INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING, VOL.

2, 217-248 (1988)

ROBUST MODEL REFERENCE ADAPTIVE CONTROL FOR

MULTIVARIABLE PLANTS

G . TAO AND P . A. IOANNOU

Department of Electrical Engineering-Systems, University of Southern California, Los Angeles, C A 90089-0781,
U.S.A.

SUMMARY
A complete procedure for generating and analysing robust model reference adaptive control schemes for
multivariable plants is developed. The procedure consists of two parts: the first part involves the
characterization of the integral structure of the modelled part of the plant, and the associated
parametrization of the controller structure; and the second part involves the development of a general
robust adaptive law for adjusting the controller parameters so that the closed-loop plant is globally stable
despite the presence of unmodelled dynamics and bounded disturbances. The use of dominantly rich
signals for improving convergence and the bounds for the tracking and parameter error is also analysed.

K E Y WORDS Model reference adaptive control Parametrization General adaptive law Robustness

1. INTRODUCTION
As in the single-input, single-output (SISO) case, the model reference adaptive control (MRAC)
problem can be divided into two parts: the first part deals with the parametrization of the
modelled part of the plant and the controller, and the second part with the development of the
adaptive law and the resulting stability problem.
A crucial assumption for stable MRAC of SISO plants is that the relative degree n* of the
modelled part Go@) of the plant transfer function is known. The extension of this assumption
to the multi-input, multi-output (MIMO) case is that the Hermite normal form H(s)”’ or the
interactor matrix t ( s ) 3 - 5 of the modelled part To(s) of the plant transfer matrix is known. The
prior knowledge about TO@)which is necessary for H(s) or t ( s ) to be known is very crucial
in adaptive control since the parameters of To@) are assumed to be unknown. If H(s) or ((s)
is diagonal and the relative degree of each element of To(s) is known, then H(s),t(s) can be
completely specified without any prior knowledge about the parameters of To(s). In the case
where H(s) or ((s) is not diagonal it is shown in Reference 6 that a diagonal stable dynamic
precompensator W,(s) can be found, using only the knowledge of the relative degree of each
element of To@),such that for most cases To(s)W,(s) has a diagonal H(s) or ((s). However,
such a precompensator cannot always be found and therefore in some cases H(s) or &s) cannot
be specified unless sufficient prior knowledge about the parameters of To(s) is available. It is
however possible to expand the class of plants for which MRAC schemes can be designed

This paper was recommended for publication by editor R . Bitmead

0890-6327/88/0302 17-32$16.00 Received 23 November 1987

0 1988 by John Wiley & Sons, Ltd. Revised 18 May 1988
218 G . TAO A N D P . A. IOANNOU

