LUND UNIVERSITY PO Box 117 221 00 Lund +46 46-222 00 00
Adaptive Control -- A Way to Deal with Uncertainty
Åström, Karl Johan
1987
Document Version:
Publisher's PDF, also known as Version of record Link to publication
Citation for published version (APA):
Åström, K. J. (1987). Adaptive Control -- A Way to Deal with Uncertainty. (Technical Reports TFRT-7345).
Department of Automatic Control, Lund Institute of Technology (LTH).
Total number of authors:
1
General rights
Unless other specific re-use rights are stated the following general rights apply:
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal
Read more about Creative commons licenses: https://creativecommons.org/licenses/
Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
CODEN: LUTT.D2/(TFRT -7 545)lI-zz/(IsB7 )
ADAPTIVE CONTROL
-A way to deal with uncertainty
I(arl Johan Aström
Department of Automatic Control
Lund Institute of Technolory
Febmary 1987
Language English
S ecu rity class ificat io n
Number of pages 22
Author(s)
Karl Johan Å.ström
Department of Automatic Control Lund Institute of Technology
P.O. Box 118
5-22100
Lund
SwedenISSN and key title -TSBN
Recipient's notes inform,ation
S up pleme nt ary b iblio gra phical
Classifrcation systern index terms (if
Key w'ords Abstræt
This paper was presented at the DFVLR International Seminar "IJncertainty and Control', in Bonn, F.R.G., 1985, and published in J. Ackerman (Ed.), "Uncertainty and Control', Lecture Notes in Control and Infor- mation Sciences, No. 70, Springer Verlag, 1g8b, pp. 191-1b2.
The paper approaches the uncertainty problem from the point of view of adaptive control. The uncertainty is reduced by continuous monitoring of the ïesponse of the system to the control actions and appropriate modifications of the control law. It is shown that this approach makes it possible to deal with uncertainties that cannot be handled by high gain robust feedback control.
Title and subti¿le
Adaptive Control-A way to deal with uncertainty
Sponsonn6 o r ganisat io n Superwisor
CODEN : LUTFD2/(TFRT-7345) / I-22 / (1.es7)
Document Number Date- of issue
February 1987 Document name
INTERNAL REPORT
?'he teport rnay be ordered îrom the Departtnent of Automatic Control or bonowed througb tlrc university Library 2, Box 7070, 5-221 03 Lund, .Sweden, Telex: SB24B lubbis lund.
C ODEN: LIIIIT'D Z(TFRT- 7 3 4 Ð n -22 / (79 87 )
ADAPTIVE CONTROL
-A way to deal with uncertainty
IGrt Johan Åstuöm
Reprint from J. Ackermøn (Ed), oUncertainty and Controlo, Lecture Notes in Control and Informati,on Scíences, No. 70, Sprínger-Verlag, 7985, pp. 13 1-L52.
Department of Automatic Control
Lund Institute of Technology
February 1987
¿{datr>tir¡e Corrtrol
¡A. \Â/alz to Þeal ra¡it-h Lfncertaint:z
Karl Johan Âström
Department of Automatic Control Lund Institute of Technology Box 118, S-22L 00 Lund, Sweden
Abstract. This paper approaches the uncertainty problem from the point
of
view of adaptive control. The uncertaintyis
reducedby
continuous monitoringof
the resPonseof
the systemto
the control actions and appropriate modifications of the control law.It is
shownthat
this approach makesit
possibleto
dealwith
uncertaintiesthat
cannot be handledby
high gain robust feedback control.1.
INTRODUCTIONThe problem
of
reducing the consequences of uncertalnty has always beena
central lssue in thefield of
automatÍc sontrol. Black's invention of the feedback amplifier was motívated by the desireto
make electroniccircuits
less sensitiveto
thevariabitity of
elestronic tubes.The development of modern instrumentation technology has
similarly
made use of feedback inthe
formof the force
balanceprinciple, to
makehigh quality
instruments whichare
only moderately sensitiveto
variations intheir
components.Feedback by
itself
has theability to
reduce thesensitivity
of a closed loop systemto
plant uncertainties. Although this was one of the original motivationsfor
introducing feedback, the idea was keptin
the background during the intensive developmentof
modern control theory.Lately the problem has received renewed interest.
It is
now avery
active researchfield
andseveral new
schemesfor robust control have recently
been developed. Such shcemestypically result in
constant gain feedback controls, whÍchare
insensitÍveto
variations inplant
dynamics. Thepossibilities
andlimitations of
constant gain feedbackare treated
in Section2.
The purposeis to find out
whena
constantgain
feedback canbe
designed to overcome uncertainty in process dynamics and whenit
can not.An integrator where the sign
of
the gainis
not knownis
a simple example which can not be handledby
constant gain feedback. This examplewÍll
be used es anillustration
throughout the paper.The main goal of the paper
is
to approach elimination of uncertainties from the pointof
view of adaptive control. When the plant uncertainties are sucbtaht
they can not be handledby
a constantgain robust control law it is natural to try to
reducethe
uncertaintÍes byv
r
eFiqure 1. Simple feedback sYstem
experimentation and parameter estimation. Auto-tuning
is
a slmple technique' whfch has theattractive feature that an appropriate input sÍgnal to the process is
generatedautomatically. The method has the additional benefit
that
parameter estimation and control design are extremely simple to do. This is discussed in Section 3.Auto-tuning
is
anintermittent
procedure. The regulator has a special tuning mode, which is invokedon the
requestof an
operatoror
basedon
some automatíc diagnosis. Adaptivecontrol is a
method which allows continuous reductionof the
uncertalnties.An
adaptive regulatorwíll
continuously monitor the systems responseto
the control actions and modífythe regulator appropriately.
