Operator Tuning in Fuzzy Production Rules Using Neural Networks

MIYOSHI Tsutomu, T A N 0 Shun'ichi, KAT0 Yasunori, Thierry ARNOULD

Laboratory for International Fuzzy Engineering Research
89-1 Yamashita-cho, Naka-ku, Yokohama, 231, JAPAN

Absfract -In production rules, the total matching degree of

the condition part Is calculated from the matching degree of each
condition by aggregation operators. In ordinal production 11. PRODUCrION RULE
systems, simple loglcal "and" and "or" functions are used as
aggregation operators because the matching degrees ore crisp Let's think about the expansion of production rules using
values fuzzy technology. By using the ability to handle uncertainty,
In the case of fuzzy production rules, there are several we can obtain advantages such as fewer rules and easy
promising approaches for handling the uncerhlnties of matching maintenance. Furthermore, by using approximate reasoning, it
degrees: 1) fuzziflcation of the matching degree of each
condition, 2) expansion of aggregation operators, etc. Many becomes possible to get approximate answers to the inputs
studies have been conducted on tuning l), that use membership that are not assumed.
functions, but few have been done on tuning 2). Most of the To handle fuzzy production rules, there are several
systems use "mln" and "max" functions as the aggregation promising approaches to expand the production system: 1)
operators in spite of their strong influence. fuzzification of matching degree of each condition, 2)
In this paper, we consider automatic operator tuning of expansion of aggregation operators, 3) expansion of
parametric T-norms and T-conorms whose characteristics can implications, 4) expansion of combinations, etc.
be modified by parameters.
A. Membership Functions
I. INTRODUCI-ION By using membership functions, we can fuzzify the
matching degree of each condition which will become a real
In production rules, the total matching degree of the number or an interval of [O, 11 instead of the crisp values of 0
condition part is calculated from the matching degree of each or 1.
condition by aggregation operators. In ordinal production
systems, simple logical "and" and "or" functions are used as B . Aggregation Operators
aggregation operators because matching degrees are crisp
values. The total matching degree of the condition part is
In the case of fuzzy production rules, there are several calculated from the matching degree of each condition by
promising approaches for handling the uncertainties of aggregation operators. Most systems use simple logical "and"
matching degrees: 1) fuzzification of the matching degree of and "or", or "min" and "max" functions as the aggregation
each condition, 2) expansion of aggregation operators, 3) operators. To choose aggregation operators, we can adjust the
expansion of implication,4) expansion of combination, etc. matching degree to match human intuition.
At the Laboratory for International Fuzzy Engineering
Research, we have developed an expert system called FOREX C . Implications
(FOReign exchange trade support Expert system), where
expert knowledge are represented by more than five thousand Implication functions express the relationship between a
fuzzy production rules [l] . We found it very difficult and condition part and an action part. To choose implication
important to tune aggregation. implication, and combination functions, we can adjust the matching degree to match human
functions as well as membership functions. intuition.
Many studies have been conducted on tuning 1) which use
membership functions, but few have been on tuning 2). Most D . Combinations
systems use "min" and "max" functions as aggregation
operators in spite of their strong influence. Combination functions express the relationship between
In this paper, as one step, we consider automatic operator rules that are candidates for firing. To choose combination
tuning for the parametric T-norms and T-conorms whose functions, we can adjust a conclusion to match human
characteristicscan be modified by parameters. intuition.
0-7803-0614-7/93$03.00 (D1993IEEE 641
 where e, de and du are linguistic variables for the process
error, the change in error and the change in the process input,
and their values are fuzzy sets such as large, small or zero,
Investigations have been conducted on the automatic denoted by Ei, dE;. and dui. f(*,*,')
is a general
tuning of membership functions using neural networks [2,3] . representation of the implication function. Now we use
However, if a complex relationship exists between the Mamdani's implication function, so the overall fuzzy relation
conditions, tuning methods are not able to adjust for errors by R is given by
tuning the membership functions. If similar characteristics
exist in the relationships in some rule blocks, we must be able
to efficiently tune the rule blocks by tuning the aggregation
operators. But, there are few studies on tuning of aggregation
operators or tuning of implications. where T[ ] represents a T-norm and T*[ ] represents a T-
From this standpoint, as one step, we considered tuning the conorm.
aggregation operators. In this paper, we consider automatic If the actual process error, the actual change in error, and
operator tuning of parametric T-norms and T-conorms whose the actual change in the process input take on the fuzzy values
characteristics can be modified parametrically. E', dE', and dU', then the actual change in the process input
Using typical T-operators, Gupta and Qi showed that the dU can be obtained by:
performance of the fuzzy controller for a given plant depends
on the choice of the T-operators [4] . There are various
parametric T-operators whose characteristics can be modified
parametrically. So we will consider automatic tuning of the
fuzzy system by parametric T-operators using back- We use a simple fuzzy controller whose response is a
propagation. function of the T-operator. The following set of control rules
is used for the controller.
If e is negative and de is negative, then du is negative.
Iv. SYSTEM STRUCTURE If e is negative and de is positive, then du is zero.
If e is positive and de is negative, then du is zero.
The system consists of two networks: the fuzzy controller If e is positive and de is positive, then du is positive.
network and the plant network. The system structure is shown The following fuzzy sets and membership functions (see

