You are on page 1of 6

Pergamon

ooo5-1098(95)00115-8

Auromarrca. Vol. 32, No. 2. pp. 261-266. 19%


Copyright 0 1996 Elsevier Science Ltd
Prmted m Great Britam. All rights reserved
om-lOY8/%
$15.tM+ 0.00

Brief Paper

Adaptive

Internal Model Control: Design and


Stability Analysis*
ANIRUDDHA

Key Words-Internal
modelling errors.

DATTA1_ and JAMES OCHOAt

model control; on-line adaptation;

Abstract-This
paper shows how adaptation
can be
combined with an internal model control structure to obtain
an adaptive internal model control scheme possessing
theoretically provable guarantees of stability. The internal
model controller structure is first reviewed in the context of
the YJBK parametrization of all stabilizing controllers, and
its appropriateness for the control of open-loop stable plants
is discussed. Using a series-parallel identification model, it is
then shown how, for a stable plant, one can adapt the
internal model on-line and guarantee stability and asymptotic
performance
in the ideal case, i.e. in the absence of
modelling errors. This is achieved without having to check
any strictly positive real conditions or requiring any
persistent excitation. By robustifying the adaptive law using
standard approaches from the robust adaptive control
literature, the robustness of the scheme to the presence of
modelling errors, such as unmodelled dynamics, is also
established. The theoretical developments here complement
the reported industrial successes of internal model control
schemes.

stability and performance;

robustness

to

As already mentioned, one of the features of IMC is that it


requires an explicit model of the plant to be used as part of
the controller. When the plant itself happens to be unknown,
or the plant parameters vary slowly with time owing to aging,
no such model is directly available a priori, and one has to
resort to identification techniques to come up with an
appropriate plant model on-line. Several empirical studies
(e.g. Takamatsu et al., 1985: Sooer et al.. 1993) have
demonstrated the feasibility of such an approach. However,
what is lacking in the process control literature is the
availability of results with solid theoretical guarantees of
stability and performance.
The principal aim of this paper is to take a step in this
direction by designing an adaptive IMC scheme with
provable guarantees of stability and asymptotic performance
in the ideal case. Appropriate modifications of the scheme to
ensure robustness to modelling errors will also be discussed.
As will be seen in the sequel, once the properties of the IMC
structure are well understood, the development of such an
adaptive IMC scheme, involving on-line adaptation of the
plant model, can be carried out using standard techniques
from the adaptive control literature -(Ioannou and Datta,
1991). Thus it appears that the main reason for the absence
of any such results thus far in the literature is the fact that
most adaptive control theorists are perhaps not familiar with
the IMC structure, while most process control engineers lack
familiarity with the subtle intricacies of adaptive control
theory. One of the objectives of this paper is to narrow this
gap between theory and applications so that the tremendous
theoretical advances made in the area of adaptive control
during the last two decades or so can better find their way
into the realm of industrial applications.
The entire presentation in this paper will concentrate on
the single-input single-output (SISO) case. Although most
process control applications are multi-input multi-output
(MIMO), we believe that a thorough understanding of the
SISO case is a necessary prerequisite for developing the
MIMO theory.
The paper is organized as follows. Section 2 contains some
standard tools used for analyzing robust adaptive control
schemes. In Section 3 we describe the IMC controller
structure for stable plants and emphasize its connection to
the YJBK parametrization
(Youla et al., 1976) of all
stabilizing compensators. This controller structure is then
combined with a standard gradient update law in Section 4 to
obtain an adaptive IMC scheme, guaranteeing stability and
asymptotic performance in the ideal case. In Section 5
normalization (Praly, 1984) and parameter projection are
introduced into the gradient update law, and the robustness
of the resulting adaptive IMC scheme to unmodelled
dynamics is established. Section 6 concludes the paper by
summarizing the main results and outlining directions for
future research.

