Professional Documents
Culture Documents
Japan: Department of Systems Science, Tokyo Institute of Technology, 4259 Nagatsuda, Midori-Ku, Yokohama 227
Japan: Department of Systems Science, Tokyo Institute of Technology, 4259 Nagatsuda, Midori-Ku, Yokohama 227
North-Holland
The problems of structure identification of a ~zzy model are formulated. A criterion for the
verification of a structure is discussed. Using the criterion, an algorithm for identifying a
structure is suggested. Further, a successive identification algorithm of the parameters is
suggested. The proposed methods are applied to an example.
1. Inb'odu~on
2. F ~ z y model
As the expression of a fuzzy model we use the fuzzy implications and the fuzzy
reasoning method suggested by Takagi and Sugeno [1]. A fuzzy implication is of
the following form:
U: If xl is A~, x2 is A ~ , . . . , xp is A~,,
then yi = do + c~xl + c~x2 + . •. + %xp,i (1)
0165-0114/88/$3.50© 1988, Elsevier Science Publishers B.V. (Non,h-Holland)
I6 M. Sugeno, G.T. Kang
/
L2: If x I is L / - __, x 2 is ~ _ then y2=-O, lx1+4 • Ox2+l. 2
3 9 4 13
x2
13
//
L1
~ L2 "
Xl
o 9 11 18
Fig. 2. Fuzzy subspaces.
Structure identification of fuzzy model 17
Y* - wiy i wi (2)
"= ti=l
where q is the number of fuzzy implications, y~ is calculated for the input by the
consequent equation of the i-th implication, and the weight wi implies the overall
truth value of the premise of the i-th implication for the input calculated as
Wi = ~ Al(x
i 0). (3)
jffil
Yho
2~
xI
1
premise parameters identification
1
--=-* choice of consequent structure
l
consequent parameters identification
1
calculation of UC, the criterion of
verification of model
1
verification of consequent structure
3, $ ~ ~ iden~ea¢on
the premises. The other is that we have ~.~ find an optimal fuzzy partition of the
inputs space, which is a problem peculiar to fuzzy modelling. For exampie,
suppose that a fuzzy model of a three inputs- one output process is represented
by three fuzzy implications such as
LI: If xt is SmaUt and x3 is Sinai;2 th~n y~ = 0.5 + 1.2x2 - 0.~3, ~,
The variable xz does not appear in the all premises and L 3 has the only one
variable xl i~l its premise. A fuzzy implication represents an input-output relation
holding in a fuzzy subspace of the inputs space represented by its premise. For
example, L 2 represents the input-output relation in the fuzzy subspace defined by
xl = Small1 and x3 = Large2. Therefore, the premise structure identification is the
same as the problem to find out the fuzzy partition of the inputs space, where the
number of ~.he fuzzy subspaces corresponds to that of the necessary implications.
The premise parameters are those of the fuzzy variables of premises.
The second part is concerned with the consequent structure identification. We
have to find which variables are necessary in the consequent of an implication.
For example, L x has two variables x2 and x3 in its consequent, while L 3 has x~
and x3.
In addition, we need to find a criterion for the verification of an obtained
structure.
UC =
"
( y e S _ y+a.)2+ ~] ( y p _ y n.),
]'° , (7)
I-i=1 i=X
where na is the number of the data set NA, y ~ the estimated output for the data
set NA from the model identified by using the data set NA, and y~n the estimated
output for the data set NA from the model identified by using the data set Na.
As a fuzzy model is constructed by partitioning the inputs space, the data for
the identification should be distributed uniformly over the space where the fuzzy
model is applied. Therefore, the data of NA and Na should be also distributed
uniformly over the space. When the number of the observed data is small, we
make NA and NB have some data in common.
