Anda di halaman 1dari 7

IBEB T R A N S A ~ O N SON AUTOMATIC c o m o ~VOL

, AC-25,NO. 3, JUNE 1980

433

Global Stability of Parameter-Adaptive


ControlSystems
A. STEPHEN MORSE, SENIOR MEMBER, IEEE

-- -

controller for which global stability could be proved for


any n* > 1. Unfortunately,
- . the complexity of this controller grows rapidlywithincreasing i*
significantly
, kte
ing
its
practical
usefulness
for n* greater
than say 4.
In 1978 an important stepforward was made when
t
h
n
e
P
r
o
v
w
n
g
g
l
o
b
a
l
s
y
s
t
e
m
s
t
a
b
i
l
i
t
y
.
Goodwin, Ramadge,and Chines [8] showedthat a relative
uncomplicated adaptive algorithm,. applicableto discreteITHIN THE past several years there have been
a time systems with n* > 1, would provide global stability.
number of importantresultscontributing
to the Somewhatmorecomplicatedalgorithms,forboth
conrapidly developing theory of parameter-adaptive control. tinuous and discretesystems,wereindependentlyproThe most significant of these apply to singleinput single- posed and studied by Egardt [9], but the stability analysis
output processes admitting linear discrete- or continuous- given in [9] is somewhat involved. An alternative stabiliztime models, with unknown parameters. In 1974 Monop- ing adaptive algorithm for discretetime systems has just
oli [l] proposed an adaptive algorithm for a continuous- been developed by Lin and Narendra [lo]. To our knowltime system of known relative degree, the relatiw degree edge this is the status of the (deterministic)adaptive
n* of alinearsystembeingthedifferencebetweenthe
stability problem at this time.
number of poles and number of zeros of the systems
In this paper we examine two relatively uncomplicated
transfer function. A somewhat similar algorithm was sugalgorithms, each intended to adaptively control a singlegested by h t r o m and Wittenmark [2] in 1973, for applicainput single-output contimum-time linear system with untion to discrete-time linear systemswith known delays;
known parameters. We show that each algorithm is capathe deZay of a discrete linear systemas defined in [2], is the
ble of providing global closed-loop stability while at the
same as the systems relative degree.
sametimecausingtheprocess
output to approach and
Monopolis algorithmincorporatescertainconcepts,
As with all of the
track a suitable defined reference signal.
most notably an augmented error and a special paraalgorithms mentioned above, it is assumed that only the
metric structure, which at present seem to be essential
process input and output can be measured, and the concomponents of anyadaptive control system for which
trollerisrequired to beadifferentiator-freedynamical
global stability can be assured. These concepts
are imsystem.
plicit in the ktrom-Wittenmark algorithm and in all the
The first of the two algorithms considered, differs from
algorithms which we discuss below, including those of the
thestabilizingalgorithm of [3]intwoways.First,the
present paper.
U used in [3] to
One of the long standing problems of adaptive control (complicated)auxiliaryfeedbacklaw
ensure stability, is not used here. Second, to generate the
theory has been to show that algorithms such as [l] and
augmented error e,, the present algorithm incorporates
[2] yield globally stable closed-loop systems. For the case
when n* = 1, the problem provesto be quite simple and a a quadratic feedback term +iQ+,Z, [cf. (15a), (15b)], not
used inthealgorithm of[3]. Apart from these two difnumber of stability proofs exist, e.g., [3], [4]. For n*> 1,
ferences,thealgorithm of[3], and thefirst of the two
the problem appears to be much more difficult-at prealgorithms discussed in the sequel,are exactly the same.
sent it is not known if Monopolisalgorithmprovides
In Section I we redevelop in state-spaceterms, the
stability [5], exceptfor thevery special case when only one
process
model parameterization upon
which the controller
process model parameter is to be estimated [6], [7J. In [l],
algorithms are based, and we give explicit descriptionsof
Monopoli also proposed for the case
n* = 2, an alternative
eachalgorithm. In Section I1 we derive for each cone.g., see
algorithm which is known to yield global stability,
troller, a related error system to be used for stability
[3]. In an effort to generalize this algorithm, Feuer and
analysis. The material in these two sections closely follows
Morse[3]wereleadin1977todevelop
an adaptive
thework of Feuer 161. In Section I11 we establishthe
stability propertiesoi the error systems, and then interpret
ManuscriptreceivedMarch9,1979;revisedNovember19,1979.
Paper
by J. L. Speyer, P a t Chairman of the Stocktic these properties in
Of Original
closed-loop adaptive
Control Committee. This work wassupported in part bytheU.S.
Air systems. The mathematical concepts needed to underForce Office of Scientific Research under Grant 77-3176.
m e author js
the Department of Engineering and Applied
stand the proof do not exceed those
taught in a
Science,
Yale
University,
New
Haven,
C
T 06520.
basic senior level
course on linear
systems.
AU-W
paper examks two reistively l~~xllllplieated
adaptipel~eontrdling a Stnglelopnt sinple-ontpnt process admitting a
~0-w
& unknown
k
Of
a elosed.~
to
approachandtrackasuitablydefhedreferenceslgeslwblleatlhesame
for

