You are on page 1of 6

Adaptive Inverse Control: A Signal Processing Approach, Reissue Edition

by Bernard Widrow and Eugene Walach


Copyright 02008 the Institute of Electrical and Electronics Engineers.

Appendix C

A Comparison of the
Self-Tuning Regulator of
blstrom and Wittenmark with
the Techniques of Adaptive
Inverse Control

The best-known adaptive control methods are based on the self-tuning regulator of Astrom
and Wittenmark. Their 1973 paper [2] has had great influence worldwide in the field of
adaptive control. Chapter 3 of their book entitled Adaptive Control [ 13 summarizes their
work on the self-tuning regulator. Figure C. 1 is a generic diagram of the self-tuning regulator, based on Fig. 3.1 of Adaptive Control.
Process parameters

Design

Estimator

Regulator

Process

Figure C.l The self-tuning regulator based on Fig. 3.1 of K.J.ASTROM, and B . W I T T E N M A R K , A & ~ ~ ~ ~ ~
Conrrol (Reading, MA: Addison-Wesley, 1989).

The system of Fig. C.l is linear and SISO, and it works in the following way. The
process or plant is excited by its input u. Its output is y . This output contains a response

363

Comparing the Self-Tuning Regulator with Adaptive Inverse Control

364

App. C

to u, plus plant disturbance. An estimator, receiving both the signal input and signal output
of the plant, estimates the plant parameters. These estimates are fed to an automatic design
algorithm that sets the parameters of the regulator. This regulator could be an input controller, or a controller within the feedback loop, or both. For convenience, we have redrawn
the diagram of Fig. C. 1 in Fig. C.2.

uc

Plant
disturbance
Input
controller

Figure C.2

An alternative representation of the self-tuning regulator.

In Chapter 8, we demonstrated that the adaptive disturbance canceler of Fig. 8.1 minimizes the plant output disturbance power. In fact, we have shown that no other linear system, regardless of its configuration,can reduce the variance of the plant output disturbance
to a level lower than that of Fig. 8. I . Comparing the self-tuning regulator of Fig. C.2 with
the adaptive disturbance canceler of Fig. 8.1, the question is: Can the feedback controller of
the self-tuning regulator be designed to cancel the plant disturbance as well as the adaptive
disturbance canceler? Another question that arises is: Can an input controller be designed
for the self-tuning regulator so that when it is cascaded with the plant and its feedback controller, the entire control system will have a transfer function equal to the transfer function of
a selected reference model? It is not obvious that the self-tuning regulator and the adaptive
inverse control system will deliver performances that are equivalent to each other.

C.l

DESIGNING A SELF-TUNING REGULATOR TO BEHAVE LIKE


A N ADAPTIVE INVERSE CONTROL SYSTEM
To address these issues, we redraw Figs. C.2 and 8.1 as Figs. C.3 and C.4, respectively, in
order to simplify and bring2ut essential features of these block diagrams. For simplicity, we
have drawn Fig. 8.1 with P ( z ) approximated by P ( z ) . Also, we included a necessary unit
delay z-' within the feedback loop of the self-tuning regulator that will be necessary only if
there is no delay either in the plant or in the feedback controller. We will assume that P ( z )
is stable. If the plant is not really stable, let it be stabilized by a separate feedback stabilizer
and let P ( z ) represent the stabilized plant. This creates no theoretical problems for the selftuning regulator, and as demonstrated in Appendix D, creates no theoretical problems for
adaptive inverse control. To compare the two approaches, we need to first show, if possible,
that the transfer function from the plant disturbance injection point to the plant output is

Sec. C.l

Designing a Self-Tuning Regulator

365

the same for the self-tuning regulator as for the adaptive disturbance canceler. For the selftuning regulator of Fig. C.3,this transfer function is

Plant disturbance

Wz)

2
Unit delay

Input
controller

FC(z)

Feedback
controller
Figure C.3

Another representation of the self-tuning regulator.

Plant disturbance
N(z)
Plant
output

Input

l-$~--j7+-1
Figure C.4

Another representation of the adaptive plant disturbance canceler

For the adaptive disturbance canceler of Fig. C.4, this transfer function is
In order for these transfer functions to be equal, it is necessary that

To obtain F C ( z ) , we need Q ( z ) and P ( z ) . In practice, both wouldbe readily available (to


a close approximation) from adaptive processes already described. So there would be no
problem in getting a good expression for F C ( z ) .

366

Comoarina the Self-Tunina Regulator with AdaDtive Inverse Control

ADD. C

If the feedback controller given by Eq.(C.3) were used in the self-tuning regulator,
would the resulting system be stable? The answer is yes. Referring to Fig. C.3, we may
note that the transfer function through the feedback loop to the plant output is

This transfer function is stable since both P ( z ) and Q ( z ) are stable. SO far SO good.
To control the dynamic response of the system, to make it behave like the dynamic
response of a selected reference model, we multiply the transfer function ((2.4) by ~ C ( Z )
and set the product equal to M ( z ) :

M ( z ) = ZC(Z) . Z - I

P (z) (1 - z-' P ( z ) . Q ( z ) ) .
9

(C.5)

Accordingly,

- M ( z ).FC(z)
Z-I

P (z)* Q ( z )'

For the entire self-tuning regulator to be stable, it is necessary that ZC(z)be stable. Stability
has already been established for the rest of the system. M ( z ) is stable. Since P ( z ) and Q ( z )
are stable, they would not cancel any unstable poles of F C ( z ) that may occur. It is necessary,
therefore, for F C ( z ) to be stable in order for ZC(z) to be stable, although this is not sufficient
for stability. I C ( z ) will be unstable if either P ( z ) , Q ( z ) ,or both are nonminimum-phase.
How would one build a self-tuning regulator if its feedback controller F C ( z ) were
unstable? There are two possibilities. One possibility would be to choose an F C ( z ) that is
stable but not optimal for plant disturbance canceling. The other possibility would be to use
the optimal, unstable F C ( z ) inside the feedback loop and build the input controller ZC(z)
having the indicated poles and zeros but allowing components of ZC(z) to be noncausal
as required for stability, The noncausal filter could be realized approximately with an appropriate delay. The entire system response would be a delayed version of the response of
M ( z ) . These difficulties are not encountered when the optimal F C ( z ) is used and the input
controller ZC(z) is stable.