The charateristicsof
suchcontrol
schemesare
discussed in Section4.
Two categoriesof
adaptívecontrol
laws,direct
andindirect, are
discussed in somedetail.
Sometheoretical results on the stabílity of adaptive control
systems arereviewed
in
SectÍon 5.It is
foundthat
the standard assumptions usedto
prove thestability of direct
adaptivecontrol
schemesare
suchthat robust high gaiin linear control
couldequally well be
apptied. Adaptivecontrollers are
nonlinear feedback systems. There are other types of nonlinear feedback systems, which also can dealwith
uncertainties. One typeis
called universalstabitizer.
Such a systemis briefly
dissussedin
Section 6.Its
capability of dealingwith
an integratorwith
unknown gain is demonstrated.Stochastic control theory
is
a general method of dealing with uncertainties. In Section 7it
is shown how adaptive control laws can be derived from stochastic control theory. The example wÍth the integrator having unknown gainis
worked outin
some detail.2.
LIMTTATIONS OF CONSTANT GATN FEEDBACKConventional feedback can deal
with
uncertaintyin the
formof
disturbances and modelingerrors.
Before discussingother
techniquesfor dealing with uncertainty it is useful
to understandthe possibilitíes
andlimitations of
constantgain
feedback.For thÍs
purposeconsider the simple feedback system shown
in
Figure 1.Let
GO be the nominal loop transfer function. Assumethat true
loop transfer functionis
G = GO(l+L) dueto
model uncertainties.Notice
that
l+Lis
theratio
between the true and nominal transfer functions.Im (t.L)
Re (t*t)
Fiqure
2. The ratio
betweenthe true transfer function G(s) and the
nomÍnal transfer function GO(s) must bein
the shaded regionfor
those frequencies where GO(io) is large.The effect of uncertainties on the
stability
of the closed loop systemwill first
be discussed.The closed loop poles are the zeros of the equation
1+G (s) +
6(s)L(s) -
Oo o
Provided
that the
nominal systemis stable it follows from
Rouche's theoremthat
the uncertaintieswill
not causeinstability
provided thatlL(s) I
Slr*cor"l
lF- (2. r,
on a contour which encloses the
left
half plane. The consequencesof this
inequalitywÍIl
nowbe discussed. For large loop gains (2.1) reduces to
ll.(s) I 3 I
This
meansthat the relative uncertainty l+L
mustbe in the
shaded areain
Figure2. It
follows from Figure 2 that
if
the uncertainty in the phase of the open loop systemis
less thang
in magnitude i.e.larg(l+L) I S
<9then the closed loop system is stable provided that the magnitude of the
relative
uncertainty satÍsfieso<
For
these frequencies wherethe loop gain is high it is thus
necessarythat the
phaseuncertainty
is
less than 90o.At
the crossover frequency crreduces to
/2, t-coÉr
.p"
where the loop gain GO(io") hasunit
magnitude equation (2.1)lL (
ir¡
e)ts
)(2.2)
m
where
g* is
the phase margin.At
higher frequencies where the loop gainis
less than one the inequality (2.1) can be approximated bylL(s)
SThis means
that
large uncertaínties san be treated where the loop gainis significantly
less than one.Stability is only a
necessary requirement. To investigatethe effect of
uncertainty on the performanceof the
closedloop
system considerthe transfer
functionfrom the
commandsignal to the output i.e.
lr læt
+G G
o 1+L
G o o
o
1+G
1+GO+GOL oGo 1+G
1 +GOL
The
error in
the closed loop transfer functionis
thusL
(2.3'
c
1+GO+GOLThis error
canbe
made smalleither by
havinga
small openloop
uncertaÍnty(L) or
byhaving a high loop gain (GO).
Equations (2.1) and (2.3)
give the
essenceof
high gain robustcontrol.
The open open loop gain GO can be made largefor
those frequencies where the phase uncertaintyis
less than 90degrees.
At
those frequencies the closed loop transfer funstion can be madearbitrarily
close to the specifications by choosing the gainsufficiently
large. For those frequencies where the uncertaintyin
the phaseshift is larger
than 90o thetotal
loop gain must be made smaller than one in order to maÍntain robustness.At
the crossover frequency where the loop gain hasunit
magnitude the allowable phase uncertaintyis
givenby
Q.21. The allowable unsertainty dependscritically
on the phase marging*.
Assumingfor
examplethat it is
desiredto
have enerror in
the closed loop transfer function ofat
most tO% of. the crossover frequency. The allowable phase margin is given in Table 1.GG
LL
Table 1
-
Maxfmumerror
in the open loop transfer function whlch glveat
most 10 %error
oÍ the closed loop transfer function at the crossover frequency.9n 10 20 30 43 60
max lL I a. o77 o. o34 o. o52 o. 076
o.
