Вы находитесь на странице: 1из 10

3EEE TRANS.4CTIONS ON AUTOMAFIC CONTROL, VOL.

AC-18,KO. 5,

OCMBER

419

1973

Comparative Studies of Model Reference


Adaptive Control Systems
CHANG-CHIEH HANG

AND

Absfracf-A brief but up-to-date survey on existing methods of


of model reference adaptive control systems is
designingaclass
given in this paper. A comparison of the merits of the various design
rulesis also made with particular attention to the M.I.T. rule
and the Lyapunov synthesis technique. Subsequently a systematic
performance comparison of the various designs, with deterministic
as well as stochastic inputs, is presentedusing the computer simulation studies of two simplegain adjustment schemes. Fromthe
dimensionless characteristic graphs obtained the Lyapunov schemes
over other designs.
are found to exhibitsuperiorperformance
These graphs also uncover the strange adaptive response of those
designs based on gradient methods in that the performance index
may increaseor decrease with increasing system gain parameters.

I. I~TRODUCTIOS

HE model reference adaptive control (MRAC) technique has been a popu1a.r approa.ch t.0 the control of
systems operating in the presence of parameter a.nd environmental variations. I n such a scheme, t,hedesirable
dynamiccharacteristics of the pla.nt are specified in a
reference model and the input signal or the cont.rollablc
Parameters of t,he plant are adjusted,continuously or discret.rly, so t>ha.t.its response will duplica.te that of t.he
model as closely as possible. The identification of the plant,
dyna.mic perfornmnce is not necessary and hence a fast
adapt,ation can be achieved.
This paper is concerned 1 ~ t . the
h continuous paramet.er
adapt,ive schemes. Generallyspeaking,t>here
are t.wo
approachcs to thesynt,hesis of this class of NRAC systens.
One is based on t.he minimization of a performance index
[ l ] and the other on a L p p u n o v function [ 2 ] . Each of
theseapproacheshas
its own meritsandlinlitations,
alt,houghmany
modificat,ions ha.ve been suggest.ed t,o
improve them further. A direct contrast. of the merits of
t,hcse drsigns has been briefly ment.ioned in t.he literature
[2], [8] but, a rigorous comparison, especially that from a
perfornlancc viewpoint,, has not. been reported. Hence a
comparative study of the various design rules will be of
great, int.crcst. to the designers who have long been faced
with the difficulty of select,ing a suitable one for certain
applications.
Since there are a1rea.dg some ddailed accounts of the
various design rulesin thc litera.t,ure [17], [lS], only a

PATRICK C. PARKS

brief but up-to-date survey is conducted in t.hispaper.


Attention will then be concentrated on single-inputsingleoutput plant. gain a,djustment systems. Some of t.he more
p0pula.r rules are criticallyanalyzed t o point outtheir
relative merits wit,h regards to the stability, realizat,ion,
and adaptive response, which will also be supported by
some simulation results. Subsequently a systematic performance comparison based on some well-known criteria is
att,empted t,hroughsimulationstudies.Determinist,ic
as
well as st,ochastic inputs are employed. Sensit.ivit,iesof the
performa.nce t.0 the input frequencybands are also examined. The interesting anduseful performance charact.eristics are presented in the form of similitudes [20].

11. A BRIEFSURVEY
OF

THE

DESIGIC
RULES

The MRAC system was first designed by t,he performance index minimization met,hod proposed by Whitaker
[ l ]of theM.I.T.InstrumentationLa,boratoryandhas
since t,hen been referred as the M.I.T. design rule. The
performance index is the integral squa.red of the response
error. This rule has been very popular due to itssimplicity
inpracticalimplement,ation,a.lthough
it mayrequirea
largenumber
of semitivitg filtersformult.iparameter
a,djustment.s.An improved design rule with respect to the
speed of response has then been proposed by Donalson
[3], who used a more general performance index than that
of Whitaker, but addit,ional filtersand themeasurement, of
t,he state vectors are required. The need of t,he sensitivity
filterscanbeavoided
by agradient met,hod developed
later on by Dressler [4], or by an accelerated gradient
methodsuggestedbyPrice
[5]. The latter is easier t o
implement. and is capa.ble of achieving faster adaptations
compared with ot.her gradient techniques. Another
cont,ribution to the simplification of the design comes from
the applicat.ion of sensit.ivity analysis by Kokotovic et al.
[29], [25], resulting in a design similar to the M.I.T. rule.
Here,withfurther
approximat,ion,only one sensitivity
filterisrequiredformultiparamet,eradjustment.s.
For
some other particular applications, Winsor [17] has also
modified the 3I.I.T. rule t o reduce the sensit,ivit.y of the
response t,o theloop gain, at the expense of addit,ional instrumentation. All the design rulesmentioned a.re not,
Manuscript received October 24,1972. Paper recommended by however, globally &able and hence the adaptive gain that
G. X. Saridis, Chairman of the IEEE S C S Adaptive and Learning governs the speed of response is limited. A good comproSystem, Pattern Recognition Committee. The work of C. C. Hang
was supported by the Royal Commission for the Exhibition of 1851 mise between t.he &ability and the speed of adaptation
and the British Council.
will have t o be decided by laborious simula.t,ion studies.
The authors are with the Inter-Inst.it,ut,eof Engineering Control,
A recent, cont.ribution by Green [6] hasextended the
University of Warwick, Covent,ry, England.