1. Introduction
The internal model controller

(IMC) structure continues to


be a very popular one in process control applications (Morari
and Zafiriou, 1989; Garcia et al., 1989). This structure, in
which the controller includes an explicit model of the plant,
is particularly appropriate
for the design and implementation of controllers for open-loop stable systems. The
fact that many of the plants encountered in process control
happen to be open-loop stable possibly accounts for the
immense popularity of IMC among practicing engineers.
As discussed in Garcia et al. (1989), the IMC structure is
inherent in most model predictive control schemes such as
model algorithmic control (Richalet et al., 1979) and dynamic
matrix control (Cutler and Ramaker, 1980). Of course, it
should be pointed out that there is another class of model
predictive schemes, whose primary objective is adaptive
control, which are not really related to the IMC structure.
Peterkas (1984) predictor-based self-tuner, Ydsties (1984)
extended horizon adaptive controller and the generalized
predictive control of Clarke et al. (1987a, b) are typical
examples in this category. The book by Bitmead et al. (1990)
contains an interesting discussion of the connections between
some of these schemes and adaptive linear-quadratic
Gaussian (LQG) outimal control. Since the focal uoint of this
paper is . adaptive IMC, the above schemes, though
important for adaptive process control, will not be discussed
any further.
* Received 18 January 1995; revised 17 May 1995; received
in final form 29 June 1995. This paper was not presented at
any IFAC meeting. This paper _was recommended
for
publication
in revised form bv Editor C. C. Hans.
Corresponding author Professor A. Datta. Tel. +l 409 8&
5917; Fax +l 409 845 6259.
t Department of Electrical Engineering, Texas A & M
University, College Station, TX 77843-3128, U.S.A.

2. Tools for analyzing robust adaptive control schemes


In this section we give some definitions and lemmas that
are by now standard in the stabiity and robustness analysis of
261

Brief Papers
adaptive control schemes (Ioannou and Datta, 1991). These
results are included here for later use while analyzing the
robust adaptive IMC scheme of Section 5.
Definition 2.1. For any signal x: [0, =)-+ Iw, x, denotes the
truncation of x to the interval [0, t] and is defined as
x(r)
r,(r) = O

if r5t,
otherwise.

Definition 2.2. For any signal x: [0, cc)+ [w, and for any
6 2 0, t 20, ]]x,]]$is defined as

Remark
2.1. Here
]](.),]]: represents
the exponentially
weighted L2 norm of the signal truncated to [0, r]. When
6 = 0 and t = x, ]1(.),]]; becomes the usual L, norm, and will
be denoted by ]].]12.It can be shown that ]].]]$ satisfies the
usual properties of the vector norm.

Definifion 2.3. Let H(s) be a transfer-function


entries are stable and proper. Then

matrix whose

X(s)N,(s)

IIWs)llz 5 sup
{A,,,[H*(j,)Hcjw)l}2,
(0

(3)

where A,,,[.] denotes the largest (necessarily real here)


eigenvalue and H* denotes the conjugate transpose of H.
Here ]].(s)]]~ is the so-called Hz norm widely used in robust
control (Francis, 1987).
Definition 2.4. Let H(s) be a transfer-function

matrix whose
entries are proper and analytic in Re [s] 2 - @. Then
IIH(s)Il: a IIH(s - :S)llx.

(4)

If, in addition, the entries of H(s) are strictly proper then


II H(s)ll$ a (&

(=_ tr [H*(jw - f6)H(jw


I

Definition

2.5. Consider

y: [0, =)+

Iw and the set

S(y)=(x:[O,

=)-+oI

the

- fs)] d-)*.

signals

(5)

x: [O. =) -+ [w.

I ~+~~T(~)X(T)dT~li_y(~)d~+c]
,

for some c 2 0 and Vt, T 2 0. We say that x is y-small in the


mean if x E S(y).
The following lemmas play a key role in the robustness
analysis of adaptive control schemes.
Lemma 2.1. Let
z=H(s)[u]:=H*u,

where
proper
entries
and u

H(s)

is a transfer function matrix with causal and


entries and * denotes time-domain convolution. If the
of H(s) are analytic in Re [s] 2 - $6 for some S 2 0
E Lzc, then
Ilz,ll26~ IIWs)ll:

Ilu,Ili.

(6)

If, in addition, the entries of H(s) are strictly proper then


lz(r)ls

IIH(s)ll: ll~,ll~.