100Design techniques which can
deal with
uncertaintyare given in
Gutman (1963), Horowitz andSidi
(1973), Leitmann (1990,1983), Kwakernaak (1985), discussion of the multivariable case is given by Doyle and Stein (1981).(19?9), Horowitz
çrüuet
(198s). AIt is clear
fromthe
discussion abovethat in order to
use robusthigh gain control it
is necessarythat
the transfer functionof
theplant
hasa
phase uncertainty less than 90o forsome frequencies. Some examples which
illustrate
the limitations of high gain robust controlwill
now be discussed.Example 2.1
-
Time DelaysConsider a linear plant where the major uncertainty
is
dueto
variationsin
the time delay.Assume
that
the time delayvaries
between Tmin"td T*"*.
Furthermore assumethat it
isrequired
to
keep the variationsin
the phase margin less than 20o.It
then followsthat
the cross-over frequency oc must satisfy(,ù
o.35
Tmex TmLn
The unsertainty
in
the time delay thus induees an upper boundto
the achievable cross-overfrequency.
ctExample 2.2
-
Mechanical ResonancesMechanical resonances are associated with transfer functions of the type 2
G(s)
2È 2f,,r^lOs
+
+
(¡)owhere the damping normally
is very
small. The phaseof
G changesrapidly
from 0to
-1800around oO. The gain also changes
rapidly
around oO.It
increases from oneto
approximately L/2E
andit
increases^" ,f;trz with
increasingo.
Variationsin !,
ando will
thus give substantial phase uncertainty. To achieve robust linear controlit is
then necessaryto
makesure that the loop gain
is
low aroundr0.
Thisis typically
achieved by a notchfilter.
trc
to
2
Example 2.3
-
Integrator Whose Sign is UnknownAn integrator whose sign
is
not known has elther a phaselag of
90oor
2?Ao. Sucha
system can not be controlled using high gain robustcontrol.
c:3.
AUTO-TUNINGWhen the uncertafnty ls such that robust htgh gain feedback cannot be applted
it ls
natural totry to
reducethe
uncertaintyby
experimentation. Auto-tuningis a
methodologyfor
doingthis
automatically. The principles are straightforward.A
modelof
the process dynamics is determinedby
making an identification experiment where aninput
signalis
generated andapplied
to
the process. The dynamicsof
the processis
then determined from the results ofthe
experiment. Thecontroller
parametersare
then obtained from some design procedure.Since the signal generation, the identification and the design can be made
in
many different ways there are many possible tuning procedures of this kind.Auto-tuning
is
also usefulin
another context. Thereare
ceses whereis is
much easier to apply an auto-tuner thanto
design a robust high gain controller Simple regulatorswith
twoor three
parameters canbe
tuned manuallyif there is not too
muchinteraction
between adjustmentsof different
parameters. Manualtuning is,
however,not
possiblefor
morecomplex
regulators. Traditionally tuning of
suchregulators
havefollowed the route
of modelingor identification
and regulator design.This is
oftena
time-consuming and costly procedure which canonly
be appliedto
important loopsor to
systems whishare
made in large quantities.Most adaptÍve techniques can be used
to
provide automatic tuning.In
such applications the adaptation Ioopis
simply switched on and perturbation signals may be added. The adaptiveregulator is run untíl the
performanceis satisfactory. The adaptation loop is
then disconnected and the systemis left
runningwith fixed
regulator parameters. Below wewill
discuss
a
specific auto-tuner which requiresvery little prior
information and also has theinteresting property that it
generatesan
appropriatetest signal automatically. This
is discussedfurther
Ín Âstrõm and Hägglund (1984a). A nice featureof
the technique dessrÍbed belowis that
an input signalis
generated automatically andthat the
parameter estimation and the control design arevery
simple. The input signal generatedis
automatically tuned to the characteristicsof
the plant.It will
haveits
energy concentrated around the frequencies where the plant has phase lagof
1800.The Basic Idea
A wide class of process control problems ean be described in terms of the intersection
of
the Nyquist curveof
the open loop systemwith the
negativereal axis,
whichis traditionally
T
0
-I
0 10 20 30 40
Fiqure
3.
Input and output signalsfor
a linear system underrelay
control. The system hasthe transfer function G(s) = 0.5(1-s)/s(s+1)(s+1).
described
in terms of the critical gain k, and the critical period Tc. A
method for determining these parameters was describedin Ziegler
and Nichols (1943).It is
done asfollows: A proportional regulator
is
connected to the system. The gain is gradually increaseduntil
anoscillation is
obtained. Thegain
whenthis
occursis the critical gain
and the frequencyof the ossillation is the critical
frequency.It is,
however,diffÍcult to
performthis experiment in such a way that the amplitude of the oscillation
is
kept under control.Relay feedback
is
an alternativeto
the manual tuning procedure.If
the process Ís connectedto
a feedback loop therewill
be anoscillation
asis
shownin
Figure3.
The periodof
theoscillation is
approximatelythe critical period. The
processgain at the
corresponding frequencyis
approximately given by2n p T
G tt-
l =-
etr4d(3.