C.2

SOME EXAMPLES
Specific examples will help to clarify some of the issues. Suppose that the plant disturbance
is constant, that it is a Dc offset or bias. The question is: How well do the adaptive disturbance canceler and the self-tuning regulator handle this disturbance? For the adaptive
disturbance canceler, the transfer function from the plant disturbance injection point to the
plant output is given by (C.2). We would like this transfer function to have a value of zero
at zero frequency, that is, at z = 1. This is easily accomplished as follows:
1

1 - P ( 1 ) . Q(1) = 0, or Q(1) = -

W)'

(C.7)

Sec. C.3

Summary

367

Any form of Q ( z ) would allow perfect canceling of the constant disturbance as long as the
value of Q at z = 1 is the reciprocal of the value of P at z = 1.
So the adaptive plant disturbance canceler will perfectly eliminate constant plant disturbance. What will the self-tuning regulator do with this disturbance? Its transfer function
from plant disturbance injection point to plant output point is given by (C.1). This transfer
function should equal zero at zero frequency to eliminate the DC plant disturbance. Accordingly,

Assuming that the plant transfer function is well behaved at z = 1, it is clear that F C ( z )
must be infinite at z = 1. One way to accomplish this would be to let F C ( z ) be a digital
integrator,
1
FC(z)= 1---I'
(C.9)
~

giving it a simple pole at z = 1. The input controller has the transfer function given by (C.6):

( C .10)

P ( z ) and Q ( z ) are both stable and finite at z = 1. If in addition M ( z ) has a finite value
at z = 1, I C ( z ) will have a pole at z = 1 and will thereby be unstable making the entire
system unstable. A noncausal realization of I C ( z ) would not be useful in this case. The
only possibility would be to choose an F C ( z ) having its pole slightly inside the unit circle,
sacrificing some disturbance canceling capability for a stable I C ( z ) .
The same kind of result would be obtained if the plant disturbance were a constantamplitude sine wave. The adaptive disturbance canceler would adapt and learn to eliminate it perfectly. The self-tuning regulator would either be unstable or, if stable, would give
somewhat less than optimal disturbance reducing performance.

C.3

SUMMARY
The self-tuning regulator has an easier job of disturbance reduction and dynamic control
than adaptive inverse control when the plant is unstable with a pole or poles on the unit circle
or outside the unit circle. Feedback used by the self-tuning regulator has the capability of
moving the plant poles inside the unit circle to stabilize the plant, reduce disturbance, and
control its dynamics. For adaptive inverse control, the first step would be to stabilize the
plant with feedback. The choice of feedback transfer function for stabilization would not be
critical and would not need to be optimized. The only requirement would be to somehow
stabilize the plant. Then adaptive inverse control could be applied in the usual manner. A
discussion of initial stabilization for adaptive inverse control systems is given in Appendix
D.
A difficult case for both approaches occurs when the plant has one or more zeros on
the unit circle. The inverse controller tries to put the FIR equivalent of a pole or poles on top

368

Comparing the Self-TuningRegulator with Adaptive Inverse Control

App. C

of the zeros. The inverse impulse response does not die exponentially but instead persists
forever. An optimal FIR inverse filter cannot be constructed. The self-tuning regulator of
Fig. C.3 also cannot cope with such a problem. Its feedback cannot move around the zeros.
It too would try to choose ZC(z) to put poles on top of these zeros, but this would make
ZC(z) unstable and thereby make the entire system unstable. A remedy that would work
for both approaches would be to cascade with the plant a digital filter having poles matching
the unit circle zeros of the plant and to envelop the cascade within a feedback loop. If, for
example, the plant had a single zero on the unit circle at z = 1, the cascade filter would be
a digital integrator. The loop around the cascade would need to be designed to be stable for
adaptive inverse control.
When the plant is nonminimum-phasewith zeros outside the unit circle, we have seen
in Chapters 5,6 and 7 how adaptive inverse control can readily cope with such a plant. Dealing with this kind of plant with a self-tuning regulator is difficult, much more difficult than
dealing with a minimum-phase plant. The literature is not clear on how this can be done.
Comparing adaptive inverse control with the self-tuningregulator reveals cases where
one approach is advantageous, and cases where the other approach is advantageous. Also,
there are many cases where both approaches give equivalent performance although the system configurations and methods of adaptation are totally different.

Bibliography for Appendix C


[ l ] K.J.ASTROM,and B. WITTENMARK,
Adaptive control, 2nd ed. (Menlo Park, CA:
Addison Wesley, 1995).
[2] K.J. ASTROM,and B.WITTENMARK,
On self-tuningregulators,Aufomafica,Vol. 9,
No. 2 (1973).

This situation is not so clear. Refer to simulation examples in Chapter 12.

You might also like