146 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1, NO.

2, MAY 1993

Stable Adaptive Fuzzy Control of Nonlinear Systems

Li-Xin Wang, Member, IEEE

Abstract- An adaptive fuzzy controller is synthesized from adaptive fuzzy controller is a standard fuzzy logic controller
a collection of fuzzy IF-THEN rules. The parameters of the [8] used in most fuzzy control systems, plus a supervisory
membership functions characterizing the linguistic terms in the control which fires only when the state hits some (large)
fuzzy IF-THEN rules change according to some adaptive law for
the purpose of controlling a plant to track a reference trajectory. boundaries. The initial adaptive fuzzy controller is constructed
These fuzzy IF-THEN rules are either collected from experienced from the fuzzy IF-THEN rules provided by human experts and
human operators or generated automatically during the adapta- some arbitrary rules; an adaptive law is then used to update
tion procedure. In this paper, a direct adaptive fuzzy controller is the parameters of the adaptive fuzzy controller during the
designed for a general higher-order nonlinear continuous system adaptation procedure. If the fuzzy IF-THEN rules from human
through the following three steps: first, define some fuzzy sets
whose membership functions cover the state space; then, use experts provide good control strategies, then the adaptation
fuzzy IF-THEN rules from human experts and some arbitrary procedure will converge very quickly; on the other hand,
rules to construct an initial adaptive fuzzy controller in which if there are no linguistic rules from human experts, then
some parameters are free to change; finally, develop an adaptive our adaptive fuzzy controller becomes a regular nonlinear
law to adjust the free parameters based on a Lyapunov synthesis adaptive controller, similar to the radial basis function adaptive
approach. It is proved that 1) the closed-loop system usiog this
adaptive fuzzy controller is globally stable in the sense that all controller [ 131, [ 141, the neural network adaptive controller
signals involved are bounded and 2) the tracking error converges [ 121-[ 141, etc.
to zero asymptotically if the minimum approximation error of In Section 11, we discuss some general issues concerning
the fuzzy controller to an optimal controller is squared integrable fuzzy control and adaptive fuzzy control and outline the way
along the state trajectory. Finally, we apply this direct adaptive in which adaptive fuzzy controllers are classified. In Section
fuzzy controller to control an unstable system and a chaotic
system. 111, we present a detailed description of fuzzy logic systems
which are used as principal components of our adaptive
fuzzy controller. In Section IV, we design a stable direct
I. INTRODUCTION adaptive fuzzy controller in a constructive manner based on the

F UZZY logic controllers [8] are generally considered appli- Lyapunov synthesis approach. In Section V, the detailed design
cable to plants that are mathematically poorly understood steps of the direct adaptive fuzzy controller are presented,
and where experienced human operators are available for and its properties (boundedness of variables, convergence of
providing qualitative “rules of thumb.” Although achieving tracking error, etc.) are analyzed. In Section VI, the direct
many practical successes, fuzzy control has not been viewed as adaptive fuzzy controller is used to control an unstable sysiem
a rigorous science due to a lack of formal synthesis techniques and a chaotic system. Section VI1 concludes the paper.
which guarantee the very basic requirements of global stability
and acceptable performance. Although there has been some
11. A GENERAL DISCUSSION OF FUZZY CONTROL
research on the stability analysis of fuzzy control systems
(e.g.. [6] and [7]), it has been assumed that the mathematical Fuzzy control is by far the most successful application
model of the plant is known, which contradicts the very basic of fuzzy sets and systems theory to practical problems. The
premise of fuzzy control systems, i.e., to control processes present interest in fuzzy theory is largely due to the successful
that are poorly understood from a mathematical point of applications of the fuzzy logic controllers to a variety of
view. In fact, if the mathematical model of a plant is known, consumer products and industrial systems. Why has fuzzy
then conventional linear and nonlinear control methods should control been so successful?
be given higher priority. Fuzzy control should be useful in
situations where 1) there is no acceptable mathematical model A. Why Fuuy Control?
for the plant and 2) there are experienced human operators
It may be helpful to divide the reasons for why fuzzy control
who can satisfactorily control the plant and provide qualitative
into two categories: theoretical and practical reasons.
control rules in terms of vague and fuzzy sentences. There are
I) Theoretical Reasons for Fuuy Control:
many practical situations where both 1) and 2) are true.
The adaptive fuzzy controller developed in this paper is As a general rule, a good engineering approach should
suitable for these situations. The basic architecture of our be capable of making effective use of all the available
information. If the mathematical model of a system is
Manuscript received June 3, 1992; revised September 5 , 1992. This work too difficult to obtain (this is true for many practical
was supported by the Rockwell International Science Center. systems), then the most important information comes
The author is with the Department of Electrical Engineering and Computer
Sciences, University of California, Berkeley, CA 94720. from two sources: 1) sensors which provide numerical
IEEE Log Number 9205330. measurements of key variables and 2) human experts
1063-6706/93$03.00 0 1993 IEEE
WANG: STABLE ADAPTIVE FUZZY CONTROL I41