without much prioi knowledge about the parameters of To(s) by using the notion of both the
modified left interactor (MLI) and the modified right interactor (MRI) matrix defined in Section
2. As shown in Section 2, in some cases the MLI matrix is not diagonal whereas the MRI matrix
is diagonal and vice versa. Therefore by considering both the MLI and MRI matrix, we can
either choose the one which is diagonal, if any, or use a precompensator W,(s) in the case of
the MLI or a filter Wf(s) in the case of the MRI so that either To(s)W,(s) has a diagonal MLI
matrix or Wf(s)To(s) has a diagonal MRI matrix.
Once the parametrization issue is resolved and a controller structure is developed, an adaptive
law has to be designed for updating the unknown controller parameters such that the closed-
loop plant is globally stable and the plant output tracks the output of the reference model
asymptotically with time. Such a procedure is developed in References 2 and 4 for continuous
time plants in the ideal case of no plant uncertainties. The stability proof for the overall closed-
loop adaptive control system however is omitted since it was claimed that such a proof should
follow directly from the SISO case. A closer look at the result of MRAC in References 2 and
4 indicates that the extension of the stability proof in SISO MRAC to the MIMO case is not
at all clear in the case of unknown high-frequency gain matrix.
The objectives of this paper are as follows. First, we give a complete picture of the
parametrization issue in the MIMO case which is important for designing MRAC. This will
involve a survey of previous results as well as the presentation of new ones related to the notion
of the MLI and MRI matrix. Secondly, we present appropriate parametrizations for the plant
and controller and develop an adaptive law for updating the controller parameters so that the
overall MRAC scheme is robust with respect to unmodelled dynamics and bounded
disturbances. And thirdly, we give a complete stability proof of the performance and
robustness properties of the developed MRAC schemes. The achievement of these objectives
will enable researchers and engineers to read about MRAC for MIMO plants from a single paper
where issues related to parametrization, controller structure, adaptive law, stability and
robustness are discussed and analysed together.
The paper is organized as follows. In Section 2 we give the definitions of the MLI and MRI
matrix together with the associated plant parametrization and discuss the prior plant
information required to specify the MLI and MRI matrix. The plant and control objective
together with the assumptions about the modelled and unmodelled part of the plant are
presented and discussed in Section 3 . In Section 4 we develop the controller structure for model
reference control (MRC) based on the MLI and MRI matrix by assuming that the parameters
of the modelled part of the plant are known. We then show that these MRC schemes are robust
with respect to bounded disturbances and unmodelled dynamics which are sufficiently small in
the low-frequency range. Thus Section 4 establishes the ideal robustness properties of the MRC
structures which can be achieved if the parameters of the modelled part of the plant are known
exactly. The problem of how to maintain these robustness properties in the case of unknown
plant parameters is considered in Sections 5 and 6 where a general robust adaptive law (GRAL)
for MIMO plants is developed. The GRAL is used to update the controller parameters of the
MRC scheme on-line so that the closed-loop plant is globally stable despite the presence of
unmodelled dynamics and bounded disturbances. The GRAL can be used to generate a wide
class of robust adaptive laws for MIMO plants by simply choosing a certain design function
in the GRAL appropriately. Using our parametrization techniques of Section 3 , we clarify the
prior knowledge about the high-frequency gain matrix of the modelled part of the plant to be
controlled. We show that robust adaptive laws for SISO plants employing a dead-zone, 7 7 8
fixed-09, or switching-a11”2modifications can be extended to the MIMO case, and their
multivariable versions are shown to be special cases of the GRAL. The effect of dominantly
 MODEL REFERENCE ADAPTIVE CONTROL 219

rich reference input signals on the convergence rate and the bounds for the tracking and
parameter error is analysed in Section 6 .
In summary, the main contribution of this paper is the development of a complete procedure
for designing MRAC schemes for continuous time MIMO plants. In doing so, some known
results related to parametrization are extended and combined with new ones in an effort to give
a full account of the parametrization issue in MRAC for MIMO plants. Furthermore, a general
robust adaptive law (GRAL) (an extension of the GRAL in the SISO case of Reference 13) is
developed and a complete stability and robustness analysis is presented which extends most of
the recent robustness global results for SISO plants to the multivariable case. In addition, the
effect of dominantly rich signals recently analysed in Reference 14 for SISO plants in the
presence of unmodelled dynamics is extended to the MIMO case where bounded disturbances
in addition to unmodelled dynamics are allowed to be present.

2. SYSTEM INTEGRAL STRUCTURE AND PARAMETRIZATION

An important concept for designing MIMO MRC schemes is the plant integral structure. This
can be characterized by the Hermite normal form1*2*6or by the interactor matrix.3-5
Depending on the characterization of the integral structure, a convenient parametrization for
the MIMO plant can be given which is useful for designing and analysing MRC. In this section
we present two different characterizations of the integral structure which are based on the
modified interactor matrix, together with the plant parametrization associated with each
characterization. The use of the Hermite normal form for characterizing the integral structure
is very similar to that of the interactor m a t r i ~ ' , ~and
, ~ ,is~ omitted.

2. I . The modified /eft inferactor matrix

The concept of the relative degree of a transfer function in the SISO case can be generalized
to the MIMO case by using the concept of the interactor matrix. The left interactor matrix
introduced in Reference 3 for a proper rational transfer matrix is defined in the following
lemma.

Lemma 2. i 3
For any N x N proper rational full rank transfer matrix TO@)there exists a unique lower
triangular polynomial matrix E l @ ) , defined as the left interactor matrix of TO@),of the form
0 ... ... 0 1

where hri,(s), j = I , ..., N - 1, i = 2, ..., N , are polynomials divisible by s; and li 2 0, for

i = 1, ..., N , are integers, such that

with KI, non-singular.

It is easy to see that the inverse of the left interactor matrix &(s) is not stable. Since in the
220 G. TAO AND P . A. IOANNOU

design of MRC ti'(S) is required to be stable, we need to modify &(s) so that the modified
left interactor matrix has a stable inverse. Such a modification is given in Lemma 2.2.