(7)

The proof can be bound in Tsakalis (1992) and Ioannou


and Datta (1991).
Lemma 2.2. If u E S(F)
L, vs>o.

for some

p 20

In this section we derive the internal model control


structure as a special case of the YJBK parametrization
(Youla et al., 1976). Although similar connections have been
pointed out earlier in the process control literature (see e.g.
Garcia et al., 1989; Maciejowski, 1989), we believe that it is
of sufficient importance to warrant a self-contained coverage
here. In this context, it should be pointed out that the
non-familiarity of most adaptive control theorists with the
IMC structure is probably the single most important reason
why many of the theoretical advances of adaptive system
theory have not found their way into the realm of process
control applications.
Consider the standard unity feedback configuration shown
in Fig. 1. Here u and y represent the plant input and output
signals respectively, r is the command signal, and P(s) and
C(s) are the rational proper transfer functions of the plant
and the compensator respectively. The YJBK parametrization seeks to parametrize the set of all compensators C(s)
that guarantee the stability of the closed-loop system. To do
so, we consider a coprime factorization N,(s)/D,(s) of the
plant P(s), where N,,(s) and D,(s) are stable rational
transfer functions. Let X(s) and Y(s) denote a solution of
the Bezout identity

then

l]u,l]$ E

+ Y(s)D,(s) = 1

C(s) =X(s) + &(s)Q(s)

J(s) - &(s)Q(s)

(9)

where Q(s) varies over the set of all stable rational transfer
functions.
3.1. IMC structure for stable plants as a special case of the
YJBK parametrization.
Now suppose the plant P(s) is
open-loop stable. Then a possible choice of coprime factors is
/VP(s) = P(s) and Dr(s) = 1, so that the Bezout identity (8)
holds with X(s) = 0 and Y(s) = 1. Then it follows from (9)
that every stabilizing compensator C(s) is given by

() =1 -

Q(s)
P(s)Q(s)

(10)

where Q(s) varies over the set of all stable rational transfer
functions. This compensator when inserted into the setup in
Fig. 1 leads to the Youla-parametrized system shown in Fig.
2. Keeping the external input-output behavior the same, the
above system can be successively transformed into the
configurations shown in Figs 3 and 4 respectively.
The configuration in Fig. 4 is referred to as the internal
model control (IMC) structure, because the controller in this
configuration incorporates an explicit model of the plant.
Thus the popular IMC structure for stable plants is a special
case of the YJBK parametrization
of all stabilizing
compensators. Since Q(s) is the Youla parameter, it follows
that the IMC configuration in Fig. 4 is stable iff Q(s) is a
stable transfer function.
Remark 3.1. The implementation of the IMC structure as in
Fig. 4 is appropriate only for open-loop stable plants. Indeed,
an open-loop unstable plant in this configuration will
continue to be closed-loop unstable, since perfect matching
between the plant and the model will ensure that there is
effectively no feedback signal for stabilization purposes.
Nevertheless, the IMC structure can be used for design
purposes even for open-loop unstable plants (see e.g. Morari
and Zafiriou, 1989) and in this case the parameter Q(s)
corresponds to the famous Q parameterization (Zames, 1981:
Zames and Francis, 1983). To guarantee closed-loop stability,
Q(s) must not only be stable but also satisfy certain

The proof can be obtained by adapting the proof of Lemma


3.6 in Ioannou and Datta (1991).
3. The internal model controller structure (IMC) for stable
plants and the YJBK parametrization

(8)

in the ring of stable rational transfer functions. Then the set


of all stabilizing controllers C(s) is given by (Vidyasagar,
1985)

Fig. 1. A standard unity feedback system.

Brief Papers

Fig. 2. Youla parametrized

system (stable plant).

constraints imposed by the unstable poles of the plant.