who provide linguistic descriptions about the system and software tools available for designing fuzzy controllers.
control instructions. Fuzzy controllers, by design, provide Thus, fuzzy control is an approach that has a high
,a systematic and efficient framework for incorporating performancekost ratio.
linguistic fuzzy information from human experts. Con-
ventional controllers, however, cannot incorporate the
linguistic fuzzy information into their designs. If in some B. Why Adaptive Fuzzy Control?
situations the most important information comes from Fuzzy controllers are supposed to work in situations where
human experts, then fuzzy control is the best choice. there is a large uncertainty or unknown variation in plant
Fuzzy control is a model-free approach; i.e., it does parameters and structures. Generally, the basic objective of
not require a mathematical model of the system under adaptive control is to maintain consistent performance of .
control. Control engineers are now facing more and a system in the presence of these uncertainties. Therefore,
more complex systems, and the mathematical models advanced fuzzy control should be adaptive.
of these systems are increasingly difficult to obtain. What is adaptive fuzzy control? Roughly speaking, if a
Thus, model-free approaches have taken on added im- controller is constructed from adaptive fuzzy systems (an
portance. Conventional control also has some model-free adaptive fuzzy system is a fuzzy logic system equipped with
approaches, e.g., nonlinear adaptive control and PID a training (adaptation) algorithm), it is called an adaptive
control. Fuzzy control provides yet another model-free fuzzy controller. An adaptive fuzzy controller can be a single
approach. adaptive fuzzy system or can be constructed from several
Fuzzy control provides nonlinear controllers, which are adaptive fuzzy systems.
well justified by the universal approximation theorem in How does an adaptive fuzzy controller compare with a
[19], [23]; i.e., these fuzzy logic controllers are gen- conventional adaptive controller? The most important advan-
eral enough to perform any nonlinear control actions. tage of adaptive fuzzy control over conventional adaptive
Therefore, by carefully choosing the parameters of the control is that adaptive fuzzy controllers are capable of incor-
fuzzy controllers, it is always possible to design a fuzzy porating linguistic fuzzy information from human operators,
controller that is suitable for the nonlinear system under whereas conventional adaptive controllers are not. This is
control. especially important for the systems with a high degree of
uncertainty, e.g., in chemical processes and in aircraft, because
In the above theoretical reasons, fuzzy control is viewed as
although these systems are difficult to control from a control
a theory and is evaluated from a theoretical point of view.
theory point of view, they are often successfully controlled
We know that theory and practice have different emphasis.
by human operators. How can human operators successfully
For example, theory emphasizes generality and rigor, while
control such complex systems without a mathematical model
practice emphasizes applicability to particular problems. So,
in their minds? If we ask the human operators what control
why is fuzzy control desirable from a practical point of view?
strategies they follow, they may just tell us a few control
2 ) Pructical Reasons for Fuzzy Control:
rules in fuzzy terms and some linguistic descriptions about
It is easy to understand. Because fuzzy control emulates the behavior of the system under various conditions which
human control strategy, the underlying principle can be are, of course, also in fuzzy terms. Although these fuzzy
easily understood by those who are not control specialists. control rules and descriptions are not precise and may not be
During the last two decades, conventional control theory sufficient for constructing a successful controller, they provide
has been using increasingly advanced mathematical tools. very important information about how to control the. system
This is needed in order to solve difficult problems in a and how the system behaves. Adaptive fuzzy control provides
rigorous fashion; however, this also results in a dimin- a tool for making use of the fuzzy information in a systematic
ishing number of practical engineers who can understand and efficient manner.
the theory. Therefore, practical engineers who are in the How are adaptive fuzzy controllers classified? We classify
front line .of designing consumer products tend to use adaptive fuzzy controllers according to whether the adaptive
approaches that are simple and easy to understand. Fuzzy fuzzy controller can incorporate fuzzy control rules or fuzzy
control is just such an approach. descriptions about the system. We detail this classification
It is simple to implement. Fuzzy logic systems, which next.
are at the heart of fuzzy control, admit a high degree
of parallel implementation. Many fuzzy VLSI chips have
been developed which make the implementation of fuzzy C. Direct and Indirect Aduptitve Fuzzy Control
controllers simple and fast. In the conventional adaptive control literature, adaptive
It is i n e ~ e n s i v eto develop. From a practical point of controllers fall into two categories [ 121: direct and indirect.
view, the development cost is one of the most important In direct adaptive control, the parameters of the controller
criteria for a successful product. Because fuzzy con- are directly adjusted to reduce some norm of the output error
trol is easy to understand, the time necessary to learn between the plant and the reference model. In indirect adaptive
the approach is short; i.e., the “software cost” is low. control, the parameters of the plant are estimated and the
Also, because fuzzy control is simple to implement, the controller is chosen assuming that the estimated parameters
“hardware cost” is also low. What is more, there are represent the true values of the plant parameters.
1

148 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1, NO. 2, MAY 1993

-
In fuzzy control, linguistic information from human experts
can be placed into two categories:
Fuzzy control rules, which set forth the situations in which
certain control actions should be taken (e.g., we often use
the following fuzzy IF-THEN rule to drive a car: “IF the
speed is slow, THEN apply more force to the accelerator,”
where the “slow” and “more” are labels of fuzzy sets).
Fuzzy IF-THEN rules, which describe the behavior of
the unknown plant (e.g., we can describe the behavior
of a car using the fuzzy IF-THEN rule: “IF more force
is applied to the accelerator, THEN the speed of the -
car will increase,” where the “more” and “increase” are
characterized by fuzzy membership functions). Engine
Interestingly enough, adaptive fuzzy controllers which make
use of these two classes of linguistic information just corre-
spond to the direct and indirect adaptive control schemes, re-
spectively. More specifically, direct adaptive fuzzy controllers Fig. 1 . Basic configuration of fuzzy logic systems.
use fuzzy logic systems as controllers; therefore linguistic
fuzzy control rules can be directly incorporated into the con-
defined in the product space U x R. Based on generalizations
trollers. On the other hand, indirect adaptive fuzzy controllers
of implications in multivalue logic, many fuzzy implication
use fuzzy logic systems to model the plant and construct the
rules have been proposed in the fuzzy logic literature. Here
controllers assuming that the fuzzy logic systems represent
we quote four commonly used fuzzy implication rules [SI:
(approximately) the true plant; therefore fuzzy IF-THEN rules
describing the plant can be directly incorporated into the Min-operation rule of fuzzy implication:
indirect adaptive fuzzy controllers. Formally, we have the P F ; x ...x F;-+Gl (7,Y) PGl (Y)]; (2)
= min[pF; x ...x FA (21,
following definitions:
If an adaptive fuzzy controller uses fuzzy logic systems as Product-operation rule of fuzzy implication:
controllers, it is called a direct adaptive fuzzy controller. (3)
PF; ... F; +cl (7,Y) = PF; x ...x F; ( ~ P G(9);
I
A direct adaptive fuzzy controller can incorporate fuzzy
control rules directly into itself. Arithmetric rule of fuzzy implication:
If an adaptive fuzzy controller uses fuzzy logic systems
as model of the plant, it is called an indirect adaptive pFi x .. . x F; +Gi (g,Y) = min[l , l-’lF,” x ...x F +pG’ ( Y)] ;
A (g)
fuzzy controller. A indirect adaptive fuzzy controller can (4)
Maximum rule of fuzzy implication:
incorporate fuzzy descriptions about the plant (in terms
of fuzzy IF-THEN rules) directly into itself.
P F ; x ... xF;+Gl ( 7 , Y ) = m m [min ( P F ; x ... XF: (71,Pcl (
Y),
In this paper, we will develop a direct adaptive fuzzy
controller. Indirect adaptive fuzzy controllers were developed
in [20]. (5 1
where pF; x . . . (7) is defined by
111. DESCRIPTION
r
OF FUZZY LOGICSYSTEMS
FF;x ...x F ; ( g ) = PF:(z1) *‘..*pF;(2n). (6)
Fig. 1 shows the basic configuration of the fuzzy logic
systems considered in this paper. The fuzzy logic system Here the symbol “*” denotes the t-norm [8], which corresponds
performs a mapping from U c R” to R. We assume that to the conjunction “and” in (1). The most commonly used
U = U1 x . . . x U,, where Vi c R, i = 1 , 2 , .. . , n. We now operations for the t-norm are
present a detailed description of each of the four blocks in the min(u, U) fuzzy intersection
fuzzy logic system of Fig. 1. algebric product (7)
The fuzzy rule base consists of a collection of fuzzy
IF-THEN rules:
max(0, U +U - 1) bounded product.