Lemma 2.2
For any N x N strictly proper rational full rank transfer matrix To(s) there exists a lower
triangular polynomial matrix t;"(s), defined as the modifred left interactor (MLI) matrix of
TO@),of the form
r dT;(s) 0 ... ...

where h & ( s ) ,j = 1, ..., N- 1, i = 2 ,..., N,are poiynomials; and dF(s) = sI I + a'& + ..-+ a;,,
i = 1,..., N,are arbitrary monic Hurwitz polynomials such that

lim ti" (s)To(s)= K lmp (4)

5 + m

with KYi non-singular. Furthermore, hE'(s) can be chosen as

h&(s)= the polynomial part of s'l(h/tj(s)+ afs-'h/ij(s)+ + a;,s-'ih/;,(s)) (5)
for j = 1, ..., N - 1, i = 2 , ..., N.For this choice of h$(s) we have that

lim E?' @)TO

(s) = K/,
S+cO

Proof, The proof is a simple extension of the proof in Reference 3 for the left interactor
matrix and is omitted.

Remark 2. I
Since the polynomials dF(s) in (3) are Hurwitz, the MLI matrix [/"(s) has a stable inverse
and therefore can be used for the design of MRC schemes. The MLI matrix can be easily
obtained from the knowledge of the left interactor matrix, It should be noted that the MLI
matrix is not unique. Given the polynomials dF(s), i = 1, ..., N,however, the MLI matrix is
uniquely specified provided the polynomials hG(s) are chosen according to (5). Furthermore,
det [Kip] = det [Klmp] but K/, # Klmp in general. If, however, the polynomials hQ(s) are chosen
as in (9,then K/, = KC.
The following lemma employs the notion of the MLI matrix to give a parametrization for
the plant which will be used in Section 4 to design MRC schemes.

Lemma 2.3
The MIMO LTI plant
 M O D E L REFERENCE ADAPTIVE CONTROL 22 1

can be represented as
Y(S) = fr(s)(E/n(s))-'Yf(s) Yf(s) = T/(s)u(s) (8)
where T/(s) is an N x N transfer matrix whose MLI matrix is fi(s)I; f i ( s ) is an arbitrary
Hurwitz polynomial of degree dr and d/ 2 the maximum degree of the elements of &!"(s).
Furthermore, the high-frequency gain matrix of Tl(s) is equal to Klmp.

the proof is complete. H

2.2 The modifred right interactor matrix

A similar concept to that of the left interactor matrix is the right interactor matrix which was
introduced in Reference 15 for the discrete time self-tuning control problem. In this section we
extend this concept to the continuous time case and introduce the MRI matrix in the following
lemma, whose proof is similar to that for the MLI matrix and is omitted.

Lemma 2.4
For any N x N strictly proper rational full rank transfer matrix To(s) there exists an upper
triangular polynomial matrix EF(s), defined as the modijed right interactor (MRI) matrix of
To(s), of the form

... ... ...

... ... 0 d%(s)]
where hzj(s), i = 1, ..., N - 1 , j = 2 , 3 ,...,N , are polynomials; and dZ(s) = sr' + bisrl-' + ...
+ bi,, i = 1, ..., N , are arbitrary monic Hurwitz polynomials such that

with Kr"p non-singular. Furthermore, h@(s)can be chosen as

+ b{s-'hrij(s)+ * * *
hZj(s) = the polynomial part of SrL(hrij(S) + b&"/hrij(s)) (13)
for j = 2, ..., N , i = 1, ..., N - 1. In this case we have that
lim .$"(s)To(s) = K,., (14)
s+m

It is easy to see that the inverse of the MRI matrix EF(s) is stable and therefore can be used
in the design of MRC.
The following lemma employs the notion of the MRI matrix to give a parametrization for
the plant which is useful in designing MRC schemes.
222 G . TAO AND P . A . IOANNOU

where T,.(s) is an N x N transfer matrix whose MRI matrix is f,.(s)I; f,.(s) is an arbitrary
Hurwitz polynomial of degree d,. and d , 2 the maximum degree of the elements of EP(s).
Furthermore, the high-frequency gain matrix of T,(s) is equal to KY,.
The proof of Lemma 2.5 is similar to that of Lemma 2.3 and is omitted.