0018-9286/S0/~00-0433$00.75 0 1980 IEEE

434

IBBB TRANSACTIONS ON

Notation: In thesequel,primedenotestranspose.
If
M ( t ) and N ( t ) are two m X n matrices defined on [0, GO),
we write M = N (E) if each element of the matrixM - N is
alinearcombination of decayingexponentials; we also
write M = N ( p ) if each element of M - N is uniformly
bounded. The Euclidean norm of M ywritten 11 M 11, is the
squareroot of the sum of the squared-values of the
components of M. E2 isthespace of squareintegrable
scalar-valued functions a(t) on [O,oo).

I.

SYSTEM s

mum

AUTQbIATIC CONTROL, VOL AC-25, NO. 3, JUNB

1980

(l/g,)&, where $, is a row vector of implicitly defined


gains and @,(t) is a reference input. To generate such an

estimate, use is made of sensitivity function filters of the


form

d, =A,@, + bou

(5)

dy =A,@, + boy

where bo is as defined earlier. The algebraic structure of


the adaptive lawwhichwe will use is motivated by the
following observation.
Lemma 2:

We assume that the single-input single-output process


to becontrolled can be modeledbya canonical linear
system with strictly proper transfer function

f , x p ( t )= W

+ k;@y(t>

(E)

(6)

where k, = F'b-'bpgp, kz,, = F'B -'hp, B is the controliabiiity


matrix of (A,,bd), and F is the obseroability matrix of
(.&,Ad.
Proof: If E,=[B,,AoB,,...,A,"-'8,]B-'
and Ey=
where g, is a nonzero constant, and %(s) and P,(s) are [ e y , A o ~ y- ,--, A , " - ~ B , ] B - ~then
, from ( 5 ) E,=A,-,.E,+I~
monic, coprime polynomials.In addition, as in [3], and for and Ey=A,Ey+Iy. Thus, if x = Eubpgp+ Eyhp, then R =
the same reasons, we assume that %(s) is a strictly stable
bpgpu Q. From this and (2) It followsthat x,(t) polynomial, and we take as known, therelativedegree
x(r) = eAd(xp(0)- x(0)) and thus that x,(t) = x ( t ) (e).
n* =deg( p,) - deg(aJ, a bound n satisfying n > deg(
Clearly, f,x,(t)=f,x(t) ( E ) . Therefore,
and the sign of gp, which without loss of generality we
assume to be positive.
f,xp=[f,euyf,~oeuy~~~yf,~,"-~eu]~-l~~~~
Using just the assumptionthat n is known, it is possible
to describe the relationship between process input u and
+ [ ~ e y , f , ~ o e y , . . . , f , ~ ~(e).- l e ~
output y by ann-dimensional,observable,stabilizable
system of the form
This expression can be rewritten as &xp = 8:F'B -'b,g, +

YJ?/

ip=(A,+ hpco)x, + bpgpu


Y =Cox,

(2)

