3146)/1-38/(1977). Wieslander, J . (1979a): Interaction in computer sysaidedanalysis and design control of tems. PhD thesis. DeptofAutomaticControl.

Lund Institute of Technology. Lund, Sweden. Report CODEN: LUTFDZ/(TFRT1019)/1-222/(1979). Weislander. J . (1979b): Designprinciplesfor computer aided design software. Preprints, IFAC Symposium on CAD of Control Systems. Zurich. 393. Wieslander. J . (1980a): Interactive programGeneral guide. Dept of Automatic Control. Lund Institute of Technology, Lund. Sweden. Report CODEN: LUTFD?/(TFRT3156)/1-30/(1980). Wieslander. J. (1980b): IDPAC commandsUser's guide. Dept of AutomaticControl. Lund Institute of Technology. Lund. Swe-

den, Report CODEN: LUTFD?/(TFRT3157)/1-108/(1980). conman&Wieslander, J . (1980~):MODPAC User'sguide. Dept Automatic of Control, Lund Institute of Technology. Lund. Sweden. Report LUTFDZRTFRTCODEN: 3158)/1-81/(1980). Wieslander. J . (1980d):SYNPACcommandsUser'sguide. Dept Automatic of Control. Lund Institute of Technology. Lund, Sweden. CODEN: Report LUTFD?/(TFRT3159)/1-130/(1980). Wieslander. J. and Elmqvist, H . (1978): INTRAC, A communicationmodule for interactiveprograms.Language manual. Dept of Automatic Control. Lund Institute of Technology, Lund. Sweden.ReportCODEN: LUTFD2/(TFRT-3149)/1-60/(1978) Wilkinson. J . H. and Reinsch. C. (1971):Linear Algebra. Springer-Verlag. Berlin.

Winston. P. H . and Horn, B. K. P. (1981): LISP. Addison-Wesley, Reading, Mass.

Karl Johen Astrom was born in Ostersund, Saeden on August 5 , 1934. He was educated at the Royal Institute ofTechnology (KTH) in Stockholm, Sweden. Heheld has various teaching positions at KTH,and he has worked for the Research Institute of Swedish Defense, and for IBM. In 1965hewas appointed to the chair of Automatic Control at Lund Institute of Technology (LTH). Hismainresearchinterests are stochastic control theory, identification, adaptive control, and computer aided design of control systems.Astromhasalso worked with industrial applications of of inertial automatic control in the fields guidance, paper mills, flight control, and ship steering. Apart from his professional work he and his family enjoy skiing and sailing.

A Simple Adaptive Smith-Predictor for Controlling Time-Delay Systems

A Tutorial, by A. Terry Bahill
Biomedical Engineering Program, Department Electrical Engineering, of Carnegie-MellonUniversity, Pittsburgh, PA 15213
ABSTRACT: This heuristic paper presents several simple techniques for analyzing the f stability o time-delay systems. It explains the Smith predictor control scheme for time-delay systems and shows how errors in modeling the plant parameters can causeinstability. Then anadaptiveconto the Smith predictor troller is added system; this pedagogical example offers a complete derivation of a simple adaptive control system. Finally, a new control scheme is discussed that allows zerolatency tracking of predictable targets by a time-delay system.
new system has good performance characteristics, but it tracks input signals with a timedelay. In some circumstances is it possible to design time-delay systems that trackpredictabletargets with nolatency 171, 181. The examples of this paper treat timedelay systems, the Smith predictor, and an adaptive control system. The examplesare complete and the derivations are explicit; no steps are omitted. Many research papers discuss adaptive control systems, but most of them are too complicated for the novice to understand; fewtextbookshaveincorporated simple examples of adaptive control systems. One purpose of this paper is tofillthisgap.Thispapershowssome simple techniques that can be used to gain insight about time-delay systems, explains theSmithpredictorcontrolscheme, and presents a complete, but simple, example of an adaptive control system. times (such as those required for making a chemical composition analysis, cortical processing of a visual image, analyzing a TV picture by a robot, or evaluating the output of a digital control algorithm) [I], [3], [7-lo]. Most elementary control theory textbooks slighttime-delay systems, because they are more difficult to analyze and design. For example, in time-delay systems initial conditions must be specified for the whole interval from -0 to 0, where 0 is the timedelay.For simplicity, in this paper I only discuss steady-state behavior, or equivalently I assume the initial conditions are zero. A unity-feedback, closed-loop control system with KGH = K/(Ts+ 1) has a transfer function of