(stage 3), Suppose that the input variable xj is found to be put into the premises at
the stage 2. Then ST[2] iS as shown in Figure 5(a). At this stage, each premise
structure consists of three fuzzy subspaces and t!~,~creare two ways in constructing
a premise structure. One is that the range of x s itself is divided into three fuzzy
subspaces as shown in Figure 5(b). The other is that another variable, say x~, is
put into ST[21 and its range is divided into two fuzzy subspaces as shown in
Figures 5(c), (d)o The fuzzy model with the premise structure of Figure 5(b) is
LI: Ifx s is 'Small' t h e n . . . ,
L2: If xj is 'Medium' t h e n . . . , (9)
L3: If x s is 'Large' t h e n . . . ,
and the fuzzy mode~ with the premise structure of Figure 5(d) is
LI: If x# is 'Small1' t h e n . . . ,
L2: If x# is 'Large1' and xi is 'Small2' t h e n . . . , (lo)
L3: If xj is 'Large1' and xi is 'Large2' then . . . .
__ xj ~ ~~_____ xj
(a) (b)
xl i
xj
(c) (d)
Fig. 5. Premise structures.
,22 M. Sugeno, G.T. Ean s
The variables which are put into the premises at stage i (i ~>3) are, however, not
all the k input variables, which will be later explained.
For each premise structure which can be constructed at this stage, a fuzzy
model is identified and its UC is calculated. Among those fuzzy models, one with
the least UC is picked up. Its premise structure and UC are similmly written as
STEa! and UC[3I, respectively. I f UC[3I is larger-than UC.I2j, the process is
terminated and the optimal premise structure is found to be ST[21. Otherwise the
process is proceeded to the next stage 4.
%
The stages after the stage 3 are similar to the stage 3. That is, at stage i (i >~3)
fuzzy models consisting of i fuzzy implications are constructed in two ways. One is
that one of the fuzzy subspaces of ST[~_q is divided with respect to one of the
premise variables of STt~_I! as illustrated in Figure 5(b). The other is that another
new variable is put into ST[~_I], and one of the fuzzy subspaces of ST[~_~] is
divided with respect to the new variable just as in Figures 5(c), (d). ST[z] and
UC[~i are determined in the sense of the least UC. Whether the process is
terminated or not is decidedby comparing UC[a] with UC[~_~].
When the number of input variables is large, the number of the possible
structures at each stage becomes combinatorially large. Therefore, it is necessary
to heuristically find which variables are possibly put into the premises. At the
stage 2 we divide the one-dimensional space of each premise variable into two
fuzzy subsp,aces: a fuzzy model consists of two implications. If a model at this
stage has a larger UC than the UC[~] (UC of a linear model), its structure is not
adequate a~d this implies that its premise variable should not appear in the
premises at the stages after the stage 2.
Structure number 1 2 3 4
Premise variable xa x2 x3 x4
UC 5.4 3.5 3.3 4.6
can be formed at this stage. Table 1 shows the UC and premise variable of each
structure. Since the structure (1) with x3 in the premise gives the least UC,
it is adopted as STI2I, and UCl21 is 3.3. The fuzzy model with STI21 is shown in
Figure 6
The variables xl and x4 should not be put into the premises at the stages after
stage 2 because the structures with those variables in the premises give the larger
UC's than UC m.
(stage 3) Three premise structures having three fuzzy subspaces are formed, as
shown in Figure 7, by dividing a fuzzy subspace of the STI21. Since the UC of the
structure (3) in Figure 7 is minimal, this structure is adopted as $T|31, and UC|31 is
2.8. The fuzzy model with STI3i is shown in Figure 8. We proceed to the next
stage since UC[3I i$ smaller than UCt2l.
(stage 4) The structures of this stage are shown in Figure 9. STHI is the structure
(3) in Figure 9, and UCt4! is 3.4. The fuzzy model with STI4! is shown in Figure
10. Since UCH]iS larger than UC[3I, the optimal premise structure is found to be
STt31.
where n is the number of data and y~ is the estimated value of Yi. Let J1 be the
value of J for the identification data, and J2 that for the verification data. At the
L1 : If x 3 i s _ ~ - ~
0 1.1 2.2
/
L2: If x 3 is ~ / o
0 1.1 2.2
x2 x2
_ ///. x3 x3
L1 : If x 3 ts e ~
0 2.3
L2: I f x 3 Is / / / ~ , . and x2 i s ~ ,
0 3.3 0 1.5
L3:
/
I f X3 i S ~ . and x 2 I s ~ /
0 3.3 0 2.3
\
'\
x2
I x2
7//.
