OPTIMIZING THE SETTLING TIME WITH ITERATIVE FEEDBA...

14th World Congress ofIFAC

Copyright © 1999 IFAC 1-3b-08-3

14th Triennial World Congress~ Beijing, P.R. China

OPTIMIZING THE SETTLING TIME WITH

ITERATIVE FEEDBACK TUNING 1

Olivier Lequin * Michel Gevers ** Lionel 1:riest **

... Solvay S.A_, rue Solvay 3D, B-51()O Jemeppe-sur-Sambre,

Belgium
Email: [email protected]
** Centre for Systems Engineering and Applied j"1echanics
(CESAAfE)
Universite Catholique de Louvain
B-1348 Louvain-La-l\leu've, Belgium
Email: ge·[email protected]

Abstract: '\le present a variant of the Iterative Feedback Tuning (1FT) method in
which time weightings are used in the minimization criterion. A particularly useful
application of this idea is when zero weightings are put on the transient phase of
the step response of the system, thereby focusing on a rapid tracking of the desired
reference change rather than on the shape of the transient response. By varying the
size of this ~~zero weighting window", or mask, one can optimize the settling time of
the closed-loop system. Copyright © 1999lFAC

Keywords: Control system design, Iterative methods, Settling times

1. INTRODUC1.'ION mechanical systems and chemical plants, can be

found in (Hjalmarsson et aI., 1998).
rrhe Iterative Feedback Thning method is a model-
In the classical formulation of 1FT, as developed in
free technique for the optimization of the param-
(IIj almarsson et al., 1994), the linear closed loop
eters of a controller of fixed structure using only
system of Figure 1 was considered, where P is the
signal information on the closed-loop system. The
unknown plant to be cOlltrolled, C ~ [Or, C y } is a
method was initia11y derived in (Hjalmarsson et
two degree of freedom controller, v is an unknown
aI., 1994) and has quicJdy proved its efficiency in
disturbance, and r, U, and y are the reference t the
both laboratory and industrial applications: see
control signal and the output signal, respectively.
e.g. (Hjalmarsson et al~, 1995), (Lequin, 1997).
It ha..."3 also given rise to a number of exten-
sions, notably for the tuning of nonlinear feed-
back loops (DeBruyne et al., 1997), (Sjoberg and
Agarwal~ 1996), (Hjalmarsson, 1998)~ A recent y
and rather complete presentation of the theory,
as well as applications to controller tuning for

1 This paper presents research results of the Belgian Pro-

gramme on InteruniveI'sity Poles of Attraction. initiated
by tlle Be-Jgian State, Prime Minister's Office for Science~
Technology- and Culture. The scientific responsibility rests
with its anthors. Fig. 1. Actual closed loop system

4659
Copyright 1999 IFAC ISBN: 0 08 043248 4
OPTIMIZING THE SETTLING TIME WITH ITERATIVE FEEDBA... 14th World Congress of IFAC

The cont.roller has a given structure in which some signaL ThtlS, in this paper the following criterion
free parameters need to be tuned to achieve a is examined:
desired objective. The vector of controller param-
eters is denoted p~ In (Hjalmarsson et aL~ 1994)~
the following quadratic criterion was adopted : J=(p) = 2~E [~(LYYt(p))2
+ ), t(LuUt(P)V] · (4)