The fuzzy inference engine performs a mapping from fuzzy

R(’):IF z1 is Fi and . . . and 2, is FA; sets in U to fuzzy sets in R, based upon the fuzzy IF-THEN
THEN Y is ~ l , (1) rules in the fuzy rule base and the compositional rule of
inference [25]. Let A , be an arbitrary fuzzy set in U ; then,
where 7 = (21, ..., 2,) T E U and Y E R are the input each R(‘)of (1) determines a fuzzy set, A , o R(‘)),
in R based
and output of the fuzzy logic system, respectively, F: and on the following sup-star compositional rule of inference:
G’ are labels of fuzzy sets in Vi and R, respectively, and
2 = 1 , 2 , . . . ,M . Each fuzzy IF-THEN rule of (1) defines a ( ~ )= s u P , ~ u [FA, (7)* P F ; X . . . x ~ ; - -(7,
P A , ~ R (Y) . ~Y
)l] ,
fuzzy implication [SI F: x . . . x FL + G’, which is a fuzzy set (8)
WANG: STABLE ADAPTIVE FUZZY CONTROL 149

where * is the t-norm (7), and p F ; x F ~ - G l (g,y ) is deter- from U C R" to R, because within each block there are many
mined by the fuzzy implication rules of (2)-(5). The final fuzzy different choices, and many combinations of these choices can
set,,A, o ( R ( l )... . , R("')),determined by all the A4 rules in result in useful subclasses of fuzzy logic systems. We now
the fuzzy rule base is obtained by combining / L ~ , ~ ~ ( L ) of ( Y ) consider one wbclass of fuzzy logic systems which will be
.
(8) for 1 = 1 . 2 , . . . A4 using fuzzy disjunction: used as building blocks of our adaptive fuzzy controller.
Lemma 1: The fuzzy logic systems with center-average
/ J . 4 L o ( R i l ) , ,.,R(,\.'))(Y)= p A , o R ( l ) (?/I+
' ' . + p A z o R i A f l(?/I.defuzzijier ( 1 3 ) , product inference (1 l), and singleton fuzzi$er
(9) are of the following form:
where + denotes the t-conorm [8]; the most commonly used
operations for + are
max(u. w) fuzzy union
u+w =
{+ U *ti - uw
+
'algebraic sum
Inin( 1,U U ) bounded sum.
If we use the product operation (3) and choose * in (6) and
(10)
where jj' is the point at which pGl achieves its maximum
value, and we assume that /iGl(y') = 1.
(8) to be an algebraic product, then the inference is called Proof: Using the center-average defuzzifier (1 3), we have
product inference. Using product inference, (8) becomes

&4,oR!') (!/) = s"PgeC [ P A , ( z ) j L F ;(%1)' . ' pl?;>( - c n ) p G ' ( Y ) ]'

( 1 1)
The fuzzijier maps a crisp point g = ( X I , . . . , z,)~ E U where / i , 4 r o n (ij~j ' )) is given by the product inference (1 1).
into a fuzzy set A, in U . There are (at least) two possible If we use the singleton fuzzifier, we have p . ~ ~ ( g=' ) 1 for
choices of this mapping: z' = g (g is the input crisp point to the fuzzy logic system)
-
i) A, is a fuzzy singleton with support g;i.e.,/LAz(c') = 1 and (g') = 0 for all other 2' E U ; therefore, the "sup" in
for z'= g andpAr(2')= 0 for all other g' E U with g' # g. (1 1) is achieved at "1:' = 2,and (1 1) can be simplified to
ii) p.4z(2)= 1 and pa4=(2')decreases from 1 as g' moves
d-,)T(&g)
away from g,e.g.,pA, (&) = exp[- (- - g 2
is a parameter characterizing the shape of p~~ (g').
1, where i r 2 /L.4zoR!lj (r/') = n71

i=l
/"F: ( x i ) (17)

In the literature, it seems that only the singleton fuzzifier

i) has been used. We think that the nonsingleton fuzzifier ii) (we assume that /iG! ( j j ' ) = 1).Substituting (17) into (16), we
may be useful if the inputs are corrupted by noise. obtain (15). Q.E.D.
The defuzzijier maps fuzzy sets in R to a crisp point in R. If we fix the / i p ( . c L ) ' sand view the j j " s as adjustable
There are (at least) three possible choices of this mapping: parameters, then (Is') can be written as
i) Maximum defuzzijier, defined as
y(.r) = HT<(.).
- (18)
?/ 1 ) (Y'))
argsupgl,R (PAzo ( R ( ' ),...,R("f . 2,
where .
is a parameter vector, and E(&)
= (;VI.. . . jj-'J)T - =
where I".~,~(R~I),...,~~.,'))(Y')
is given by (9). ( < I ( & ) . . . . . <."(.IJ)~
is a regressive vector with the regressor
ii) Center-average defuzzijer, defined as (which is called fuzzy basis function in [ 2 3 ] )defined as
('(2)