2.3. Prior information f o r specifying the interactor matrix

One of the basic assumptions in the design of MRC is that the interactor matrix is known
exactly. The interactor matrix is unique whereas the modified interactor matrix depends on the
choice of its diagonal elements which can be arbitrary. If, however, the interactor matrix is
known, then the modified interactor matrix can be specified in a straightforward manner. One
of the advantages of using the modified interactor matrix, as is clear from Lemmas 2.3 and 2 . 5 ,
is that the plant transfer matrix can be parameterized to have a simple modified interactor
matrix, i.e. fi(s)I or fr(s)I. Such a parametrization is helpful in solving the non-commutativity
problem in the design of robust MRAC, as will be shown in the sections to follow.
The prior information about the modelled plant transfer matrix To@)required to deduce the
interactor matrix is very crucial in adaptive control because the parameters of To@) are
assumed to be unknown. From ( 3 ) (resp. (11)) it is clear that if the relative degree of each
element of To(s) is known and Ef"(s) or &(s)(resp. &"'(s) or &(s)) is diagonal, then
&!"(s)(resp. ,$"(s)) can be specified without any a priori knowledge about the parameters of
To(s). Furthermore, in this case K,, = KY, (resp. K/, = K E ) will not depend on the parameters
of To@) other than the high-frequency gains of the elements of TO@).If, however, [r"(s)(resp.
&"'(s)) is non-diagonal, then the knowledge of the relative degree of each element of TO($)may
not be sufficient to determine t:"(s) (resp. &"'(s)). Furthermore, K,, (resp. KI,) may depend on
more parameters of To@) in addition to the high-frequency gains of the elements of To(s). In
most of these cases the following lemma may be used to design a precompensator W,(s) or filter
W&) so that the modified plant transfer matrix To(s)W,(s) or Wf(s)To(s) has a diagonal
interactor matrix and therefore can be specified without the knowledge of the parameters of
To@).

Lemma 2.6
Let To(s) be an N x N rational matrix with unknown parameters, but with elements whose
relative degrees are known exactly. Then there exists a stable, diagonal and non-singuIar
dynamic precompensator W,(s) (resp. filter Wf(s)) such that To(s)W,(s)(resp. Wf(s)To(s))
generically (in almost all the cases) has a diagonal left (resp. right) interactor matrix iff To(s)
is non-singular.
The proof of Lemma 2.6 is a long one. It follows exactly the same steps as that of Reference
6 for the Hermite form and is omitted.
 MODEL REFERENCE ADAPTIVE CONTROL 223

We use the following example to illustrate the use of Lemma 2.6 for designing the
precompensator W,(S) or filter W&).

Example 2. I
Consider
1
( s + OI)(S+ a2) S + b
(17)

(S + CI)(S + ~ 2 s) +d
where a,, c,, i = 1,2, b, d , k , h are unknown constants. The MLI matrix of TO@)is of the form

where PO,41, q o E R t are arbitrary known constants. Hence ti"(s) is not completely specified
unless h is known. The high-frequency gain matrix given by

K g = [k!h h(b-d)'I
depends on b , d in addition to the high-frequency gains k , h.
It can be shown that for h # k a precompensator of the form

can be used for To(s) given in (17) such that To(s)W,(s) has a diagonal MLI matrix with the
high-frequency gain matrix given by

which is independent of the parameters of TO@)other than the high-frequency gains of its
elements.
Similarly, it can be shown that for
1 k
( s + a l ) ( s + a21 (s+ a3)(s+ l74)
To@) = 1 h ]h$k (22)
S + l75 S + a6

the MRI matrix is not diagonal and cannot be completely specified unless h is known. In this
case a filter of the form
224 G. TAO A N D P. A . IOANNOU

can be used for TO@)given in (22) such that Wf{s)To(s) has a diagonal MRI matrix with the
high-frequency gain matrix given in (21).
In some cases the use of precompensator or filter can be avoided by considering both the MLI
and MRI matrices. The following example illustrates such a case.

Example 2.2
Consider the transfer matrix To(s) given in (17); its MRI matrix can be chosen as

E,?’(s)=
r2+ p:+ p z s +O 42 I
where p1, pz,42 E R+ are arbitrary known constants. Furthermore, the high-frequency gain
matrix is given by (21) which is dependent only on h and k. Hence the MRI matrix for TO@)
given in (17) is completely specified without knowing the plant parameters whereas the MLI
matrix is not.
Similarly, it can be shown that for the transfer matrix To@) given in (22) the MLI matrix
is diagonal and therefore completely specified whereas the MRI matrix is not.

Remark 2.2
In examples 2.1 and 2.2, if h = k , it can be shown that the MRI or MLI matrix of To(s) given
in (17) or (22) cannot be specified without the knowledge of some of the parameters of TO@).
Furthermore, in this case the use of a precompensator or filter does not lead to a compensated
or filtered To(s) with a diagonal and therefore fully specified MLI or MRI matrix.

