc (8) and Ra
(9) at A = u. and B = u,. Figure 3(c) indicates the consequence C by (11). In
a similar way, we can obtain consequences C at A = u. and B = u,, by other
fuzzy implications in Table 1 from (7) and (10) by letting pA(
pB(u,,) = a
or pA( z+,).~,( uO) = a in Figure 2.

VARIOUS

FUZZY REASONING

135

METHODS
C ,*
*

(a>

(b)

Fig. 3. Inference results C at pA(u,,) = 0.8 and pe(uo) = 0.6: (a) pc,(w) = (pA(
AC&W) of (8); (b) ~c,(w)=lA[l-(~,(u,)A~,(u,))+~~c(w)l
of (9); (4

~P~~O~~PS~~O~l~PC~~~
of (11).

pB(uo)]
PC,(W)=

136

MASAHARU

MIZUMOTO

As a generalized form of approximate reasoning of (2), we shall consider


approximate
reasoning with several fuzzy conditional propositions combined
with else:
Arul:
If x is A, and y is B, then z is C, else
Ant2:
ifxisA,andyisB,thenzisC,else
. .. . .... .. .... .. ... ... .. .. ... .. . ... . .
Antn:
Antn +l:

if x is A,, and y is B,, then z is C,.


x is A and y is B.

Cons:

z is C.

For example, the consequence C by Mamdanis method


follows by interpreting else as union (u) and from (5):
C=(Aand
=[(A~oA~

B)o[((A,

and B,) -C,)u

...

u((A,

(12)

Rc is given as

and B,,) -C,)]

-C,)n(BoB,+C,)]

u . . . u[(A~A,+C,)n(B~B,+C,)].

(13)

Note that else is also interpreted as union for the fuzzy implications Rp,
Rbp, and Rdp in Table 1, and the above equality holds for these implications.
When A = u,-, and B = u,, the consequence C by the method Rc is given as
c=c;uc;u

-0.

UC,,

(14

where for i=l,...,n,


(15)
Namely,

from (8) and (13)

In the same way, C is obtained

from Rp, Rbp, and Rdp as in (14).

VARIOUS

FUZZY

REASONING

METHODS

137

For the fuzzy implications Ra, Rm, Rb, R* , R#, Rs, Rg, and RA in Table 1,
else in (12) can be interpreted as intersection (n). Thus, the consequences C
for these fuzzy implications are defined as
C=(Aand
c

B)e[((A,

and B,) +C,)n

[(AoA,~c,)u(BoB,-,c,)]

.-. n&4,

and B,) -C,)]

n .*a n[(AoA,~c,)u(BoB,-,c,)].
(16)

It is noted that the consequence


C is not equal to but contained in the
intersection
of fuzzy inference results [(A 0 Aj + C;.) U( B 0 Bi + C,)] (i =
1 9.1.) n). In the following discussion, however, we shall assume that C is given
as the intersection of the individual fuzzy inference results, for simplicity in the
calculation of C.
When A = u,, and B = u,, the consequence C, say, by the method Ra is
given as

where each C).l, ill,...,

C=c;nc;n

... nc;,

n, is represented

from (9) as

(17)

cc;(w) =l~[l-(Pri(~o)~~,,(OO))+Pc,(W)].
(18)
In the same way, we can have C by Rm, Rb, R* ,R#, Rs, Rg, and RA as in
(17).
To obtain
fuzzy set C,
which has the
the following
gravity of the

a singleton % which is a representative point for the resulting


several methods have been proposed. For example, the point
largest membership grade of C is taken as a desired singleton. In
discussion, the method is employed which takes the center of
fuzzy set C, as a desired singleton, that is,

Jwc44 dw

w=
/pc#(w)dw .

(19)

3. FUZZY CONTROLS UNDER VARIOUS


FUZZY REASONING
METHODS
We shall consider a system with first order delay as a simple plant model
which is represented by a differential equation Tdh/dt + h = q, with T being a
time constant.

MASAHARU

138

MIZUMOTO

TABLE 2
Fuzzy Control Rules e, Ae + Aq [4]
e

1 Ae-NB

NB
NM
NS
zo
PS
PM
PB

NM

PM

PB

NS

ZO

PS

PM

PB

PS

PB
PM
PS
zo
NS
NM
NB

NS

NM

NB

Let e and Ae be input variables of a fuzzy controller which represent


error and change in error, and let Aq be an output variable representing
change in action, where e and Ae are defined as
e = Ah = (present value of h) - (set point),
Ae=e(k)-e(k-l),
and the actual action q(k) to be taken at time k is given as
q(k)

=q(k-l)+Aq.