420

work of Dwssler t o form astablemaximumdescent


method. Though the problem of global stability is solvcd,
this adaptkc. rulc is not attractive from a practical viewpoint because the first derivatives of the state vrctors arc
often requircd fur its inqhnrntation.
On t he other hand.in the Lyapunov synt
hesis approach!
thc ndaptivcrule is obtained by selecting the design
equations t o satiefy conditions derived from Lyapunovs
second method, so that thv system stability is guarnntwd
for all inputs. Butchart and Shackcloth [i]first suggested
the use of a quadraticLyapunov funct.ion, which \vas
employed later on by Parks [2] t o redesign systems formerly designed by the 3I.I.T. rule. The use of a different
Lyapunov function by Phillipson [SI and Gilbart et al. [9]
has resulted in the introduction of feedforward loops that
would improve the damping of theadaptive response.
The main disadvantage of the Lyapunov method is that
the entire state
vector must be available for measwenlent.
which is not often possible. Recent efforts in t h r application of the idea of positive real transfer function, notably
that by lIonopoli, have allo-wd onr t o eliminate or reduce
the number of differentiat ors required for implemcwting
the design rule [2]. [lo]. Among other possiblc solutions t o
avoid the use of derivative net.worlis: Currie and Stcar
[I1 ] h v c envisaged thc use of a Kalman filter. which
would also handle the measurement noise problem, while
thc use of state obsrrvers [12] t o estimate the state5 of
an unlmown timevarying plant. is still an open question.
Anothcr &sadrantage is that the Lyapunov
design rulc
may not be applicable t o cases where the plant paralneters
cannot be directly adjusted. Such acase \vas nmltiorlc~db?Winsor et a/. [13] and a solution. though quite complex,
was offered by Gilbart, et al. [9].
All these methods may be extended
to trcat both the
single-dimension and mult.i&mension systems. For the
latter case. another synthesis technique [14] based on the
concept. of hyperstabilityhas
been recently proposed.
The resuking design is asymptotically hyprrstablr andt h r
structure of the adaptive controller is simple. The dran-back of this scheme is that a series of differcntiating net\vorl<s is often necessary. In the single-dimcwsion caw.
it results in a design identic.al t o tlw Lyapunov redcsign
due to Parks [I.
From this brief survey it is observed that many design
rules for MRhC systems areavailable. So far as the singledimension case is concerned, the M.I.T. rulc has bcen the
most. popular although it has a severe stability
problem.
The Lyapunovsynthesis,inavoiding
the problem of
stability, is difficult t o realize inpractice because the
measuremmt of the entire sta.te vector i5 not often possible. While the degree of these considerations n-ould vary
with the particular process t o br controllcd. another importantfactor. t h e performance. such as the. qwcd o f
ada.ptation. has t o be considered. It is thr intrntionof this
investigation to attempt to supplement some useful informationregardmg
therclatiw
performancc of the
various de:,-1TlS.

IEEE TRrnSACTIONS ON AGTOMSTIC COSTROL, OCTOBER

1973

111. A CRITICAL CONPARISOX


OF THE DESIGX
RULES
Tllc following analysis is based on the aggregate of
l m d e d g e scattered in thc litemture. This information is
reviewed here and st.udied by means of simulations. We
shall first compare the J1.I.T. rule [ l ] and the Lyapunov
synthesis [2]. [9] through the design of a gain adjustment
loop of a linear system as sho\m in Fig. 1. Follon-ing this
w $hall csaminc~dcsign rulesdue t o Dressler [4]. Price
[>I, and AIonopoli [lo]. The block diagrams of these
designs are sho~vnin Fig. 2 .
A .

J1.I.T. Rule and Lyapunol: Synthesis

The notation used below is that s11ow1 in Figs. 1 and 2.


The performance index used in the 1I.I.T. rule [ I ] is
sei? dt and the parameter adjustment law is

38

In this case the sensitivityfunctionisproportional


dk,
8, and hence the above equation becomes

k,

Be16,

to

(2)

u-hcre the constant B is the adaptive gain.


TheLyapunovs-nthesisis
basedon
the use of a
Lyapunovfunction.Tllc
most successful form of this
function 1- used to date is t.hat proposed by Gilba.rt et al.
[9]. This is briefly described in the Appendix, Section A .

eTPe

+ X ( s + YELT~VZIZ)~

(3)

BeTPbr

(4

K - K,K,

( 3

where

and the time derivative


of 1; is given by

Thcsr result in the stable adjustmcwt laTv

d l e r c y is a proportional constant that is chosen t o provideadditionaldamping


if requircd. Putting y = 0
rcsults in thedesign rulr used by Parks [2].
Equations f2) and (7) will be compared in the following
manner.
l j Stahilify: A stability anal;\-sisof @) is w r y difficult.
The doubtabout
possible instabilityhas been demonstrated by Parks [2] for a vcond-order system with step
inputs. Evcn for a first-ordw system n-ith a sinusoidal input, . J a n m [ l 5 ] hasobtained
a complicatedstability
domain in the parameter space. Hence, extensivc simulations during the design stage arcnecessary t o establish the
region of stableoperations. On the otherhand: (7) is

42 1

HAXG AND PARKS: NRAC SYSTEMS

P
Om

CM.I.T.)