1-
inFig. 1. Fig. 2 through Fig. 4) are used to represent fuzzy rules.
NE: negative error
E y . -Controller
rI du(n) Plant Y(r
PE:
NdE:
positive error
negative change in error
Network PdE: positive change in error
Network NdU: negative change in process input
ZdU: zero change in process input
PdU: positive change in process input
Fig. 1. Syslem Structure.

There have been many investigations on neural networks

learning about real plants. First, the plant network learns
about targei plant, then the fuzzy controller network learns
using the plant network. It is assumed that the parameters of
plant network are stable while the fuzzy controller network is
learning. Fig. 2. Membership Function for E.

x
A. Fuzzy Controller
In general, a fuzzy controller has N rules. The i-th rule can
be represented as the fuzzy relation R; and its membership
function is given by

~ de, du) = f (PEi ( e ) ,P a i ( d e ) ,PdUi (du 1)

P R (e* ( 1) Fig. 3. Membership Function for dE.

642
 We will show how to tune T and T* in this network using

X"
Fig. 4 . Membership Function for dU.
the back-propagation algorithm.
The system is evaluated by the following error. G u is
scaling factor.

E, = x - (1e ( n ) 2 + G U x d u ( n ) 2 )
2
The actual error and the actual change in error at any
sampling instant are calculated as follows: = x i ( ( x ( n ) - y ( n ) ) 2 + G Uxdu(n)2)
n (6)
e ( n ) = [.(.I-y(n)] GE, The minimization of this error represents the quicker
(4)
de(n)=[y(n-l)-y(n)]xGdE, response with a minimum input of energy. We can transmit
the error through the plant network as it is done in neural
where y(n) is the process output at the n-th sampling, GE and "Qrk learning.

>yy
GdE are scaling factors, and x(n) is the reference input.
e(n) and de(n) are always crisp values, therefore, we obtain
the following fuzzy process input dU. Inner Product

PdU ( d u )= T* [T[PivE(e),PNdE(de),PNdU(du),],
W
[T[PNE(e)3PPdE
( d e ) * P n U(du),],
(5)
[T[C(PE( e ) ,P N d E ( d e ) ,PZdU (du),],
Y2
[T [ P ~ E ( e ) ~ ~ P d E ( d e ) , P(du),]].
PdU SUm
The crisp change in the process input du(n) is calculated by Fig. 6. Output Value of Each Node.
the Center of Gravity method from d U .
Let's assume that the outputs of each node of the network
B . Error Transmission are y1 through y5, z and w (see Fig. 6). The values of yo
through y5 and w are:
The network representing the fuzzy controller is shown in
Fig. 5. The normal lines indicate crisp signals and the bold
lines indicate fuzzy sets in Fig. 5. Refer [3] for more
information.
NE
e (9)

de U

Note that the values of yo through y5 and w can be

dU differential. Then we can obtain the following from
differential equations for each output.

Fig. 5. Network of Fuzzy Controller.

643
 Thus, the automatic tuning of the system becomes possible 1
by using these differential equations.

V. SIMULATION
Gupta and Qi have produced simulation examples using P > 0.
typical T-operators [4] . Based on their paper, we showed in if p --> 0, then T = T5 and T* = T*5
this paper, that the performance of the fuzzy controller for a if p = 1, then T = T4 and T* = T*4
given plant depends on the choice of the T-operators. We used ifp--> -,thenT=T1 andT* =T*l
two kinds of plants, four types of typical T-operators,and two Note that the Dombi T-operator encompasses all the
types of parametric T-operators. First, we will introduce the operators from the logical operator to the drastic operator.
plants and operators, then present the results of automatic
tuning. Hamachar T-operator

A. Plants

The target plants used for tuning are:

