A Theoretical Justification For Nonlinear Control Property

Of A Class Of Fuzzy Logic Controllers

Bijan Sayyarrodsari Abdollah Homaifar Wesely E. Snyder

Dept. of Elec. Eng. Dept. of Elec. Eng. Dept. of Blec.&Comp.
North Carolina A&T North Carolina A&T Eng. , North Carolina
State University, State University, State University,
Greensboro, NC 27405 Greensboro, NC 27405 Raleigh,NC 27695-7911

Abstract: Later, viewing fuzzy systems as three-layer

In a previous work [l], a hybrid feed forward network(TLFFN) Wang and
implementation of fuzzy and conventional Mendel [3] developed a back propagation
PID controller was introduced and its algorithm to train the TLFFN fuzzy system
application to a 24egree of freedom robot to match desired input output pairs. Wang
manipulator arm was examined. and Mendel [4], used a "series expansion" of
A theoretical justification for that fuzzy basis functions to represent a fuzzy
approach, based on the fact that "fuzzy system. Then, they proved that a linear
systems are universal approximators" is combination of the fuzzy basis functions
presented in this paper. approximates any real continuous function
on a compact set to arbitrary accuracy.
Keywurdr: fuzzy logic controllers, universal Buckley [5] showed that a class of
approximation, universal control. fuzzy controllers, which he termed Sugeno
type controllers, are both universal
I. Introduction: approximatorsand universal controllers. The
Recently there has been a number of term "Universal approximator" in Buckley's
reports [2-61 about fuzzy systems being work has the same meaning as in Wang's.
universal approximators and/or universal By a universal controller he meant "given
controllers. Due to space limitations, we any process that can be controlled by a
restrict our literature survey to those directly continuous controller it can also be
related to the approach presented in this controlled by a Sugeno type controller." For
paper. a complete literature survey of the research
In an attempt to answer the very activities that led to these conclusions see
fundamental question why does a fuzzy
It
[5-61.
system work?" Wang [2] proved that fuzzy In this paper the results of previous
systems, viewed as a mapping from a works [2,5] have been used to establish a
nonfuzzy universe of discourse uc R"to R theoretical justification for the nonlinear
(set of real numbers), could be used to controlling property of the hybrid fuzzy-PID
approximate any given real continuous controller proposed in [l]. Figure 1 shows
function on the compact set uc R" to any a block diagram of the hybrid fuzzy-PID
degree of accuracy, if they satisfy certain controller.
criteria. He viewed his results as an This paper is organized as follows.
existence theorem of an optimal fuzzy Section II provides the necessary background
system for a wide variety of applications. information. Section 111 discusses the
theoretical justification in several steps.