(DRESSLER)

Fig. 1. ,4basic MRAC gain adjustment syst,em.

assured t o be stable for all input.s such t,hat,e


0 asymptotically,withthe
assumpt,ion that K , is slowly time
varying.When thisassumption is severely violat,ed, a
stabilit;y problem similar t.0 that of (2) may exist. Eventual stability, however? can be assuredbyusing
a
theoremdue t.o L a d e and Ra.t,h if t,hetime-varying
function K , belongs t o a cert,ain class [16].
2) Physical Rea1iza.tim-z: Equat.ion c2) can be easily inlplemented and it. is t.his distinct) advant.age t,hat has made
MRAC a popular adaptive control strategy. Equat,ion (7)
however, requirest.he estimat.ion of thc completestat.e
vector, which is not oft,en asailable and, hence, necessitates the me of different.iat,ing networks causing a noise
nmplificat,ion problem.
3 ) Response: The speed of a.dapt,ation of bot,h equations
depends on t>he magnitudes of t.he adaptive gain B and
t,he input signal R. A large B isalways necessary t o
maintain a high speed of adapt,ation. However: as B and
R vary, the damping of t.he response will a.lso mry. Root
locus plots of these equations for a second-order system
[SI, [9] would shon- tha.t>whcn B P is la.rgc: the 3I.I.T.
design will br underdamped while t,he Lyapunov design
will be adequately damped with suit,able values of -/.

B. Other D e s i p Rules
We shall next examine the follom-ing rules.
[dl: The pammeter adjustment,law is

1 ) Dressier.

K c = Belr..

(8)

The resulting controller is very simple a.nd no sensitivity


filter is required. The disadvantages a.re that, t.he da.mping
of the response suffers at. larger loop gains and that t.hc
global sta,bility is not guaranteed. Its stability problem is
simi1a.r to that of the 3I.I.T. rule.
2 ) Price [SI: Thc paramctcr adjustment,lam is called the
acce1erat.cd gradient method.
K c =

Belr.

+ yc cltd (BelT)
-

whcrc, yc is a conat,a,nt..The controller is sinilar to that of


Dresslcr except for the a,ddition of t h r fcedforward term.
This t m n has t hr rffcc t, of improving thc damping and t hc
stability o f the rrsponsr. This stabilizing cffcct would br
inqmired as thc ordvr of the yi-stcm increases- and gcncrally t.he global Ftability cannot be guaranteed.
3) Jfonnpdi [ I O ] : This is ba.sed on a. modification of the
Lyapunov scheme. A differentiating block (Z(s)) is used

(MONOPOLI)

Fig. 2. Various designs of t,he adaptive loop.

t,o modify the plant t,ransfer function such t,hat Z ( s ) N ( s ) /


D(s) is positive real, and a Balman-I\Ieyer lemma [26],
[27] is used t o eliminate t.he error derivatives as required
in (7). Hence,

where y is t.he modified input signal t o t.hc plant. and is


obtained by passing t,he original signal through afilter
(l/Z(s)). For annth-order plant with nz zeros, t,he orderof
Z ( s ) is (n - m - 1). Global stabilit,y of the adaptive loop
will be guaranteed while the number of derivative network
required is reduced t o (n - m - I), or (n - m - 2) if the
extra damping loop is not in use. This technique can be
easily extended to the case of a general time-varying gain
[lo] and also t o t.he adjustment of ot,her parameters [28].

C . Sinru.lation Results
A simulation study of the adaptive response of these
designs has shown t.hat very often the Lyapunov designs
could achieve excellent performance not. attainableby
other rules. -4s an example consider a second-order plant
n-hose gain is t o be adjusted. Referring t o Figs. 1 and 2,
the following values are assumed:

1,

IC,

= 2,

Kc(to) = 0.2.

~
eTPb = el P1.
From thc Appendix, Srction 8 , w ( obtain
R c shall choosc Z ( s ) = 2s
2 (a.s in [lo]) and shall limit

the values of y and y c t.0, say, 50 percent. because in actual


use thry may have t o be quitc small t o reduce t.hc effect. of
any noise a t t.he plant. output. and the excessive transient

422

IEEE TIWNSACTIONS

ON AUMMATIC CONTROL, OCTOBER

1973

overshoots due t o large initial errors. Some of the typical


ada.ptive responses of the various designs are depicted in
Fig. 3 for strp as ~ 4 as1 sinusoidal inputs. The responses
shon-n for thc 3I.I.T., Dressler, andPricc designs have
been optimized withresprct to the convergence time.
The responses shown for the Lyapunov and 3Ionopoli's
schemes are, however, not opt.imized, i.e., they can still be
further improved if required by increasing the ada.ptivc
gain. From t.his simulation study. the 3I.I.T. design is
found t.o be unst.ablcfor B' > 1. Even \\-hen it is stable at
lox\-er values of B', t.he response is slow! the convergence
time being ~vell over fivc systemtimrconstants.The
response of t.he Dresslcr scheme t o a step input is similar
to that of the 31.1.T. scheme. However, for a sinusoidal
input, theDressler scheme shows a. steady-stat e parameter
error that is dependent on the loop gain as well as the input,
signa.1frequency. The design due t o Pricc shows a, better
damping and stability, x=,-hichimprove as yc is increased.
On the other hand. the Lyapunov design is always stable
and the damping and convergence can be improved systematically by varying B' and y. A convergencr time of
even less than onesystemtimeconstant
can bc easily
achieved. The design duc t o 3Ionopoli, u-hich docs not
require any &ffcrentiator in this caw,cshibita quite a fast, Fig. 3. Responses of the second-order s y s t e m : (i) hI.1.T.; (ii)
response. Although its damping would sufl'cr at higher B':
Lyapunov
= 0.1); (iii) Ylonopoli; (iv)Ihessler;
Price
(Ye - 0.5).
the system sta.bilit- would alu-ays be maintainrd. Thcse
results support and subst,antiate the foregoing theorrtical
analysis.
r2 = KKzB'R2T
(1I.I.T. design)
It. is appreciatrd that a, further study on the relative
= K,B'R?T
(Lyapunov design)
performance of these rival designs is a complex problem.
This is pursued in the next section.
= T,/T
(5 percent T , criterion)
(A,