3 . PLANT AND CONTROL OBJECTIVE

Consider the MIMO LTI plant with N inputs and N outputs

where G(s),the plant transfer matrix, is strictly proper; To@) is the dominant or modelled part
of C(s); w ( t ) is a bounded output disturbance; pA,(s) and pAa(S) are the multiplicative and
the additive unmodelled plant perturbations respectively. Without loss of generality and for
clarity of presentation, A,(s) and A,@) are rated by the same scalar I” > 0.
For the unmodelled part of the plant we make the following assumptions:
(Al) A,(s), A,@) are rational transfer matrices, and A,@) is strictly proper.
(A2) Let f,(s)(resp. fr(s)) be a monic Hurwitz polynomial of degree d , (resp. dr), the
maximum degree of the MRI matrix &!“(s)(resp. the MLI matrix ti”)of To@), with all
its roots in Re [s] < - q, and define
1
D, = lim Aa(S)S D,,, = lim -A,,,(s)s (26)
s- = f(s)
s-‘m

where f ( s )= f,(s)(resp. f , ( s ) ) ;then there exist constants k,, k , > 0 such that
II Da I1 ka II D m II km (27)
 MODEL REFERENCE ADAPTIVE CONTROL

Remark 3.1
We should note that in (25) A , ( s ) is allowed to be improper as long as C ( s )and therefore
To(s)A,,(s)are strictly proper. Hence we may have
p 11 A,(jw)\l -+ ~0 as w --t co
for any p > 0.

Remark 3.2
We should stress that assumption (A2) could be relaxed by augmenting (25) with an input
filter, i.e.
Y (s) = G ( s > u f ( s ) Uf@) = Wf(S)IU(S)

Choosing

assumption (A2) can be replaced with

Furthermore, (Al) can also be relaxed to include A m @ ) Aa(s)

, which are not rational, as shown
in Reference 16 for the SISO case.
The control objective may be stated as follows. Given the reference model
Y ~ ( s=) Wm(s)r(s) (29)
where W , ( s ) is a strictly proper stable transfer matrix to be selected and r ( t ) E R" is a piecewise
continuous uniformly bounded known reference input signal, find the control input u(t) such
that the output y ( t ) of the plant tracks yrn(t) as close as possible and all the signals in the closed-
loop plant are uniformly bounded for any bounded initial conditions. To achieve such an
objective, we need to make some assumptions about the modelled part of the plant and
reference model (29). These assumptions depend on the particular parametrization of the plant
and are given below.
If the MRI matrix is used, then we assume that:
(A3) The MRI matrix &!"(s) of To(s) is known.
(A4) An upper bound vr on the observability index of f;'(s)To(s)&!"(s) is known.
(A5) The matrix K,. such that the matrix K,.KFPis positive definite is known, where Kg is the
high-frequency gain matrix associated with the MRI matrix Ef"(s).
(A6) The degree d,. of f ( s ) is chosen so that f(S)Wm(S) is proper.
226 G. TAO AND P. A. IOANNOU

If the MLI matrix is used, then we assume that:

(A3') The MLI matrix Ei"(s) of TO@)is known.
(A4') An upper bound V I on the observability index of fl-'(s)Ef"(s)To(s)is known.
(A5') The matrix KI such that the matrix KIK,", is positive definite is known, where KG is
the high-frequency gain matrix associated with the MLI matrix Ei" (s).
(A6') The degree 4 of f i ( s ) is chosen so that fr(s)W,(s) is proper.
In addition to assumptions (A3)-(A6) or (A3 ' )-(A6') we also assume that
(A7) To(s) is non-singular and has stable zeros.