However, if the YJBK parameterization is used instead, i.e.
Q(s) is the Youla parameter, then even in this case the
closed-loop system is stable iff Q(s) is stable.
Remark 3.2. Since for an open-loop

stable plant, the stability


of the configuration of Fig. 4 is guaranteed for any stable
Q(s), one can now choose Q(s) in an optimal fashion to
meet some performance objective. Indeed, it has been shown
(Morari and Zafiriou, 1989) that choosing Q(s) to minimize
an Hz performance objective for certain first-order plants
yields familiar PID controller structures.
4. Adaptive internal model control: stability and asymptotic
performance in the ideal case
In order to implement the internal model controller of Fig.
4, the plant must be known a priori, so that the internal
model can be designed. When the plant itself is unknown,
the internal model controller cannot be implemented. In this
case the natural approach to follow is to retain the same
controller structure as in Fig. 4, with the internal model
being adapted on-line based on some kind of parameter
estimation mechanism. This results in an adaptive IMC
scheme. The idea of combining adaptation with an IMC
structure to obtain an adaptive IMC scheme is not new.
Indeed, such an idea has been used in several applicaions
(e.g. Takamatsu et al., 1985; Soper et al., 1993). What is
lacking in the literature, however, is the presence of any kind
of theoretical guarantees for the stability, convergence,
robustness, etc. of such schemes beyond what is observed in
empirical simulations. The main aim of this paper is to
develop such a result.
To this end, we assume that the stable plant P(s) is
described by P(s) = Z,(s)/R,(s),
where R,(s) is a manic
Hurwitz polynomial of degree n and Z,(s) is a polynomial of
degree m, with m <n. The coefficients of Z,(s) and R,(s),
except for the leading coefficient of the latter, are all
assumed to be unknown. The problem of adaptive IMC is to
estimate the coefficients of Z,(s) and R,(s) on-line and use
these estimates for constructing the internal model; in other
words, the internal model is adapted on-line using a standard
identifier.
Although the IMC structure in Fig. 4 suggests the use of a
parallel identification model (Ioannou and Sun, 1994), leading
to an adaptive law based on the output error method, we note
that historically such schemes have not been very successful,
as far as establishing closed-loop stability without strictly
positive real (SPR) conditions or persistent excitation (PE) is
concerned. Consequently,
we shall use a series-parallel
identification model based on the equation error method
(Ioannou and Sun, 1994). The design and stability analysis in
this section will concentrate on the ideal case.
4.1. Design of the adaptive law. The design of the adaptive
law is carried out using a standard approach from the
identification literature. We start with the plant equation

+$I,7

Fig. 4. The IMC configuration.

Filtering both sides by l/A(s), where A(s) is an arbitrary


manic Hurwitz polynomial of degree n, we obtain
Y= *\jsyJ

[y] +

[u],

or

;ra,o
y=e1A(s) [yl+

eY$u1,

03)

where
a,(s) = [Y-l
a,(s)

= [P

s-~

llT.

sm-

llT,

and Of and et are vectors containing the coefficients


A(s) - R,(s) and Z,(s) respectively. Defining
*(s)

e*= [et e:]*,

4=

[4T 4T],

the plant in (13) can be rewritten as


y = e*.d

e= y&4,

y>o,

&=y-y^,
p =

eT4,

(15)
(16)
(17)

and O(t) is the estimate of 8* at time t.


The adaptive IMC is obtained by replacing the internal
model in Fig. 4 by that obtained from (17). This results in the
control law
u = Q(s)[r - 4,
(18)
which is depicted schematically in Fig. 5. From here, it is
clear that the internal model in the case of adaptive IMC is
driven not only by the plant input u but also the plant output
y through
the regressor
(b. Such a series-parallel
arrangement results in a scheme whose stability properties
can be easily established using well-known techniques from
the adaptive control literature.
4.2. Analysis of stability and asymptotic performance. The
following theorem describes the stability and performance
properties of the adaptive IMC scheme.
Theorem 4.1. Consider the plant (11) subject to the adaptive
IMC control law (15)-(N). Then, for any bounded r, a11

(11)

= Z,(s)bl.
Y

AUTO 32-2-I

(14)

which is in the form of the standard linear parametric model,


for which a large class of adaptive laws can be developed. In
particular, using the gradient method (Ioannou and Datta,
1991), we obtain the adaptive law

Fig. 3. Equivalent feedback system.

of

42 = a,(s)[u],

an(s)
41 = ,(,)[Yl,

or

Ms)[yl

(12)

Fig. 5. Adaptive IMC (ideal case).