First-order plant: y' + ay = bu
Second-orderplant: y" + ay' + by = cu
where a, b, and c are constants.
pro
B. T-operators if p = 0, then T = T4 and T* = T*4
Typical T-operators are shown below: if p = 1, then T = T2 and T* = T*2
if p --> =, then T = T5 and T* = T*5
TABLE 1 Note that Hamachar T-operator encompasses all the
TYPICAL
T-OPERATORS operators from the non-parametric Hamachar operator to the
drastic operator.
C. Result for the First-order plant
x+y-xy
max(x + y -1.0) min(x + y , 1)
X Y /(x + Y - X Y ) (x + Y - ? X Y ) / ( X - XY) Fig. 7 shows the results for a fuzzy controller that uses
x,lfy=l x,dy=O typical T-operators applied to the first-order plant. In Fig. 7,
y. i f x = 1 y,ifx=o
0, otherwise 1, oherwise m l represents Tl(x, y) and tcnl represents T*l(x, y) of Table
1. This result is the same as the result presented in [4] .
Those typical T-operators are ordered as follows: Fig. 8 shows the result of automatic tuning. In Fig. 8, tcnD
represents the Dombi T-conorm and tcnH represents the
T s < T 3 < T z <T4 < T i , T ; < T i < T i <T; <T;.(14) Hamachar T-conorm. Fig. 9 is a close-up view of Fig. 8. To
make the figure easy to read, a part of the tn3-tcn3 line is
Various kinds of parametric T-operators are introduced in reduced.
[4] ,we selected the Dombi T-operator, that covers a wide The Dombi T-conorm is better than the Hamachar T-
range of strengths, and the Hamachar T-operator, that requires conorm because of the range it covers. The asymptotic limit
less computer power. of the parameter for the Dombi T-conorm is 3.063. This
means that the strength of the Dombi T-conorm is situated
Dombi T-operator between T*l and T*4. The parameter for the Hamachar T-
conorm under flows to 0.0. This means that the strength of the
Hamachar T-conorm is T*4 and the suitable strength is
weaker than T*4.

644
 D . Results for the Second-order plant
Fig. 10, 1 1 , and 12 illustrate the results for the second-
order plant. The asymptotic limit of the parameter of the
Dombi T-conorm is 1.3345 and the Hamachar T-conorm
under flowsto 0.0. The same interpretations for the result of
the first-order plant apply here.

3.5

0 10 20 30 40 50 2.5

Fig. 7. Resulls for Typical T-operators. 1 .5

0.5

0
0 10 20 30 40 50

-.- tnl tcnl

-1n2 tcn2 Fig. 10. Results for Typical T-operators.

-'-tn3 lcn3

-tn4 tcn4

- -tn2 tcnD
4.5 T
4
(p-3.063)

-In2 tcnH 3.5

(p-0.0)
3
___c___
0 t o 20 30 40 50 2.5 -1n2 tcn2

2 - -In3 tcn3
Fig. 8. Results for Parametric T-operators. -tn4 tcn4
1.5

1
-- tn2 tcnD

3.27

3.1 1 .i 0.5 In2 lmH

3 0
0 10 20 30 40 50
2.9

2.8
Fig. 1 1 . Results for Parametric T-operators.
2.7
0 10 20 30 40 50

Fig. 9. Close-up view of Fig. 8.

645
 Through the simulations, we found that not only the
3.2 characteristics of the T-operators, but also the characteristics
3.1 of differential equations of the T-operators are important in
automatic tuning.
3
We will study the preferred characteristics for parametric
2.9 T-operators and their differential equations, and the
2.8
relationship between membership functions and aggregation
operators from the perspective of automatic tuning.
2.7 We plan to expand the system proposed here to a total
0 10 20 30 40 tuning environment that can manually and automatically tune
membership functions, aggregation operators, and
Fig. 12. Close-up view of Fig. 11. implications.

REFERENCES
VI. CONCLUSION
[I1 H. Yuize, T. Yagyu, M. Yoneda, Y. Katoh, S . Tano, M.Grabish. and S.
To get the optimum strength for the aggregation operator, it Fukami, "Decision Support System for Foreign Exchange Trading -
is not sufficient to use only typical T-operators, because the Practical Implemenlation -",IFES.91, Yokohama, JAPAN, pp. 971 -982,
improved strength for the target plant often falls between 1991.
typical T-operators. Some of the parametric T-operators are 12] S. Horikawa, T. Furuhashi, S. Okuma, and Y. Uchikawa. "A Fuzzy
Controler Using A Neural Network and Its Capability to Leam Expert's
advantageous to improve the strength of the operator. Control Rules", International Conference on Fuzzy Logic & Neural
We proposed a system that consists of a fuzzy controller Nelworkr (IIZUKA'90). lizuka. JAPAN, Vol. I, No. 2, pp. 103-106,
network and a plant network and presented a method for 1990.
automatic tuning of parametric T-operators. 131 A. Maeda. R. Someya. S . Yasunobu, and M. Funabashi, "A Fuzzy-
Through computer simulations, we obtained the optimum based Expert System Building Tool with Self Tuning Capability or
strength of T-operators for each kind of plant through Membership Functions", World Congress on Expert Systems, USA, pp.
automatic tuning of the parameter of the T-operator. 639-647, 1991.
We already presented the results of two classes of 141 M.M. Gupta and J. Qi, "Design of fuzzy logic controllers based on
parametric T-operators, and will investigate the possibility generalized T-operators", Fuzzy Sets and Systems, Vol. 40, pp, 473-489.
and the sensitivity of all known kinds of parametric T- 1991.
operators.

646