1)where d
is
therelay
amplitude and ais
the amplitude of the oscillation.A simple
relay
control experiment thus gives the desired information about the process. This method has the advantagethat it is
easyto
sontrol the amplÍtudeof
thelimit cycle by
anappropriate choice
of
therelay
amplitude.A
simple feedback from the output amplitude tothe relay
amplitude makesit
possibleto keep the output
amplitudefixed during
the experiment. Noticealso that an input signal
whichis
almostoptimal for the
estimation problemis
generated automatically. This ensuresthat
thecritical point
can be determined accurately.When
the critical point
onthe
Nyquistcurve is
known,it is
stralghtforwardto apply
the classÍcal Ziegler-Nichols design methods.It is
also possibleto
devise manyother
design schemesthat
are based on the knowledgeof
one point on the Nyquistcurve.
The procedure canbe
modifiedto
determineother points on the
Nyquistcurve. An integrator
may be connectedin
the loopafter
therelay to
obtain the point where the Nyquist curve ÍntersectsRelay
PID
Process
-1
I
A u
T
Fiqure
4.
Block díagramof
an auto-tuner. The system operates as arelay controller in
the tuning mode (T) and as an ordinary PID regulatorin
the automatic control mode (A).the negative imaginary axis. New design methods, which are based on such experiments, are described
in
.4ström and Hägglund (1984b).Methods
for
automatic determinationof the
freguency andthe
amplitudeof the
oscillationwill be given to
completethe description of the
estimation method.The period of
anoscillation can be determined by measuring the times between zero-crossings. The amplitude may be determined
by
measuringthe
peak-to-peak valuesof the output.
These estimation methods are easyto
implement becausethey are
based on counting and comparisons only.More elaborate estimatÍon schemes
like least
squares estimatlonand
extended Kalmanfiltering
may also be usedto
determfne the amplitude and the freguencyof
thelÍmit
cycle oscillation. SÍmulations and experiments on industrial processes have indicatedthat little
isgained
in
practÍceby
using more sophisticated methodsfor
determiningthe
amplitude and the period.A block dÍagram
of
a control systemwith
auto-tuningis
shownin
Fígure4.
The system can operatein
two modes.In the
tuning modea relay
feedbackis
generated as was discussed above. When a stablelimit cycle is
establishedits
amplitude and period are determined asdescribed above and
the
systemis
then switshedto the
automaticcontrol
mode where aconventional PID control law
is
used.Practical Aspects
There
are several practical
problemswhich
mustbe solved in order to
implement anauto-tuner. It is e.g.
necessaryto
accountfor
measurementnoise, level
adjustment, saturation of actuators and automatic adjustmentof
the amplitudeof
theoscillation. It
maybe advantegeous
to
use other nonlinearities thanthe
purerelay. A relay with
hysteresisgives a system which
is
less sensitiveto
measurement noise.Measurement noise may give errors in detection of peaks and zero crossings. A hysteresls in the
relay is
a simple wayto
reduce the influence of measurement noise.Filtering is
anotherpossibility.
The estimatÍon schemes based on least squares and extended Kalmanfiltering
can be made less sensitive to noíEe. Simpte detection of peaks and zero crossíngsin
comblnatÍonwith
an hysteresisin
therelay
has workedvery well
in practice. See e.g. Aström (1982).The process output may be
far
from the desired equilibrium condition when the regulator is switched on.In
such casesit
would be desirableto
havethe
system reachits
equilibrium automatically. Fora
processwith finite
low-frequeney gain thereis
no guaranteethat
the desired steadystate will be
achievedwith relay control
unlessthe relay
amplitude issufficiently
large. To guarantee that the output actualty reaches the referencevalue' it
maybe necessary to introduce manual or automatic reset-
It is
also desirableto
adjust therelay
amplitude automatically. A reasonable approachis
torequire that the oscillation is a given
percentageof the
admissible swingin the
output signal.Auto-Tuninq with Learninq
Auto-tuning
is a
simple wayto
reduce uncertaintyby
experimentation.In
many cases the characteristicsof a
process may depend onthe
operating conditions.If it is
possible to measure somevariable
whích correlateswell with the
changing process dynamicsit
is possibleto obtain a
systemwith interesting
characteristicsby
combiningthe
auto-tunerwith a tabte look-up function.
Whenthe operating condition
changesa new tuning
is performed on demand fromthe
operator. Theresulting
perametersare stored in a
tabletogether
with
the variable which characterizes the operating condition. When the process has been operated over a renge covering the operating conditions the regulator parameters can be obtained from the table. A new tuningis
then reguired only when other conditions change.A system of this type
is
semi-automatic because the decision to tune restswith
the operator.The system
will,
however, continue to reduce the plant uncertainty.4.
ADAPTIVE CONTROLAdaptive
control is
another Ìvayto
dealwith
uncertainties.A
block-diagramof a
typical adaptive regulatoris
shownin
Figure5.
The system can be thoughtof
as composedof
two loops. The inner loop consistsof
the process and an ordinary linear feedback regulator. The parameters of the regulator are adjusted by the outer loop, which is composed of a recursive parameter estimator anda
design calculation.To obtain
good estimatesit
mayalso
beProcess þarameters
Regulator þarameters
U6
Fiqure 5. Block diagram of an adaptÍve regulator
necessary
to
introduce perturbation signals.This
functionis
omittedfrom the fÍgure
forsimplicity.
Noticethat the
system may be reviewed as an automationof
process modeling and design wherethe
process model andthe control
designis
updatedat
each samplingperiod.
The block labeled
fregulator
design"in
Figure 5 represents an on-line solutionto
a design problemfor
a systemwith
known parameters. This underlying design problem can be solvedin
manydifferent ways.