where ,ql is the point in R at whichpG(l)(y) achieves its

maximum value, and p U d z o R i ~ is
) ( ygiven
) by (8). The case where /iF; ( x i ) ' s are also adjustable is studied in 1201.
iii) Modijied center-average defuzzijier, defined as There are two main reasons for using the fuzzy logic
systems ( 1 5 ) as basic building blocks of adaptive fuzzy con-
trollers. First, it was proved in [19] and [23] that the fuzzy
logic systems in the form of (15) are universal approximators;
i.e., for any given real continuous function g on the compact
where m' is a parameter characterizing the shape ofPG1 (y) set U , there exists a fuzzy logic system in the form of (15)
such that the narrower the shape of/LGl(y), the smaller 0 ' . such that it can uniformly approximate !J over U to arbitrary
For example, if pG((y) = e ~ p [ - ( + ) ~ ] , thenal is such a accuracy. Therefore, the fuzzy logic systems (15) are quali-
parameter. , fied as building blocks of adaptive controllers for nonlinear
Note that if we use the center-average or modified systems. Second, the fuzzy logic systems (IS) are constructed
center-average defuzzifiers, we do not need to calculate the from the fuzzy IF-THEN rules of ( 1 ) using some specific
/ L ~ , ~ ( ~ (, R ~ () h, f ) ) ( y of
) (9); we only need to calculate the fuzzy inference, fuzzification, and defuzzification strategies;
11 A T O R ! ' ) (y) of (8) in the fuzzy inference engine. therefore, linguistic information from human experts (in the
From the above we see that the fuzzy logic systems of form of the fuzzy IF-THEN rules of (1)) can be directly
Fig. 1 comprise a very rich class of static systems mapping incorporated into the controllers.
150 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. 2, MAY 1993

Iv. BASICIDEAS OF CONSTRUCTING STABLE

DIRECTADAPTIVE FUZZY CONTROLLERS
In this section, we first set up the control objectives, and
then show, in a constructive manner, how to develop a direct
adaptive fuzzy controller to achieve these control objectives.
Consider the nth-order nonlinear systems of the form

,(n) = f(2,i,. . . ,dn-l)) + bu, y = 2, (20)

where f is an unknown continuous function, b is a positive

unknown constant, and U E R and y E R are the input and - 0 1 0 0 ... 0 0 -
0
output of the system, respectively. We assume that the state 0 0 1 0 ... 0 0
vector g = ( ~ 1 ~ x .2. ,, z,)~
. = (2,2 , . . . ,x ( " - ~ )E) R"
~ is A,= . . . .. . ; ,bc=[;].
available for measurement. In the spirit of the nonlinear control 0 0 0 0 ... 0 1
literature [5], [ 171, these systems are in normal form and have -k, - k,- . .. . .. . . . . .. -h-
the relative degree equal to n. The control objective is to force
y to follow a given bounded reference signal, ym(t), under
the constraint that all signals involved must be bounded. More
specidcally, we have
Control Objectives: Determine a feedback control U =
u(:lO) (based on fuzzy logic systems) and an adaptive law
e
for adjusting the parameter vector such that the following
conditions are met:
i) The closed-loop system must be globally stable in the
sense that all variables, :(t),@(t)and u(:I@), must be uni-
formly bounded; i.e., Ig(t)l 5 M, < CQ, lO(t)l I Me < 00,
andlu(:l6)l 5 Mu < 03 for all t 2 0, where M,, Me and Mu
are design parameters specified by the designer.
ii) The tracking error, e ym - y, should be as small as
possible under the constraints in i).
We now show the basic ideas of how to construct a direct
adaptive fuzzy controller to achieve these control objectives.
To begin, let g = ( e ,e , . . . , e(n-l))T and = ( k n , . . . , I C I ) ~E
Rn be such that all roots of the polynomial h ( s ) = S" +
klsn-l+. . .+k, are in the open left-half plane. If the function
f and the constant b are known, then the control law
-f(a)
):/I+ +IC Tg ]
(21)

applied to (20) results in

e(,) + kle("-') + . . . + kne = 0, (22)

which implies that limt,, e ( t ) = 0-a main objective of

control. Since f and b are unknown, the optimal control, U * ,
of (21) cannot be implemented. Our purpose is to design a
fuzzy logic system to approximate this optimal control. 1
Suppose that the control U is the summation of a fuzzy I --eTQe
2- 5 0.
control uC(:lO) and a supervisory control u,(g): Therefore, using the supervisory control U , of (30), we always
have V, 5 V. Because P > 0, the boundedness of V, implies
U = 4al@) + %(a)> (23)
the boundedness of g , which in tum implies the boundedness
where uC(g[f9)is a fuzzy logic system in the form of (15) (or of 2.
equivalently (18)), and U,(:) will be determined later in this From (30) we see that the U, is nonzero only when the
section (for the reasons why U, is called a supervisory control, error function V, is greater than the positive constant V. That
see the discussion later in this section). Substituting (23) into is, if the closed-loop system with the pure fuzzy controller
(20), we have U , is well behaved in the sense that the error is not big (i.e.,
V, I V), then the supervisory control U, is zero; on the other
hand. if the svstem tends to be unstable (i.e., V , > V , . then
WANG: S W B L E ADAFTIVE FUZZY CONTROL 151

the supervisory control U , begins operating to force V, 5 V .

In this way, the control U , is like a s ~ p e r 7 ~ i s othis
r ; is why
we call the U , a supervisory control.
If we choose 1; 5 1 in (30), then from (31) we can
guarantee to have not only the boundedness of the state vector
but also the error e converging to zero. We do not choose this I I 1 I I
Fuzzy Controller uc(xle)
strategy because the U , is usually very large. Indeed, from
in the form of (15)
(30) we see that the U,, is proportional to the upper bound
f '-, which is usually very large. Large control is undesirable 4
because it may increase the implementation cost. We therefore
choose the U , to operate in the above supervisory fashion. determined Adaptive Law
Next, we replace the uc(gl@)by th'e fuzzy logic system (18)
and develop an adaptive law to adjust the parameter vector e. by linguistic
information
Define the optimal parameter vector:

H* a r g I ~ ~_i-r l ~ ~[~ <

~ . ~q r ,~ ~I W~ ( S~ I ~t -
) r ?L*I]
, (32)

and the ''minimum approximation error":

