Sie sind auf Seite 1von 10

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

php

SlAM J. CONTROL AND OPTIMIZATION


Vol. 21, No. 2, March 1983

1983 Society for Industrial and Applied Mathematics


0363-0129/83/2102-0006 $01.25/0

A NEW CLASS OF STABILIZING CONTROLLERS FOR UNCERTAIN


DYNAMICAL SYSTEMS*
B. R. BARMISH?, M. CORLESS

AND

G. LEITMANN

Abstract. This paper is concerned with the problem of designing a stabilizing controller for a class of
uncertain dynamical systems. The vector of uncertain parameters q(.) is time-varying, and its values q(t)
lie within a prespecified bounding set Q in R p. Furthermore, no statistical description of q(.) is assumed,
and the controller is shown to render the closed loop system "practically stable" in a so-called guaranteed
sense; that is, the desired stability properties are assured no matter what admissible uncertainty q(.) is
realized. Within the perspective of previous research in this area, this paper contains one salient feature:
the class of stabilizing controllers which we characterize is shown to include linear controllers when the
nominal system happens to be linear and time-invariant. In contrast, in much of the previous literature
(see, for example, [1], [2], [7], and [9]), a linear system is stabilized via nonlinear control. Another feature
of this paper is the fact that the methods of analysis and design do not rely on transforming the system
into a more convenient canonical form; e.g., see [3]. It is also interesting to note that a linear stabilizing
controller can sometimes be constructed even when the system dynamics are nonlinear. This is illustrated
with an example.

Key words, stability, uncertain dynamical systems, guaranteed performance

1. Introduction. During recent years, a number of papers have appeared which


deal with the design of stabilizing controllers for uncertain dynamical systems; e.g.,
see [1]-[7]. In these papers the uncertain quantities are described only in terms of
bounds on their possible sizes; that is, no statistical description is assumed. Within
this framework, the objective is to find a class of controllers which guarantee "stable"
operation for all possible variations of the uncertain quantities.
Roughly speaking, the results to date fall into two categories. There are those
results which might appropriately be termed structural in nature; e.g., see [1]-[3], [6].
By this we mean that the uncertainty cannot enter arbitrarily into the state equations;
certain preconditions must be met regarding the locations of the uncertainty within
the system description. Such conditions are often referred to as matching assumptions.
We note that in this situation uncertainties can be tolerated with an arbitrarily large
prescribed bound. A second body of results might appropriately be termed nonstrucrural in nature; e.g., see [4] and [5]. Instead of imposing matching assumptions on
the system, these authors permit more general uncertainties at the expense of
"sufficient smallness" assumptions on the allowable sizes of the uncertainties.
This work falls within the class of structural results mentioned above. Our
motivation comes from a simple observation. Namely, given a theory which yields
stabilizing controllers for a class of uncertain nonlinear systems, it is often desirable
for this theory to have the following property: upon specializing the "recipe" for
controller construction from nonlinear to linear systems, one of the possible stabilizing
control laws should be linear in form. It is of importance to note that existing results
do not have this property. Upon specialization to the linear case, one typically obtains
controllers of the discontinuous "bang-bang" variety; e.g., see [1] and [2]. One can
often approximate these controllers using a so-called saturation nonlinearity; e.g., see
* Received by the editors March 4, 1981, and in revised form January 15, 1982.
Department of Electrical Engineering, University of Rochester, Rochester, New York 14627. The
work of this author was supported by the U.S. Department of Energy under contract no. ET-78-S-01-3390.
Department of Mechanical Engineering, University of California at Berkeley, Berkeley, California
94720. The work of these authors was supported by the National Science Foundation under grant
ENG 78-13931.