Design methods basedon on
phase-and
amplÍtude margins, pole-placement, minimumvariance control, linear quadratic
gaussiancontrol and
other optÍmization methods have been considered, see Astrõm (1983). Robust design techniques can of course also be used.The adaptive regulator also
containsa recursive
parameterestimator.
Many different estimation schemes have been used,for
example stochastic approxfmation,least
squares, extended and generalizedleast
sguares, instrumental variables, extended Kalmanfiltering
and the maximum liketihood method.
The adaptíve regulator shown in Figure 5
is
calledindirest or explicit
because the regulator parametersare
updatedindirectly via
estimationof an explicit
processmodel. It
issometimes possible to reparameterize the process so that
it
can be expressedin
termsof
the regulator parameters. Thisgives a significant
simplificationof the
algorithm because thedesign calsulations ere eliminated. In terms of
Figure5 the block labelled
design calculations disappears andthe
regulator parametersare
updateddirectly.
The scheme isthen called a direct
scheme.Direct and indirect adaptive regulators have
different properties which isillustrated
by an example.Design
Regulator
Estimation
Process u
v
Example 3.1
Consider the dissrete time system descrlbed by
y(t+l) + ay(t) = bu(t) + e(t+l) + ce(t) l=... -1,O, 1,... (4'
1) wherete(t)l is
a sequence of zero-mean uncorrelated random variables.If
the parameters a,b and c are known the proportÍonal feedback
u(t) = - Oy(t) = i: Y(t) (4.2)
minimizes the variance of the output. The output then becomes
y(t) = e(t) (4.3)
This can be concluded from
the
fotlowing argument. Considerthe
sit-uationat time t.
Thevariable
e(t+l) is
independentof y(t),
u(t) ande(t).
The outputy(t) is
known and the signalu(t) is at our
disposal. Thevariable e(t)
canbe
computed frompast
inputs and outputs.Choosing the variable u(t) so that the terms underlined
in
equation (4.1) vanishes thus makes the variance ofy(t+l)
as small as possible. This gives (4.2) and (4.3). For furtherdetails'
see .{ström (1970).Since
the
process (4.1)is
characterizedby three
parametersa straightforward explicit
self-tuner would require estimationof
three parameters. Estimationof the
parameterc
isalso
a
nonlinear problem. Notice, however,that the
feedbacklaw is
characterizedby
oneparameter
only. A
self-tuner which estimatesthÍs
parameter can be obtained based on the modely(t+1)=Oy(t)+u(t) (4.4'
The least squares estÍmate of the parameter
6
in this modelis
given byT y(k) ty(k+1) - u(k)l
e(t) k=l (4.5)
T
Y2(k)
k=1
and the sontrol law
is
then given byu(t) = - e(t)y(t) (4.6)
The self-tuning regulator given
by
(4.5) and (4.6) has some remarkable properties which can be seenheuristically
as follows. Equation (4.5) can bewritten
ast
t
t t t
y(t+1)y(t) I I t 2
t t t te(t) - 9(k) ly (k)
k=1 k=1 k=1
Assuming that
y is
mean square bounded and that the estimate O(t) converges ast
+o
we getI T (4.7'
t y(k+l)y(k) =
Qk=1
The adaptive algoríthm
(4.5),
(4.6)thus
attemptsto adjust the
parameterê so that
thecorrelation of the
outputat lag
oneis zero. If the
systemto be controlled is
actually governedby
(4.1)it follows from
(4.3)that the
estimatewill
convergeto the
minimum variance control law under thegiven
assumption. Thisis
somewhat surprising because the structureof
(4.4) which was the basisof
the adaptive regulatoris
not compatiblewith
the true system (4.1). More details are givenin
Âström and Wittenmark(1973,1985)
c¡IndÍrect AdaptÍve Control
An advantage of indírect adaptive control
is
that meny dífferent desígn methods can be used.The key issue
in
analysisof the indirect
schemesÍs to
showthat the
parameter estimates converge. Thiswill in
general requirethat
the model structure usedis
appropriate and that the input signalis persistently exciting.
To ensurethis it
may be nesessaryto
introduce perturbation sígnals. Provided that proper excitationis
provided there are nodifficulties
in controlling an integrator whose gain may have different sign.Direct Adaptive Control
The
direct
adaptive control schemes may workwell
evenif
the model structure usedis
notcorrect as wa
shownin
Example 3.1. The direct
schemeswill,
however,require
other assumptions. Assume e.g. that the process to be controlled can be described byA(q)y(t) = B(q)u(t) + v(t) (4.8)
where u
is
the input,y
is the output,v is
a disturbance and A9q) and B(q) are polynomials in the forwardshift
operator.Stability
of adaptive controlof
(4.8) have been givenby
Egardt (1979)' Fuchs (L979r, Goodwinet al.
(1980), Gawthrop (1980), de Larminat (1979), Morse (1980), and Narendraet al.