8;F'B -'hp (E), from


which
(6) follows.
The expression for &xp in (6)suggests that to estimate
$xp + (l/g,)@,, one should use a control law of the form

where (A,, b,co) is any n-dimensional canonical system,


preselected so that A , is strictly stable, and $ and bp are
vectors of unknown parameters. This is adirect consequence of the following lemma.
where L,(t), b(t),and i r ( r ) are vector-valued estimates of
Lemma I : For each monic, strictiy stable poiynomial rS, k,, I$, and I/g,, respectively.
of degree n - deg( p,), there exist vectors $ and bp such that
To complete the description of the adaptive controller,
we must 1) explain how& is implicitly defined, 2) indicate
co(~l-A~-~co)-'bpg,=T,(s)
( 3 4 how e, is generated, and 3) specify parameter adjustment
and G., Wenow addresseach of these
char. pol.(A, + $c0) = /3,j3".
(3b) laws for k,,

4,

issues.

Prooj Since (c,,A,-J is observable and sincedeg

At the present time, to implicitlydefine4 requires exact


knowledge of the relative degree n* defined earlier (cf. [3,
Since (c,, A,+ h,c,) is clearly observable,
Proposition ID. With n* known, choose &(s) to be any
n*th degree, monic, strictly stable polynomial. Since n* =
n-deg(Q,) (cf. Lemma
l), it follows that deg(QoflP,)=
n. This and the observation that ,Bo is the uncontrollable
where d(s) is a row-vector whose entries form a basis for polynomial of (A,+$c,b,gp) (cf. Lemma l), imply that
thevectorspace of polynomials of degreelessthan n; the equation
% B O is an element of this vector space since deg(Qd)<
(8)
char. p o l . ( A , + h , c , + b p g ~ ) = g P o S ,
deg( p,@,)= n. Thus, there exists a vector
b, such that
can be satisfied for some f , and wetake this to be the
d(s)b,= crp(s)po(s).From this, (1) and (4) it follows that
0 implicit definition of 4.
(3a)
With& and thus u so defined, we can use the fact that
The adaptive control signalto which u will be set equal
(6) to
is an estimate of a state-feedback law of the form $x, + (A,+ $co + bps$) is strictly stable together with

( Bpp0)= n, there exists a vector h, such that (3b) is true.

435

where

rewrite (2) as
\ I

~,=(Ao+~co+bpg~)x,+b,g,(ke+(l/gp)B,)

(e)

Y = cdcp

i=[i;,i;,i,I
and

(15b)

q1= P+,- CIZl

(154

i,=A,z,+b,Pe

(154

81= -&,e,

(154
(150

Cp;= c , H ,

(10)

k(t)=i(t)-k,.

(154

+ i141-CpiQCp1e1

61

(9)
where 8=[8:,8;,8,!l, k,=[Il,,k;, l/gpl,
k is the parameter error vector

e, = 0 , + e

k,=A,H,+b,B.
(15g)
To indicate how 8, is generated, we assume that y,-the
signalwhich y is ultimatelysupposed to track-is the Here e, is an augmented error (cf. [ID, C
p
l is a vector of
output of alinearsystemwithstrictlystabletransfer
filteredsensitivityfunctions, (c,,A,,b,) is a canonical
function T,(s)and uniformly bounded, piecewiseantinu- realization of l/&(s), and Q and 4 are any preselected
ous reference input r. Under the reasonable assumption
positive definite matrixand scalar, respectively. The adap
(cf. [3l)that the relative degreeof T, is no smaller than n*, tive controller defined by this adjustment law, with the
the transfer function T(s)=&(s)T,(s)is both proper and term Cp;QCp,e, deleted from 15(b),is essentially the same as
strictly stable. Thus,if 8, is now defined as theoutput of a theoneproposedearlierbyMonopoli
[l]; thepossible
linear system with input r and transfer function T(s),then inclusion of such a term has been suggested previously by
8,(t) is uniformly bounded and
Egardt [9]. Discrete-time parameter adjustment lawssimilar to (15) have been proposed by Ionescu and Monopoli
[1 I], Lin and Narendra [ 121, and others.
2) The secondparameteradjustmentlaw
to beexamined is described by the equation
where (A,b,c)is any canonical realization of (1/,8,).