IV. A SI-STEMATIC
PERFORMASCE
CONPARISOX

($7)

- K2R?R'TS e l f rlt
(HE criterion)
Some commonly used performance criteria [19], which
include t h r settling t.inw (T,)? integral of squarederror
1
- -~
Jtlel! dt (ITAE criterion)
(ISE). integralof time absolutc error (ITXE). and integral
KRT?
of timr squared error (ITSE)will bc employed t o compare
t.he rcsponsc of thc various designs against tllcir y,-stcm
parameters. This will be studied expcrimentall\- through
computersimulations of twogain adjustmrnt schemes.
The parameters that cannot, be groupedinto the above are
The rrsults will be presented in the form of sinditudrs
fixed at frequency of sinusoidal input = 2.5 c/s, K,(fo)= 0 ,
byapplyinga
dimensional a,nalysis [20] to the system
y = 0 and 0.1.
differential equations such that the quantitics to be inThe performance characteristics obtained are shown in
vestigatedare
expressed in dimensionlcs groups. The
Figs. 4 and 5 . For step inputs, in n-hich case the 3I.I.T.
dimensionlws perfornlance criterion i? dcnotcd bJ- rl
design is always stable, the T , criterion shows aregion
d i l e the dimcnsionlcss system pa.ran1ctc.r is dcnotcd by
where this design is unfavorable since rl may increase or
r?. The performancecha,ractcristics arc dcfinrd inthis
decrease with anincrement in asswhereas the same type of
connection as theplots of a1 against r4.
uncertainty does not appear in the Lyapunov design with
y = 0.1. For sinusoidal inputs, all the four Characteristics
A. First-Order Systems ( X ( s ) j D ( s ) = 1 / ( 1
ST))
for the 1I.I.T.design possess regions of uncertainty over a
In this casr, the dcsigps due t o Drcssler and Price are wide range of r2.Furthermore, ithas already been ensured
identical to thc L ~ ~ ~ U schemes.
I I O V Also, tlw latt cr docs that within the parameter ranges tested, that is a? < 2 5 ,
not. requireanydifferentiators.Hencc
we onl\- need t o this design isoperated below the region of conditional
stabilityaspointedoutbyJames
[15]. These findings
conlparc the 31.1.T. and the Lppunovdcsigns.
extensivesimulationstud>would be
I ) Detern~inisticI?z.pts: Step and sinusoidal inputs arc suggest thatan
enyloJ-rd. I.'rom the dimensional analysis s l ~ n v nin the. necessary in order t o determine a safe and economic value
of a2 t o achieve any specific r1 even though the system is
Appendix. Section B: the follon-ingare dcfincd.

423

HANO AND PARKS: M R A C SYSTEMS

TL

here, indicate that inM.I.T.


the
schemesystem
the
Derformance is verysensitive t.o the cha.nge infrequency
whereas in the Lyapunov scheme it is almost insensitive
to the frequency especially at higher gains.
2) Stochastic Inputs: The above experiment is repeated
witha
band-limit.ed Gaussian n7hit.e-noise input.This
stochast.ic signal is obtained by spacing a digitally generated, zero-mean,
Gaussianly
distributed sequence of
pseudorandom numbers, by
a.n interval of h s and with
linear interpolations. The varianceof the signal is denoted
h N 2 and its power spectrum, which is a.pproximately fla.t,
possesses a cutoff frequency of 1/2h Hz. T o reduce the
complexity of thisinvestigation, only the ISE criterion
will be studied in detail.
The dimensionless quantities are

20

10

(b)
TI1

1.0

20

IO

lT2

(d 1
x xr. (M.I.T.)

...

<UAFUNOV, J ' 0 )

A A A

(LIRPUNOV, I = a0

Fig. 4. Performance characteristics of firsborder systems with step


inputs. Crit.eria: (a) T , ; (b) ISE; (c) ITAE; (d) ITSE.

Fig. 3. Performance characterisbies of first-order systems with


sinusoidal inputs. Criteria: (a) T,;(b) ISE; (c) ITAE; (d) ITSE.

operatrdinthestable
region. On t,he other ha.nd, the
similitudes for the Lyapunov designs show a monotonic
decrease of rln7it.h increasing T?.This isa desirable feature.
I n additionthis design can achievevalues of rl not
attainable by the3I.I.T. design. Exanlinat.ions of t,he
effect of changing t.he input signal frequency have also
been conducted. The results, which are too long t o show