"

246

STABILIZING CONTROLLERS FOR UNCERTAIN DYNAMICAL SYSTEMS

247

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

[7]. Such an approach leads to uniform ultimate boundedness of the state to an


arbitrarily small neighborhood of the origin; this type of behavior might be termed
practical stability.
Our desire in this paper is to develop a controller which is linear when the system
dynamics are linear. By taking known results (such as in [3]) which were developed
exclusively for linear systems, one encounters a fundamental difficulty when attempting
to generalize 2 to a class of nonlinear systems; namely, it is no longer possible to
transform the system dynamics to a more convenient canonical form. The subsequent
analysis is free of such transformations.

2. Systems, assumptions and the concept of practical stability. We consider an


uncertain dynamical system described by the state equation

2(t) =f(x(t), t)+ Af(x (t), q(t), t)


+[B(x(t), t)+ AB(x(t), q(t), t)]u(t),
where x(t)R is the state, u(t)R is the control, q(t)R p is the uncertainty and
f(x, t), Af(x, q, t), B (x, t) and AB (x, q, t) are matrices of appropriate dimensions which
depend on the structure of the system. Furthermore, it is assumed that the uncertainty,
q(.):R-->R p, is Lebesgue measurable and its values q(t) lie within a prespecified
bounding set Q cR for all R. We denote this by writing q(.)M(Q).
As mentioned in the introduction, given that "stabilization" is the goal, we must
impose additional conditions on the manner in which q(t) enters structurally into the
state equations. We refer to such conditions as matching assumptions.
Assumption 1. There are mappings

h(.):R"xRPxRR" and E():R"xRPxR-R


such that

Af(x, q, t) B (x, t)h (x, q, t),


AB(x, q, t)= B(x, t)E(x, q, t),
liE (x, q, t)ll < 1
R.

for all x R q O and


We note that this assumption can sometimes be weakened. For example, in [9]
a certain measure of mis-match is introduced and results are obtained under the
proviso that this measure does not exceed a certain critical level termed the mis-match

threshold.
Our second assumption reflects the fact that the uncertainties must be bounded
in order to permit one to guarantee stability.
Assumption 2. The set O c R" is compact.
Our next assumption is introduced to guarantee the existence of solutions of the
state equations.
Assumption 3. The mappings f(. ): R x R
E(.) (see Assumption 1) are continuous. 3

R n, B (): R x R

R mxn, h (.) and

This notion is not to be interpreted in the sense of Lasalle and Lefschetz [12] but as defined
subsequently.
That is, one begins with a linear control law for a linear system and generalizes the controller in
such a way that is applies to a class of nonlinear systems.
In fact, one can modify the analysis to follow so as to allow mappings which are Carath6odory and
satisfy certain integrability conditions. See, for example, Corless and Leitmann [7]. All the results of this
paper still hold under this weakening of hypotheses.

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

248

B. R. BARMISH, M. CORLESS AND G. LEITMANN

In order to satisfy our final assumption, one may need to "precompensate" the
so-called nominal system, that is, the system with Af(x, q, t)=--O and AB(x, q, t)----0;
e.g., see [2]. Thus, prior to controlling the effects of the uncertainty, it may be necessary
to employ a portion of the control to obtain an uncontrolled nominal system
(UC)

(t)=f(x(t),t)

that has certain stability properties embodied in the next assumption.


R and, moreover, there exist a C function
Assumption 4. f(0, t)- 0 for all
V(.):R xR [0, o0) and strictly increasing continuous functions y(.), y2(),
y3(): [0, 00) [0, 00) satisfying4 yl(0) y2(0)--y3(0) 0 and limr_. yl(r)
limr_. TE(r) lim_.oo ya(r) o, such that for all (x, t) e R" R,

(2.2)

v (llxll)