711 E U,(.C\@*) - U * . (33)

FIg. 2. The overall scheme of direct adaptive fuzzy control
The error equation (26) can be rewritten as

= A,. + ,,[~,(gl@*)- ~,(gl@)]- bc7~,(g)- ~ , U I Assumption 2: There are L fuzzy control rules from human
experts in the following form:
A,P, + bCgT<(:rr) b , ~ , - - bcw, (34)
R p ) : IF :xl is A;' and . . . and s, is is C".
THEN 11,
where U', E @* - @ and <(g)is the fuzzy basis function (19). (41)
Define ?he Lyapunov function candidate where '4; and C" are labels of fuzzy sets in R, and I' =
1 b 1 . 2 , . . L.
' '

v = -eTPe.+
2-
-$T$
27- -'
(35) We allow L = 0, which means that there are no linguistic
control rules; therefore, Assumption 2 is not necessary. We
where Y is a Positive COnstant. Using (34) and (281, we have make this assumption for the purpose of emphasizing that our
direct adaptive fuzzy controller can directly incorporate these
v. = --eTQ
1
2-
+ g =pbc($T <(z) - U, - W) + -4
7-
T '
4.
-
(36) fuzzy control rules (if there are any) into its design. The details
are given next.
Let U,,be the last column of P ; then from (27) we have
v. DESIGNAND STABILITY ANALYSIS
OF
gTPb, = gTpnb. (37) THE DIRECTADAPTIVEFUZZY CONTROLLER

Substituting (37) into (36), we have In this section, we first present the detailed design steps
of the direct adaptive fuzzy controller, and then study its
. 1 b
v= -- C T Q ~ + - ~ ~
2- y-
[yg7pn<(g)+ 41 - ~ T P ~ , ~ L , - ~ T P ~ , ~ . properties.
(38)
If we choose the adaptive law: Design of Direct Adaptive Fuzzy Controller
Step 1 ) Off-Line Processing:
8 = 7eTpllg.).
- (39) Specify the k . 1 . . . . . k,, such that all roots of S ~ ~ + A : ~ S " - ~ +
then (38) becomes . . . + k f , = 0 are in the open left-half plane. Specify a
positive definite 71 x 71 matrix .)C
v' 5 --eTd)P,-
I
2-
gTPbcw, (40)
Solve the Lyapunov equation (28), e.g., using the method
of [24], to obtain a symmetric P > 0.
Specify the design parameters M O , and ,U,, based
where we use,the facts gTPb,u, 2 0 (cf. (30)) and = 4 -6. on practical constraints.
This is the best we can achieve. In order to guarantee 5
.4!l0, we will use a moiection algorithm to modify the basic Step 2 , Initial Construction:
- " I

adaptive law (39); the details are given in Section V. Define 7 n ; fuzzy sets F;' whose membership functions
To summarize, the overall scheme of our direct adaptive j L F 2 1 c uniformly cover tTL, where 1, = 1 . 2 . . . . . rrL, and

fuzzy controller is shown in Fig. 2. 'i = 1 . 2 . . . . . T I , . The I' = U1 x . x U,, is usually chosen
Finally, we specify the linguistic information which can be to be {x E R'l : 11-1 5 h f z } .We require that the F,'"s
directly incorporated into the direct adaptive fuzzy controller. include the Ay's in (41).
152 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1. NO. 2, MAY 1993

Construct the fuzzy rule base for the fuzzy logic system and
uc(gl@),which consists of ml x m2 x ... x m, rules
whose IF parts comprise all the possible combinations of
the Fit'sfor i = 1,2, . . . I n. Specifically, the fuzzy rule
base of uc(gl0)consists of the rules

R~"-~"-)
: IF z1 is F,"'and . . . and z,is :F for all t 2 0, where Amin is the minimum eigenvalue of P ,
W ) T .
THEN U, is G(ll,..'>ln) (42) and ym = ( ~ m , Y m , , . . , Y m
ii)
where li = 1 , 2 , .. . ,mi, i = 1 , 2 , . . . , n, and G('"'''>'n)
are fuzzy sets in R, which are specified as follows: if
the IF part of (42) agrees with the IF part of (41),
set G('li.,,,'n) equal to the corresponding C' in (41);
otherwise, set G('1i"',*n-)arbitrarily with the constraint for all t 2 0, where a and c are constants, and w is the
that the centers of G('l).,.i'n) (which correspond to the minimum approximation error defined by (33).
y' parameters) are inside the constraint sets {e
: 181 I (iii) If w is squared integrable, i.e., Iw(t)12dt < 30,
M O } .Therefore, the initial adaptive fuzzy controller is then limt,, le(t)l = 0.
constructed from the fuzzy control rules (41) and some Proof of this theorem is given in the Appendix. We now
arbitrary rules. make a few remarks.
Construct the fuzzy basis functions Remark I : For many practical control problems, the state
z and control U are required to be constrained within certain
-