In viewof the well-known fact that 4 is a state-feedback invariant, it is clear that if $, satisfies (S), then after
left-half plane pole-zero cancellations

This and (9) imply that the relationship between


(1/gp)8, and y can alternatively be writtenas
i p

=Azp + bgp(Ke+(l/gp)Or)

where

i =- @#,e2
-

(16)

k8+

(1

y =CZp.

From this and (11) it follows that the relationship between


and Q, 4, and X, are as before. The adaptive controller
k8 and the output tracking error
defined by this adjustment law is similar but not quite
4 t ) = ~ ( t-yr(t)
)
(12) analogous to the discrete-time adaptive controller studied
in [SI.Related adjustment laws have been studied in [9].
can be expressed as
Remark: Equations (15f) and (15g) associatedwith
ie=Axe+bg,K8
adjustment law l), and (17e) and (17f)associatedwith
adjustment law 2) define linear filters with input 8 and
e = cxe
outputs (PI and +2, respectively. By choosing the characterThere remains the problem of specifying a parameter isticpolynomials of A , and A to bedivisors of the
adjustment law for k^ which yields a globally stable sys- characteristic polynomialof A,, and by making use of the
tem. Two alternative laws w
l
ibe considered.
fact that thefirst 2n components of 8 are themselves,
/?,
to be of theform
1) To speafy thefirst,take
outputs of linear filters (5) forced by u andy, it is possible
(s+X,)P,(s) where X,> 0 and PI is monk and strictly to construct filters (forced byu and y ) of lower dimension
stable. Then k^ is adjusted accordingto the rule
Cp,
than (15f),(15g) and (17e), (17f) which also generate
and Cp2, respectively [6].Although these modifications lead
i =- QCp,e,
(14) to somewhat simpler controllers requiring fewer integratorstorealize,they
do notseem tosimplifystability
analysis and, in fact, they actually appearto complicate it.
Ifdeg(&)=O, then n*= 1, a case already treated in the literature [3], For this reason, such modifications w
li not be pursued
by relatively straightfornard means. In the sequel we therefore assume
when discussing adjustment law 1) that n* > 1.
here.

436

lEEE TRANSACTIONS ON AuToMAnC CONTROL, VOL

11. ERRORSYSTEMS
Much of the mystery of parameter adjustment law1)
[4] that it is
disappears as soon as it isrealizedasin
possible to write
= k'+, - C , Z
where
i = A , z blk'8.

I),

AC-25,NO. 3, JUNE 1980

For future reference, we now summarize the pertinent


equations for either type of system using a single uniform
notation. By an error system with adjustment law 1) [respectively 2)], we mean a nonlinear dynamical system of
the form
i=Ax+b(k'O+r)
(224

(Zb)

e=Cx+rir+z(t)

a=A H + b8'

This and (13) imply that

(22c)

+'= CH
i=Az+bk'8
l/=k'+-cz

From this and the


definitions of e , and ul in (15) it follows
that
i~=-(Ao++~Q+,>e,+g~k'+l+rCIlg+e(t)
(18)
where g =6 - gp and E ( t ) = 0 ( E ) . In addition, sincekp and
gp are constant, (14) and (15e) can be rewritten in "error
form" as
(19)

k=-Q*

g = -&e
where for adjustment law
;
=
-(h,++'Q+)>+gpk'$+&+E(t),

(23)

while for adjustment law 2)

Using similar reasoning it is easy to see that parameter


adjustment law 2) leads to the alternative equations

k'cp,- cz

Here Z(t) and E(t) both equal 0 ( E ) , r is a-piecewise-con--tinuous,uniformlybounded