Moreover, defining the Lyapunov derivative o(" ): R

R by

a 0 V(x, t)
o(X, t) =+VxV(x,t)f(x,t),

(2.3)
where

"R

at
denotes the transpose of the gradient operation, we also require that

t) -<_- -w(llx II)


for all pairs (x, t)eR R. This assumption, in effect, asserts the existence of a
Lyapunov function for the uncontrolled nominal system (UC). Consequently, the
origin, x 0, is a uniformly asymptotically stable equilibrium point for the uncontrolled
nominal system (UC).
The stability concept employed in this paper differs slightly from the traditional
Lyapunov-type stability. To motivate this change of definition, consider the following
very simple example of a system satisfying (2.1) and the associated assumptions:
(t) x(t)+q(t)+ u(t), with initial condition X(to) 1 and uncertainty q(. such that
[q (t)] -<_ 1. Furthermore, suppose the control is a linear feedback of the form u (t) kx (t),
with k <-1. Then, if a state x(t) <-1/(1+ k) is reached, an admissible uncertainty
q(t) =-1 results in motion of the state away from zero. Hence, although we cannot
guarantee uniform asymptotic stability (using a finite gain), we can nevertheless drive
the state to an arbitrarily small neighborhood of the origin. 5 The following uniform
ultimate boundedness-type definition captures this notion.
DEFINITION 1. The uncertain dynamical system (2.1) is said to be practically
stabilizable if, given any _d >0, there is a control law pg(.):R x R R for which,
given any admissible uncertainty q(.)eM(Q), any initial time to eR and any initial
state x0 R the following conditions hold
(i) The closed loop system

(2.4)

(t) f(x(t), t) + Af(x(t), q(t), t)


+[B(x(t), t)+ AB(x(t), q(t), t)]pa_(x(t), t)

possesses a solution x(. ): [to, tx] R n, X(to)


4

Xo.

The limit condition on 33( can in fact be removed at the expense of a somewhat more technical
development; e.g., see [7].
This is not to be confused with Lyapunov stability, because the required gain k depends on the size
of the neighborhood to which we wish to drive the state.

STABILIZING CONTROLLERS FOR UNCERTAIN DYNAMICAL SYSTEMS

249

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(ii) Given any r>0 and any solution x(.):[to, tl]-->R n, X(to)=Xo, of (2.4) with

Ilxoll <--r, there is a constant d(r)> 0 such that


IIx (t)ll <-- d(r) for all

[to, tl].

(iii) Every solution x(. ): [to, t]--> R can be continued over [to, oo).
(iv) Given any d >_- _d, any r > 0 and any solution x (.)" [to, oo) --> R ", x (to) Xo, of
(2.4) with Ilxoll--<r, there exists a finite time T(a, r)< oo, possibly dependent on r but
not on to, such that Ilx (t)ll-<- a7 for all >_- to + T(a, r).
(v) Given any d ->_d and any solution x(): [to, oo)-->R", X(to) =Xo, of (2.4), there
is a constant 6 (a)> 0 such that Ilxoll-<- (a7) implies that

IIx (t)ll--< a7

for all

>= to.

3. Controller construction. We take _d > 0 as given and proceed to construct a


control law p _e(" which will later be shown to satisfy conditions (i)-(v) in the definition
of practical stabilizability.
Construction of pa_(). The first step is to select functions AI(.) and A2(.)R"
R --> R satisfying

(3.1)

A(x, t)-->_max Ilh(x, q, t)ll,

(3.2)

1 > A).(x, t) >-max liE(x, q, t)l[.


qQ

The standing Assumptions 1-4 assure that there is a A2(x, t) such that
1) A2(x, t)<l can be satisfied for all (x, t)eR"R;
2) A(. and A2(.) can be chosen to be continuous; e.g., see [10, p. 116].
Now, one simply selects any continuous function y(. ):R" R --> [0, oo) satisfying

(3.3)

"g(x,

where

a(x,t)

t)>-4[1_ AE(x t)][C2-Cl.o(X, t)]

C and C2 are any (designer chosen) nonnegative constants such that

a) C1< 1;
b) either C1 # 0 or C2 # 0;
(3.4)
c) C2 # 0 whenever limx_,0 [A(x, t)/o(X, t)] does not exist;
d)

C (
<
v v)(_d).
1
-C

Note that these conditions can indeed be satisfied because of continuity of the 3(.)
and the fact that lim_,o y(r)= 0 for
1, 2, 3.
This construction then enables one to let

(3.5)

Pa(X, t)a---y(x, t)B(x, t)VV(x, t).