regions. For given constraints M, and Mu, we can specify the

design parameters ICl MO,Q, and V , based on (47) and (48),
Ll,%. . . t l, such that the state g and control U are within the constraint
and collect them into a ny==, mi-dimensional vector - [(g)
sets. To do this, we need to know some fixed bounds of lyml
and IfU(g)I.Since these functions are known to the designer,
in a natural ordering for 11 = 1 , 2 , . . . ml . ,I, =
1 , 2 , . . . ,m,. Collect the points at which the pG(t1,. . , l n ) ' s it should not be difficult to determine these bounds. After these
achieve their maximum values, in the same ordering as bounds are determined, we can specify the values of the right-
e.
[(c),into a vector The u,(gle) is constructed as
-
hand sides of (47) and (48) by properly choosing the design
parameters. Note from (28) and (27) that Amin is determined
by and Q; therefore we can specify the and Q to achieve
the required Amin.
Step 3) On-Line Adaptation: Remark 2: From step 2 we see that the fuzzy control rules
Apply the feedback control (23) to the plant (20), where (41) are incorporated into the adaptive fuzzy controller by con-
U , is given by (44) and U , is given by (30).
structing the initial controller based on them. If the linguistic
Use the following adaptive law to adjust the parameter rules (41) provide good control strategy, then the initial U ,
vector e: should be close to the optimal control U * ; as a result, we can
hope that the closed-loop system behaves approximately like
(22). If no linguistic information is available, our adaptive
fuzzy controller is still a well-performing nonlinear adaptive
controller, in the sense of having the properties (i)-(iii) of
Theorem 1. In summary, good linguistic information can help
where the projection operator P{*} is defined as [4], [9]: us to construct a good initial controller so that we can have
a fast adaptation. We will show an example in Section VI to
illustrate this point.
Remark 3: From (iii) of Theorem 1 we see that in order for
the tracking error e ( t ) to converge to zero, we require that the
The following theorem shows the properties of this direct "minimum approximation error," w,defined by (33) be small
adaptive fuzzy controller. (in the sense of squared integrable). Based on the universal
Theorem I : Consider the nonlinear plant (20) with the approximation theorem of [19], [23], we can say that if we
control (23), where U , is given by (44), and U , is given by (30). use a sufficient number of rules to construct U,, the w should
Let the parameter vector be adjusted by the adaptive law (45) be small.
and let Assumptions 1 and 2 be true. Then, the overall control
scheme (shown in Fig.2) guarantees the following properties:
i) le(t)l I Me7 VI. SIMULATIONS
In this section, we apply the direct adaptive fuzzy controller
(47) developed in the last section to control an unstable system
(Example 1) and a chaotic system (Example 2).
WANG. STABLE ADAPTIVE FUZZY CONTROL IS?

-3’ ” ” ” ’ ’
I
0 2 4 6 8 10 12 14 16 18 W
t

Fig. 3. Fuzzy membership functions de6ned over the state space. Fig. 4. Closed-loop system state . r ( t ) using the direct adaptive fuzzy
controller for the plant (50) without incorporating any fuzzy control rules.

Example 1: In this example, we use our direct adaptive

fuzzy controller to regulate the plant 3

1
1
to the origin; i.e., y771 E 0. It is clear that the plant (50)
is unstable if without control, because if P L ( ~ ) E 0, then
. lpe-’ -1 L
l + r , - ‘ > 0 for 3: > 0, and :i: = l + e -
.I‘ 1 ~
< 0 for -1

.I’ < 0. We choose y = 1, Ibfr = 3, MO = 3, b~ = -2 t

lpe-s(f)
0.5 < 1 = b, and j”‘- = 1 2 l + c - r ( t l . In step 1, we
~

1
3O 2 4 6 8 IO 12 14 16 18 W
define six fuzzy sets over the interval [-3,3], with labels
n’3. AV‘2.1Vl. P1. P2, and P3, and membership functions
+
11 .,-:3 (.I‘) = 1/ ( 1 exp(5(:c + + ’
2) ) ) . p A V 2(:c) = exp( - ( z 1.5) ),
Fig. 5. Closed-loop system state x ( t ) using the direct adaptive fuzzy
controller for the plant (50) after incorporating the fuzzy control rules (51)
/i.~l(.r= ) exp(-(.r + 0 . 5 ) ’ ) . p p ~ ( . r . )= exp(-(:r - 0.5)’). and (52).
p p 2 ( , 1 ‘ )= exp-(:1;-1.5)~) and , u P ~ ( : J ;= ) l/(l+exp(-5(:c-
2 ) ) ) , which are shown in Fig. 3. In step 2 , we consider two
cases: 1 j there are no fuzzy control rules and the initial 0;(0)’s
are chosen randomly in the interval [-a. 21; and 2) there are
two fuzzy control rules:

R2 : IF .I’ is P2. THEN II is N B .

where p p ~ ( ’ u=) e ~ p ( - ( u - 2 ) ~ )and , p . v ~ ( u= ) exp(-(,ii+

These two rules are obtained by considering the fact
that our problem is to control :I:( t ) to zero; therefore, if z is
negative, then the control ‘u(:c) should be “positive big” (PB)
so that it may happen that .i; > 0 (see (SO)). On the other hand,
if :I: is positive, then the control * I ~ , ( : I ; )should be “negative big”
(NB) so that it may happen that :i; < 0 . We used the MATLAB parameters the same as in the simulation shown in Fig. 4 (i.e.,
command “ode23” to simulate the overall control system. We without using the fuzzy control rules). The simulation result
chose the initial state ~ ( 0 =) 1. Figs. 4 and S show the z ( t )for for this case is shown in Fig. 6. We see that the supervisory
the cases without and with the linguistic control rules (SI) and control U, did force the state to be inside the constraint set
(S2), respectively. We see from Figs. 4 and S that 1) our direct 1x1 I 1.5.
adaptive fuzzy controller could regulate the plant to the origin It is interesting to observe how this bounded control was
without using the fuzzy control rules and 2) by incorporating achieved. From Fig. 6 we see that as soon as the state hit
the fuzzy contrpl rules, the speed of convergence becomes the boundary, the supervisory control U, began operating
much faster. We also simulated for other initial conditions, to force the state back to the constraint set. Since as soon
and the results were very similar; we do not show them in as the state was inside the constraint set, the supervisory
order to keep the clear the figures and make the comparison control stopped operating, which resulted in the state hitting
easier. the boundary again. The continuation of this kind of back-and-
In the simulations shown in Figs. 4 and 5 , the state I:(/,)did forth operation resulted in the “holding” of the state around the
not hit the boundary 1x1 = 3; therefore the supervisory control boundary, as shown in Fig. 6 in the intervals t E [0.5, 2.51 and
U , , never fired. Now we change M r = 1.5 and keep all other t E [5.5. 71 (approximately). That is, the supervisory control
154

8 -
4-
6 -

4-
2-
2-

r 0 -
2-

4- 2-

i
-6-

8-

-101
4 - 3 - 2 - 1 0 1 2 3 4
I I I
d4 -3 -2 -1 0 1 2 3 4
X1
X I

Fig. 7. Trajectory of the chaotic system (53) and (54) in the ( ~ 1 ~ x phase
2) Fig. 8. Closed-loop system trajectory ( z l ( t ) z, 2 ( t ) ) using the direct adap-
plane with u ( t ) G 0 and zl(0)= zz(0) = 2. tive fuzzy controller for the chaotic system (53) and (54) for the Abfz = 10
case.