1
E [ S e ' dt]
K26A72T
rz = K K , B ' C ~ ~ ~ T(M.I.T. design)

r1 =

K,B'hN2T

(Lyapunov design)

where E [ ] denotes the expect,ation (i.e., ensemble average) operator.The


fixed parametersare
12 = 0.002,
K,(to) = 0.0, y = 0 or 0.1.
The result.s obtainedareplot,tedin
Fig. G(a). The
similit.udes show that. both t,he h.1.I.T. and Lyapunov designs exhibit the desirable characteristics that r1 decreases monotonically wit,h increasing rz.The latter also
achieves a much lower rl,which cannot. be reached by t.he
former. Another import,ant pr0pert.y t,hat has been noted
is thatthe
variances a.bout, t.he expectedvalues
are
Merent, in each case. From the plot shown in Fig. 6(b),
one observes that t.he variances in the N.1.T. design a.re
verymuchlarger
t,hanthose in the Lyapunovscheme.
This indicatesthat in theformer scheme there may exist' a
considerable degree of uncertainty about. it,s performance.
This is confirmed by studying tJhe ensemble members of
t,he random process. One of these is shown here in Fig. 7.
Also shown are ensembles of t,hecorresponding results
usingt,he othertwointegral
criteria.. Thesesimilihdes
reveal that the M.I.T. scheme possessw the undesirable
property that r1 may increase or decrease n4th increasing
r2 while that in the Lyapunov scheme shows an almost
monotonic decrease.
In addition to thecase just reported, ot.her experiments
have been carried out. Thefinding is t.hatwhen the power
spectrum of t,he input signal (proportional t.o l j h ) is
reduced, t.he performance of theM.I.T. designwould
deteri0rat.e whereas t,hat of tlhe Lyapunov design xould
improve.

B. SecmdOrder System ( N ( s ) / D ( s ) = 1 / ( 1

+ als + a&))

The five designs described inSection I11 d l be examined here. It is not.ed t,hat while the Lyapunov design
requires one different,iat,or,that. due t o Monopoli docs not
need any.
1 ) Deternzinistic Inputs: The following dimensionless
parameters are defined :

424

IEEE TRANSACTIONS O S BUTOMATIC COSTROL, OCTOBER

1973

!
(LOG.

SCRiE)

Fig. 6. Performancecharacteristics
of first-order s y s t e m with
stochastic inputs. Crit.erion: ISE.

Fig. 7.

Asample of t.he characteristics. Criteria: (a) ISE; (b) ITa=;


(e) ITSE.

T I ! and the latter is


more crit.ica1 than the former. The
design by Price shows an unfavorable performance in that
the uncertaintyas discussed inthe first-ordersystems
occurs. The Lyapunov and the 3Ionopoli designs, however, still maintain the desirable performance characteristics similar to that with step inputs.
The performance of these designs with different frequencies of the sinusoidal input signal has also been examincd. The same range of 1r2 is used. The general Conclusion is that the Lyapunov and 1lonopoli des1'gns are

Iws sensitive t o the signal freyuwcy nith regards t o bot,h


the stabilitJ- and theconvergence rate. The 1L.I.T. system
always pouscsses a nlinimunl r1 at some valur of n2 t.hat
increases n-ith the frequency; at l o w r frequencies. more
than one nlininlunl point may be obscrvrd. The conwrgcncc rate decreases withincrrasing frequrncies. The
Drcsslrr system is unstable at highrr frrqurncics; atlower
frrquencirs thc system is stsblr for a small rangc of rB
but
this range may increaw or drcrcasc u-it11 decreasing f r c
qurncies. The design by Price improrcsat lower frequrncics. in that the fluctuationin rl reduccs. but deteriorates rapidly at frequencies higher than the resonant
frcqucnc>-of the plant and evcntually becomes unstable.

HAWG

AND PARRS: MFCAC SYSTEMS

-rr, k
15.

IO

CI -

..,

x X x (n.1.7.)
ooo

,.

(MONOPOLI)
OR

(LIRPUNOV, '6

(PRICE,
(CHRNGE

=o)

sC= 0.51

...
,..

(LIRWN3V, 1 = dl!

( DRESSLEU)

IN S C ~ L I N G )

c, o((

(CHRNGE

IN SCdLINo)

Fig. 8. Performance charact.eristics of second-order systems Kith


step input&. Criteria: (a) T,s;(b) ISE; (c) ITXE; (d) ITSE.

Fig. 9. Performance characterist.ics of second-order systems wit,h


sinusoidal inputs.Criteria: (a,) T,;(h) ISE; (c) ITAE; (d) ITSE.

2') Stochastic I q m t s : The sameexperiment as in t.he


first-order case is repeated. The main results with jz. = 0.1
are shou-n inFig.
lO(a) and(b).To
summarize,t,he
Lya.punor a.nd Monopoli designs exhibit. monotonic
decrease of rlwith increasing a2and theva.riances of r1 are
small; the other designs exhibit one or more minima in a1
andthe variances are also largeindicatingseriousuncertainty as mentioned in the previous section. Different.
spectra of the input,signal h a w also been used. The
general observat.ion is t.ha.t.the 3I.I.T. and
Dressler designs
cshibit worse performance %Then thebandffidth of t.he
signal is reduced, n-hile the other designs show improved
performance.

effective ga.in varies due t o different attenuation by the


syst,em at, different frequencies. The possible outcome of
these tn.0 disadvantages is instability, poor damping, or
poor convergence of the adapta,tion. It is unfort>unatethat
in trying tocompensate for t.he change in environment., the
adapt.ive syst.cm may become sensitive t o it.s o ~ paramn
et ers.
2 ) The performance of those designs that are assured t o
be globally &able improves as the gain paramet.er az increases. I n addition, they ca.n be ma.de less sensit,ivc t o t,hc
input, signal magnitude and frequency content b)- operat,ing a.t larger values of az..
3) Among the t>hreeschemes based on gradient. methods,
the Dressler design eihibit,s theworst perfornlance characteristics, especially when the input is sinusoidalorstochastic. The 3I.I.T. design is quite acceptable if the performance specification is not very strict.. The
design by
Prier performs better than t.he 1I.I.T. system with stepor
stochastic input.s but is infcrior n i t h sinusoidal inputs.
4) The two designs based on stability consideration may
a.chieve low values of rl not attainable by other designs.
Bet.ween the t.m-o, the Lyapunov scheme is better as it.
requires a lower value of a2to sa.tisf- the
same performanee
specification.

