Professional Documents
Culture Documents
1978
I. INTRODUCTION
HE theory of adaptive observers,
which
simultaTneously estimate the state and parameters of an unknown linear time-invariant system, is well understood at
the present time. The original observers were proposed in
[l] and [2] and were later generalized in [3]. The extensions to multivariablesystems [4],[5], the analysis of
related stability questions [6], [7] and the study of the
speed of response of the observer [8] have provided valuable insight into the design of such systems. Further, it
has been shown that all the results can be extended in a
straightforward manner to discrete systems [9].
Contrary to the expectations of research workers in the
field, the solution of the control problem did not follow
directly from the solutions obtained for the adaptive observer. The principal difficulty lies in the assumption that
the input and the output of the observed plant are uniformly bounded. As described in [lo], this is not assured
in the control problem, since it is precisely the question of
stability of the overall feedback system that has to be
resolved in this case. Further, the structure of the observer
can be chosenby
the designer so that thenecessary
signals can be fed into any part of the observer to generate the desired estimates. In contrast to this, in the control
problem, only the input to the unknown plant can be
affected and this poses a serious constraint in its resolution.
In this paper, the adaptive control problem for the case
of an unknown plant with a single-input and a single-output is considered. The plant is assumed to be linear and
time-invariant and a differentiator-free controller is to be
Manuscript received June 22, 1977; revised December 30, 1977. Paper
recommended by R. Monopoh, Past Chairman of the Adaptive, Learning Systems, Pattern Recognition Committee. This work was supported
by the Office of Naval Research under Contract N00014-67-C-0017.
The authors are with the Department of Engineering and Applied
Science, Yale University, New Haven, CT 06520.
NARENDRA AND
57 1
y, = h 'xp
(1)
(4)
e , = h'e
'Throughout the paper both differential equations and transfer functionsare used in the arguments and, depending on thecontext, "s'' is
used as a differential operator or the Laplace transform variable.
512
r-30
4, AUGUST 1978
In the development of the solution to the control problem, an operator P,(@ is found toplay an important role.
If L(s) is a Hurwitz polynomial of degree n, in the
differential operator s and 8(t) is a bounded differentiable function of time, the operator PL(8) is defined as
PL(e) g ~ ( s ) e ( t ) ~ - l ( s ) .
V ( e , + ) = 1Z [ e T P e + + T I - l @ ] with P = P T > O .
(9)
PL(0) - 8- PL(+)
-+.
iii-a) If ~ ( s ) = s + a , P , ( e ) = [ e + B ~ - ~ ( s ) l ;
iii-b) If L(S)= s2 + US b, pL(e)=[e + 8(s ~ / 2 ) / y s )
(s a/2)8/L(s)];
P(e,+)= 1
-3eT(qq+~)e<O.
iii-c) If L(s) is any Hurwitz polynomial of degree n,,
PL(@
can be expressed as
Hence, the equilibrium state is uniformly stable and e(t)
and +(f) are bounded if e(to) and +(to) are bounded. Since
n,
w ( t ) is bounded, from [T, lim IIe(t)ll=O. This, in turn,
t+m
pL(e)=e+
[ B , ( ~ ) ~ ( ~ ) A , ( ~ ) ] L (io)
-~(~)
yields from (7)
i= 1
+ +
lim II&t)ll=O
and
t+30
Iim +(t)w(f)=O.
2-30
t-30
Il+(t)ll =os
n,-i
B,(s)=
II (s+z/);
j=2
A,(s)=
II (s+zj)
i=1,2,---,n1
j=1
and Bl(s)=l.
iv) If W(s)is a stable transfer function with n poles and
m zeros and n, < n - m - 1, W(s)P,(B)is a linear bounded
operator. This follows directly from the facts that L(s) is
Hurwitz, 8 is bounded and W(s)L(s) is a stable transfer
function.
C. Lemma2
;J=Aw(t)+ be,
(8)
0
1
1
4,-
'
4n-I
1
4,-1
40
40
0
0
0
1
1 L - 1
C=
0 1
0
0
0 0 0
4l- I
0
0
0
1
to
to
Il
0
0
0
0
0
tm- 1
'm-2
tm-k
m-1
where 0 < k < m,is nonsingular. This is assured if Q and T
are relatively prime.
Let M be the arbitrary polynomial of degree ( < )2n - 1.
Then
n-1
differential equations
w(I)
We then have
(MA+NT)Q+(MB-NQ)T=M
SYSTEM
where R is of degree ( 9 ) n - 1.
IV. THESTRUCTURE
OF THE
= cTu(')
[ko(t),c,(l), . . . , c n-,(t),do(t),d,(t)...dn-l(t)]
Li [ko(t),cT(t),do(t),dT(f)]. (1 1)
(1 2 4
(Fl)
d(2)=Au(2)+byp
d(1)
MAQ+MBT=M.