could prevent the system from being unstable, but could not
regulate the state to the origin. The encouraging observation is
that during these “holding periods” the adaptive law adjusted
the parameters of the fuzzy controller U,, and, finally, the fuzzy
4t 1
control “recovered” and finished the control task-regulating
the state to the origin.
We may view the fuzzy controller U , as a lower-level
operator and the supervisory control U , as a higher-level
supervisor. If the operator can perform successful control, the
supervisor just observes and does not take any action. If for
-5 I
some reason the operator cannot control the system well, the 4 - 3 - 2 - 1 0 1 2 3 4

XI
supervisor takes actions to prevent catastrophic consequences.
Fig. 9. Closed-loop system trajectory ( z l ( t ) ~
, ( t )using
) the direct adap-
During this period, the operator learns to correct hisher tive fuzzy controller for the chaotic system (53) and (54) for the .If, = 3
mistakes and regains control. After the system is back to case
normal, the supervisor becomes an observer again. From Fig.
6 we see that our adaptive fuzzy control scheme operated in VII. CONCLUSIONS
exactly the way described above.
In this paper, we developed a direct adaptive fuzzy con-
Example 2: In this example, we consider the Duffing forced
troller which 1) does not require an accurate mathematical
oscillation system:
model of the system under control, 2) is capable of incor-
XI = 2 2 (53) porating fuzzy control rules directly into the controllers, and
j.2 = -0.122 - 2: + 12cos(t) +
u(t). (54) 3) guarantees the global stability of the resulting closed-loop
system in the sense that all signals involved are uniformly
Without control, i.e., u ( t ) 0, the system is chaotic. The bounded. We provide the specific formula of the bounds so that
trajectory of the system with ~ ( t )0 is shown in the (21, z2)
controller designers can determine the bounds based on their
phase plane in Fig. 7 for initial condition zl(0) = z2(0) = 2 requirements. We used the direct adaptive fuzzy controller to
and time period t o = 0 to t f = 60. We now use our direct regulate an unstable system to the origin and to control the
adaptive fuzzy controller to control the system state z1 to track Duffing chaotic system to track a trajectory. The simulation
the reference trajectory ym ( t ) = sin(t ) .In the phase plane, this results show that 1) the direct adaptive fuzzy controller could
reference trajectory is the unit circle: y i +&
= 1. We choose perform successful control without using any fuzzy control
k1 = 2, kz = l , ? = 2, Q = diag(l0,lo), Me = 30, b~ = 1,
rules and 2) after incorporating some fuzzy control rules into
+
and f U = 12 Izl 13. We use the six fuzzy sets shown in Fig. the controllers, the adaptation speed became much faster. We
3 for z1and x 2 , and assume that there are no fuzzy control also showed explicitly how the supervisory control forced the
rules. We simulated two cases: M , = 10 and M, = 3. We state to be within the constraint set and how the adaptive fuzzy
directly integrated the differential equations of the closed-loop controller learned to regain control.
system and the adaptive law with step size 0.02.
The closed-loop trajectories for the M , = 10 and M , = 3 APPENDIX
cases are shown in Figs. 8 and 9, respectively, where the PROOF OF THEOREM 1
initial condition zl(0) = z2(0) = 2 and the trajectories were i) To prove 5 Me, let VO =
from t o = 0 to t f = 60. We see that in both cases our is true, we have either le1 < Me or
If the first line of (45) $eT!.
=~ g ~ p ~ @ I ~ <0 ( z ) ve
adaptive fuzzy controller could control the system to track when I@(= Me; i.e., we always have 5. Me. If the second le(
the reference trajectory. In Fig. 8 the trajectory never hit the line of (45) is true, we have
boundary xf+zg = 10 so that the supervisory control U , never
= Me and Ve = rgTp,,eT<(g) le1
fired. Fig. 9 shows how the supervisory control U, forces the -ye E,
T IslZBTg..)= 0; i.e., (@I 5 M . Therefore, we always
lglz
trajectory to be within the constraint set zf zg 5 3. + have l@(t)(5 Me,b? 2 0.
WANG: STABLE ADAPTIVE FUZZY CONTROL 155