Yamazaki and Sugeno [4] give fuzzy control rules for a system with first
order delay as in Table 2. This table shows 13 fuzzy control rules interpreted as
Rl:

ejsNB

and AeisZO

Aq is PB,

R2:

eisNM

and AeisZO

AqisPM,

eisZ0

and AeisPB

Aq is NB.

R13:

-6

-5

-4

-3

-2

-1

(20)

Fig. 4. Fuzzy sets of fuzzy control rules in Table 2.

VARIOUS

FUZZY

REASONING

139

METHODS

where NB (negative big), NM (negative medium), NS (negative small), ZO


(zero), PS (positive small), PM (positive medium), and PB (positive big) are
fuzzy sets in [ - 6,6] as shown in Figure 4.
When e = e,, and Ae = Ae, are given to a fuzzy controller as a premise of
(20), the change of action Aq = Aq,, is obtained as the center of gravity (19) of
the fuzzy set which is aggregated from the fuzzy sets inferred from each of fuzzy
control rules of (20) given e, and Ae, by use of (14) or (17).
EXAMPLE
ity:

3. We shall consider the following three fuzzy controls for simplic-

eisNS

and AeisZO

A9 is PS,

eisZ0

and AeisZO

Aq isZ0,

and AeisPS

-,

Aq is NS.

eisZ0

(21)

When Mamdanis method of (15) is used, the change in action AqO is obtained
as in Figure 5. In the same way, AqO is given as in Figure 6 by the method of
Ra of (18).
Figure 7(a) shows AqO at e = e0 and Ae = Ae, when using all 13 fuzzy
control rules in Table 2 by Mamdanis method Rc. Figure 7(b) and (c) show Aq,
according to Ra and Rg, respectively.
Using the above methods, we shall first indicate control results for a plant
model G(s) = e- 2/(1 + 20s) with first order delay and dead time under various
approximate
reasoning methods in Table 1. In this experiment, we use the
following expression:

clc:(Aq)= [pa,(edApB,(Ae,)]
+puc,(Aq)

(22)

[see (7)-(g)], where and in (20) is interpreted as A ( = min), and A,, B,,C,
(i = 1,. . . ,13) are fuzzy sets shown in Figure 2 and Table 2. It is found from the
computer simulation in Figure 8(a)-(c) that all of the approximate reasoning
methods except Rm, Rg, Rs, and RA obtain good control results. In particular,
Rc, Rp, Rbp, and Rdp obtain the best results. Note that these methods are
based on fuzzy products known as t-norms. Similar control results are observed
in other computer simulations not shown in this paper.
In the case of Mamdanis method Rc, which gets a good control result, it is
found from Figure 7(a) for Aq,, at e, and Ae, that we have AqO = 0 at e, = 0
and be, = 0 (indicated by a dot in the center of the figure) and that AqO
decreases to minus when e, and/or
be, increase to plus in the area of
e, = AqO + 0. On the other hand, for the method Ra [see Figure 7(b)], the rate

140

MASAHARUMIZUMOTO

VARIOUS

FUZZY

REASONING

METHODS

NI
. .

I
i

.- :=. .
-.

-_

.- .-

c-

i.

--..

--_

w
I

Q11
.

_. . .--

.-

-._p--.
-_

a?

_Ei _-_. .
:
:
_.-I
*.
--.

*.

--.

141

MASAHARU MIZUMOTO

142

-6

e0

(a>

-6

+eo

(b)
Fig. 7. Aq, at e, and Ae, by fuzzy control rules in Table 2: (a) Mamdanis method Rc (14):
(b) Zadehs method Ra (18); (c) Rg based on Gadelian logic in Table 1.