function, ( A , b ,C,d) is an
unknownbutstrictlystable,canonicallinearsystem,
( A ,b, c) is a known strictly stable canonical linear system
with transferfunction of theform l/P(s), p beinga
k= - Q+,e2
monic polynomial of degree m >O; q and &, are positive
constants, and Q is a positive definite matrix.
8= - &2%
Remark I : The fact that ( A ,byc ) canonically realizes
To complete the descriptionof the error system for both
1//3(s) implies that cA'-'b-O for i { l , - - - , r n - l } and
cases, we show that 8 can be expressed as the output of a
that CA"'- 'b= 1. This, in turn, together with (22c) implies
strictly stable h e a r system forced by k'8 and 8,. For this,
that cA'-'fi = cA'H, i E (1,2, *
, m - l} and that
we use (6) to rewrite (7) as
CA"'- = CA"'H+ 8'. Use will be made of these relations
1
in the sequel.
u = k ' 8 + f , x p + -8,
(E).
412=

i=Az+bk'8

--

la

gP

By substituting this expression into (5) and then using (9)


we obtain

e= Cz+le,
x'=ZZ+b(k'8+(1/gp)8,)

(21)

----

where ( A , bCy
y d) is a canonical realization of the strictly
stable linear system

lr 1
eu

111.

STABILITY ANALYSIS

Webegin ourstabilityanalysis byshowing that no


matter which adjustment lawis used, a solution to the
error equations must exist for t > O and certain components of the solution must be uniformly bounded.
Proposition 1: For any time to > 0, any prescribed state
(xo,Ho,zo,kwgo,Q [respectiuely (xo,Hoyzo,
ko,go)l and any
uniformb bounded, piecewise continuour input r(t), t > 0,
there exists a solution to (22), (23) [respectively (22), (24)J
on [0,03)which passes through the prescribed state at t = to.
Along such a solution e, k , and g (respectiuely k and g) are
bounded uniformly and I/ell and I l l ( / are in
Proot Since theright-handsides
of (22)-(24) are
continuousfunctions of theirarguments,thereexistsa
nonemptysubinterval [tl,tJ of [0, 03) containing to on
which thesolutionto (22),(23) [respectively (22),(24)]
with the prescribed state at t = to exists. If the derivatives
of the nonnegative time functions

e*.

with inputs (k'8+ l/gp8,) and 8,, and output 8.

437

MORSE: PARAMETER-ADAPTNE CONTROL SYSTEMS

and

are evaluatedalongthesolution
to (22),(23) and (22),
(24), respectively, there results for i = 1,2

Since iji(t)< 0,vi(t) is nonnegative, monotone nonincreasing and thus bounded by vi(tl). This together with the
definition of vi provesfor(22),(23),[respectively(22),
(24)] that Z, k , and g (respectively k , g ) are bounded on
[t,,tJ by a constant not dependingon t,.
Now observe that (22a)-(22e) can be viewed as a linear
system on [tl,tJ with uniformly bounded inputs r and Z
and uniformlyboundedcoefficients
k . From this it
follows that if t2< 03 the system's state ( x , H , z ) is
bounded on [tl,ta.Thus, either for (22), (23) or for (22),
(24),system state is bounded on [tl,t2) by aconstant
possibly depending on t , and t,. Since this is true on any
finite interval[t,,tJ containing to on which solution exist,
it followsfrom standard existencetheorems {e.g.,see
[13]}, that a solution existson [O, m), and that the uniform
boundedness of 2, k , g for (22), (23) or just k , g for (22),
(24) holds on [0, 00).
Next observe that since vi(t) is a nonnegative, monotone nonincreasing function, bounded by v,(O), the limit
ui(co) exists. Clearly,

bounded
input
O,(t), t 2 0, the
state
response
(xpy~u,8y,xryk^yz2y~2y
H2) of the adaptive control system defined by (2), (5), (7), (II), (It), (16),and (27) is un$ody
bounded and the tracking error e approaches zero as t+m.
-In- -the sequel,* denotes convolution product
and R(t) =
CeAtb;since A is stable, R(t - T) is an exponentially stable
weighting pattern. The proof of Proposition 2 depends on
four lemmas.
Lemma 3:
Hk=z+eA'*Hk

(28)

(E).

Proo$ From (22c)


d
-(Hk)=AHk+bKB+Hk
dt

so Hk= eA'b*k'8+ eA'*Hi

(E).

Since (22e) implies that

z = eAtb*k'8,it follows that (28)


true.
is

Lemma 4: For each constant row vector f ,


(fH)'=R*(fHk-feAf*H&)

(p).