(1980). Sofar the stability
proofsare
availableonly for
some simple algorithms. The following assumptions are crucial:1
T
(.A1)
(A2t
I re(r tyz<*t - u(k)y(k)
rthe
relative
degree d = deg A-
deg Bis
known,the sign of the leading coefficient bO of the polynomial B(q)
is
known,t
lin
(á,3) (.â'4)
the estlmated model
ls
at least of the same order ag the processt the polynomlal B hasall
zeros inslde the unlt disc.The assumption A1 means that the time delay
is
known with a precision, which corresponds toa
samplingperiod. This is not
unreasonable.For
continuoustime
systemsthe
assumption means that the Togetherwith
assumption (42)it
also meansthat
the phaseís
knownat
high freguencies.tf this is
the cese,it is
possibleto
design a robust high gain regulatorfor
the problem, see Horowitz (1963), Horowitz andSidi
(1973), Leitmann (1979) and Gutman (1979).For many systems
like flexible aircraft,
electromechanical servos andflexible
robots, the maindifficulty in
controlis
the uncertaintyof
the dynamÍcsat
high frequencÍes, see Stein( 1980).
Assumption A3
is very restrÍctive,
sinceit
implies that the estimated model must beat
least as complex as thetrue
system, whích may be nonlinearwith distributed
parameters. Almostall
control systems are in fact designed based on strongly simplified models. High freguency dynamics are often neglected in the simplified models.Assumption A4
is
also crucial..It
arises from the necessityto
havea
model, whichis
linear in the parameters in thedirect
schemes.5. ROBUST ADAPTIVE CONTROL
For a long time the research on
stabilÍty of
adaptive control systems focussed on proofs ofglobal stability for all
valuesof the
adaptationgain.
Theresults
obtained under such premises are naturally quÍterestrictlve.
To get some insightinto
thÍs consider a continuous time system described byY =
G(P)u(5.
1)where u
is
the input,y
is the output, G Ís the transfer function of the system and p =d/dt
is the dífferential operator. Consider also the model reference adaptive control law given byu=S
TI
(5.2)
e=v-v'
'mwhere
y 'm is
the desired model output, e theerror
and O a vectorof
adjustable parameters.The components of the vector
g
are functionsof
the command signal. In a simple case, where the regulatorÍs
a combinationof
a proportional feedback and a proportional feedforwardr g becomesde
ãT= -
l{9eg=[r-yl
Twhere
r
Ís the reference signal.It
follows from (5.1) and (5.2) that$?- kerc(plsrer = ksy, (5.3)
This equation gives insÍght into the behavior
of
the system. Assume that the adaptation loopis
much slower than the process dynamics. The parameters then change much slower than the regression vectorg
and the term OtplgTein
(3.3) can then be approximatedby its
averaçJei.e.
etplgTe
^stG(plgTtel :e (5.4)
where
rt-rt
denotes time averages. Noticethat the
regressionvector g
depends on the paremeters. The following approxÍmation to (5.3) Ís obtained(5.5)
This
is the
normal situation becausethe
adaptive algorithmis
motivatedby the fact
the parameters change slower thanthe other
variablesin the
system underthis
assumption.Notice, howerer,
that it is not
easyto
guaranteethat the
parameters changeslowly
by choosing ksufficiently
small.Equation (5.4)
is
stableif
kgfCtplqTl is
positive. Thisis
true e.g.if
Gis strictly
positivereal and
if
the input signalis
persistentlyexciting.
However,if
the transfer function G(s) isstrictly positive real it is also
possÍbleto
designa robust high gain
feedbackfor
the system. We thusarrive at
the paradoxthat
the assumption requiredto
showstability of
the adaptive systemwill allow the
designof a
robust feedbask. The assumptionthat
G(s) isstrictly
positive realis,
however, not necessary asis
shown by the following example.Example 5.1
Consider a system where only a feedforward gain
is
adjusted andlet
the command signal be a sum of sinusolds i.e.og+
- ke(€)t'(pleTrelte o
ksYmf aosin
(oot
)n
r(t)
k=1
Using the model reference algorithm given
by
(5.2) the parameter estimates satísfyAssuming
that the gain ls small and uslng
everagêswe find that the
estímates areapproximately glven by
g9=krtl-el6(p)r
ctt
q9 = katl
ctt
- el (5.6)
where
n
(5.7
)2 k=1
The equation (5.6)
is
stableif a Ís
posítive. Considerfírst
the caseof a
slngle sinusoÍdal, D = 1r the equationis
then unstableif
the frequencyof
the command signalis
chosen so that G(it¡_) hasa
phase-shiftlarger
than 90o.If the input
containsseveral
frequenciesit
isn
necessary
that
the dominating contributionto
(5.7) comes from frequencíes where the phaseof G(it¡)
is
less than90o.
ct6. UNIVERSAL STABILIZERS
Adaptive
control
systemsare
nonlinear systemswith a
specialstructure.
Theyare
often designed based on the ideaof
automating modelÍng and desÍgn.It is
naturalto
askif
there are other typesof
nonlinear sontrols whÍch also can dealwith
uncertaintiesin
the process model.A special
classof
systems wêre generatedas
attemptsof solving the
following problem which was proposedby
Morse (1983). Consider the systemü. dt =aY+bu
where a and b are unknown constants. Find a feedback law of the form
u = f (9, y)
dg
m = g(0, y)
which stabilizes the system
for all a
andb.
Morse conjecturedthat
there are norational f
andg
whichstabilize the
system. Morse's conjecture was provenby
Nussbaum (1983) whoalso
showedthat there exist
nonrationalf
andg which stabilize the
system,e.g.
the following functionsa 1
T
é-2l<eos[arg G(it^rn)I
à A.