V. DISCUSSIOK
AND CONCLUSION
The extensive comput.ersimu1at)ionstudy of the various
MRXC designs reveal many interesting properties regarding the perfornmncc of the adaptive systems at different
loop gains and under different input signals. These may be
summarized as follows.
1) Thcl dvsigns t.hat are not assured t o be sta.blc globaily
behaw vcr;v differently when the gain parameter a?varies.
They arc' also found t.o be sensitive to the frequency band
of the input signal; one reason for this is that t.he total

426

IEEE TRANSACTIONS ON AUMVSTIC

CONTROL, OCTOBER

1973

preferred by many designers due to the ease of t.heir implementations should be used wit.h great. caution.

APPENDIXA
A . Lyapunov Design [9]
T h e d a t eequat,ions of the plant, and model are
plant,:

+ bK,K,r
= A0, + bKr.

6,

model: 8,

(A.1)
(A.2)

Define

0, - 0,, X = K -

KX,.

We obtain

Ae

+ bXr.

(A.3)

Choose a V function

eTPe

+ x(X + yKzm)?

(4.4)

where

m = B'eTPbr.

(A.5)

The time derintiveof 1' is

Fig. 10. Performance characterist,ics of second-order systems with


stochastic inputs. Criterion: ISE.

+ P,4)e + 2eTPbXr
+ 2X(X + yK8m)(X+ yK&).

eT(ATP

then V beconles

V
Although the designs investigRted in this paper are only
lox--order gain adjustment schemes! it seems pussiblc that

t.he general obscrvations may hold for higher order systems and for the adjustment of other parameters. At least
for those schemes basedon
thestabilitytheory,the
L p p u n o v function alwa.ys assures global stabilityand
t.hat the convergence rate reduces as the nda.pti5-e gain
increases (the nlasinlunlvalue of -'d/c gives an est.imate of
the convergence rate, see r21]). This paper also assumes a
noise-free system. The important. problem of noisebiasing
action is briefly discussed in the Bppendis.Section C.
On interchanging the roles of the plant and themodel. the
case studied becomes an identification system. Hence this
investigation also reveals the shortconings of those model
reference identification schemes ['Z], [23] based on
gradient methods.
We suggest that the globally stable design rules should
be g i r m prinw considerations on account of their stability
and performance properties. There are many cases xherr
thesc dwign rulcs arr not applicablc and furthcr rcscarch
t.0 rclicvthissituation
i:: neccssary. 1Icanwhile. the
hI.1.T. rulc and other gradient. methods that. have been

(AS)

--=&e - 2XyKn2n1?.

P a.nd Q are positive definite symmetricmatrices


satisfy the Lya,punov matrix equation
ATP

+ PA

[-'1

-2

For example, if

'1

-Q.

(A.8)
that
(A.9)

.I![

Let

[i

Solving (A4.9)?
we obtain

P
Hence, eTPb = el

[;

;I.

+ e*.

B. Diwzensional Analysis [20]


The equations describing the first order lI.I.T. system
arc

427

HANG AND PARKS: MRAC SYSTEMS

Te

+e

X,

(K

K,K,)r

(A.10)

BeB,.

(A. 11)

Define the following dimensionless variables :


E

e/(KR)

(A. 12)

rn

r/R

(A.13)

ym

O,/(KR)

(A.14)

( K - K&)/K

(4.15)

t/T.

(A.16)

Substituting these int,o (A.3) and (A.4), we obtain


(A.17)

dX
cl T

-(KK,BR2T)~.y,

Hence the required dimensionless parameter


~2

KK,BR2T.

(A.18)
~2

is
(A.19)

Repeat.ing the same process for the Lyapunov design


with the same dimensionless variables, we obtain
(-4.20)

C
IAT _
- - (K,BRT)c T,.
dl-

(A.21)

Hence,
~2

K,BRT.

(-1.22)

The dimensionless performance indices are obta.incdby


using the dinlensionleas error E and the dimensionless time
variabIe T . For instance, t o obt;ain T I using t.he ISE
criterion, we h a w

Likewise, dimensionless parameters and performanceindices for ot,her systems and for other inputs arederived.

C. Effects of Xoise in Adaptive Control [SI ]


I t can be readily shown that t,he noise, which is present
in t.he pla.nt states measurement or enters t.he plant as an
additivedisturbance,has
a desta,bilizing effect, on the
adapt,ive syst.em3 considered so far [30].I n part.icular although the stat,e error vect.or is bounded, the paramet,er error ma.y become unbounded duet o t,he noisc biasing action.
To see this t.he following equation, which representsa
general form of parameter adjustment, is considered.
(AT))

h ( e )Y2(em,e,).