0-7803-1896-X/94 $4.00 01994 IEEE 3 82

Finally section IV includes the conclusions minimum of all fuzzy membership values in
and a brief summary. the "if-side" for the rule being evaluated,
and clips the corresponding output
II. Background: membership at this level. The product
For the convenience of the reader, in operator scales the output membership as
this section we will briefly introduce the key opposed to clipping them. Interpreting
concepts which will be referred to "also" as a triangular co-norm operator, Max
throughout the paper: in most cases combines the individual output
membership functions generated by each rule
11-1.Fuzzy controllers: and produces the final output fuzzy set.
Fuzzy controllers have been The result of the inference engine.
developed as an application of a logical must now be defuzzified. Defuzzification
conceptual framework which is based on serves to provide the final crisp output of the
fuzzy set theory [7-lo]. Aiming to convert fuzzy logic controller. Two methods of
a linguistic control scenario into an defuzzification are most often used:
automatic control strategy, a fuzzy controller Maximum membership and centroid or
consists of four principal components [ 11- center of gravity [ll-121. The former
121: fuzzification interface, fuzzy knowledge chooses the output value corresponding to
base (consisting of membership functions the maximum degree of membership in the
and linguistic control rules), fuzzy inference output fuzzy set. One problem associated
engine and a defuzzification interface. with this method is that typically many
Fuzzification is simply the process by output values will have the same
which crisp, real world input values are membership level, particularly when the
converted into fuzzy linguistic values. These Min-Max inference is used. The second
fuzzy values can be labels for fuzzy sets. problem with the maximum criterion method
Each fuzzy set is characterized by an is that much of the information in the output
appropriately defined membership function. set is ignored and lost. The centroid
The membership function definition defuzzifier, on the other hand, incorporates
substantially affects the performance of the all the information of the output fuzzy set in
fuzzy controller and reflects "expert" the defuzzification process. An average, or
knowledge of the application domain. weighted sum of the output values is
Fuzzy rules consist of a premise with calculated, yielding a single crisp value.
one or more antecedents , and a conclusion
with one or more consequences. The 11-2. Sugeno Type Controllers [5,13-141:
individual rules in the rule base are A fuzzy controller with the collection
connected through the operator "also". of the fuzzy if-then rules of the following
Given the fuzzy rule base and the input type:
values, the fuzzy controller then applies
some type of inference operation. The Ri : IF x1 is A,' and x2 is A,' and . . . and
inference engine performs two functions: x, is A,,' THEN y=Pi (al' , q i, . . . , a,');
determination of applicable rules for a set of 15 i IK.
inputs, and inference of output fuzzy set(s).
Several inference operators have been [where:
developed, but the two most common types (i) xi , (an error signal, its time
are Min of Mamdani and prcduct of Larson derivative or integral), is the crisp input to
[ll-121. The Min operator takes the the fuzzy controller . It is worth mentioning

383
that the dimension of the input vector x=(x, continuous membership function for input
,x,,...,a is a function of the complexity of fuzzy sets and a continuous T-norm,
the system under control and is a subjective probabilistic product for instance ( k,=
design parameter. Ali(xl). A;(xJ. ... . A,,'(xJ )* for rule
(ii) A,,,' is the label for the fuzzy set evaluation procedurt. For further discussion
corresponding to the input variable x,,. on Sugeno type controllers see [5].
(iii) Pi is a polynomial in x, ,x2 ...
,x,, given by: 11-3. Stone-Weierstrass Theorem [Z]:
"Let Z be a set of real continuous
functions on a compact set U. If 1) 2 is an
algebra, i.e. the set Z is closed under
addition, multiplication and scalar
multiplication; 2) 2 separates points on U,
i.e, for each x1 , x2eu ,x, # x2 there exists
f~ 2 such that f(xl) f f( xz); and 3) Z
vanishes at no point of U, i.e. for every
XEW there exists f~ 2 such that f(x) # 0;
where Ci (al , a, , ... , a, ) are real numbers then the uniform closure of 2 consists of all
.
and a,',q', . . ,Ki are constant positive real continuous functions on U, i.e. (Z,dJ is
integers.] is called Sugeno type controller. dense in (qu1,dJ".
Note that in original Sugeno's Notes:
controller [13-141 the conclusion, y, was a ,
1) d-cf fi) is defined as "sup Xru
linear function of input variables and I I fdx) - f A x ) I where fl fie 2 '
)'I 9

therefore Buckley's d e f ~ t i o nof a Sugeno 2) qu] is the set of all continuous

type controller, equation (l), could be real valued functions defined over .the
viewed as a generalization to that of compact set U
Sugeno.
To evaluate the rules, given input
vector x=(xl ,x2 ,...sh)first let: III. Theoretical Justification:
In this section the theoretical basis
xi =T(A[(X,),A;&), ...,A$JJ) (2) for the following statement is provided: I'