So
èJè)
a L
\ a (J
I
0
0
*¡È c¡
'Èò
S
-1
2
-4
-5
0 e- 2
L
\)\)
\3¡-
s
I
0123 0 0r23
Fiqure
6.
Simulation of an integrator with Nussbaum's control law.f (g, y) = yêzeos I
g(o, y)
v2This correspond to proportional feedback with the gain
k = e2sog
gFigure 6 shows a simulation
of this
control law appliedto
an integratorwith
unknown gain.Notice
that
the regulatoris initÍalized
sothat
the gain has the wrong sign.In
spíteof
thisthe regulator recovers
and changesthe gain appropriately.
Nussbaum'sregulator is
of considerableprincipal interest
becauseit
showsthat the
assumption A2is not
necessary.The
control law is,
however,not
necessarilya
goodcontrol law in a practical
situation becauseit mãy
generatequite violent control actions. The inÍtial conditions for
the simulation shown in Figure 6 were chosen quite carefully.7.
DUAL CONTROL THEORYUncertaÍnties can also be captured using nonlinear stochastÍc control theory. The system and
its
environmentare
then describedby a
stochastis model.To do so the
parameters are introducedas state variables and the
parameteruncertainty is
modeledby
stochastic models. An unknown constantis
thus modeled by thedifferential
eguationUç
u
Fiqure
?.
Block diagram of an adeptive regulator obtained from stochastic control theory.v
3?=0
with an initial distribution that reflects the
parameteruncertainty.
Parameterdrift
is modeledby
adding random variablesto
theright
hand sides of the equations. Acriterion
is formulated as to minimize the expected value of a loss function, which Ís a scalar functÍon of states and controls.The problem
of
fíndinga control,
which minimizesthe
expected loss function,is diffÍcult.
Under
the
assumptionthat a
solutionexists, a
functional equationfor the optimal
loss function can be derived using dynamic programming, see Bellman (1957,1961). The functional equation, whichis called the
Bellman equation,is a
generalÍzationof the
Hamilton-Jacobi equationin
classical variational salculus.It
can be solved numericallyonly in very
simple cases. The strustureof
the optimal regulator obtainedis
shownin
Figure 7. The controller can be thoughtof
as composed of two parts: a nonlinear estimator and a feedback regulator.The
estimator generatesthe
conditíonalprobability distribution of the state from
the measurements.ThÍs dístribution is called the
hyperstateof the
problem.The
feedbackregulator is a
nonlinear function,which
mapsthe
hyperstateinto the
speceof
control variables. This function can be computedoff-line.
The hyperstate must, however, be updated on-line. Thestructural simplicity of
the solutionis
obtainedat
theprice of
introducing the hyperstate, whichis
a quantityof very
high dimension. Updating of the hyperstate requiresin
general solutionof a
complicated nonlinearfiltering
problem. Noticethat there is
nodistinction between the parameters and the other state variables
in
Figure 7. This means that the regulator can handlevery
rapid parameter variations.The optimal
control law
hasinteresting properties
whieh have been foundby solving
anumber of specific problems. The control attempts
to drive
the outputto its
desired value,but it will
also introduce perturbations (probing) whenthe
parametersare
uncertain. This improves thequality of
the estimates and the future controls. The optimal control gives the correct balanse between maintaining good control and small estimation errors. The name dualNonlinear function
Hyþerstate
Process
Calculation
of
hyþerstate
control was coined
by
Feldbaum (1965)to
expressthis property.
Optimal stochastíc controltheory also offers other possibilities to obtain
sophlsticatedadaptive algorithms,
seeSaridis (L977r.
It is interesting to
comparethe regulator in
Figure7 with the self-tuning regulator
in Figure5. In
the adaptive regulator the states are separatedinto two
groups,the
ordinarystate
variablesof the
underlying constant parameter model andthe
parameters which are assumedto vary slowly. In the
optimal stochasticregulator there is
no such distinction.There
is
no feedback from the varianceof the
estimatein the
adaptive regulator althoughthis
informationis
avaílablein the
estimator.In the
optimal stochastic regulator there is feedback from the condÍtionaldistribution
of parameters and states. The design calculations in the adaptive regulatoris
madein
the same way asif
the parameters were known exactly.Finalty there are no attempts in the adaptive regulator
to
introduce the estimates when they are uncertain. In the optimal stochastic regulator the control lawis
calculated based on the hyperstate which takesfull
account of uncertainties. This also introduces perturbatÍons whenestimates
are poor. The
comparison indicatesthat it may be useful to add
parameter uncertainties and probingto
the adaptive regulator.A
simple exampleíllustrates the
dual control law and some approximations.Example 7.1
Consider a discrete time version of the integrator with unknown gain
y(t+1) =y(t) +bu(t) +e(t),
where u
is
the control,y
the output and e normal (0ro") white noise. Let thecrÍterion
be to minimize the mean square deviationof
the outputy.
ThisÍs a
special caseof
the system in Example 3.1with
a = 1 and c = 0. When the parameters are known the optimal control law is givenby
(3.2) i.e.(7. L'
(7.2' u(t) y(t)
bIf
the parameter b is assumed to be a random variable with a Gaussianprior distribution,
the conditionaldistribution of b,
given inputs and outputs upto
timet, is
Gaussianwith
meanb(t) and
standarddeviation d(t). The
hyperstateis then
charaeterizedby the triple (y(t)rb(t),o(t)).