I-Iere i}t (p) is the error between the output Yt (p) \\re say in such case that a mask of length to is put
of the actual system controlled by the controller on the transient response of the tracking error.
C(p) ~ [C.,..{p), Cy(p)] and a desired output sig- The rnotivation for the u..~e of such masks is a.~
nal Yd, Ut(p) is the control signal, L y and Lv. follows.
are frequency weighting filters, ..\. expresses the
relative importance of the penalty on the control
signal versus the tracking error, N is the num- One of the frequent practical uses of controller
ber of data points, and E stands for expected design is to tune a controller of fixed structure
value~ The output of the actual system~ Yt(p), and (for example a PlO controller) in such a way that
the control signal, Ut(p), are explicitly shown to the step response of the closed-loop systern has
depend on the control parameters p. The main a minimal settling time with a small overshoot.
contribution of (Hjalrnarsson et al.) 1994) was to The objective in such applications is to move the
show how to compute the minimum of this cost output of the closed-loop system quickly from
function with respect to these control parameters one reference value to another one; however, the
p without knowledge of the system, Le. by itera- particular shape of the transient response from the
tive computations of the gradient *f;
and the use initial reference value to the final value is of no
of a stochastic approximation algorithm for the importance, provided that it does not have large
update of the controller parameter vector Pi: overshoots. In addition, without knowledge of the
actual system (which is a major reason for using
_1 8J IF~r) it is not known in advance how fast a settling
Pi+l == Pi - JiRi 8p (Pi). (2)
time can be achieved for this particular systenJ
,,~i th this particular controller structure.
Here Rt is some appropriate positive definite rna-
trix~ typically a Gauss-Newton approximation of
By imposing the entire response of the closed-
the Hessian of J ~ while "''fi is a positive real scalar
loop system through a specific choice of a de-
that determines the step size. The sequence {'Yi}
sired response Yd, rather than just the endpoint
must obey some constraints for the algorithm to
of this transient response, the classical 1FT crite-
converge to a local minimum of the cost function
rion leads to controller parameters that realize a
J(p): see (Hjalmarsson et al. 1 1994).
compromise between fitting the transient response
In this paper we present a variant of this criterion and fitting the new reference value, even though
in which the signals Yt (p) and Ut (p) (or their the user does not care about the exact shape of the
frequency weighted versions) are tirne weighted transient response. Instead, by imposing a mask
by weightings wy(t) and wu(t), respectively. This on the transient response, the criterion will tune
idea waS" initially suggested in (Lequin) 1997). the controller parameters in such Cl. way as to
fThus, the criterion (1) is replaced by: achieve the new desired reference value without
focusing on a particular pre-imposed transient
response that is perhaps not naturally achieved
by the closed loop system. In other words, by
imposing a mask on the transient response one
does not waste the available degrees of freedom
(3) in the controller parameters on the matching of a
specific and entirely arbitrary transient response~
Instead one can focus these parameters entirely on
where w y (·) and lV u (-) are any nonnegative num-
achieving a fast settling time. The cost achieved
bers. rrhe flexibility offered by the time weight-
after the lllasked interval is always snlaller than
ings wy(t) and 7LJu (t) is that they allow one to
when no mask is used: see Section 2 below.
put different weightings on different parts of the
time responses. A particularly interesting applica- These same observations can of course also be
tion, that is discussed in this paper, is when zero made about any model-based control design me-
weightings are put on the transient response of the thod. For example, in classical (Le. model-based)
output response to a step change in the reference LQG design, one often chooses a control design

4660
Copyright 1999 IFAC ISBN: 0 08 043248 4
OPTIMIZING THE SETTLING TIME WITH ITERATIVE FEEDBA... 14th World Congress of IFAC

criterion that requires the OUt.pllt of the controlled • a white noise of variance a 2 :::= 0.0025, fil-
system to track a desired response Yd, even though tered through a shaping filter H(s) == 3~1'
in applications where a setpoint change has to was added to the output of the closed loop
be made, one is not particularly interested in system;
the specific shape of the transient response. It • an industrial PID controller structure was
so happens that, with 1FT, this modification of used in which the integral action is applied to
the criterion (Le. the imposition of masks) is the tracking error, but the derivative action is
extremely easy to handle. applied to the output signal only. In addition,
a first order filter was used to limit the
Our first contribution in this paper will be to
illustrate, through a few examples, the typica.l
high frequency gain of this derivative action.
These modifications of the "academic" PID
advantages that can be gained by imposing zero
weightings on the tra.nsient response, in terms of controller, typical of industrial applications,
result in a two degree of freedom controller
achieving a faster settling time. Our second and
perhaps most useful contribution will be to pro- as ShOVlll in Figure 1;
pose a procedure whereby this settling time can • the maximum step size "'li in the iterative
be minimized even when the system is unknown, algorithm (2) was set at Imax =: 1; a line
search procedure was used to compute the
as is assumed with IF'T. This procedure consists
of imposing initially a rather large mask (Le. a local minimum in the descent direction, with
rather large zero-weighting time interval), and successive divisions of the step size by 2 until
1'i reached 2 x 10- 4 ;
then to progressively reduce the size of the mask
• a Gauss-Newton approximation of the Hes-
until oscillations start to appear in the transient
response. Thjs allows one to chooze the mask sian was used for R ~ as recommended in
(Hjalmarsson et a1., 1998).
of appropriate length~ and hence to design the
controller that achieves the smallest settling time
without oscillations. The procedure will again be
illustrated with an example. 2.1 Example 1