Remark 3.3
Assumption (A3) or (A3') is equivalent to the assumption of known relative degree in the
SISO case. As we indicated in Section 2, if the relative degree of each element of To@)is known
and the MRI matrix or the MLI matrix is diagonal, then the MRI matrix or the MLI matrix
is completely specified without the knowledge of the parameters of TO@).If neither the MRI
matrix nor the MLI Matrix is diagonal, then their off-diagonal elements may depend on some
of the parameters of To@)and therefore cannot be completely specified unless these parameters
of To@) are known. However, as stated by Lemma 2.6 in Section 2, generically a known
precompensator or filter can be used so that the compensated or filtered transfer matrix has
a completely specified diagonal MRI matrix or MLI matrix.
The knowledge of the upper bound or V I for the observability index in (A4) or (A4' ) is
equivalent to the upper bound for the order of the transfer function in the SISO case, which
is used to parametrize the MRAC system. This knowledge is necessary to achieve the exact
model matching in the MRC system when w ( t )= 0, p = 0 in (25). Assumption (A4) (resp. (A4 ' ))
requires an upper bound for the observability index of f; '(s)To(s)4f"(s) (resp.
f;'(s>&"'(s)To(s)) which involves only the dominant part of the plant, To(s). However, the
observability index of G ( s ) is allowed to be arbitrarily large and unknown.
Assumption (A5) or (A5 ') is the generalization of the assumption that the sign of the high-
frequency gain is known in the SISO case. As shown in Section 5 , such knowledge of the high-
frequency gain matrix is used of develop a stable adaptive law for estimating the controller
parameters.
Since MRC involves the cancellation of the zeros of the plant, these zeros are assumed to
be stable. In our analysis we assumed (A7) that only the modelled part of the plant has stable
zeros. This, however, does not imply that the overall plant given by (25) has stable zeros. This
is, the unstable zeros of G ( s ) can be treated as part of the plant uncertainty paln@),pA,(s)
provided these zeros are fast, i.e. O(l/p), as illustrated by the following example.

Example 3.1
Consider

G ( s )=

i
1

1
&I
s(s- 1) s(s - 2)

1-P
s(s - 3) s(s - 4)
 MODEL REFERENCE ADAPTIVE CONTROL 227

which has an unstable zero for any g > 0. A model reference control law cannot be designed
for G(s) since such a law will involve the cancellation of the unstable zero and will lead to
internal instability. If we use the parametrization (25), G(s) can be expressed as

r
where

~~
1 1
s(s - 1 ) s(s - 2)
To@) = G(s)l,=o =
~~
1 1)
s(s - 3) s(s - 4)

S(S - l)(s- 3)
I
That is, the unstable fast zero is treated as part of the multiplicative uncertainty, which, as we
will show in the following sections, can be handled by the MRC law designed for the minimum
phase dominant part To(s).

4. MRC: KNOWN PARAMETERS

In this section we assume that in addition to (Al)-(A7) the parameters of To@), the modelled
part of the plant, are known and design the control input u ( t ) . Since the controller structure
will be different for different parametrizations of TO@),each case is considered separately. The
controller structure will be designed based on the modelled part To(s) of the plant. However,
the robustness and stability properties of the control design will be analysed when applied to
the actual full order plant.

4.1. MRC based on the MRI matrix

We use Lemma 2.3 to express the plant (25) as
Y (s) = Gr(S)v(s) + w u(s) =f;'(s)EP(s)v(s) (34)
228 G . TAO AND P. A . IOANNOU

where
c,(s,e;)=e;lsy~-2+ ..-+e;,,-,
D r ( S , 0 5 ) = 0 $ 1 S " - ' + * * . + eiv,-l+ e i , N , ( s )
e;= [e;,,...,e ~ , T- ~ i e i = [eil,..., eiy,-I,e'z,,~ 7

and Nr(s) = diag ( n r ( s )1 ; n, (s) is an arbitrary monic stable polynomial of degree Y , - 1 and Kb
is a constant matrix.
Using a similar procedure as in References 2 and 4, it can be shown that there exist constants
matrices Or*, ei* and K6* such that the modelled part of G,(s), i.e. T,(s) = R,(s)P;'(s)
satisfies the matching condition
R,(s) [ (N, (s) - C r ( S , ei*))Pl(s)- D, (s, ei*)R,(s)l - ' N J(s)Kh*f, (s) = I (37)
Since the parameters of R,(s) and P,(s) are known, Or*, ei*and Kh* can be calculated from
(37) and used in (36) to form the desired control input for MRC, i.e.
v = Cr(S, ei*)N;'(s)v + D,(s, &*)N;'(s)y + K6*f,(s)W,(s)r
u = f;'(s)tf"(s)v (38)
which guarantees that I/ y ( t ) - ym(t)ll + 0 exponentially fast provided that w ( t ) = 0, p = 0.