Introduction
If a time delay is introduced into a well

tuned system, the gain must be reduced to

maintain stability [l]. The Smith predictor control scheme can help overcome this limitation and allow larger gains [2], but it is critical thatthemodelparameters exactly match the plant parameters [3-51. An adaptive control system [6] can be added to theSmithpredictor to change themodel parameters, so that they continually match thechanging plant parameters [3]. This ReceivedJune 22. 1982:revisedOctober 25, 1982:revisedJanuary 18, 1983. Accepted in revised form by Technical Associate Editor
F. Aminzadeh.

Y(S) -- -

K
7 s -I- 1

R(s)

fK

Why time-delay are plicated?

sysrems more com-

Time delays occur frequently in chemical, biological, mechanical, and electronic systems.Theyareassociated with travel times (as of fluids in a chemical process, hormones in the blood stream, shock waves in the earth, electromagnetic or radiation in space), or with computation
0272-1 70818310500-0016$1.0001983IEEE

This is stable for - 1 < K.If a time delay of the form e-@ is introduced in the forward path, stability is no longer guaranteed. The transfer function of such a system is Y(S> -Ke-Se (1)

R(s) -(TS f 1 f KeCSe)

The stability limits are not obvious. The does not exponential in the numerator

1 6

control systems magazine

Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

bother us, therefore,it will be left undisturbed. The exponential in the denominator will be approximated by an algebraic expression.The following four approxima; tion techniques have been suggested.
1. By mathematician, a series expansion:

the Taylor

Fig. 1. Block diagram for a typical time-delay system.

(2)

2. Byprocess a control engineer, Pade approximation:

the

lL
Fig. 2. System performance be could improved the if B could be fed back instead of the output Y.
Y S) K(2 - so) -( - R(s) 0 r s 2 + ( 2 r + 0 - K 8 ) s + 2 ( K 1 ) +

3. Bya digital controlengineer,the transform equivalent:

,-so

z-

-2 -

-nh

fictitious variable

where h is the system sampling period, n is an integer, and nh = 8.

The Smith Predictor

If a time delay were introduced into an optimally tuned system, the gain would have to be reduced to maintain stability. TheSmith predictor algorithm [2] avoids this reduction of gain and consequent poorer performance. The following development of the Smith predictor algorithm is based on Despande andAsh [I]. The block diagram for conventional control of a time-delay system is shown in Fig. 1. For simplicity, I will use theshorthand notation of Marshall [3]: R represents the system input, R(s); C represents the controller, C(s); L represents potential load disturbances,L(s); G, represents plant dynamics, G,(s); T, represents the plant time delay, T,(s); and Y, represents the plant output, Y,(s). For a simple first-order plant with a pure time delay G, = Kp/(rps 1) and T, = e -S*P. As shown in Fig. 1, the process can be conceptually split into delay free system dynamics and a pure time delay. I f the ficticiousvariable B could be measured, we could connect it to the controller, as shown in Fig. 2. This would move the time delay outside the control The loop. signal Y, would be the same as the signal B after a delay of 8,. Since there would be no delay in the feedback signal, the response of the system would be improved. Of course, this cannot be done in a physical system, because the time delay is probably distributed-not lumped-and there is no a priori reason to place the timedelay after the plant dynamics rather than before it.