Brief Papers
signals in the closed-loop
system are uniformly bounded and
the internal model output p(t) converges
to the actual plant
output y(t) as I+ 3~. In other words, the adaptive
internal
model controller
behaves asymptotically
as the non-adaptive
one, which could be implemented
if the plant were perfectly
known.
Proof: the proof
follows by combining
the results of a
standard
Lyapunov-like
argument
with the properties
of the
IMC structure.
Consider the positive-definite
function

Then,

along

the solution

of the estimated
parameters
can be achieved
by using a
normalized
adaptive
law as in the next section, instead of
the unnormalized
adaptive law considered
here.
5. Robust adaptive internal model control: stability in the
presence of unmodelled dynamics
The adaptive
IMC of the last section
was designed
assuming that the plant is ideal. However,
it is well known
from the adaptive
control literature
that such schemes can
exhibit unstable behavior in the presence of modelling errors
such as unmodelled
dynamics
or bounded
disturbances.
Accordingly,
in this section we shall design a robust adaptive
law (Ioannou
and Sun, 1994) and combine it with an IMC
structure to obtain a robust adaptive IMC scheme.
To this end, we consider the plant equation

of (15), we have

ri = ere4
-E

(since E = --fi*+

from (14), (16) and (17))

5 0.

(19)

Therefore
V and hence 8 and 0 are uniformly
hounded.
Also, since V(r) is monotonplly
decreasing
and bounded
V(e(r)) = V, exists. Hence, integrating
from below, lim,,,
both sides of (19) from t = 0 to I = x, we obtain
*
r*(r) dr = v, - V,,
u;, $ V(@O)).
(20)
I0
i.e. E E Lz. Since Q(s) is stable and r E L,, E E Lz, it follows
from Fig. 5 that the control input u = ui + uz, where U, E L,
and a2 E L,. Since P(s) is strictly proper and stable, it follows
that y E L, (Desoer and Vidyasagar,
1975). Since

where u and y are the plant input and output signals as


before, Z,(s)/R,,(s)
represents
the modelled part of the plant
and is assumed to be stable and strictly proper, and PA,,(S)
is
a
multiplicative
such
stable
uncertainty
that
[Z&)/R,(s)]A,,(s)
is strictly proper.
We next present the
design of the robust adaptive law, which is carried out using
a standard
approach
from the robust
adaptive
control
literature
(Ioannou and Datta, 1991).
5.1. Design of the robust adaptive law. The plant equation
(22) can be rewritten as
R<,(s)[yl = Z<,(.s)]nl + &,(.~)Z<>(s)]r~l.
Filtering
both sides by l/A(s).
where A(s) is an arbitrary,
manic, Hurwitz polynomial
of degree n. we obtain

y = MS)- R<>(s)
w)
,yl

are strictly
we obtain
from (18)
closed-loop
To show
from (14)

proper and stable, it follows


from (17) and (16) that 9, E
that u E L,, so that all
system are bounded.
asymptotic
convergence
of
(16) and (17) that
g = $4

that r$ E L,. Hence


E L,. It now follows
the signals
in the
9 to y, we observe

,ijT&

where

r
4s)

H(S)

MS)
O-

is a proper, stable matrix.


follows from (21) that 6
implies that E-+ 0 as r-+
Sun, 1994), and therefore

-%,(s)
MS)

ll(.s)

LLll,

(23)

A(s)

as

e*4 + ~~11,

y =
where

8* and 4 are as defined

in Section

(24)
4.1 and

(21)

we
have already
established
that
0 = YE+, so that E E L, j f3 E L,. Again

yaw,

This can be rewritten

z,,(s) u + 1~4,hP<,(s)

8, I$ E L,.

Also,

Equation
(24) is exactly in the form of the linear parametric
model with modelling error, for which a large class of robust
adaptive
laws can be developed.
In particular,
using the
gradient
method with normalization
and parameter
projection, we obtain the rollowing robust adaptive law (Ioannou
and Datta, 1991):

Hence U, y E L, 3 4 E L,. Thus it


E f.,, which, together
with E E L2,
x (Barhalats
lemma; Ioannou and
0
the proof is complete.