VARIOUS

FUZZY

REASONING

METHODS

143

-6
-6

c)

+e

Fig. I. Continued.

of decrease of Aq, is observed to be smaller than that of Aq, by the method Rc.
Therefore, the convergence on the set point of the control result by the method
Ra becomes slower than that of the method Rc [see Figure 8(b)].
It is noted that the methods Rg, Rs, and RA show the worst control results,
as in Figure 8(c). We shall analyze why the method Rg, which is based on the
implication
rule of Giidelian logic and which can get reasonable inference
results in fuzzy reasoning [2], cannot get a good control result. As is seen from
Figures 7(c) and 9, the rate of decrease of Aqo is zero (flat) at e,, be,, 8 0, that
is, Aq,, = 0 in the area. Thus, no change is made in the control action q, and so
the same action continues to be taken. More precisely, it is seen from Figure
8(c) that the control result of Rg converges on the point h = 58.3 (not at 60). In
our computer simulation we use the expression

e, = -h-40
40

x6

40 = set point,

6 = scale factor,

to obtain the error e, from the output h of the plant model. For example,
have e, = 3 at h = 60. We shall show what value of Aq,, can be obtained
e, < 3 and Ae, = 0, which corresponds to the case of the conuol result of
converging to h = 58.3. When e, < 3 and Be, = 0, the fuzzy control rules

we
at
Rg
in

MASAHARU

MIZUMOTO

(b)

Fig. 8. Control results under various approximate reasoning methods [,,and is interpreted as
A ( = min)]: control results by (a) Rc, Rp, Rbp, and Rdp; (b) Ra, Rm, Rb, R*, and R#;
(c) Rs, Rg, and RA. Time constant = 20, dead time = 2, (scale factor of Ae) = 1.2, (scale factor
of Aq) - 2.5.

VARIOUS

FUZZY

REASONING

145

METHODS

Fig. 8. Continued.

Table 2 to be used are the five rules below:


eisPS

and AeisZO

Aq is NS,

eisPM

and AeisZO

Aq isNM

eisZ0

and AeisZO

Aq is ZO,

eisZ0

and AeisNS

Aq is PS

eisZ0

and AeisPS

Aq is NS.

(*l),

(*2),

The right-hand members NM and PS of the rules (* 1) and (* 2) are disjoint


fuzzy sets from Figure 4. The fuzzy sets NM and Ps which are inferred from
therules(*1)and(*2)ate=e,(<3)andAe=Ae0(=O)byusing(22)and
Fig. 2(xi) are also disjoint fuzzy sets. Since the intersection of disjoint fuzzy sets
is empty, the resulting fuzzy set, which is given by taking the intersection [see
(17)] of the fuzzy inference results derived from the five rules given above, is
also an empty fuzzy set. Thus, the center of gravity (19) of the empty fuzzy set,
i.e. Aq,,, is not defined. In this experiment it is assumed that Aq,, = 0 when the
resulting fuzzy set is empty. Hence, the control action q to be taken at e,, < 3

MASAHARU MIZUMOTO

146

Fig. 9. A9o at - 3.0 4 e, 4 3.0 and - 3.0 6 be, s 3.0 by fuzzy control rules in Table 2 under
the method Rg.

and Ae, = 0 does not change, since AqO= 0, so that the control result converges
to h - 58.3.
It is noted that if e,, becomes exactly equal to 3 (h = 60) and Ae, = 0, the
fuzzy control rules to be used are
eisPS

and AeisZO

AqisNS,

eisPM

and AeisZO

AqisNM,

and the center of gravity of the resulting fuzzy set is given as - 3 (indicated by
0 in Figure 9). Thus, AqO= - 3 when e, = 3 and Ae,,= 0, though Aq,, = 0 at
e, < 3 and Ae, - 0. The same holds for the method RA. For example, the
control result by the method RA in Figure 7(c) does not change until time 95,
when h reaches 60, and then it decreases suddenly, since Aqc changes suddenly
from 0 to - 3.
We shall next show control results in Figure 10 when and is interpreted as
the algebraic product (e) in the following [see (lo)]:

p&W = [&h&W]

+p&d

(23)

VARIOUS

FUZZY

REASONING

METHODS

147

40

Fig. 10. Control results under various approximate


( = algebraic product)].

reasoning

methods

rand

is interpreted

as

In this computer simulation, we have similar control results to those in Figure


8. Rc, Rp, Rbp, and Rdp also achieve best control results. The control results of
Rs, Rg, and RA are omitted because they are the worst, as in Figure 8(c).
It is concluded from these experiments that Mamdanis method Rc as well as
Rp, Rbp, and Rdp, which are all special cases of the t-norm, can get better
control results than the methods Ra, Rm, etc., which are based on the
implication rules of many-valued logic.