4.
,-se

And by aclassicalcontrol engineer:

=

1
(1

+ s0/n)"

where n is a large number. The first technique implies that the original system has an infinite number of poles that can be reduced by using an approximation. If 8 is small so that se << 1 for frequencies of interest, then we can derive a single pole solution, by using e-@ = 1 - st?, in equation (1). Thetransfer function becomes

All denominator coefficients will be positive and the system will be stable if - 1 < K < 1 + 2718. These three approximations yield different stability limits. This should be expected because they are approximations. For the sake of comparison, we can use the Nyquist criterion [ 111 to derive stability constraints for equation (1). These constraints depend on the relationship of 8 to 7:
for 8 < < r stability requires
-1

< K < v' 1 + (1.57~/e)';

for 8 = r stabilityrequires

-- W S )

Ke-Se
( - K0)s 7

R(s)

-E (K -- I ) I
-1

< K < V' + (2r/eI2; I

and for 8

By the Routh-Hurwitz criterion this transfer function is stable if all denominator the coefficients are positive. This implies - 1 < K < 710. The following two-pole solution can be derived by substituting the first three terms of ( 2 ) into (1):

>> r

stability requires

-1

<K <

+ (nT/e)2.

Y(s) -- R(s)
KO's'

2KePse

f 2(7- K0)s

+ 2(K

1)

All denominator coefficients will be positive and the system will be stable if

0 < K < r/8.

Using the Pade approximation in equation (1) produces a third transfer function.

The Taylor series approximations err on the safeside. The Pad$: approximation could indicate stability for some unstable systems. The introduction of a time delay makes it moredifficult to access the stability of a system. The approximation methods shown here are not? in general, good methods for assessingthe stability of asystem.Sometimes they yield bizzareresults,as shown on p. 33 of [3]. However, they do allow us to make the following generalization. Large gains can only be used in timedelay systems if the plant time delay, 8 : issmall compared to the plant time constant, r .

may 1983
Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

17

I /

To improve the design let us model the plant as shown in Fig. 3. G , represents the model of the plant dynamics, T, represents the model of the plant timedelay, and E represents the error between the output of the model and the output of the plant. For the previous exampleof a f i t order pcg;s G , = K,/(T,s + 1) and T, = e Althoughthefictitious variable B is unavailable, B, can be used as the feedback signal, as shown in Fig. 3. This arrangement controls the model well, but not the overall system. The control of the system output, Y,, is open loop: it will not accommodate either load disturbances or inaccurate models. To compensate for these errors a second feedback loop is implemented using E, as shown in Fig. 4. This is the Smithpredictorcontrol strategy. The controller C is a conventionru PI (Proportional plus Integral), PD (Proportional plus Derivative), or PID (Proportional plus Integral plus Derivative) controller,which can be tuned more closely because the effect of the time delay in the feedback loop has been minimized. Sometimes the Smith predictor is drawn as shown in Fig. 5, which is equivalent to Fig. 4. The closed-loop transfer function of this system, for L = 0, is
Yp(s) -- CGPTP
1

lL

Fig. 3. Preliminary form of the Smith predictor.

IL

I
Fig. 4.

iI

____~
~

The complete Smith predictor control scheme.

R(s)

+ CG,

-CG, T,

+ CG,Tp
(4)

If G , = G , and T, = T, this reduces to CGpTp Yp(s)

R(s)

+ CG,

(5)

The effects of thetime delay have been removed from the denominator of the transfer function, and the system performance has been improved. However, it tracks input variations with a time delay. Assuming perfect model matching, the transfer function between load disturbance and system output becomes

~~

simplicity, we will not consider load disturbances again in this paper.

CG,T,

-1, or

- GpTp [ 1 + GmC(11 +CG,

Stability of the Smith Predictor Control System

T , ) ]

(6)

Once again there is no time delay in the denominator. However, the system tracks disturbances with a time delay. Thesystem has dynamic poor response unless load disturbancesare restricted to frequencies below 2/0, Hz, where 0, is the magnitude of the planttimedelayin seconds.For

The dashed box labeled G,, in Fig. 5, calIed the Smith predictor controller, is a stable closed-loop feedbackcontrol system with a transfer function of

The condition for

instability is CG,-

Clearly the magnitude the of righthand side is never one. Therefore, the controller G,, isstable.However, when this controller is used in a closed-loop system, the result may be unstable. What happens if the model does not match the plant exactly? Following the development of Marshall [3], let us set
control systems