The equationsfor
updatingthe
hyperstateare the
sameas the
ordÍnary Kalmanfiltering
equations, see Âström (19?0) and (1978).Introduce the loss function
t+N
VN
-2 min
E\-2
Lv Yt
=ct eu
k=t+
I
(k) (7.3'
$rhere
Y,
denotesthe data
avaltabteat tlme t l.e. (y(t), y(t-l)r...). By
lntroducing the normallzed varlablesn - y/6., I = b/cÍ,
tr-ur./y (7. 4'
it
can be shownthat V*
depends onq
andp only.
The Bellman equationfor
the problem can be wrÍtten asVr(q, p) = min
UT(n, Ê, p)(7.3'
where
v
(q,Ê) =
o oand
U
(n,
Ê, p) ( n-pn )2+1+
T
where
g is
the normalprobability
density andy=rì-pn+e / 1+
tË"1b=
unep +Ft1+
tË"12tË"1' * f vr-r(v, b)e(e)de 0'6'
qt
- att
2
e
pl(n,P) = arg
min (n-pn)
2+1+
sÉ.
n
see .{strðm (1978}. When the minÍmization
is
performed the control law is obtained aspT(n,P) = âFg min U, (n,Ê,t¡)
The minimization can be done
analytically
for T = 1. We get(7.7 '
tË"]'] _L
L*þ2
Transforning back
to
theoriginal
variables we getu(t) (7.8)
This
control
lawis
called one-step controlor
myopiccontrol
becausethe
loss function V, only looks one step ahead.=- I
b(t)
û2ttr
bt(t)+6'<t)
ôâtFor T > 1 the optimization can no longer be made
analytically.
lnstead we haveto resort
to numerical calculations. For large values of T the solutlon can be approxfmated byo. s6Ê
+9,
21.99
p(n,
þ) 2 + 4n>
2.2+0.089+p q(1.7 +
PThe control law
is
an odd functionin
r¡ andp,
see Âström and Helmersson (1983).Some approximations
to the optimal control law will also be
dÍscussed.The
certainty equivalence controlu(t) - y(l)/b (7-9)
is
obtained simply by takÍng the control law Q.24)for
known parameters and substitutÍng the parametersby their
estimates. The self-tuning regulator can be interpreted asa
certainty equivalence sontrol. Using normalized variables the control law becomest¡
I (7.9',)
The myoplc control law (7.8)
Is
another approximation. Thisis
also called cautlous control, becausein
comparisonwith
thecertainty
equivalence controlÍt
hedges and uses lower gain when the estimates are uncertaln. Notice thatall
control laws are the samefor
largep i.e if
the estimate
is
accurate. The optimal control lawis
closeto
the cautious controlfor
largecontrol errors. For
estimateswith poor
precision and moderatecontrol errors the
dual control gives larger control ections than the other control laws.A simulation of the dual control law
for
an integratorwith
variable gainis
shownin
Figure g. Noticethat
the gainvaries by
an orderof
magnitudeÍn
size andthat it
changes sign at T = 2000. In spiteof this
the regulator havelittte difficulty in
controllíng the process. Also noticethat
the regulator does probingwell
before the gain changes tÍme andthat it
jumpsbetween cautÍon and probing when the gain pesses through zero.
0
5
0
à a.
o
è)
o L
() s
<-s -si
s I
0
0 1000 2000 4000
Fiqure
8.
Simulation of dual control law applled to integrätor wlth variable gain.REFERENCES
Âström, K.J. (1970). Introduction to Stochastic Control Theory. Academic Press, New York.
Âström, K.J. (1978). Stochastic Control Problems. In Coppel, W.A. (Ed.). Mathematical Control Theory. Lecture Notes in Mathematics' Springer-Verlag' Berlin.
Âström, K.J. (1982). Ziegler-Nichols auto-tuners. Report CODEN: LUTFD2/TFRT-3167, Dept. of AutomatÍc Control, Lund Institute of Technology, Lund' Sweden.
Âström, K.J. (1983). Theory and applícations of adaptive control
-
A survey. Automatica, vol.19, pp. 4lt-487, t993.
Âström, K.J., and
T.
Hägglund (1984a). Automatic tuningof
simple regulators. Proceedings IFAC 9th World Congress, Budapest, Hungary.Âstrõm, K.J., and T.
Hägglund (1984b).Automatic tuning of simple regulators
with specifications on phase and amplitude marglns. Automatica,vol.
20, No. 5, Special Issue on Adaptive Control, pp. 645-651.Âström, K.J., and B. Wittenmark (1973). On self-tuning regulators. Automatica,
vol. 9,
pp.185-199.
Aström, K.J., and B. Wittenmark (1985). The self-tuning regulators
revisited.
Proc. ?th IFAC Symp. IdentifÍcation and System Parameter Estimation, York, UK.BeIIman, R. (195?). Dynamic Programming. Princeton University Press.
Bellman, R. (1961). Adaptive Processes
-
A Guided Tour. Princeton University Press.Doyle, J.C and G. Stein (1981). Multivariable feedback design: Concepts for
aClassical/Modern Synthesis. IEEE Trans. Aut. Control,
vol.
AC-ãO, pp. 4-16.2
T
0
-1
10