(29)

Proof: From (22c)


(fH)'=
feAfb*8

(E).

In addition,since
is strictlystable and r and E are
uniformly bounded, we can use (224 and (22b) to obtain

8 = R*k'O

(p).

Thus, (fH)'=feAtb*R*PO (p), and since R and feA'b are


stationary,

O< -~m
I t.ji(t)dt=u(0)-u(m)<m.

(jH)'= R*feArb*JIO
From this and the structureof t.ji in (27) it followsthat IlZll
and
IICll are in @.Since
from
(22g),
11&112=
+'Q2+Z2< X+'Q+Z2 (A=maximumeigenvalue of Q), it
must be that ll&llEwell.as
0
With existence now established, we
turn to our main
technical result.
Proposition 2: For any initial state, the state response of
error system (22) together with either adjustment law (23) or
(24) is uniform& bounded on [0,co)and E and 4 approach
zero as t+W.
Proposition 2 implies for i E { 1,2}, that 8, k, g,and Hi
are uniformly bounded and that e, and +
b
approach
i
zero.
Since g = g + g p and k = k + k p , g and k areuniformly
bounded, Examination of (9), (15d), and (17c) shows that
xp,z,, and z2 are uniformly bounded as well.
For parameter adjustment law l), (15b) together with
the fact that e, and +blapproach zero, allows us to conclude that is uniformly bounded and approaches zero.
From this and (15a) it follows that e approaches zero as
well. For parameter adjustment law 2), (17a) and the fact
that ~,b~and e2 approach zero, imply that e approaches
zero. These observations can be summarized as follows.
Theorem: For eachinitial
state andeach
ungormly
bounded
input
O,(t), t > 0, the
state
response
( x p , 8 ~ , 8 y , x ~ , f f , ( r , , z l y of
~ l ,the
H I )adcqDtive controlsystem
defined by (2), (5), (7), (ll), (12), (14), and (15) is unifomz@
boundedandthe
tracking error e(t) approaches zero us
t+m, Similarly, for each initial state and each unijormly

!
2
'

(I,

(p).

(30)

Using (22e),fz- feAfb*k'O ( E ) ; hence, by Lemma 3


feAfb*KO=f H k -feA'*Hi

(e).

Substitutinginto(30)thusyieldsthedesiredresult.
0
Lemma 5: If w(t,7) is an exponential@ stable weighting
pattern and if p(t) is uniformly bounded, then there exist
constants c1> 0 and c2 > O such that for m y piecewise
continuous input v(t),

Prm$ Since p(t) is uniformly bounded and w(t,r) is


exponentially stable, there exist constants dl > 0, d2> 0,
and X > O such that p(t)9dl and Ilw(t,~)11<4e-h('-').
Hence,

IEEB TRANSACTIONS ON AUTOMATIC CONTROL, V O L AC-25, NO. 3, JUNE 1980

438

Setting c1=2d: and c2= 2G2/2/U yields the desired result.

By (22d) and Lemma 6, (35) is true for i = 1. Suppose the


claim is true at fixed i < m . Using Lemma 4, we can write

Lemma 6: With i? defined either by (23) or (24) there


exist nonnegative
constants
c1 and c2that
such

(cAiH)'=R*(cAiHk-cA'eA'*HI;)

(p).

From Lemma 3, (220, and (22d)

+= ceAf*HI;

(33)

(E).

We now claim that if Z is defined by either (23) or (24),


then

+= - R * - (1c H i + g ( c e A ' * H i ) )

(p).

(34)

gP

Now
R*cAi-'-(Hk)=SrR(t-T)cAi-'-(H(T)k(T))d-.
dt
d
d
0
rh

Integration by parts thus yields


R*cA i-'

If 5 is defined by (23), then using (22g),

(Hk)=R(O)cA'-'H(t)k(t)
dt
- R(t)cA'-'H(O)k(O)

1 (;+&Z-P#r&)
k'+= -

(E).