G / ( s )= Ti@)[ I + pAfn(s)l + PA',@)

T/(s) = J-'(s)Ei"(s)To(s) = R/(s)Pl'(sf
A',@) = Aln(s) AL(s) =fi-'(s)Ef"(s)
Aa(s) w 4 s ) =fi-'(s>ti"(s)w(s) (40)
and R/(s) and P/(s)are right coprime polynomial matrices of dimension N x N and Pi@) is
column proper.
We propose the following controller structure:
u = C / ( S0:)NT'(s)u
, + D/(s,e:)NF'(s)Yf+ KbEP(s)Wm(s)r
Yf(S) =fi-'(s)E;"(s)y(s) (41)
where
c,(s,e:)= e:lsv/'-2
+ ... + eL-l
D/(s, e',)= C~',IS''-* + *.* + 0L- I + 0:J%(s)
e: = [ e L..., e:e,-l] e: = [e21,
...,eiv,-,,
/
eLi T
and N/(s) = diag( n / ( s ) ) ; n/(s) is an arbitrary monk stable polynomial of degree V I - 1 and Kb
is a constant matrix. In a similar manner as in the case of the MRI matrix, it can be shown
that constant matrices 0[*,0? and K6" exist such that the modelled part of G , ( s ) , i.e.
 MODEL REFERENCE ADAPTIVE CONTROL 229

Tl(s)= R/(s)Pi (s) satisfies the matching condition

R/(s)[ (N/(s) - G(s,e;*))P/(s)- D/(s, &*)Ri(s)I -'N/(s)Krfl(s) = I (42)
Since the parameters of R/(s)and P/(s)are known, e:*, &* and KI* can be calculated from
(42) and used in (41) to form the desired control input for MRC, i.e.
u = G ( s , Oi*)Ni'(s)u + D/(s, &*)Ni'(s)yf + KrEf"(s)Wm(s)r '

Yf(S) =.f-l(s)E/m(s)Y(s) (43)

which guarantees that )I y ( t ) - Ym(t)ll -+ 0 exponentially fast provided that w ( t ) = 0, p = 0.

Remark 4.1
The controller structures in both the MRI and MLI matrix cases are similar; however they
use different signals in the control laws, i.e. v(t) in the MRI matrix case is an internal fictitious
variable used to generate the control signal u ( t ) , while yf(t) in the MLI matrix case is the filtered
plant output which is a physical variable. The implementation of (36) or (41) in the time domain
is the same as that in References 2 and 4.
In the following analysis we will omit the subscripts 'r', 'I' in n(s),f ( s ) , v, K, KE, etc. and
the superscripts 'r', ' r , 'm' in 81,&, Ko,K,, etc., and use them only where needed for
clarification.

4.3. Robustness analysis

Let us now apply the desired control input (38) or (43), which meets the control objective
when p = 0, w ( t )= 0, to the actual plant (25) with p # 0, w # 0.
The stability and robustness properties of (38) and (43) when applied to (25) are given by the
following theorem.

Theorem 4.1
For some p* > 0 and v p E [0, p*) all signals in the closed-loop plant (25), (38) or (25), (43)
are bounded for any bounded initial conditions. Furthermore, the tracking error
e l ( t ) = y ( t ) - ym(t) converges exponentially to the residual set
S" = (el I II el 11 < + wo)) (44)
where k E R + and 1-0, wo are the upper bounds for 11 r(t)ll and 11 w(t)ll respectively.
The proof of Theorem 4.1 is given in Appendix I.

Remark 4.2
Theorem 4.1 guarantees that the desired control laws (38) and (43) are robust with respect
to unmodelled dynamics and bounded disturbances. Furthermore, the residual tracking error
depends on the bound for the output disturbance w ( t ) and the strength of the unmodelled
dynamics (represented by p ) in such a way that as these plant uncertainities disappear, the
tracking error reduces to zero. The robustness and stability properties described by Theorem
4.1 are obtained under the assumption that the parameters of the modelled part of the plant
are known exactly. If these parameters are not known, the control laws (38) and (43) cannot
230 G. TAO AND P. A. IOANNOU

be implemented. In this case, instead of (38) (resp. (43)), (36) (resp. (41)) is used together with
an appropriate adaptive law for updating the unknown controller parameters. A class of such
adaptive laws will be presented and analysed in Section 5 . The objective in the unknown plant
parameters case is to develop adaptive control schemes which guarantee that the robustness and
stability properties of the closed-loop plant are as close as possible to those described by
Theorem 4.1 for the MRC with known parameters.

5. A ROBUST ADAPTIVE LAW FOR MIMO PLANTS

The stability and robustness properties of MIMO MRC schemes are established in Section 4
by assuming that the parameters of TO@),the modelled part of the plant, are known exactly
so that the controller parameters can be calculated. If the parameters of To(s) are unknown,
the controller parameters cannot be calculated and therefore the MRC schemes of Section 4
cannot be implemented. In this case a suitable estimation scheme or adaptive law has to be
developed for estimating the controller parameters on-line, and combined with the MRC
structure in such a way that the robustness properties of the MRC schemes presented in Section
4 are preserved.
A wide class of robust adaptive laws can be developed by examining the properties of a
certain augmented measured estimation error el ( t ) . To express such an error explicitly, we first
give the following lemma.