BM
-=N+-
(F,)
(12b)
11.
kokpZp(s)Ns)
[ N W + C(s)]%(s)+k,Z,(s)[D(s)+doN(s)I
(14)
2 [ r(t),~(')(t),y,(t),u(~)(t)]
4, AUGUST 1978
,..
Fig. 2.
yp = h 'xp.
Denoting 8(t)= 8* +Q(t) the state equations (15) can also
be written as
f=A,x+b,[ k,*r++'(t)u]
where
[ Ap+d,*bphT
bpc*T
e, = h,Te
- -
I bpd*T 1
(18)
(19)
+
-
A, =
---
+ b,k,*r
(17)
NARENDRA
AND
575
-$ e, ( t l
w(1)
L-'
L(S)
WM (s)
Fig. 3.
L ( s )= (s + p).
3 are given by
u(t)=
V. THECONTROL
PROBLEM
2 [e;(t)+B,~-~]~~(t).
i=
1
e= - r e , ( t ) t ( t )
r = rr>o
where
L - ' o ( t ) = < ( t ) , the error e,(t)+O as t+m.
Proof: The proof of stability of the overallsystem
and the convergence of le,(t)l to zero follow along lines
very similar to those in case i).
2n
kP.
e , = A,'.
i=l
input is fed into the plant, it can also befed into the
as described in the
section.
Comment 2: For this
is case it seen that 2n additional
576
Ym
MODEL
SIGN
AA
UL
X.
GENERATOR
Ei
[ i =2,
4, AUGUST 1978
...n)
ti ( i = n + l ,
... 2 n l
--
Since +- P~(+)=+-L+L-'=L[L-~+-+L-'],
if L(S) is
chosen to be a polynomial in "s" of degree n - m - 1 so
that
{h,T(sI-A,)-'b,}L(s)= W , ( s ) L ( s ) = {l2:(SI-Ac)-'bpR}
is strictly positive real, the auxiliary signals can berealized
without the use of differentiators.
The model state equations (24) are modified to
~m,=Acxm,+bck~r+b~RIL-'+(t)-+(t)L-l]'w(t).
If
[L-L~;i(t)-+;(t)L-']w,(t)=S;.(t)
(25)
~=A,e+b,PL(+)Tw(t)
e, = h,'.
which are identical to (22) in case ii). From (26) it follows
that the adaptive law
+= -rel(t)E(t)
--
577
whose resolution is essential to the completion of the suggested in this paper. We discuss this question here and
adaptive control problem, is discussed later in this section make a conjecture whose verification will complete the
solution to the adaptive control problem.
in greater detail.
The proof of boundedness of the output error e l ( t ) is
Care iv): Wp(s) has m( < n - 3) Zeros and n Poles; kp relatively straightforward in all the 4 cases discussed
Unknown
earlier. In i) and ii), since the output of the model is
bounded, the output of the plant yp(t) (=y,(t)+e(t)) is
In iii) the gain kp was assumed to be known. Here, since also bounded, in which case e,(t)+O as t+w. In iii) and
the auxiliary signal corresponding to each parameter Q t ) iv) the introduction of auxiliary signals to the model
is fed into the model (whose gain kM is different from kp), precludessuch an immediate conclusion. If it canbe
additional compensatory gains k,(r) ( i = 1,2; ,2n)are
shown that the plant output is bounded, then the error
needed in the adaptive controller. While the plant equa- el(t) will tend to zero asymptotically and, hence, the
tions remain the same as in iii) the model state equations auxiliary signals also tend to zero as t - x a . The model
are modified to
output then evolves to the desired output and is followed
by
the plant.
r
2n
The plant state equations (29) and the system error
k,*r+k, 2 L ( s ) k , ( t )
i= 1
equations (30) are
x
212
=A,x,+
b,k,*r+ b,
k,ki(t)&(t).
i= 1
2n
e = ~ c e + b p R k p [i2
=1 ~;(t)S;(t>++;(t>~i(t>
(28)
e, = hTe.
(i=1,2,-.-,2n)
\ii(t)=-elli(t)
(i=1,2;-.,2n)
where
by
(20)
and (25) L - ' o , ( t ) = & ( t ) ; [L-'+;(t)Q j ( f ) L - ' ] w i ( t ) = { i ( tand,
)
hence, 4nparameters and 4n
transfer functions L - ' ( s ) are needed in this case.
The questions of stability raised in iii) also apply here
and are considered in the following.