+l o ' ~ ~ e ~ ( ' - ' ) ~ c A i - ' H ( 7 ) k ( 7 ) d 7 .

gP

Hence, from (22d) and (33),


1

kt+= - ( ; + ~ ~ - c ~ I ; - g ( c e ~ ' * ~ (~E ) ). )


gP

Since R(r - 7 ) is exponentially stable, 2 is strictly stable,


and k(t) is uniformly bounded, for suitabIe nonnegative
constants dl, d2, d,, X with A > 0,

Since
by
Proposition
1, Z uniformly
is
bounded,
R*(l/gp(g+&a) is also. nus,if the preceding expression
for k'+ is substituted into (32) we obtain (34).
If Z is defined by (24), then using (22g) wecan write
1

k'+= -(&Z-P+-&)

(4.

gP

Hence, from (22d) and (33)

+ = p ( t ) + l t w0( t , T ) H ( T ) & ( T ) d T

where p ( t ) is uniformly bounded and w ( t , ~ is


) the weighting pattern
- W ( ~ , T ) = ~ ( ~1 - ~ ) - c + ~ r ~ 1( t - ~ ) - g ( ~ ) ~ ~where
~(uw
-~
( t ,)rd) is6 the
. exponentially stable weighting pattern

gP

gP

Since A is strictly stable, R(t - T) is exponentially stable,


and g is uniformly bounded,w(t,') is expnenmY
It follows from Lemma5 that(31) is true.
Proof of Proposition 2: We claim that there exist non-

i ~ { 1 , 2 , - - - , m ) . (35)

w ( t , ~ ) = - R ( t - ~ ) c A ~ - ' - R(t-a)cAieA("-')do

and 4
Hence,
is a

from (37) and

for somenonnegative constants c l , i + l , c z i + l .By induo


tion, (35) is thustrue.

439

MORSE: P A R A M E T W - A D r n CONTROL SYSTgMs

Since (c, A ) is an rn-dimensional observable pair, withN


the observability matrix of (c,A)

tivelyuncomplicateddiscrete-timeadaptive
control system is globally stable, then why should the same not be
true in continuous time? The question proved especially
~ l ~ l ~ = <~~ l~ ~~ I I --N H I 1I ~ 1<HI~ I ~N - ~ I I I ~ ~ A ~ - ~perplexing
H I I . at the outset, since examination of the proof
i=l
given in [8] quickly revealed that the key technical facts
upon which it was based, could not be used in continuous
From this and (35) it follows that for suitable nonnegative
time, and thus if the continuous casewere to be dealt with
constants dl and d2,
a fundamentally new approach would have to be developed. Such an approach has been presentedin this paper;
IIHIt)1I2<~'+ ~ ~ ~ I I ~ ( . r ) l l ~ l l ~ ( . r ) 1 1 2 ~ . r .
and the stability proofwe have given is 'ccanonical" in the
sense
that, with obvious modifications, it will apply to the
Thus, using the Bellman-Gronwall lemma
discrete case as well.
(IH(~)((~<~~~J.~P~II&(T)II~~.

Since
by
Proposition
1, llkllE@, H(t) is uniformly
bounded.
To show that x is bounded, first observe that by Remark 1
g f - d m - 1 k - d m H

so
9fk=cAm-1Hk-cAmHk
=cAm-'$(Hk)- cAm-lHk-cAmHk.

Since H and k are uniformly bounded and llill E2 and


since R(t) is the weighting pattern of astrictlyproper
exponentially stable system, R*(cA m- ' d / d t ( H k ) cAm-lHk -(cAmHk) must be uniformly bounded. Thus,
R+(k'B) is uniformly bounded. From this and (22a) (22b)
it follows that x and 9 are uniformly bounded. Examination of(22e)nowshows
that z must be uniformly
bounded as well.
To show that $+O, we use (22d)and (22f) together with
Lemma 3 to concludethat
ceAt*H i (E).Since Ilk11 EE2
and H is uniformly bounded, +(t)-.O as t+m.
If i is definedby(23), then is uniformlybounded.
Z(t)+O as t+m.
Since IISll is bounded and in @,
If i isdefinedby(24),then
it is clearlyuniformly
bounded.Differentiation of Z in (24) shows that is a
continuousfunction of uniformlyboundedfunctions.
Thus, since llZll P , Z((t)+O as t+m.
Remark: The key step in the proof of Proposition 2,
and thus main result of the paper, is the simple inductive
argument leading to (35). Once this inequality is established, the remainder of the proofproceedsalongstandard lines. We also make use in the second part of the
E e2;the proof of Proposition 1
proof of the fact that ~lill
shows that this is a direct consequence of the inclusion in
(23) [or (2411of the quadratic term +'@.