Remark. In fact, (3.3) and (3.5) describe a class of controllers yielding practical
stability. It will be shown in 5 that this class includes linear controllers when the
nominal system happens to be linear and time-invariant.
4. Main result and stability estimates. The theorem below and its proof differ
from existing results (see [1], [2] and [6]) in one fundamental way: The control p_a()
which leads to the satisfaction of the conditions for practical stabilizability degenerates

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

250

B. R. BARMISH, M. CORLESS AND G. LEITMANN

into a linear controller whenever the nominal system, obtained by setting


Af(x(t), q(t), t)=--O and AB(x(t), q(t), t)=-O in (2.1), is linear and time-invariant. This
will be demonstrated in the sequel. In fact, even for certain nonlinear nominal systems,
the controller turns out to be linear. This phenomenon will be illustrated with an
example of a nonlinear pendulum. Central to the proof of the theorem below is one
fundamental concept" a system satisfying Assumptions 1-4 admits a control such that
the Lyapunov function for the nominal system (UC) is also a Lyapunov function for
the uncertain system (2.1).

THEOREM 1. Subfect to Assumptions 1-4, the uncertain dynamical system (2.1)


is practically stabilizable.
Proof. For a given _d>0 and a given uncertainty q(.)M(Q), the Lyapunov
derivative (.)Rn R R for the closed loop system obtained with the feedback
control (3.5) is given by

(x, t)a_ o(X, t)+ VV(x, t){Af(x, q(t), t)


+[B (x, t)+ aB (x, q(t), t)]pa_(x, t)}.

(4.1)

By using the matching assumptions in conjunction with (3.5), (4.1) becomes


q(x, t)= f0(x, t)-3/(x, t)llB(x, t)VxV(x,

t)ll

+ VV(x, t)B(x, t)[h(x, q(t), t)


-y(x, t)E(x, q(t), t)B(x, t)VxV(x, t)].
Letting

(.)" R

R --> R be given by
&(x, t) a---B(x, t)VxV(x, t),

and recalling the definition of AI(. and Az(. ), a straightforward computation yields

(x, t)_-< 0(x, t)-[1-A2(x, t)]y(x, t)ll (x, t)l[ =


+ +/-(x, t)ll (x, t)[I.
Now there are two cases to consider.
Case 1. The pair (x, t) is such that Al(x, t) 0. It then follows from the preceding
inequality that

(x, t) <=o(X, t).


Case 2. The pair (x, t) is such that A(x, t) 0. Then it follows from (3.3) that
,(x, t) > 0. Moreover, in view of (3.3) and the conditions on the Ci,
o(x, t)_<-.0(x, t)-[1-A2(x, t)]y(x, t)ll (x, t)ll2 + A(x, t)ll (x,
Lo(X, t) +

a(x,t)
4V(x,t)(1-A2(x,t))

A(x, t)
+/-(x, t)ll (x, t)ll-2(1_
t)[
a=(x
t))3,(x, t)
-x

(1-A(x, t))y(x,

A(x, t)
- <- o(X, t) + 4y(x, t)(1- Az(x, t))
<-- (1-- C1)o(X, t) + C2.