In Section IV we proved that V, 5

V ; therefore, [3] D. Dubois and H. Prade, Fuzz! Sets und Sysretns: Theo? und Applira-
r i m s . Orlando, FL: Academic Press, 1980.
$ ~ ~ ~ 5~$ gl T gp g 1I ~V , i.e., lgl
5 Since g = 141 G. C. Goodwin and D. Q. Mayne, “A parameter estimation perspective
we have 141 5 Ig,l + le1 I lgmLl
2, -z, + (E)’/’,
which of continuous time model reference adaptive control,” Automatica, vol.
23, pp. 57-70, 1987.
is (47). Finally, we prove (48). Since u,(glO) is a weighted [SI A. Isidori, Nonlinear Control S\..sterns. Berlin: Springer-Verlag, 1989.
e,
average of the elements of we have Iu,(gle)lI 5 MO. [6] J. Kiszka, M. Gupta, and P. Nikiforuk, “Energetistic stability of fuzzy
dynamic systems,” IEEE Truns. Syst., Man, Cybern., vol. 15, no. 5 , pp.
Therefore, from (23) and (30) we have (48).
783-792, 1985.
ii) From (38), ( 4 9 , and (46), we have [7] G. Langari and M. Tomizuka, “Stability of furzy linguistic control
systems,” in Proc. IEEE Conf: Decision and Control (Hawaii), 1990,
pp. 2185-2190.
’
181 C. C. Lee, “Fuzzy logic in control systems: Fuzzy logic controller, parts
I and 11,” IEEE Trims. Syst.. Mtrn. Cyhern., vol. 20, no. 2, pp. 404-435,
where 1 1 = O( 1) if the first (second) line of (45) is true. Now 1990.
[9] D. G. Luenberger, Lineur und Nonlitiecrr Progrrrmmirig. Reading, MA:
we show that the last term of (AI) is nonpositive. If 11= 0, the Addison-Wesley, 1984.
conclusion is trivial. Let 1 1 = 1, which means thatlei = MO 101 W. T. Miller, R. S. Sutton, and P. J. Werbos, Eds., Neural Nemorksfbr
Control. Cambridge, MA: MIT Press, 1990.
and gTp,lflTE(.c) > 0; we have 4Tf = (@* - = I I ] K. S. Narendra and A. M. Annaswamy, Srcihlr Adcrptii<e Systems.
$[18*12 - 1812 - le - @ * 1 2 ] I 0 , since = M O 2 [@*I. Englewood Cliffs, NJ: Prentice-Hall, 1989.
121 K. S. Narendra and K. Parthasarathy, “Identification and control of dy-
Therefore, the last term of (Al) i5 nonpositive, and we have namical systems using neural networks,” IEEE Truns. Neural Nehrorks,
1 vol. I , no. I , pp. 4-27, 1990.
V I --eT@ - gTPb,u, - cTPh,w. (A2) 131 M. M. Polycarpou and P. A. loannou, “Identification and control
2- nonlinear systems using neural network models: Design and \tability
From (30) we have grP4,u, 2 0; therefore, (A2) can be analysis,” USC EE-Report 91-09-01, 1991.
further simplified to 141 R. M. Sanner and J. E. Slotinc, “Gaussian networks for direct adaptive
control,” in /‘roc. Americun Control Cor$. 1991, pp. 2153-2159.
1 ( 151 S. Sastry and M. Bodson, Adcrptivr Control; Stahiliry, Convcrgence. trnd
V I - - e T Q g - gTPbcw Robu,srness. Englewood Cliff\, NJ: Prentice-Hall, 1989.
2-
[ 161 R. M. Tong, “Some properties of furry feedback systems,” IEEE Trtms.
1
I- 2
XQniin - 1
lgl2 - - [lei’
2 -
+
2cTPb,w IPb,(u1*] + Syst., Mcrn, Cyhern., vol. SMC-IO, no. 6, pp. 327-330, 1980.
(171 J . E. Slotine and W. Li, Applied Nodineirr Control. Englewood Cliffs,
1 NJ: Prentice-Hall, 1991.
+ ,/Pbrw12
L
(181 A. L. Schwartr, “Comment\ o n ‘Fuzzy logic for control of roll and
moment for a flexible wing aircraft‘,” IEEE Conrro/ S y s t e m Mtrgucine,
Feb. 1992.
I 191 L. X. Wang, ”Fuzzy system\ are uni\cr\al approximatorc,” in Proc.
IEEE Int. Cor$ Fuzzy S w t e m (San Diego), 1992, pp, 1163-1170.
where X Q is ~the minimum
~ ~ ~eigenvalue of Q.Integrating 1201 L. X. Wang, “Stable indirect adaptive fuzzy control, with application to
both sides of (A3) and assuming that XvlnLn > 1 (since Q is inverted pendulum tracking” wbmittcd to IEEE Trclns. Fuzz! S w . .
(211 L. X. Wang and J . M. Mendel, ”Generating fuzzy rules by learning from
determined by the designer, we can choose such a Q ) , we have examples,” IEEE Trans. S ~ s t . Mtrn.
. C,hni., vol. 22, no. 6, 1992.
‘l (221 L. X. Wang and J. M. Mendcl, “Back-propagation f u / ~ ysystems as
nonlinear dynamic system identifiers,” in /‘roc. IEEE Inr. Cor$ FULL!
S~sternv(San Diego), 1992, pp. 1409-1418.
[23] L. X. Wang and J. M. Mendel, “Fuzzy basis function\, universal
approximation, and orthogonal least squares learning,” IEEE Trcrns.
Nercriil Nehrorks, vol. 3, pp. 807-814. 1992.
(A41 (241 L. X. Wang and J. M. Mendel. “Three-dimensional structured networks
Define U = &[lV(O)l + Iv(t)l]and (. =
for matrix equation solving,” IEEE Trms, Comjxct., vol. 40, no. 12, pp.
1337-1346, 1991.
AQ,!,%C1
IP&,1*. Equation (A4) becomes (49) (note that 12.51 L. A. Zadeh, “Outline of a new approach to lhe analysis of complex
systems and decision processe\.“ IEEE Trtiris. Svst.. M m , Cyhern., vol.
supt>,, IV(t)l is finite because and q5 are all bounded). SMC-3. no. I , pp. 3 - 4 4 . 1971.
iiijIf iii E L 2 , then from (49) wehave p E L2. Because
we have proved that all the variables on the right-hand side of
(34) are bounded, we have E L,. Using Barbalat’s lemma
[15] (if g E L2 n L , and- E L,, then liIiitix Ig(t)l = 0 ) , Li-Xin Wang (S’90-M’93) received the B S and
M.S. degreesfrom Northwestern Polytechnical Uni-
we have liiiit-% lg(t)l = 0. Q.E.D. versity, Xian, People’s Republic of China, in 1984
and 1987, respectively, and the Ph.D. degree from
ACKNOWLEDGMENT the University of Southern California, Los Angeles,
in 1992, all in electrical engineering.
The author would like to thank Prof. J. M. Mendel and From 1987 to 1989 he was with the Department of
Prof. L. A. Zadeh for their encouragement and the reviewers Computer Science and Engineering, Northwestem
Polytechnical University, Xian, People’s Republic
for their construcJive comments. of China. From the fall of 1989 to the spring
of 1992 he was a research/teachine assistant with
REFERENCES the Department of Electrical Engineering-Syqtems, Univcrsity of Southem
Califomia, where he was working toward the Ph.D. degree. In the summer
[ I ] R. E. Bellman and L. A. Zadeh, “Local and fuzzy logics,” in Modern of 1992 he became a postdoctoral fellow in the Department of Electrical
Uses of Multiple-Valued Logic, J. M. Dunn and G. Epstein, Eds. Engineering and Computer Science\. University of California at Berkeley,
Dordrecht, Netherlands: Reidel Publ., 1977, pp. 103-165. where he is working with Prof. Lot6 A. Zadeh. His research interests include
(21 S. Chiu, S. Chand, D. Moore, and A. Chaudhary, “Fuzzy logic for fuzzy systems, intelligent control and signal processing, and neural computing.
control of roll and moment for a flexible wing aircraft,” I € € € Control Dr. Wang received a Phi Kappa Phi Student Recognition Award in 1992
Sysrerns Magazine, vol. 1 I. no. 4, pp. 42-48, 1991. for his work on fuzzy systems.