The StabiliQ of f h e Adaptive Loop
578
{>
;{>
;
IEEE TRANSACTIONS
ON
AC-23,NO. 4,
Yp(t)
AUGUST
19'78
YP(t)
PLANT
w,( f )
6(t)
t(t)
(a)
(b)
Fig. 6.
corresponding points l', 2', 3' and 4' in the model are also
accessible. The open-loop transfer function W,,(s) between points l and 4 is equal to W,.,(s) in the model. The
aim of the adaptive procedure is to determine the rule for
updating $(t)(or equivalently b(t)) so that limt+mle,(t)l=
0. Thefour cases that we considered in the previous
section can now be summarized for this general problem
as shown below. It is worth pointing out that the transfer
function W14(s) (or W,,.(s)) is of direct interest in this
case and not the transfer functions W24(~) (or Wr4(s)).
However, for ease of exposition, we shall assume that
points 1 and 2 coincide (i.e., the output of the adjusted
parameter is f e d back to the input) so that W&)=
W,,(s). In this case all the conditions can be stated in
terms of W2.,(s), or the model transfer function. For the
more general case where points 1 and 2 do not coincide,
W,(s) in the followingdiscussion must be replaced by
{ WL+L- '} y(t). The basis for the conjecture lies in the W,&).
i) W,(s) E { SPR } : The adaptive rule
fact that y ( t ) has to belong to the class of signals generated by the feedback loop shown in Fig. 5. If al is the
&(O = - e , ( t b d t )
class of all unbounded signals defined for all t E[O,00)
such that { WL+(f)L-'}yis bounded foryES2, and a2is where o,(t) is the input to the parameter being adjusted
the set of all solutions of an nth order linear differential assures that lel(t)l+O as t+00 by the direct application of
equation with a single bounded time-varying parameter Lemma 1. The evolution of the parameter toits final
+(t),the conjecture states that a, n !& is the null set.
value and the error e l ( t ) to zero for stable and unstable
The application of the conjecture to the adaptive con- plants is shown in Fig. 7(a), (b).
trol problem may be described using Fig. 5(c). According
ii) PVW(s)(s + p ) E { S P R } : When W&) is such that
to the conjecture if el([) is bounded, so is the output of the W+,(s)(s+p) can be made strictlypositiverealby
the
plant j , ( t ) .
proper choice of a constant p, an additional stabilizing
signal is fed into the plant as shown in Fig. 4. T h e signal is
proportional to $(t) and tends to zero as the adaptation
VI. GENERALIZED
A
D
m
- PROCEDURE
progresses. The adjustment of a single parameter using
The method presented in the previous section can now this scheme is shown in Fig. 8(a), (b) for both stable and
be generalized to arbitrarily located control parameters. unstable plants.
iii) W,(s)L(s) E { SPR }: When a polynomial L(s) (of
Since the same procedure can be used for any control
parameter, we shall merely consider the case of a single degree 2 2) exists such that W,L is strictly positive real,
adjustable parameter in the plant. The following develop- the stabilizing signal ( 6 - PL(8)}wI is fed into the model
ment is motivated in part by a procedure used for the as shown in Fig. 9. The model is suitably modified so that
this can be accomplished without theuse of differentiageneration of sensitivity functions as described in [12].
Let M represent the model and P the plant in Fig. 6(a) tors. The evolution of the feedback parameter, the stabilizand let B(r) be a control parameter. It isassumed that ing signal and the output error is shown in Fig. 10(a), (b)
when 6(t)= 6*, a constant, the transfer function of the for both stable and unstable plants.
iv) W , ( ~ ) = k , / k , ~ W , ( s ) ;W , ( s ) L ( s ) E { S P R } : If the
plant is identical to that of the model. The plant with
8( t ) = 8* is denoted by P * ; if e(r)= 8*+I$([),
the plant can transfer function between points 1 and 4 in the plant is
be represented as shown in Fig.6(b) by P* together with a the same as that of the model except for a constant gain,
parameter +(t). For the purposes of analysis four points 1, an additional gain is needed in the path of the stabilizing
The evolution of the feedback
2, 3, 4 in Fig. 6(b) are of interest and we assume that the signaliniii)(Fig.11).
579
52 + 4 s + 3.75
s3+ 5 s 2 + 7 s +2 25
[PLANT) Wp(s) =
INPUT
Square
wave
Amplitude 5 units
Frequency = 1 hz
-\
INPUT: Squarewave
Amplitude . 5 units
Frequency= I hz
\A
CONTROL OF UNSTABLEPLANT
12
IO
-8
t in seconds
I\
IO
12
14
.A
Fig. 7.
(PLANT) Wp(s) =
s+ 2.5
s3+6s2+9s+ I
4l
( M O D E L )sW
+ 2y 5( s ) =
s 3 + 6 s 2 + 11s + 6
CONTROL OF UNSTABLEPLANT
ux. signal
IO
I2
14
nl
'
-11 I
t in seconds
Fig. 8.