>'//~///.
x3 _ // N ~_ ~ _ x3
x2
I (4) UC = 6.7 ( 5 ) UC - 7.2
L1 : If
0 3.2 0 3.3
L2: If x3 2
0 3.2 0 4.5
L3: If
0.9 3.2 0 3.3
then y3 = 9 . 1 6 + 2 . 4 8 x I + 7 . 1 2 x 2 - 1 . 5 9 x 3
L4: If x3 I s ~ , x2 i s ~ / ~ -
0.9 3.2 0 4.5 =
improved GMDH model of Eq. (12), J~ is 4.7% and J2is 5.7%. At the fuzzy
model of Figure 8, J1 is 1.5% and J2 is 2.1%. J~ of the fuzzy model of Figure 10 is
0.59% and better than that of the model of Figure 8 since the former model has
one more fuzzy implication than the latter. J2 of the model of Figure 10 is,
however, 3.4% and worse than that of the model of Figure 8. Therefore, we can
see that the model of Figure 10 does not have the true structure. The criterion for
verification of structure, UC, also shows that the model of Figure 8 is better than
that of Figure 10. Table 2 shows the results.
O_ 0 0
P! P2 P3 P4 P5 P6 P7 P8
= yo - ±
~=x j = l
Aj(xj)y
"= j = l
(15)
The ~igorithm for adjusting the premise parameters by using the data
(yO xO x O , . . . , xO) is as follows.
(1) For the input (x °, x ° , . . . , x°), calculate the grade of membership of each
premise fuzzy variable and the output y~ (i = 1, 2 , . . . , q) inferred from each
implication.
(2) Select the fuzzy variables to be adjusted, as follows.
For each premise variable, two fuzzy variables are selected. When there are p
variables in the premise, the number of variables to be selected is 2 x p
accordingly.
Suppose that we are to select the fuzzy variables concerned with the premise
variable x~. There are two cases according to the value of x °. One is that for x °,
two fuzzy variables have grades of membership greater than zero as seen in
Figure 12(a), In this case, the two fuzzy variables (A~' and A~ of Figure 12(a)) are
selected. The other is that for x °, one fuzzy variable has grade of membership
greater than zero as seen in Figure 12(b). Here the fuzzy variable (A~ of Figure
Structure identification of fuzzy model 27
l'c "
Akm Akn 0Ak=~ Akn
0
P Xk°
(a) (b)
Fig. 12. The fuzzy variables to be adjusted.
12(b)) and:a fuzzy variable which has parameter nearest to x~, (A~ of Figure
12(b)) are selected.
(3) Adjust the grades of membership of the selected fuzzy variables as follows.
Suppose that we are to adjust the grades of membership of A~' and A~, for x~.
For simplicity, the present values of A~(x °) and AT,(x°) will be written as w° and
w °, respectively.
By setting the error at zero, we can obtain the next equation from Eq. (15):
2~+ a3 = 0
0 ..... ~
7, "x "2
Vm+ vu-- ~
Fig. 13. w*~ and w.*.
28 M. Sugeno, G.T. Kang
When there is nolle satisfing the constraint, A 7 and A~ are aot adjusted. The new
grades of membership ( ~ / , ~'n) are calculated, using *he ;~daptive gain or, from
(w °, w °) and (w*, w*):
- w° + W°m) (19)
where o~- ~1 x c~2, and
~, - (1/h)'a, (20)
-2 = 1 / 0 + * - w°l), (21)
with h the number of all the fuzzy variables to be adjusted. The adaptive gain o~
has a significant effect on the convergence properties. The gain 0~2 is set to avoid
an excessive adjustment.
(4) Adjust the premise parameters by using the new grade of membership as
follows.