1 8

magazine

Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

G,=Gp+AGp=G+AG

T ,

= , - ~ ( O ~ + A B , ) = e-sepe-sAep

= T,ATp = TAT

Thissimplified notation gets rid of the subscripts and allows us to redraw Fig. 5 as Fig. 6. Suppose that model the of the plant dynamics isin error, so that AG is non zero, but the model of the plant time delay is correct, so that T, = T, and AT = 1. The magnitudeof the time delay affects the controller via the A G term in Fig. 6. This is a positive feedback loop T of AG. Therefore,errors in estimatingtheplant gain or timeconstants could cause instabilities. If there is an error in modeling only the time delay so that A G = 0, then the inner positive feedback loop of Fig. 6 becomes zero. We are leftwith a simple closed-loop feedback control system with feedback of (1 -AT). The time delay in this feedback loop could cause instability. Alternatively, we can study the transfer function, equation (4),to see the effects of a mismatch between model and plant parameters. Let us use a simple plant, G , = G, = l/(s + I), a PD controller, C = 4(0.5s + I ) , aplant time delay of 1 sec, and a model time delay of 0.1 sec.
-

I
1A T

l?G

Fig. 6. A rearrangement predictor the of Smith model and plant mismatches. Based on [3].

illustrating effects of

Y,(s) (6s -- R(s)

+ 14s 4- 4)Tp s2 + 18s 4- 5

This transfer function appears to be stable by use of the Pad$: approximation and the Routh-Hurwitzcriterion, although it has been shown tobe unstable by the techniques of [4], [5].

An Adaptive Smith Predictor Control System

If is oftenadvantageous to change the controller to compensate forchanges in plant parameters caused by age, wear, temperature,fatigue,disease, oceancurrents,etc.Let us try to do this for the Smith predictor of Fig. 4. Let us presume that the plant time delay changes, and design a system that will automatically change the model time delay. The easiest way to do this would be to apply a step input and measure the output. From this wecould compute the plant time delay. Normally we cannot do this because applying a step input to an operating plant woulddisturb the process we wanted to control (this would become annoying if we did it every second). Therefore, we desire a generalized on-line method of adapting the model time delay. There are many different types of adaptive control schemes [ 121. One easily understoodscheme is model the reference adaptive control scheme discussed by Landau [ 6 ] .In one form of this scheme the controlled system, which is composed of a time-varying plant with adjustable parameters,isin parallel with awelldefined model.Theparameters ofthecontrolled system are varied to make the controlled system behave like the parallel model [ 131, [ 141. This description is close enough to

that of the Smith predictor to makethis type of adaptive control applicable. TWO major items must be chosen before an adaptive controller can be designed: a performance criterion a and method of function minimization. Typical performance criteria are minimization of time, energy, monetary cost: error,or square error. Typical function minimization techniques include gradient, conjugate gradient, and Fletcher-Powell.

The adaptive control scheme

Combining examples by Landau [6] and Marshall (31, let us minimize the square error between the model output and plant output, e = (y, - y,), using the gradient method of minimization. The performance function is

+ 4)Tp 3s 4- 5 4- (2s + 4)(Tp - T m )

(2s

Using the Pade approximation for the exponentials yields

T -T
P

= ,-s - e - o * l s
-

J = 0.5 e2 dt.
Let the magnitude of the plant time delay be represented by 0 , and assume its initial value, e,,, has been changed by a small amount called AO,. In the model reference approach we would now change-gains in the controlled system to minimize the performance function. However, for our Smith predictor we wish to modify the model time delay so that it tracks the plant timedelay thus minimizingtheperformance us function. The gradient method tells AO, = -kVJ. The gain k must be selectedforeachapplication. With only one parameter the gradient is simplythe partial derivative.