IO
14
t in seconds
ta
22
580
1978
MODEL
r(t)
MODEL*
Fig. 9.
INPUT: Sqmare
wave
Amplitude : 5 units
Frequency = I hz
:I
3
L ( s ) =(s+1.5)(r+2.5)
t in seconds
Fig. 10.
(b)
parameter 8(t) to the value 8*, the stabilizing signal { ( t ) the output error Ie,(t)l, as well as thestabilizingsignal
for both IS(f>I, tending to zero at rates which are Practical. In View
and the output error is shown in Fig. 12(a),
of the fundamental theoretical questions that arise in the
stable and unstable plants.
stability analysis of the adaptive system, these simulation
VII. FURTHER
SIMULATION RESULTS
studiesprove
invaluable
pro&ing insight into the
nature of the adaptive controller. In this section we preExtensive simulations of the control problem on thesent
simulation results for a third order system for the
digital computer indicate that the procedures outlined in fourcasesdescribed in Section V. The plant and model
the previous sections do result in a robust controller with transfer functions for the four cases are given in Table I.
e)
in
58 1
MODEL
r(t)
__
P*
L-
L-9
Fig. 11.
-1
- 1.5
-2
-3
(b)
Fig. 12.
TABLE I
Case
Plant W,(s)
WdS)
i)
s2+5s+4
S3+9S2+24S+20
2(s2+4s+3.75)
s3+6s2+11s+6
+ 9 2 +24s +20
s+4
2(s+ 1.5)
s3+6s2+11s+6
1
s+9s2+24s+20
1
s3+6s2+ lls+6
2
s3+6s2+IIs+6
i
i
)
s3
iii)
iv)
+ +24s +20
s3 9 2
582
(PLANTIWp I s ) =
(MODELIWyIs)-
1978
r+4
s3+
9~2+
24s+ 20
r3+6r2+llr+6
INPUT : S w a m m e
Amplitude: 5 units
FrEqYency
I hz
Llsl :( s t 2 . 5 )
dA-l.--I-r.-I_CA-.
14
IO
12
t i n seconds
16
18
20
-I
1:
-ow
( P L A N T ) Wp
(5
1=
-2IO
12
14
t in seconds
(4
2.6
P1
a
2
S3+9S2+24S+20
II
IO
L ( s ) =(s+1.5Ks+2.5)
signal 2
(PLANT) Wp
(5)
(MODEL)WM(s)=
i
s3+952+245+20
s3+6s2+lls+6
L ( s ) = IS+1.5)IS+2.5)
7
6
0
v$
IO
t in seconds
12
14
16
aux.signal I
-I
2
1
-2
t in seconds
(4
Fig. 13.
REFERENCES
[2]
[3]
h(s)=s2+4s+3.75.
The reference input r(t) in each case was a square wave
with an amplitude of 5 units and frequencies ranging from
0.5 to 6 Hz (Fig. 13).
(4
[4]
[SI
A.
4, AUGUST
583 1978
.
,
_1
*,
- A
,&
2!j
Abstmcr-A general consistency theorem for stationary nonliwar predictlon error estimatom is presented. Since this theorem does not requjre
the existence of a parameterized systemgenerating the observationq it
applies to the practical problem of modeling complex systenrs with simple
parameterizedmodels.Inordertomeasurethequalityofftbetweenaset
of o h e d processes and a given candidate set of predictors, ttbe notion of
predictor set completeness is introduced. Several examples are given to
illushate this idea; in particular, a negative resultconcerning ttae completeness of certain sets of linear predictors is presented. Ibe relationship of
Y W s definitions of identifiability to various notions ofpredictor set
completeness is examioed, and the strong coosistency of mrurimum likelihood estimators for GaussianautoregresSme moving average s
y
s
t
e
m is
obtained via an application of our techniques. Finally,problems for fnture
research ate described.
Manuscript receivedApril15,1977;revisedMarch
12,1978. Paper
recommendedby R. K. Mehra,Past Chairman of the Identifkabon
Committee. This work was supported by the Joint Services Electronics
Pro am under Contract N00014-75-C-0648.
#e author is with the Division of Applied Sciences,Harvard University, Cambridge, MA 02138.
R EEE,
I. INTRODUCTION
N MANY identification procedures, observations are
Itaken on a complex stochastic system such as a paper
mill, a refinery, a brain, or an economy. These observations are then used to build approximate or simple
systemmodelssuch as low order autoregressive moving
average (ARMA)models, or linear state space models
with state spaces of low dimension.
The implications of the use of approximate system models for regulation havebeen
studied byChang
and
Rissanen [ 11. They consider a general least squares control
problem. In thisproblem, linear control laws are permitted to utilize only a limited set of past observations on
the input and output processes of a given linear system. It
[l] that the optimal control laws in any
is shownin
specified class can be characterized by the minimization