Suppose we are to adjust the parameters of A~ and A~ by using ~m and ~,,.
(a) When w° > 0 and w° > 0 (see Figure 12(a)): adjust the parameters of A~'
and A~ so that their grades of membership for x ° become Wm and ~,,,
respectively.
(i) When ~, ~>0.5: adjust the parameter of which the grade of membership is
one, as shown in Figure 14.
(ii) When ~, < 0.5: adjust the parameter of which the grade of membership is
zero, as shown in Figure 15.
( b ) When w ° = 0 (see Figure 12(b)): If the parameter of A~' is adjusted so that
its grade of membership for x ° is Ore, the amount of adjustment becomes
excessive. Therefore, adjust the parameter of which the grade of membership is
zero so :~hst it gets nearer to x ° (see Figure 16).
For A~, adjust the parameter of which grade of membership is one, as shown in
Figure J6.
l -'~ 1 6-.
0 . :I • •
0__ 0 __
xk ° Xk o
i - . A p n ..~
1 1
o o ;k~._ Pm Pn Xk
Fig. 15. The adjustment when ~ < 0.5. Fig. 16. The adjustmentwhen Wm
o - 0.
Structure iden~cation of.fuzzy model 29
i ..p.. - °~.,~
L"'" I . . . . ,'= ÷
0 I/5L 215L 315L 4/5L L
t
I I
I I
~ 2 A4
Pl P2 P3 P4 P5 P6 P7 P8
Xl x2
New system 2.0 5.5 3.0 6.0 5.5 8.0 6.0 9.0
3.4. Consequentstructureidentification
All the input variables do not always appear in the consequent of an
impli~tion. So we have to find which variables are necessary in the consequent of
an impfication. This is the same kind of problem as finding input variables i~" a ~
linear model. If a fuzzy model has q fuzzy implications and p input variables, and
if each consequent includes all the p input variables, the output is represented by
a linear equation with q x (p + 1) terms as shown in Eq. (4). The q x (p + 1)
terms are as follows:
(gl), (glXl), (glx2),..., (g~),
(g2), (g2x,), ( g 2 x 2 ) , . . . , (g2xp),
s
+ 0.18rt_ 9 + 0.16wt_lO
t h e n Ay = 3 . 5 - 2 . 4 y t _ 1 - 2 . 4 y t _ 7 + O . 1 3 r t _ 1 + 0 . 3 r t _ 8 - O , O l r t _ 9
L3: If w t Is _~~
9.5 10
1800
140(
~ observed values
t (year)
100C
1965 1 6 19 7 1968 1969 1970 1971
check the prediction accuracy, the mean square error A of the checking data is
calculated.
1971 / 1971
y2 (27)
t- 1965 t= 1965
where y,* is the predicted value of y,.
While A is 1.3% at the GMDH model, it is 0.93% at the fuzzy model and the
number of parameters is much smaller. Figure 20 shows the results of the
prediction by the fuzzy model.
$, C~udus|en
Refereuees
[1] T. Takagi end M. Sugeno, Fuzzy identification of systems and its appfications to modelling and
comrol, IEEE Trans. Systems Man Cybernet. 15 0985) 116--132.
Structure iden~'fication of fuzzy model 33
[2] A.G. Ivakhnenko, V.N. Vysotsldy and N.A. Ivakhnenko, Principal versions of the minimum bias
criterion for a model and an investigation of their noise immunity, Soviet Automat. Control 11
(1978) 27-45.
[3] A.G. Ivakhnenko and M.M. Todua, Prediction of random processes using self-organization of the
prediction equation-part 1. Problems of simple mediura-term prediction, Soviet Automat.
Control 5 (1972) 35-51.
[4] T. Kond0, R e ~ d G ~ H algorithm estimating degree of the complete polynomial, Trans. Soc.
I n s t ~ e n t and Control Engrs. 22 (1986) 928-934 (in Japanese).
[5] H. Akalke, A new look at the statistical model identification, IEEE Trans. Automat. Control 19
(1974) 71 723.