Any process that can be controlled by a

for each rule in the rule base. Here Li(x,,,) continuous control command can also be
denotes the membership function for input controlled by the hybrid fuzzy-PID controller
fuzzy set Li evaluated at x, and T is a of figure 1".
triangular-norm properly selected in the
design procedure.(see [12] for different types (HI-1)
of T-norms) Buckley[S] used the stone-Weierstrass
The final output of the fuzzy theorem to show that the collection of all n-
controller is the crisp value: input Sugeno type controllers is "dense" in
K q I" 3 and therefore any continuous real
6 =Exi P~(~x~,c&.,& (3) valued function in 9111 can be approximated
to any degree of accuracy by a Sugeno type
i=l
controller. ( q P ]= P+B is the collection
To have 6 as a continuous function of the of all real valued continuous functions on
crisp input vector X it is sufficient to have a ...
I"=IxIx XI where I is a subset of the set

384
of real numbers over which all our fuzzy coefficients C,' in ( 5 ) are required to be
sets are defined. ) time varying. This means that L O is
Then he concluded "given any actually a time varying linear mapping from
process that can be controlled by a the input space I" to the output space R
continuous controller it can also be
controlled by a Sugeno type controller". (111-4)
Thus we need to prove that Buckley's For most practical control problems
Sugeno type controller can be approximated error signals, constructed by comparing
to any degree of accuracy by a hybrid fuzzy- system's interesting observable states with
PID controller. reference settings, their time derivative and
integral are inputs to the controller unit. In
(111-2) other words x,in input vector X is either an
Any continuous real valued function error signal, a time derivative of an error
in a P 1, and in particular 6 in equation signal or the integral of an error signal. In
(3) can be approximated to any degree of such cases conventional PID controllers with
accuracy by properly selected piecewise a proper setting of proportional, integral and
linear segments. To visualize, the piecewise derivative coefficients, can be used to
linear approximator is a line when 6E a I construct the piecewise linear approximator
1, a plane when 6~ q f ] and a hyper plane of equation (5).
of order n when 6E q I"] ,n 23.
To restate the above mentioned fact (III-5)
in mathematical notation, let SEI'' be an On the other hand, for a "smooth"
input vector. Define : continuous function ("smooth" being an
Si ( X I={{ 1 for X,E 1, X ~ E ), assumption which is reasonable in real world
... , X,,E R i d ) }& { 0 otherwise}} (4) physical systems) the time varying
coefficients of the linear piecewise
and: approximator of equation ( 5 ) themselves will
be real valued continuous functions over R.
This means that the proportional, integral
and derivative coefficients of the PID
controllers need to be continuous real valued
then coefficients Cmi could be found such time functions.
that for every E > 0 :
(HI-6)
Ha-WFE (6)
In a hybrid fuzzy-PID controller the
proportional, integral and derivative
The d, , defined in section 11-3, can be used coefficients of each PID controller are
to evaluate the distance function in equation outputs of a fuzzy logic controller with the
(6). control rules of the following type:

(HI-3) IFe, is A,' and d(e,)ldt is A,' and le, is A,'