+=

CONCLUDING
REMARL(s

Weweremotivatedtorestudythecontinuous-time
adaptive
stability
problem
considered
in this paper,
largelybythework
of Goodwin, Ramadge, and Caines
181. The question generatedby their results wasthis: if it is
possible to prove by relatively simple means, that a rela-

REFERENCES

R V. Monopoli, "Model reference adaptive control with an augmented error signal" IEEE Trans. Automat. &Mr., vol. AG19,
pp. 474984, Oct. 1974.
K. J. Astrom and B. Wittenmark, "On self-tuning regulators,"
Automatics, VOI. 9, pp. 185-199, 1973.
A. Feuer and A. S. Morse, 'Adaptive control of single-input,
single-output linear systems," IEEE Trans. Auto-.
Contr, voL
AG23, pp. 557-569, Auk 1978.
K. S. Narendra and L. S. Valavani, "Stable adaptive controller
designdirect control," IEEE Trans. Automat. Contr, voL AG23,
pp. 570-582, Aug. 1978.
A. Feuer, B. R Barmish, and A. S. Mow, "An unstable dynamical system associated withmodelreference
adaptive control,"
IEEE Trans.Automat. CO&, vOL AC-23, pp. 499-500, June 1978.
A. Feuer, "Adaptive control of singleinput, single-output linear
systems," Ph.D. dissertation, Yale Univ.,New Haven, C T , May
1978.
A. Feuer and A. S. Morse, "Local stabiity of parameter-adaptive
control systems," Prqrints, 1978 Johns Hophiins Conf. Inform. Sei.
%st., Mir. 1978.
G.C.Goodwin, P. J. Ramadge, and P. E. Caines, "Discrete time
multivariable adaptive control." Harvard Univ., Tech. Rep., Nov.
1978.
B. Egardt, "Stability of model reference adaptive and self-tuing
regulators," Dep. Automat.Contr., Lund Inst. Technol.,Tech.
Rep., Dec. 1978.
K. S. Narendra and Y. H.Lin, "Stable discrete adaptive control,"
Yale Univ., S&IS Rep. 7901, Mar. 1979.
T. Ionescu and R Monopoli, "Discrete model reference adaptive
control with an augmented error signal," Automticu, VOL 13, pp.
507-518, W t . 1977.
Y. H. Lin and K. S. Narendra, "A new error model for discrete
systems and its application to adaptive identification and control,"
Yale Univ., %IS Rep. 7802,
1978.
J. K.Hale, OraYnary Diflerentiai Equations. New York: Wdey-Interscience, 1969.
~

Oct.

A. Stephen Morse (S'62-M'67-SM78) was born


in Mt. Vernon, N Y , on June 18, 1939. He re
ceived the B.S.E.E. degree from Cornell University, Ithaca, N Y , in 1962, the M.S. degree from
the University of Arizona, Tucson, in 1964, and
the Ph.D.
degree
from Purdue
University,
Lafayette, IN, in 1967, all in electrical engineering.

From 1%7-1970 he was associated with the


Office of Control Theory and Application,
NASA Electronics Research Center, Cambridge,
MA. Since July 1970 he has been at Yale University, New Haven, CT,
where he is currently Professor of Engineering and Applied Science. His
main interest is in system theory, and he has done research in network
synthesis, optimal control, multivariable control, urban transportation,
and adaptive control.
Dr. Morse has served as an Associate Editor of the IEEE TRANsAcTIONS ON A ~ M A T ICONTROL,
C
and as aDirector of theAmerican
Automatic Control Council representing the Society for Industrial and
Applied Mathematics.

Anda mungkin juga menyukai