Remark 4.1. In the above adaptive


IMC scheme only the
internal model is adapted on-line, while Q(s) is any arbitrary
fixed stable transfer
function.
Since any fixed stable Q(s)
guarantees
closed-loop
stability,
one can intuitively
expect
that closed-loop
stability could still be guaranteed
if Q(s)
were to be pointwise-stable
and vary slowly with time.
This, along with the fact that any stable Q(s) is acceptable
for closed-loop
stability, opens up the interesting
possibility
of choosing
Q(s) optimally
on-line
to minimize
some
performance
objective
based
on the estimated
internal
model. Although
such issues are beyond the scope of this
paper and will not be discussed
here, we observe that the
feasibility
of such a scheme is contingent
upon the slow
variation
of Q(s), which in turn requires that the estimated
internal model parameters
vary slowly. Such slow variation

e(o) E %@,

6 = Pr [yEb],

,_v-s
m2

(27)

y = 86,
m=l+nf,
ti, = -6,,m,

(28)

nt=m,.

+ II* + y,

(26)

m,(O) = 0,

(29)
(30)

where y >O is an adaptive


gain, & is a known compact
convex set containing
0*, Pr [.I is the standard
projection
operator
that guarantees
that the parameter
estimate
e(r)
does not exit the set Y&.,and S,, > 0 is a constant chosen so
that A,,,(s) and l/A(s) arc analytic in Re [s] 2 - is,,. This
choice of 6,,, of course, necessitates
some a priori knowledge
about the stability margin of the unmodelled
dynamics,
an
assumption
that has by now become fairly standard
in the
robust adaptive
control literature
(e.g. Ioannou
and Datta,
1991).
As in the ideal case, the robust adaptive IMC is obtained
by replacing
the internal model in Fig. 4 by that obtained
from (28). This results in the control law
u = Q(.s)[r - c&l,
which is depicted

schematically

in Fig. 6.

(31)

Brief Papers
As in Ioannou and Datta (1991) we take truncated
exponentially weighted norms on both sides of (36) and (37)
and use (6), to obtain
II~,llzS~c fc II(~m*MI~,

(38)

l/u,/I$~c + c ll(~m%l~~

(39)

which, together with (35) imply that


mr(t) 5 c + c II(~m%l~.

(4)

Now squaring both sides of (40) we obtain


Fig. 6. Robust adaptive IMC.

-G(-r)e2m2m~(~) dr

5.2. Stability and performance


analysis. The following
theorem describes the stability and performance properties
of the robust adaptive IMC scheme.

(since m(t) 5 mdt))


*m:(r)

5 c +

c emsctmt)
E*(s)mi(s)exp(c[&zdr)dF
0

Theorem 5.1. Consider the plant (22) subject to the robust


adaptive IMC control law (26)-(31). Then 3~* >O such that
VP E [0, CL*) and for any bounded r, all the signals in the
closed-loop system are uniformly bounded and the error
y -9 E S(c~zn2/m*) for some c > 0.t
Proof. As in the ideal case, the proof is obtained by
combining the properties
of the robust adaptive law
(26)-(30) with the properties of the IMC structure.
From (24) (27) and (28), we obtain

Consider the positive-definite

function

Then, along the solution of (26)


(Ioannou and Datta, 1991)
V 5 erc+
= .s--Em2 + pn)

it can be shown that

(using (32))

1
<--&2m2+1Lw
_
(completing the squares).
2
2 mz
From (25)
nlm E L,.

(29)

and

(30),

using

(7)

it follows

(33)
that

Now, the parameter projection guarantees that V, 6, t3 E


L,. Hence integrating both sides of (33) from t to t + T, we

obtain

Also, from (26),

From the definition of 4, it follows using (7) that 4/m E L,.


which in turn implies that 0 E S(cp2~2/m2).
This completes the analysis of the properties of the robust
adaptive law. To complete the stability proof, we now turn to
the IMC structure. Now let S E (0, 8,) be chosen such that
Q(s) and R,(s) are analytic in Re [r] 2 - $8, and define the
fictitious normalizing signal mdt) by
m&) = I.0 + llu,ll4 + llv,ll$.

(35)

Now, the plant and controller satisfy


Y =

Q(s)[$$f

[l + dA~)l][r

u = Q(s)[r

- em2].