4. FUZZY
OF FUZZY

CONTROLS ON VARYING
CONTROL RULES

THE FUZZY

SETS

It is very interesting to investigate how control results are influenced when


the fuzzy sets of fuzzy control rules are changed. In the following, we shall
discuss fuzzy controls when the widths W of the fuzzy sets in Figure 4 are
changed as in Figure 11. In the computer simulation, we use Mamdanis
method Rc. Figure 12 shows control results on changing the width W at dead
time 2 and 5. When the width of fuzzy sets is small ( W = 0.2 and l), that is, the
fuzzy sets are completely separated, the control results are bad and do not
converge on the set point. On the other hand, when the width W is large
( W = 8, lo), control results are satisfactory, but their overshoot becomes rather

-4

-6 -5 -4

I
-6

1 NE

-2

-2

NS

0
1

(c) Width 2

-1

PS

(a) Width 0.2

0.2

zo

3
4

PM

PB

PB

-6 -5

-6

-6

NB

-4

-4

-4

-3 -2

-2

-2

1
_.--..,

(ZO

1
(f) Width 1"

-1

(d) Width 4

(b) Width 1

(NS

Fig. 11. Fuzzy sets on changing the width W.

(e) Width 6 (same as Fig.4)

-3

Nf?

PS

PM

P8

PB

VARIOUS FUZZY REASONING METHODS

149

(b)

Fig. 12. Control results on changing the widths of fuzzy sets (by Mamdanis method Rc):
(a) dead time = 2; (b) dead time = 5.

MASAHARU MIZUMOTO

150

big. When W = 4 and 6, we get good control results. It is found from the
computer simulations that good control results are obtained when the fuzzy sets
of the fuzzy control rules are not isolated and not too much overlapped. In fact,
at W = 4, the height of the crossing-point of neighboring fuzzy sets is 0.5, as in
Figure 11(d).
When the width W of fuzzy sets is small (W = 0.2,l) as in Figure 11(a) and
(b), no fuzzy control rules are used at, say, e,, = 3 and thus the resulting fuzzy
set C for bq, is empty. The center of gravity Aq, of C is not defined, but is
assumed to be 0, and no change is made to the control action q. Therefore, in
general, the control results at W - 0.2 and 1 do not converge on the set point as
shown in Figure 12.
When the width W of fuzzy sets is large, the number of fuzzy control rules to
be used is larger than in the case of small W. For example, when W = 10 [see
Figure 11(f)], the fuzzy control rules to be used at e,, = 3 and Ae, = 0 are as
follows:
eisPS,

AeisZO

-,

Aq is NS,

eisPM,

AeisZO

Aq is NM,

e isZ0,

Ae isNM

Aq is PM,

eisZ0,

AeisNS

--,

Aq is PS,

isZ0,

be isZ0

Aq isZ0,

eisZ0,

AeisPS

--,

AqisNS,

e isZ0,

Ae isPM

--)

Aq isNM,

and the center of gravity Aq, of the resulting fuzzy set C inferred from the
seven control rules is - 1.37 at e, = 3 and be, = 0. On the other hand, when
W = 4 [see Figure 11(d)], we have the following control rules at e, = 3 and
he, = 0:
e isPS,
eisPM,

he isZ0

Aq isNS,

AeisZO

Aq is NM,

and the center of gravity Aq, is - 3. Thus, the change in action Aq, at large W
becomes smaller than Aqa at small W as shown in the example, so that the
control results at large W become slower owing to the insensible A%.

VARIOUS
5.

FUZZY

REASONING

METHODS

151

CONCLUSION

We have proposed various approximate reasoning methods in using fuzzy


controls. When using Yamazakis fuzzy control rules for a system with first
order delay, Mamdanis method Rc and Larsens method Rp as well as Rbp
and Rdp, which are all r-norms, get much etter control results than the methods
Ra, Rm, Rb, R* and R#, which are based on the implications of many-valued
logics. Furthermore,
good control results are obtained when the fuzzy sets of
fuzzy control rules are fuzzily separated (that is, not separated but not too
much overlapped).
REFERENCES
1. E. H. Mamdani, Applications of fuzzy algorithms for control of a simple dynamic plant,
Proc. IEEE 121:1585-1588 (1974).
2. M. Mizumoto and H. J. Zimmerman, Comparison of fuzzy reasoning methods, Fuzzy Sets
and Systems 8 (3):253-283 (1982).
3. M. Mizumoto, Extended fuzzy reasoning, in Approximate Reasoning in Expert Systems
(Gupta et al., Ed.), Elsevier Science, pp. 71-85, 1985.
4. T. Yamazaki and M. Sugeno, Fuzzy controls (in Japanese), Systems and Conirols 28:442-446
(1984).
Received 4 May

I987

You might also like