Consider the plant

2. 1FT CONTROL DESIGN WITH A FIXED P( ) 0.5s -+- 1

MASK
s ==: 8(8s + 1) (5s + 1)
The objective is to tune a PID controller in order
It is very easy to show that a controller designed to achieve a unit step change in about 100 seconds
with a mask always achieves a srnaller cost over with no overshoot. The initial regulator parame-
the non-ma.c;;ked time interval than a controller ters were K ::::: O~l, 'L : : : : 500 and Td == 2, and the
designed without a mask. Let p* be the parameter experiTnent tilne was 500 seconds. T'~lO different
vector p that minimizes the initial criterion .J (p) criteria were minimized.
of (1), and let p:n be the parameter vector p that
minimizes the modified criterion Jrn{p) of (4) with
The first OIle v/as the classical 1FT criterion (1)
a mask of length to. Assume that, in both cases, we
with the desired trajectory Yd being the ouput
are only concerned with the cost after the initial
of the following reference model Td == (lOs~])4 i
period to~ Then, by the optimality principle, we
this reference model has a settling time of about
obviously have
100 second~. After 5 iterations of the 1FT scherne,
(5) the parameters had converged to K == 0.026495,
T i = 32441 and Td =::: 1.9165, with a cost over
the time interval [100, 500] of J = 4.3158 X 10- 4 •
This shows that, over the interval [to) oc:), the
Figure 2 shov.rs the corresponding closed loop step
controller C(p~) designed with a mask, always
response (in full) together with the desired re-
achieves a better performace than the controller
sponse of the reference model (dotted). Observe
C(p*) designed without tlle u~e of a mask.
that the required specifications concerning set-
We now compare the performances that are typi- tling time and absence of overshoot are not met.
cally acllieved for a step change in the reference,
with and without a mask, with the following ex-
The second criterion "vas the modified criterion
ample. In this and the next example, the following
Jm.(p) of (4) with a mask over the first 100 time
simulation design choices were made:
periods, Le. to -:= 100~ After 7 iterations of the
• no frequency weightings were used in the If'T 1FT scheme, the parameters had converged to
criterion, Le.. L y = L u ~ 1; K ~ 0.017929, T i =-= 18948 and Ta == o~ 71413,
• no weight was Pllt on the control energy~ Le. with a corresponding cost over the time interval
,.\ =- 0; [100,500] of J == 1.9390 X 10- 4 • Figure 3 shows the

4661
Copyright 1999 IFAC ISBN: 0 08 043248 4
OPTIMIZING THE SETTLING TIME WITH ITERATIVE FEEDBA... 14th World Congress of IFAC

corresponding closed loop step response~ which easy to implement~ We have also observed that~ if
now meets the specifications. such procedure terminates with a mask length tb,
then the corresponding closed-loop step response
is typically better than what v,rould be achieved by
directly choosing a mask of length rather than to
progressively decreasing the length of the mask.
The assillIled reason, which is in tunc with the
observations made in (Hja.Imarsson et ai., 1998)
about the choice of reference models, is that by
progressively shortening the length of the mask
(and hence the required settling time) through
the iterations, one minimizes a succession of well
behaved cost functions, whereas the direct min-
imization of the criterion with a mask of length
t8 typically results in a criterion with multiple
Fig. 2. Optirnal closed loop step response obta.ined minima. Thus~ by progressively reducing the mask
with a reference model (full), and desired length, one reduces the risk of converging to a
response (dashed) local minimum.
The use of masks of decreasing length is illustrated
by the following example.