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

STABILIZING CONTROLLERS FOR UNCERTAIN DYNAMICAL SYSTEMS

251

Combining Cases 1 and 2, and noting that C1 < 1, we conclude (as a consequence of
Assumption 4) that
(x, t) <-( -C)o(X, t)+ C <-- ( -C,)(llx II)+ C
(4.2)
for all (x, t) e R" x R. Having this inequality available now enables one to apply directly
the results of [7]. That is, the closed loop system (2.4) possesses a solution x(. ): [to, t]-->
R ", x (to) Xo, which is required by condition (i) in the definition of practical stabilizability. Moreover, in accordance with [7], if Ilx0ll--< r, one can satisfy the uniform boundedness requirement (ii) by selecting
ifr<-R,
(,y-I y2)(r) if r > R,
where

d(r)&{(y-ly2)(R)

R V (C2/1
It now follows that there is no finite escape time so that the solution is continuable
over [to, m) and hence condition (iii) holds. Again for aT>_-_d, using the estimates
provided in [7], one can define
0

T(d-, r) A

(4.3)

if r _-< (yl

y2(r) Z(]/_l
(1 C1)(y3

yl)(a),

yl "y1)(d)

"- "

"1)= C2

otherwise,

and in accordance with [7], the desired uniform ultimate boundedness condition (iv)
holds with the proviso that

(1 C1)(y3 yl yl)(a)- C2 >0.

(4.4)

Note that this requirement is implied by the satisfaction of condition (d) of (3.4) which
entered into the construction of the controller.
Finally, to complete the proof, it remains to establish the desired uniform stability
property. Indeed, let d->_d be specified and notice that if 6(d)= R, the following
property will hold" Given any solution x()[t0, )R", X(to)=Xo of (2.4) with
tlx011=<3(d), it follows (in view of the uniform boundedness property (ii) and the
requirements on the Ci) that [[x(t)ll<=d(R)<= for all t->t0. ]

5. Specialization to linear systems. The objective of this section is to show that


the flexibility permitted in choosing y(x, t) (see (3.3)) can be exploited in a "nice
way" when the nominal system dynamics happen to be linear and time-invariant; that
is, the control P4() in (3.5) can be selected to be a linear time-invariant feedback
of the state. We consider the special case when

(5.)

(t) [A + AA(q(t))]x(t)+[B + AB(q(t))]u(t)+ w(q(t)),

X(to)-Xo,
where A, AA (q (t)), B and AB (q (t)) are matrices of appropriate dimensions and w (q (t))
is an n-dimensional vector. In light of Assumptions 1 and 3 given in 2, it follows that
for all q e Q

AA(q)=BD(q),

(5.2)

AB (q

BE (q ),

w(q)=Bv(q),

E(q)ll< 1

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

252

B. R. BARMISH, M. CORLESS AND G. LEITMANN

where D (.), E (.) and v (.) have appropriate dimensions and depend continuously on
their arguments. In accordance with Assumption 4, the matrix A must be asymptotically stable. To obtain a Lyapunov function for the uncontrolled nominal system, we
select simply an n x n positive-definite symmetric matrix H and solve the equation

AP + PA -H
(5.3)
for P which is positive-definite; see [11]. Then we have
V(x, t) xPx
(5.4)
and

(5.5)
0(x, t)=-xHx.
It is clear from (5.4) and (5.5) that one can take the bounding functions yi() to be
yl(r) h min[e]r 2, "y2 (r) -& h max[.P]r 2, "y3 (r) h min[H]r 2,
(5.6)
where h max(min)[ denotes the operation of taking the largest (smallest) eigenvalue.
Construction of the controller. We take _d >0 as prescribed and construct the
controller p_d(" given in 3. Using the notation above, we define first 6
po a--maxl[O(q)[I, pz a-maxl[E(q)ll<l p a=maxllv(q)[I.
(5.7)
qO
qO
qO
Then, in agreement with (3.1) and (3.2), we may take

(5.8)

Al(x, t)

0ollxll + 0,

A=(x, t) pz.