-36s

s2
This produces

+ 22s 4- 40

Y,(s) 2Tp(s3 -I- 24s2 + 84s f 80) __3s3 - S -I- 86s 4- 200 R(s)
The negative coefficient in the denominator shows thatthis is unstable. The Pad; approximation was used here for pedagogical reasons. It is not agoodgeneral technique for analyzing stability of Smith predictor systems. Its use indicates stability for manyunstable systems. The techniques of Ioannides,Rogers andLatham [4] or Palmor [5] are much more comprehensive. For example, if we use the Pade approximation on equation (4), with G , = I/(s + l ) , G, = 1/(3s + l), C = 4(0.5s + I), and plant and model time constants of1 sec, we get
may 1983

Themathematicsbecome complicated if we allow both OP and 8, to change continuously. Therefore, although 8, changes
1 9

Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

continuously, we will change 8, onlyat discrete intervals. We will let the identification and control phases alternate. First we identify changes in 6,.

Substituting (12) into (11) we get

AO,

= -k

I ;:

e -dt

aJ Ad, = -k The rate at which plant time delay char,ges is

Weare pleased with this result because the 6y,/dB,, called a sensitivity function, can easily be computed. From equation (4) we have

The pure derivative function shown in this figure would not be physically realizable analog with components. However, with a digital computer it could be' approximatedover any givenfrequency range. The boxes labeled G, and T, in the lower Smith predictor are supposed to contain the plant values, but they are not available. Therefore, we will have to use G , and T, and update them as often as possible.

Computing the gain k

d NOW, we stillhave to evaluate the constant k in (13). From equations ( 1 1 ) and (13) we have

= -AO, dt

Using Equation (9) Forsimplicitythetransferfunction has been labeled M(s), and the complex frequency argument, s , will now be omitted. Now, we calculate the partial derivative

If theadaptation

is slow. - 'e can interchange the order of differentiation

, =-kg P

5_

a
ae,

dt

dJ

a YP ae P
-

The error, e, is a function of theplant time delay. For a small change inthe time delay, AO,, usingthedefinition of the differential and assuming the initial value of this error is zero (so that A e = e), we get

- (-sCG,Tp)( 1 4- CG, -cG,T,)R (1

+ CG, + CG,T,

- CG, T,

l2
Substituting this and (12) into (15) yields

=-k-k

ae

dt

0.5

e2dt

Using (14) twice yields

e,=-

ae2

ae,

By the chain rule

Figure 7 shows one method for producing dy,/dO,. It has been assumed the that Laplace transformofay,(t)/a8, is equal to au,(s)/ae,.

ae, ae,
A 0 can be removed from the integral and

Finally,

Now, we can compute the desired change in the model time delay as

AO,

= AO, =

/6,

dt

=-k

I ::a
e-

dt

But now how do we get aelae,? To a first approximation y, is not a function of e,. (For each specific application this should be verified by simulation as was done by Marshall [3].) Therefore,

Fig. 7. Method on 131.

of

generating the

sensitivity function

ay,lae,.

Based

Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

then both sides can divided by kA0, to be yield

If the ay,/&3, is known, the usual case, then this constant could be precomputed. Otherwise, it would have to be calculated on line. This now gives complete the algorithmforchangingthemodeltime delay.

INVERT
Fig. 8 shows this implementation. Marshall [3] has simulated the system of Fig. 8 (which encompasses the system of Fig. 7) on both a hybrid computer and on a digital computer. It was stable and had satisfactory dynamicswhen the initial model time delay was wrong by as much as 8 0 8 . Thesystemadjusted its model time delay to within 10% c f the correct value in one time delay period, and settled at the correct value in two periods. For most linear systems there is a maximum value of gain that will ensure stability. If a time delay is introduced into suchsystem, the a then gain must be reduced. Use of a Smith predictor controller will allowthe gain to be restored to itsoriginalvalue.However, the system still tracks targets with a time delay. Recently many embellishments timefor delay systems have been reported. There are ways of dealing with nonlinear timedelay systems [15]; there are ways of applying optimal control techniques to the Smith predictor scheme [3], [16]; and there are ways of making certain systems track targets with zero-latency [ 7 ] , [8]. I will discuss only the last of these techniques.

Fig. 8. An adaptive controller on a Smith predictor. Based on [3].