(A.23)

The parameter error A p will approach zero when there is no


noise. If noise is present, both e and 6, will h a w noise
components.Hence, the expectedvalue of A p , which is
now a cross correlation of two noise dependent functions,
will cont,ain a dc term. The situationbecomes worse if t.he
input is also not active enough, as A p may not, approach
zero even without the noise.
Returning t.o the various designs considered inthis
paper, all except, the Dressler scheme suffer from the noise
biasing problem. The designs based on Lyapunov sta.bility
t.heory are still quite flexible in that t.hcbound of the
parameter error could be controlled using a modificat.ion
due t o Karendra et al. [29] even in the presence of bot,h
noise a.nd insufficient frequencies. For t.he particu1a.r case
of single plant gain adjust,ment, only the 3I.I.T. design
has t,henoise biasing problem.
REFEREXES
[I] P. V. Osburn, H. P. Whitaker, and 4 . Keezer, New develop
ments in the design of adaptive cont.ro1system, Inst. Aeronautical Sciences, Paper 61-39, 1961.
[2] P. C. Parks, Lxapunov redesign of model reference adapt.ive
IEEE Trans. Attiomat. Contr., vol. AC-11,
controlsystenls,
pp. 362-367, July 1966.
[3] D. D. Donalson and C. T. Leondes, A model referenced
parameter tracking technique for adaptive control systems,
IEEE Trans, A p p l . Ind., pp. 241-262, Sept,. 1963.
[4] R. M. Dressler, An approa.ch to model-reference adaptive control SysStems, IEEE Tmns. Automaf. Contr., vol. AC-12, pp.
7340, Feb. 1967.
[5] C. F. Price, An accelerat,ed gradientmethodforadapt,ive
control, Proc. 9th IEEE Symnp. Ada.ptbe Processes Decision
and Control, Dee. 1970, pp. IV.4.1-4.9.
[6] J. W . Green,Adaptivecontrol
of multi-loop speedcontrol
systems with particular reference t o the Ward-Leonardsystem,Ph.D.dissertation,
Dep. Elec. Electron.Eng.,Leeds
LTniv., Sept.. 1969.
[7] R. L. Butchart. and B. Shackcloth, Synthesis of model reference adaptive cont.rol systenls by Lyapunova second method,
Proc. 1965 IFAC Symp. AduptitleCon.tro1 (Teddington,England), ISA, 1966, pp. 145-132.
[8] P. H. Phillipson, Design methods for model reference adaptive
systems, Proc. Inst. iifechanical Engineers, vol. 1S3, part. I,
pp. 693-706, 196s-1969.
[9] J. W. Gilbart, R. Y. Nonopoli,and C. F. Price, Improved
convergence and increased flexibility in t.he design of model
reference adaptive control system, Proc. 9th IEEE Symp.
AdupticeProcesses Decision and Control, Dec. 1970, pp. IV.3.1.3.10.
[IO] It. 1. hIonopoli, J. W. Gilbart., and W
. U. Thayer,Model
reference adapt.ive control based on Lyapunov-like techniqueq,
Proc. 2nd IFAC S y m p . Syskm Sensitivity and Adaptivity
(Dubrovnik, Yugoslavia), Aug. 196S, pp. F.24-F.36.
[11] M. G. Currie and E. B. St,ear, State space struct.ureof model
reference adaptive control for a noisy system, Proc. 9n.d Asilomar Conf. Circuits and Syslstens, 1968, pp. 401-40.3.
[12] D. G. Luenberger, Observing the state of alinearsyst,em,
ZEEE Trans. X i l . Elrctron., vol. MILS, pp. 7440, Apr. 1964.
[13] C. A. Wimorand It. J. Roy, l)&gn of modelreference
adaptive cont,rol systemsby
Lyapunovs second method,
IEEE Trans. Autonmi. Contr. (Corresp. j, vol. X - 1 3 , p. 204,
Apr. 196s.
[14] I. D. Landau, A hyperstabi1it.y criterion for model reference
ada.ptive
control
syst.ems, IEEE Trans. Autonmt. Con.tr.
(Short Papers), vol. AC-14, pp. 532-555, Oct. 1969.
[I31 D. J. G. James, Stabilityof a model reference cont.ro1system,
AZdA J., vol. 9, pp. 950-92, May 1971.
[16] B. Port,er and 3.1. L. Tatnall, Ytahilit,y analysis of a clads of
multivaria,ble model reference adaptive systems having t.imevarying-process parameters, Int. J . Conir., vol. 11, pp. 32.5332, 19,o.
[l7] C. A. Winsor, Model reference adaptive design, NASA-CR98453, Nov. 1968.