Using these choices in (3.3) and the fact that o(X, t)<-- Amin[H]llx[[ 2, one can select
3() such that
(po +
>
y(x, t)_(5.9)
2

IIx

4(1 pz)(Chmin[H]llxll + C)
with the constants Cx and C2 yet to be specified. We shall examine three possible
cases and see that in all instances one can take y(x, t) constant. Of course, this
implies that the control p_a(" is a linear time-invariant feedback; that is,
(5.10)
pa_ (x, t)=-2yoBPx,
where 30 is the constant value of 3(" ), which will be specified.
Case 1. po > O, p O. In this case, we may select C2- 0 and
sequently, we can satisfy (5.9) by choosing

C (0, 1). Con-

y(x, t)=-- yo >-

(5.11)
Case 2. po
(5.12)

Po

4(1 -pE)Chmin[H]"

O, p > O. Clearly, it suffices to take C -0 and


y(x, t)_= y0 >

02
4(1 p)C2

where C2 is required to satisfy condition d) of (3.4). Using the descriptions of the


yi(. given in (5.6), this amounts to restricting C2 by

(5.13)
with
6

X min[e]
C2 < Amin[H]_d
2
1 C1 Amax[P]

C1 0 in the above.
One can in fact use overestimates to and tSE for Po and pE as long as the inequality tSz < is satisfied.

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

STABILIZING CONTROLLERS FOR UNCERTAIN DYNAMICAL SYSTEMS

Case 3. Oo >0, p >0. Now, in order to satisfy (5.9), we select


satisfying (5.13) and
(por +p)2
y(x, t) =- yo >max
(5.14)
r0
4(1-pE)[Clhmin[n]r 2-t- C2]

253

Ca (0, 1), C2

Letting f(r) denote the bracketed quantity in (5.14) above, a straightforward but
lengthy differentiation yields
1

(5.15)

maxf(r)=
r_->0

020

4(1-Oz) Clhmin[H]

+P22}

Hence, any 3o equal to or exceeding this maximum value will be appropriate in (5.14).
6. Illustrative example. We consider now the simple pendulum which was
analyzed in [7]. However, here it will be shown that the desired practical stability can
actually be achieved via a linear control. This may seem somewhat surprising in light
of the fact that the nominal system dynamics are nonlinear. A pendulum of length
is subjected to a control moment u (.) (per unit mass). The point of support is subject
to an uncertain acceleration q(.), with [q(t)[ <_-l =- constant. Letting x denote the
angle between the pendulums arm and a vertical reference line, one obtains the state
equations

21(t)=x2(t),

(6.1)

22(t)=_a

sinx(t)+u(t)_q(t) cos x(t)

where a > 0 is a given constant. In order to satisfy the assumptions of 2 one must
assure a uniformly asymptotically stable equilibrium for (UC), the uncontrolled
nominal system. Hence, for a given _d > 0, we propose a controller of the form

(6.2)

u(t) =-bx(t)-CXE(t)+pa_(x(t), t),

where b and c are positive constants and p_a(.) will be specified later in accordance
with the results of 3. The linear portion of the controller (6.2) is used to obtain a
stable nominal system. Substitution of (6.2) into (6.1) now yields the state equation

(6.3)

2 (t)

f (x (t), t) + B [p (X (t), t) + h (x (t), q(t), t)],

where

(6.4)

f(x, t)
h(x, q, t)

-bx-cxE-a sinx
-q cos xl

A suitable Lyapunov function for the uncontrolled nominal system (with x- 0 as


equilibrium) is

(6.5)

V(x,t)=(b+gc 2\)x2

+cxxE+x +2a(1-cosx),

and, provided b is sufficiently large, the associated 3"(. are given by

3"(r)=Alr 2,
2r

/3(r) X3r

+ 4a

if r >

254

B. R. BARMISH, M. CORLESS AND G. LEITMANN

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

,5

(6.7)

P=[ b+c2 ]

(6.8)

min {/, c },

&b+amin
x

Following the procedure described in


p (.), we select first

( x) >0.
si

3 for the construction of the controller

l(X,t)=lcosxl,

(6.9)

,.

Xmax[P],

=(x, t) 0.