I

SELECTIVE
CONTROLLER

y : : E : : "

1I

7 r b

rig. 9.

The target-selective adaptive control scheme.

indicates the vector transpose operation. Thedimensions of thevectors and matrices are such that the numerator and the denominator of(1 8) are scalars. The feedmust be back vector kT and the gain K selected to achieve stability. For zero-latency tracking output the must be identically equal to the reference input: y(t) = rs(t). Applying the requirement Y(s) = R,(s) to equation (18) produces:
,

system with a time delay in the forward path. The system input, ri(t). composed is of two parts; the reference source, rs(t), and the adaptive signal, r,(t). When rs(t) is not a known target waveform, r,(t) is turned off; ri(t) then equals r,(t) and the closed-loop transfer function becomes
VS) -

Zero-latency Tracking In Time-delay Systems

In some tracking systems target the position can be predicted, and it is posTo sible to produce zero-latency tracking. do this one does not simplypredictthe futuretargetpositionandfeedthisinto the time-delay system. Instead, one computes an adaptation signal that depends on thetargetmovement,aswell as on the timedelayanddynamics of theplant. Thisadaptationsignalisappliedtothe time-delay system, and it allows zerolatency tracking [7], 181. Figure 9 shows this scheme applied a to simple state-variable feedback control

hT(sI - A)-l bKe-S'

1 kT(sI--A)-'

bKe-Se

(18)

hT(sI- A)- bKe-,'

%=[

1 kT(sI -A)-

'

'bKe-,'

The e-'* term in the numerator is a pure timedelaythatremains in spite of the feedback. The e-" term in the denominator produces the phase lag that reduces theallowablegain.Theothersymbols represent Y(s) (thescalaroutput), Ri(s) (thescalarsysteminput), I (then x n identity matrix), A (the n x n system matrix), K (the scalar gain), kT (the 1 x n feedbackcontrolvector), hT (the 1 x n output coefficient vector), and b (the n x 1 input coefficient vector). Superscript T

(R,

+ R,)
e''

1
(19)

Solving for R, yields:

Ra =

kT(d - A)+ hT(sI- A)-'b'b - 1 1 Rs

hT(sI - A)-'bK

may 7 9 8 3
Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.

LI

If the time delay 0, the gain K, the matrix b , kT, and hT are known, and if r,(t) can be estimated, then ra(t) can be computed. The output can be made equal to the input, compensating for both the time delay and the plant dynamics.Thiscontrolscheme has onlybeen studied for single-input single-output systemswithscalarKandonlyonetime delay. One reason studying zerofor such latencytracking is thatthehumaneye movementcontrolsystemseemsto use target-selective adaptive control. The human oculomotor systemhas a 150 msec time delay. When a target starts moving there is a 150 msec delay before the eye starts moving. When the target stops, the eye continues to follow the predicted target for about 150 msec. However, a human can track a predictable target, such as one the shown in Fig. 10, without latency or phase lag. The top trace shows the target position (dotted) andtheeye position (solid), the and bottom trace shows target velocity (solid) and eye velocity(dotted).Thetargetmoved f5 degrees. The time axis is labeled in seconds. The target position waveform was that of a repeated cubic segment. model A was built to help explain how the human couldovercomesuchatimedelayand track with no latency. The model performed as well as the human [7], [8]. The model required knowledge about plant dynamics, plant time delay, target amplitude,targetfrequencyandtargetwaveform.

A , andthevectors

Fig. 10. Human zero-latency tracking of a cubic target waveform.

Conclusion
To maintain stability of a control system after a time delay is introduced, the gainmustbereduced.TheSmithpredictor algorithm larger allows gains. However, it requires an exact matchingof model and plant parameters. An adaptive control loop added to the Smith predictor can automatically adjust model parameters match time to the varying plant parameters. resulting The system still tracks targets with a time delay. certain In circumstances controllers canbe designed to track targets with no latency in spite of plant time delays.