- &ml,

(36)
(37)

t In the rest of this paper, c is the generic symbol for a


positive constant. The exact value of such a constant can be
determined
(for a quantitative robustness result) as in
Tsakalis (1992) and Ioannou and Datta (1991). However, for
the qualitative presentation here, the exact values of these
constants are not important.

(using the Bellman-Gronwall

lemma)

Since em E S(p2~*/m2) and n/m is bounded, it follows using


Lemma 2.2 that 3~* > 0 such that V(ILE [0, p*), mr E L,,
which in turn implies that m E L,. Since 4/m and n/m are
bounded, it follows that I$, n E L,. Thus em2 = -eT@ + pq
is also bounded, so that we conclude from (36) and (37) that
y, u E L,. This establishes the boundedness
of all the
closed-loop signals in the adaptive IMC scheme. Since
y -9 = em2 and cm E S(~ZnZ/m2), m E L,, it follows that
y - y^F. S(c ~2n2/m2) as claimed, and therefore the proof is
0
complete.
Remark 5.1. The robust adaptive IMC scheme of this section
recovers the performance properties of the ideal case if the
modelling error disappears, i.e. we can show that if p = 0
then y - 9 ---f0 as t + 30.This can be established using similar
arguments as in Section 4, and is a consequence of the use of
parameter projection as the robustifying modification in the
adaptive law (Ioannou and Datta, 1991). An alternative
robustifying modification which can guarantee a similar
property is the switching-c modification (Ioannou and
Tsalakis, 1986).
Remark 5.2. As in the ideal case considered in Section 4, the
robust adaptive IMC scheme of this section only adapts the
internal model on-line while keeping Q(s) fixed. However,
since any stable Q(s) is acceptable, one may consider
replacing Q(s) by a time-varying operator that pointwise
optimizes some performance index based on the estimated
internal model. As long as this time varying operator is
exponentially stable, all of the arguments in the preceding
analysis will still go through. Now, by an appropriate choice
of the projection set V&(for example, by exploiting results
from Kharitonov theory; Kharitonov, 1978), it is possible to
ensure that the estimated internal model is pointwise-stable.
Also, the normalization in the adaptive law ensures that the
parameters in the estimated internal model are slowly
varying with time (see (34)). Thus, as long as the pointwise
control design rule that chooses the IMC parameter, based
on the estimated
internal model, is stabilizing and
Lipschitzian (e.g. Hz-optimal control), it should be possible
to exploit results from slowly time-varying systems (e.g.
Zames and Wang, 1991; Wang and Zames, 1991) to establish
the stability and performance
properties
of such a
pointwise-optimal IMC scheme. The precise details, however, remain to be worked out.

6. Concluding remarks
We have designed adaptive IMC schemes for open-loop
stable
plants
using the traditional
certainty
equivalence
approach
of adaptive control. Both the ideal case and robust
designs were considered,
and the stability and performance
properties
in each of these cases was established.
The design
and analysis were carried out by combining
the properties
of
the IMC structure
with well-known
techniques
from the
adatpive
control
literature.
The principal
motivation
for
undertaking
such a study was the immense
popularity
of

IMC in industrial applications,

on the one hand, coupled

266

Brief Papers

with the provable theoretical


guarantees
provided by
adaptive control research during the last two decades, on the
other. We do believe that the results of this paper will,
among other things, familiarize many adaptive control
theorists with the IMC control structure while simultaneously
arousing the interest of process control engineers in the
adaptive control literature with all its rich and elegant theory.
Thus this paper is an attempt to narrow the gap between
theory and applications.
The adaptive schemes considered in this paper only adapt
the internal model on-line, while keeping the IMC
parameter fixed. We do, however, believe that the IMC
parameter that really parametrizes the set of all stabilizing
controllers can also be optimally (pointwise) adapted
on-line, based on the estimated internal model, and that the
stability and performance of such a scheme can be similarly
established.
The precise details are currently
under
investigation.
Acknowledgement-This

work was supported


by the
National Science Foundation under Grant ECS-9210726.

References

Bitmead, R. R., M. Gevers and V. Wertz (1990). Adaptive


Optimal Control: The Thinking Mans GPC. Prentice-Hall,
Englewood Cliffs, NJ.
Clarke, D. W., C. Mohtadi and P. S. Tuffs (1987a).
Generalized predictive control-I.
The basic algorithm.
Automatica, X3,137-148.