The input vector X=(x, ,x2 ,...,x,,), in THEN proportional-coeficient is B]'. (7)
a control problem, is a functiw of the
conditions the process under control actually Therefore we need to prove that a
experiences. Therefore it is a time varying fuzzy controller(which is not a Sugeno type
quantity. With this new observation, the controller in this case) is mpable of
approximating any real valued continuous Further research in fuzzy modelling
function which PID coefficients may be of complex systems may extend the
required to adopt. proposed control strategy to cases in which
no analytical model of the process under
(HI-7) control is available.
Wang [2] proved that the set of n-
input-single-outputfuzzy systems, where the ACKNOWLEDGEMENTS
fuzzy rule base may consist of rules as in
(7), with product inference, centroid The work done by the fnst and
defuzzification and Gaussian membership second authors is partially supported by
function can approximate any real valued grants from NSF under grant number
continuous function on the compact set U E 446009. The authors wish to-thankthem for
B"to any degree of accuracy. As mentioned their financial support which made this work
earlier, Wang's result guarantees that there possible.
exists a fuzzy system to approximate any
continuous real valued function to any References:
degree of accuracy.
[l] Sayyarrodsari, B., A. Homaifar, J.E.
(111-8) Hogans, 1993, "Fuzzy Controller for Robot
Thus the collection of all fuzzy Arm Trajectory", Second International
controllers constructed by the cascade of a Conference on Fuzzy Theory and
fuzzy controller (not necessarily a Sugeno Techrwlogy, Durham, N.C., 1993
type controller) and a conventional PID
controller is dense in C[ I" 1, when the input [2] Wang, L.X.,"Fuzzy Systems Are
space satisfies the requirements of III-4 and Universal Approximators", Proc. IEEE
111-5,and can control any process that can International Cod. on Fuzzy Systems, San
be controlled through continuous control Diego, PP. 1163-1170, 1992.
commands.
[3] Wang, L.X., J.M.Mendel, "Back-
Propogation Fuzzy System as Nonlinear
VI. Conclusion: Dynamic System Identifiers", Proc. IEEE
We have shown that under certain International [email protected] Fuzzy System, San
conditions the hybrid implementation of Diego, March 1992.
bzzy and conventional PID controllers is a
nonlinear system controller. Conventional [4] Wang, L.X.,J.M. Mendel, "Fuzzy
PID controllers are the most commonly used Basis Functions, Universal Approximatien,
controllers in industry today. This approach and Orthogonal Least-Squares learning",
allows fuzzy controllers to be used in a wide IEEE Trans. on Neural Networks, vol. 3, no.
variety of industrial applications with no 5, PP.807-814, 1992.
more design effort than that needed for the
design of conventional PID controllers and [5] Buckley, J.J., "Sugeno type
convenient framework for the overall system controllers are universal controllers", Fuzzy
analysis. Furthermore, it provides the Sets and Systems 53, North-Holland
versatility required to extend the application Publishing Company, PP. 299-303, 1993.
of conventional PID controllers to new
nonlinear complex control problems. [6] Buckley, J.J., Y. Hayashi, "Fuzzy

386
input-output controllers are universal [ l l ] Lee, C.C., "Fuzzy Logic in Control
approximators", Fuzzy Sets and System 58, Systems: Fuzzy Logic Controller- Part I",
North-Holland Publishing Company, PP. IEEE Transactions Systems, Man,
273-278, 1993. Cybemetics, vol. 20, no. 2, PP. 404-418,
1990.
[7] Zadeh, L.A., "Fuzzy Sets",
Information Control 8, PP. 338-335, 1965. [12] Lee, C.C., "Fuzzy Logic in Control
Systems: Fuzzy Logic Controller- Part II",
[8] Zadeh, L.A., "Similarity Relations IEEE Transactions Systems, Man,
and fuzzy orderings", Information Science, Cybemetics, vol. 20, no. 2, PP. 419-435,
vol. 3, PP. 177-200, 1971. 1990.

[9] Zadeh, L.A., "Outline of a new [13] Takagi, T., M. Sugeno, "Fuzzy
approach to the analysis of complex systems Identification of Systems and its
and decision processes", IEEE Transaction Applications to Modeling and Control",
System Man Cybemetics, vol. SMC-3, PP. IEEE Transactions Systems, Man,
28-44, 1973. Cybemetics, vol. SMC-15, no. 1, PP.
116-13, 1985.
[lo] Zadeh, LA., "The concept of a
linguistic variable and its application to [14] Sugeno, M., G.T.Kang, "Structure
approximate reasoning I, 11, III", Identification of Fuzzy Model", Fuzzy Sets
Informution Science, vol. 8, 199-251, 301- and Systems 28, North-Holland
357, vol. 9, PP. 43-80, 1975. PublishingCompany, PP. 15-33, 1988.

- FUZZY
LOGIC
CONTROL LEA
p
b
PID
CONTROLLER
UUDER
CON r R O L -CONDITIONER

Figure 1: Overall block diagram of Hybrid Fuzzy-PID Controller

387