C. T. Leondes d al., An invest.igation study on model reference


adnptive technicplea a:: applied to altitude control system for
larmch vehicles, SASA-CR-102846, 1969.
, J. E. Gibson: .\-onlincar
dufotnafic Cvnfrol. New York:
3Id~raw-Hi11,1963, ch. 11.
I). C. Ipson, (-nif.s,Uirnmsionsand Uirrwnsivtz!rs.s .\rcmthrrs
Sew Yolk: 3I&raw-TTill, 1960, ch. 9-11.
11. E. Knlman and J. E . Bertram. Control sJ-atem analysis and
de5ipn via the second method of Lynpunov, Trans. dS.lIE, J.
Basic Eng., pp. 371-393, June 1963.
11. 3Iargolir and C. T. Leondes, On the theory of adaptive
control systems:The
learning model approach, Proc. 1st
I F A C Congr. !3Io.<cow!, 1960, pp. 556-363.
T.C. Hsia and V. \,7jndvartirh,<-In on-line technique for
system identification, IEEE Trans. =1cctornnt. Contr. (Short
Papelnj, vol. AlC-14,pp. 32-!)6, Feh. 1369.
J. L. 3IedaniC. and 1. V. Iiokotovic, Some proplems in the
development of adaptive systems uing the sewitwit>- operator, Prvc. 1.965 I F A C S y m p . Adaptice Control (Teddington,
England j, ISB, 19F6, pp. 204-212.
P. J7. Kokotovic, J. V. Medanit, 8. P. Bingulac, and 11. I.
Vn%kovi6,Sensitivitymethodin
the esperirnent,al design of
adaptive control systems, Proc. 3rd IFAC Con.gr. (London),
1966, pp. 4.3B.1-3.12.
It. E. Kalman, Lgapunov functions for t.he problem of Lnr6 in
automatic control, Iroc. S a f i o n a l dcadrn1y o j Sciences
(L7.S.A.i, vol. 49, 1863, pp. 2Ol-3J5.
K. It. lleyer, Lyapunov functions for the problem of Lur6,
Prvc. Sational .Academy oj Sciencrs (U.S.A.), vol. 33, 1963, pp.
301-.jO3.
lemma in adaptive
R. x-. ?rIonopoli, .The Kalman-~-acrlbovicll
control system design, t o be published.
Tripathi, T h e choice of adaptive
I<. %.Sarendrs and S. S.
parameters in model reference control systenw, presented at
the 5th Xsilonlar Conf., 1 9 i l .
1). P. LindorB, Effects o! incomplete adaptation and disturbance in adaptive control, in 1972 Joint Atclvrnal. Cunlr. Con!.,
Preprints.
C. C. Hang, Ph.11. dissertation Dep.Eng., Karwick Univ.,
Coventry, England, 1973, to be submitted.

Chang-Chieh Hang ( S i O ) was born in


Singapore
on
March 25, 1048. He received theB.Eng.
degree (m-ith first class
honors) it1 electrical engineering from the
University of Singapore in 1970, and t,he
3I.Sc. degree in engineering cont.rol from the
University ofLVarkvick,
Coventry, England,
in 1971.
Since 19i0 he has held an oversea3 scholarf
ship, an-arded from the Royal Commission
for the Eshibition of 1851, and a fees award
from the British Council, while working towards the Ph.11. degree in
engineering control at. the University of Warwick. H i present. research itlt.erek are in thearea of stability theory as applied to
adapt.ive control atid parameter estimation.

Patrick C. Parks \vas born in London, England, on Xovember 24, 1930. He received t,he
B.ri. and Ph.1). degrees from Cambridge Univerait.>-,Cambridge,England,
in 1932 and
1970, respectively.
Since 1972 he h.as been Professor of Control
Theory and 1)irector of the Control Theory
Centre at the Ctliversit>- of Jt-arwick,
Coventry, England. Hi- current interests inclrtde stability theor>- and its applications,
dynamics of aircrafttrailingvortices
and
economic modeling. He has been in the C.S..l. on several occa$ions,
b0t.h as an S S F and as an N h d / ~ l ! C \kiting Fello\v.
Profesor Parks is a Fellow of the Institute of _\lathematic3 and its
Applications, an Associate Fellow of the Royal Aeronautical Society
and an Aasociat.eEditor of the IFACjournal dulornalica.

An Adaptive Observer for Single-Input


Single-Output Linear Systems
ROBERT L. CAR.ROLL SXD DAVID P. LINDORFF

Abstract-A f u l l order adaptive observer is described for observing the state of a single-input single-output observable continuous
differential system with unknown parameters. Convergence of the
observer states to those of the system is accomplished by directly
changing the parametersof the observer using an adaptive law based
upon Lyapunov stability theory. Observer eigenvalues may be freely
chosen. Some restriction is placed upon the system input in that it
must be s a c i e n t l y rich in frequencies in order to insure convergence.
Manuscript received Xovember 1, 19i2. Paper recommended by

G. 1.dnridia, Chairman of the IEEE 8 C d Adaptive and Learning

Systems, Pattern Recognition Committee. This work n-assupporfed


in part by theSational .Ieronautics and Space .ldminiatratlon
Research Grant SGL-07-002-002.
The authors are n-ith the School of Engineering, University of
Connecticut, Storrs, Conn. 06268.

I . ISTRODL-CTIOX

Tille

HE Lucnbcrgcr ubservcr [1]-[3] allow extraction of

state of an obscrvablc linear s?-stcm n-hen given


I ) thc systrm input. 2 ) the sJ-stcm output, 3) the form of
t h r systrm, and 4) t h r paramrtcr values of thr s>-stem.
111 tllosr cases for u-hirh the systrm paramrtcrs :\re unknwn-11, thr state obwrvation is subject t o error. Sonw
previous investigators o f paranwtrr ignorance in obscrvers
[4].[5] alleviateto somc dcgrce thcobscrration error,
but t h q - arc unablr to guarantcc that the error vanishes
or that their computational algorithm converges when the
nngnitude o f parameter igrlorancc is large. Tllr basics of a
full order adaptive obscrvrr ~ ~ - 1 l irwgates
cl~
these disad-

Вам также может понравиться