....".'f·.. 'fl ' -1': i I

t'" t ,t·· r-
t·..,..,. ! .."_.of'" .. -.. t-J -.-.
I
.~

1- : ., j. ··,,·,·!·····,···i····· +..
t..· ···,j.._·· t···~..··t..,·.. +., . 3.1 Example 2
... ~ . ,!.....,.- .!.....-·"1·""" l····.. ,·t.. ·.. .·\'····.. ·t···.. ·.. \·········t·····.. Consider the plant
'·:·:·r·::·:,r::::::r:::::::r::::::r::::::::·r:::::::r::::::r:::::::F::: P(s) == 82
1
-t- O.ls -+ 1
-"'*..
~~~---::-!-=-~~---=~--'-.o~-~-=-!-:--~-~-~
.......·00
One wishes to tune a PID controller in order to
f"'ig. 3. Optimal closed loop step response obtained achieve a settling time of 20 seconds for the closed
with a rnask of length 100 loop systeIIJ. The initial PID parameter values are
taken as ]( = 0.025, T i ~ 2 and T d = 1. This
yields the very sluggish response shown in Figure
4.
3. If"'T CON'rROL DESIGN WI'lIH A
VARIABLE SIZE MASK
~ ~
~_ . . .! ---+-- :
f i ~ ~
One of the difficulties in choosing a reference •• Q •••••••• r' ~.r- f '•. '1- "L.+".. 'L ~
rTIodel or a reference signa.l is to decide what. is ; ; i ! ! ~

a reasonable settling time for the system to be

controlled. The same difficulty arises with the
t :: ::.::t:·':"r:.':':l·-,:···t':.·.~r.:::::j:::::::::r-:':.t:::.:t::'::.
'.-.• ~
~"lIIr····
~
...... l1· ...
~ ~~"I
~
·r··.. ~
r·f . ·.. . . ·• I 'l........_II I J .. ;"f ....-.... ,·i··....··....~'"
f ....._....
mask techniques, because the length of the mask, r . + 1I"- ..

to, is the design parameter that fixes the settling • r_ t~ -, -ii _ -i~ -~ .-_. _. ';f ~
~ ~
.. -i·· -..-t· .- ill- "•••. T ~~ -.
t -Ill I f~
t -.. -1"-. • i
~ l
i ~ ,"' I t •

t i f ~ ~ ~ f i i
time. If this length to is chosen too long, then the --:l•• Q~--;=---...-.1z-- ...............r::----'=r----~~....
;.-~r--~~--:'!"MV~.
- ~---=!

obtained response is slower than could be achieved

by the system with the chosen controller struc- F"lig. 4. Closed loop st.ep response with initial PID
ture. If it is chosen too short, then our experience parameters
shows that an overshoot and oscillations appear in
the transient response. rrhe same phenomenon of The application of the classical 1FT criterion with
course occurs Vl,rith the choice of a reference model a desired response shown in dotted line in Figure
if the step response of the reference model is cho- 5 yields the response shov-rn in full line on that
sen too short with respect to what the controlled same figure~ This response is very unsatisfactory;
system is able to achieve. this is in large part due to an unfortunate choice
of initial parameters.
Our solution to this problem, with the use of
mask8~ is to initially use nlasks of rather large With the use of a fixed mask of length to ~ 20
length and to progressively decrease the length seconds, the minimization of the modified 1FT
of the mask until an overshoot begins to occur. criterion with the same initial parameters leads to
This has shown to be extremely practical and the closed loop response of Figure 6. This response

4662
Copyright 1999 IFAC ISBN: 0 08 043248 4
OPTIMIZING THE SETTLING TIME WITH ITERATIVE FEEDBA... 14th World Congress ofIFAC

rnasks of decreasing lengths in the last simulation,

leading to a sequence of cost criteria (rather than
·.. ·····I--···---i·,,·,,·-[ ." ....1-.. ,." .. a one-shot criterion), and to a different sequence of
Pi parameter vectors than resulted with the direct

t~l.t: :I: : : : : I. ·: : :j: :·:·: l:.: ·: :I: : · : j:·: :·: I:· :· :·[: : · :·j·: ·:·:·