Clarke, D. W., C. Mohtadi and P. S. Tuffs (1987b).


Generalized predictive control-II.
Extensions and interpretations. Automatica, 23, 149-160.
Cutler, C. R. and B. L. Ramaker (1979). Dynamic matrix
control-a
computer control algorith. In Proc. AIChE
National Meeting, Houston,
TX; also in Proc. Joint
Automatic. Control Conf, San Francisco, CA (1980).
Desoer, C. A. and M. Vidyasagar (1975). Feedback Systems:
Input-Output Properties. Academic Press, New York.
Francis, B. A. (1987). A course in H, Control Theory.
Springer-Verlag, Berlin.
Garcia, C. E., D. M. Prett and M. Morari (1989). Model
and practice-a
survey.
predictive
control:
theory
Automatica, 25, 335-348.

Ioannou. P. A. and A. Datta (1991). Robust adaptive


control: a unified approach. Proc. IEEE. 79,1736-1768.
Ioannou, P. A. and J. Sun (1994). Stable and Robust
Adaptive Control. Prentice-Hall, Englewood Cliffs, NJ.
Ioannou. P. A. and K. S. Tsakalis (1986). A robust direct
adaptive controller. IEEE Trans. &tom. Control, AC-31,
1033-1043.

Kharitonov,
V. L. (1978). Asymptotic stability of an
equilibrium position of a family of systems of linear
differential equations. Differentsial nye Uravneniya, 14,
2086-2088.

Maciejowski, J. M. (1989). Multivariable Feedback Design.


Addison-Wesley, Reading, MA.
Morari, M. and E. Zafiriou (1989). Robust Process Control.
Prentice-Hall, Englewood Cliffs, NJ.
Peterka, V. (1984). Predictor-based
self-tuning control.
Automatica, 20,39-50.

Praly, L. (1984). Robust


model reference
adaptive
controllers-part
I: stability analysis. In Proc. 23rd IEEE
Conf on Decision and Control.

Richalet, J. A., A. Rault, J. L. Testud and J. Papon (1978).


Model predictive
heuristic control:
applications
to
industrial processes. Automatica, 14, 413-428.
Soper, R. A., D. A. Mellichamp and D. E. Seborg (1993).
An adaptive nonlinear control strategy for photolithography. In Proc. American Control Conf, San Francisco,
CA.
Takamatsu, T., S. Shioya and Y. Okada (1985). Adaptive
internal model control and its application to a batch
polymerization
reactor.
In Proc. IFAC
Symp.
on
Adaptive Control of Chemical Processes, Frankfurt am
Main.
Tsakalis, K. S. (1992). Robustness of model reference
adaptive controllers: an input-output
approach. IEEE
Trans. Autom. Control, AC-37,556-565.
M. (1985). Control System Synthesis:
Factorization Approach. MIT Press, Cambridge, MA.

Vidyasagar,

Wang, L. Y. and Zames, G. (1991). Local-global double


algebras for slow H, adaptation: part II&-optimization of
stable plants. IEEE
Trans. Autom.
Control, AC-36,
143-151.
Ydstie, B. E. (1984). Extended horizon adaptive control. In
Proc. 9th IFAC World Congress, Budapest.
Youla. D. C.. H. A. Jabr and J. J. Boneiorno (1976).
Modern Wiener-Hopf design of optimal coitrollem-Part
II: the multivariable case. IEEE Trans. Autom. Control,
AC-21,319-338.

Zames, G. (1981). Feedback and optimal sensitivity: model


reference transformations, multiplicative seminorms, and
approximate
inverses. IEEE
Trans. Autom.
Control,
AC-26,301-320.
Zames, G. and B. A. Francis (1983). Feedback, minmax
sensitivity and optimal robustness. IEEE Trans. Autom.
Control, AC-28,585-600.
Zames, G. and L. Y. Wang (1991). Local-global double
algebras for slow H, adaptation: part I-inversion
and
stability. IEEE Trans. Autom. Control, AC-36, 130-142.

You might also like