Elements of Computer Process Control With Advanced Control Applications. Research Triangle Park NC: Instrument Society of America, 1981. pp. 227-250. [2] 0 . J. M. Smith, Closer control of loops Chemical Engineering with time, dead Progress. vol. 53(5). pp. 217-219. 1957. [3] J . E. Marshall. Control of Time-Dela). Sysrems. Stevenage United Kingdom: Peter Peregrinus Ltd.. 1979. C . Ioannides, G . J . Rogers and V . [4] A. Latham.Stabilitylimits of aSmithcontroller in simple systems containing a time 557delay. Inr J Control. vol.29.pp. 563.1979. Stability properties of [5] Z . Palmor, Smith dead-time compensator controllers, lnt J Control, vol. 32. pp. 937-949, 1980. [6] I. D. Landau, Adaptive Control: The New York: Model Reference Approach. Marcel Dekker Inc. 1979. [7]A. T . Bahilland J . D. McDonald.The smooth pursuit eye movement system uses anadaptivecontrollertotrackpredictable targets, in Proc Int Conf Cxbern Soc.New York: IEEE, 1981.pp. 269-278. [SI J . D. McDonald, A. T. BahillandM. B. Friedman. An adaptive control model for human and movements head eye while walking. IEEE Trans Svst Man Cybern.

[I31 J. W . Gilbart and G. C. Winston, for an optical Adaptive compensations trackingtelescope, Automatica, vol.10, pp. 125-131, 1974. 1141 - - C. G . Kallstrom. K . J. h o m . N. E. Thorell, J . Eriksson and L. Sten, Adaptive Automatica, vol. autopilotsfortankers, 15, pp. 241-254, 1979. A design of [I51 M. Jamshidi, three-stage non-linear systemstimecontrol with delay, Int J Control, vol.21,pp. 753762, 1975. [I61 K. Watanabe M. A and Ito, processmodelcontrolforlinearsystemswithdelay, IEEE Trans Aurom Control, vol. AC2 6 , p p . 1261-1269,1981.

vol. SMC-13, in press, 1983.

[9]M. J. Mendel. J . Kormylo, F. Aminzadeh, J. S . Lee and F. Habibi-Ashrafi, A novel approach to seismic signal processing andmodeling, Geophvsics. vol.46.pp. 1398-1414.1981. [ 101 G . J . Agin, Real time control of a robot with a mobile camera, in 9th International Symposium on Industrial Robots, March 1979: Washington D.C.. 1979. pp. 57-70. [I ] F. L. N-Nagy and M. N. AI-Tikriti, Stabilitycriterion of linearcontrolsystems with time delay, Meas Control, vol. 3, pp. 86-87, 1970. [I21 K. S. Narendra R. and V . Monopoli. Applications of Adaptive Control. New York: Academic Press, 1980.

Acknowledgements

I thank Jeffery S. Kallman and two perceptivereviewersforconstructivecriticisms.Thisresearchwaspartiallysupported National by Science Foundation grant ECS-8121259.
References
[I]

P. B. Deshpande and

R. H.

Ash,

A . Terry Bahill (S66-M68-Sh481) i n o r ns b wa Washington. Pennsylvania, on January 31, 1946. He received the B.S.E.E. degree in electrical engineering from the University of Arizona. Tucson, in 1967; the b1.S.E.E: degree in electrical engineering from San Jose State Univesity. San Jose, California, in 1970; and the Ph.D. in electrical engineering and computer science fromthetiniversity of California, Berkeley. in 1975. He served as a Lieutenant in the U.S. Navy teaching mathematics and electrical engineering for four years to the students of the Navy Nuclear Power School, Mare Island, California. He is an Associate Professor of Electrical and Biomedical Engineering at Carnegie-Mellon University and also an Adjunct Associate Professor of Neurology at the Universitv of PittsburghMedicalSchool.Hisresearchinterests include control theory, modeling physiological systems, head and eye coordination, and computer text and data processing. He is the author of Bioengineering: Biomedical, Medical, and Clinical Engineering (Prentice-Hall, 1981).

22

control systems magazine

Authorized licensed use limited to: The University of Arizona. Downloaded on July 09,2010 at 17:46:05 UTC from IEEE Xplore. Restrictions apply.