use of a mask of length to = 20.

4.. CONCLUSION

The If'T method is a very flexible controller design

tool, lending itself to variants that arc typically
Fig. 5. Optimal closed loop step response (full) easy to implement.. Here we have studied a vari-
obtained with the classical 1FT criterion and ant, first proposed in (Lequin, 1997), in which
using the desired response (dashed) time weightings are added in the 1FT criterion.
is better than that obtained with a reference tra- "Ve have shown that, by doing so, one can easily
jectory, but still very oscillatory. design controllers that focus better on a rapid
tracking of a new reference value rather than wast-
ing some of the degrees of freedom on the tracking
of a specific transient response. In addition, we
• ~_ • "

have shoWll that the use of masks of decreasing

• I •••

length allows one to easily design controllers that

minimize this settling time..

5. REF'ERENCES
DeBruyne, It"'., B.D.O. An derson, M~ Gevers and
N. I..inard (1997). Iterative controller opti-
Fig. 6. Optimal closed loop step response obtained mization for nonlinear systems.. In: Proc~ 36th
with the 1FT criterion using a mask of length IEEE Conf. on Decision and Control. San
20 Diego, Califonlia. pp. 3749- -3754_
Hjalmarsson, Il. (1998). Control of nonlinear sys-
F'inally a mask of decreasing length was used, tems using iterative feedback tuning. In: Proc.
with an initial length of 80 seconds, and with the American Control Conference 98. Vol. 5.
same initial parameters again. At every iteration Philadelphia. pp. 2083-2087.
of the 1FT scheme, the length of the mask was Hjalmarsson, H., M. Gevers, S. Gunnarsson and
decreased by 20 seconds, until a mask of length 20 O. Lequin (1998). Iterative Feedback Thning:
was reached~ This led to the closed loop response theory and applications. IEEE Control Sys-
shown in Figure 7. tems Magazine 18~ 26-41.
Hjahnarsson, H., S. Gunnarsson and M. Gevers
(1994). A convergent iterative restricted com-
plexity control design scheme. In: Proc. 33rd
IEEE Conf. on Decision and Control. Or-
I :.: ··:,t:=:::::l·:::·::),:·::·:::(::::·:t:·::·:::j"::::'::j:':::::"j::::::j..::-.:' lando, Florida. pp~ 1735-1740.
Hjalmarsson) H., S. Gunnarsson and M. Gevers
. ··"·I.. "··..·t··.. ··,··t,.. ··.. ···~LI· ·r·········l······'.. j. ···· .. t····· ··f···· .-.
.... (1995). Model-free tuning of a robust regula-
··::,·:r::·::~,:::,::::t,::::,::r:.::,:r::,::::!,:::·'::t"::':"j::::::'::F:'::' tor for a flexible transmission system. Euro-
pean .Journal of Control!, 148-156.
..... •.. 0 _~ ,_
Lequin, O. (1997). Optimal closed loop PTn tun-
ing in the process industry with the Itera-
Fig. 7. Optimal closed loop step response obtained tive Feedback Tuning scheme.. In: CD-RO!J!
with the 1FT criterion using masks of de- of European Control Conference? paper TH-
creasing length A-H6. Brussels, Belgium.
Sjoberg, J. and M. Agarwal (1996). Model-free
Observe the dramatic improvement of the re- repetitive control design for nonIinear sys-
sponse of Figure 7 vis-a-vis the response of Figure tems~ In: Proc. 35th Conference o'n Decision
6, even though in both cases the final criterion and Control. Kobe, Japan. pp. 2824-2829.
is identical (i.e. the criterion Jrn(p) of (4) with
a mask of length t.o === 20), and even thOllgh the
starting values used for Po were identical in both
simulations. The improvement is due to the use of

4663
Copyright 1999 IFAC ISBN: 0 08 043248 4