Inequality (3.3) can then be assured by requiring


2

e(x, t)

(6. 0)

COS X

4[

Given our desire for a linear feedback, one can select


choosing

C 0

and satisfy (6.10) by

(6.11)

(x, t)v0>

=4C"

To complete the design, C must be selected to satisfy condition d) of (3.4). The


analysis must account for two cases, depending on the size of the given radius > 0.
Case 1.
h+ 4a. The required conditions on C are

a>

(6.12)

0<C<(3oa

h3

o)(g)=(h_d

-4a).

Case 2. hlNh +4a. In this case, the constraints imposed by (6.12) are
C > 0 is chosen suNciently small so that

met if

(6.13)

2a(

h2C2
+ 1-cos

<d.

Having now selected C, the controller is specified by (3.5) in conjunction with (6.11);
that is,

p(x, ) -oB(x, t)%(x, t) -o(cx: + 2x),


with the proviso that ToN/4. It is interesting to note that there is an obvious
tradeoff between the required gain constant T0 and the given radius > 0. As the
radius decreases, C decreases, which in turn implies that T0 increases. In contrast,
(6.14)

the nonlinear saturation controller of [7] remains bounded by the bound of the
uncertainty, and the radius can be decreased by increasing the nonlinear gain; i.e.,
by approaching a discontinuous control.

eels. This paper addresses the so-called problem of practical stabilizability for a class of uncertain dynamical systems. In contrast to previous work on
problems of this sort, the main emphasis here is on the structure of the controller. It
is shown that by choosing the function T( in a special way, the resultant control law
can often be realized as a linear time-invariant feedback.
REFERENCES

[1] G. LEITMANN, Guaranteed asymptotic stability for some linear systems with bounded uncertainties, J.
Dynamic Systems, Meaurement and Control, 101 (1979), pp. 212-216.
[2] S. GUTMAN, Uncertain dynamical systems, a Lyapunov rain-max approach, IEEE Trans. Automat.
Control, AC-24 (1979), pp. 437-443; correction, 25 (1980), p. 613.

Downloaded 01/03/13 to 128.148.252.35. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

STABILIZING CONTROLLERS FOR UNCERTAIN DYNAMICAL SYSTEMS

255

[3] J. S. THORP AND B. R. BARMISH, On Guaranteed Stability of Uncertain Linear Systems via Linear
Control, J. Optim. Theory and Applications, in press.
[4] S. S. L. CHANG AND T. K. C. PENG, Adaptive Guaranteed Cost Control of Systems with Uncertain
Parameters, IEEE Transactions on Automatic Control, AC-17 (1972), pp. 474-483.
[5] A. VINKLER AND I. J. WOOD, Multistep Guaranteed Cost Control of Linear Systems with Uncertain
Parameters, J. Guidance and Control, 2, 6 (1980), pp. 449-456.
[6] P. MOLANDER, Stabilization of Uncertain Systems, Report LUTFD2/(TRFT-1020)/1-111/(1979),
Lund Institute of Technology, August 1979.
[7] M. CORLESS AND G. LEITMANN, Continuous State Feedback Guaranteeing Uniform Ultimate
Boundedness for Uncertain Dynamic Systems, IEEE Trans. Automat. Control, AC-26, 5 (1981),
pp. 1139-1144.
AND I. J. WOOD, A Comparison of Several Techniques for Designing Controllers of
Uncertain Dynamic Systems, Proceedings of the IEEE Conference on Decision and Control, San

[8] A. VINKLER

Diego CA, 1979.


[9] B. R. BARMISH AND G. LEITMANN, On Ultimate Boundedness Control of Uncertain Systems in the
Absence of Matching Conditions, IEEE Transactions on Automatic Control, AC-27 (1982), pp.

153-157.

[10] C. BERGE, Topological Spaces, Oliver and Boyd, London, (1963).


[11] N. N. KRASOVSKII, Stability of Motion, Stanford University Press, Stanford, CA, (1963).
12] J. P. LASALLE AND S. LEFSCHETZ, Stability by Liapunovs Direct Method with Applications, Academic
Press, New York, (1961).

Das könnte Ihnen auch gefallen