Lemma 5. I
The desired parameter matrix 8* = [ e:', e:',Ko*] of the MRC scheme satisfies the
following equation:

where

A(s) is a strictly proper transfer matrix with stable poles, d ( t ) is bounded function which
depends on w ( t ) , and
a = u(t) Y = Yf(t) Ym(s) =fi-'(s>Ci"(s)~m(s)
if the MLI matrix is used and
a = v(t) Y = Y (f> Ym(s) = Ym(s)

if the MRI matrix is used.

The proof of Lemma 5.1 is given in Appendix 11.
In view of ( 4 3 , the measured estimation error € 1 is defined as
&l(t) = E 1 ( t ) + lLo(t)[eT(t)r(f)-f-'(s)a(t)l (46)
where e(t),#,,(t) are the estimates of 8*,Ko*-' = Kp respectively at time t. Defining
(P=e-e* 1L = $0 - Kp
 MODEL REFERENCE ADAPTIVE CONTROL 23 1

and substituting for in (46) from (45), we obtain

= K,cpT{ + $4+ pq + d (47)

where 4 = eT{ - f - ' ( s ) u is available for measurement.

Let us now consider the following cost function:

J=- + GT9 + GZ$

m2
where G l ( t ) and G z ( t ) are time-varying matrices to be designed; m ( t ) is a normalizing design
such that v/m, { / m E L,, and therefore only 9,$ can drive J to infinity. As shown
in Appendix 111, such a property is guaranteed by designing W,(s - q ) ,n - ' ( s - q ) , f - l ( s - q )
to have stable poles and choosing m ( t ) as the state of

h(t)= -6om(t) + &(I) u (1 + (I 9 (I + 1) m(O) > 61/80 (49)

where 60,61 > 0 are design parameters and 60 + 6 2 < q for some 6 2 E R'. Considering (47) and
using the steepest-descent method to search for the minimum of J with respect to 9,$, we
obtain

r-&ItT
-
m2
where K satisfies (A5) or (A5'), i.e. KK, is positive definte and I' is positive definite.
The following theorem gives a set of sufficient conditions to be satisfied by G l ( t )and G2(t)
and establishes the stability and robustness properties of the MRC schemes of Section 4 with
0* replaced by 0 ( t ) obtained from (50) when applied to the actual plant (25).

Theorem 5.1
Let G l ( t )and G2(t) be chosen so the the following conditions are satisfied:

(9 dt
(tr[9(t)KlKT9T(t)+ $T(t)l'-l$(t)l) <0
whenever 1) 9(t)ll + 11 $(t)ll > NO, V ~ [O,p?)
E and some No,p?E R'.
(ii) 9 ( t ) ,$ ( t ) E L , implies that

where k, 2 0, g ( t ) E L , and Sf,g(T) d7 < CYO + cl(t - t o ) , to 2 0 , for Some 010, CL 2 0.

(iii) 9 ( t ) ,$ ( t )E L , implies that G I( t ) , G2(t) are globally Lipschitz functions of time.
Then for some ji * > 0 , p * > 0 and vji E [ 0 , ji *), p E [ 0, p *) all signals in the closed-loop plant
(25), (36) (resp. (41)), (50) are uniformly bounded for any bounded initial conditions and the
tracking error el ( t )E S, where

y E R' and 2 > 0 is a small number.

232 G. TAO A N D P. A. IOANNOU

Corollary 5. I
If G l ( t ) and C 2 ( f ) are chosen so that condition (ii) of Theorem 5.1 holds for some
,iiE [0,f ( p ) ] where f(p) > 0 and f ( 0 ) = 0, and if kg = 0 for w(t) 2 0, then in the absence of
unmodelled dynamics (i.e. p = 0) and disturbances (i.e. w(t) = 0 ) the tracking error e l ( t )
reduces to zero asymptotically with time.
The proofs of Theorem 5.1 and Corollary 5.1 are given in'hppendix 111.
Theorem 5.1 and Corollary 5.1 establish the global stability properties of the closed-loop
plant without using any persistent excitation conditions. Moreover, Theorem 5.1 and Corollary
5 . 1 can be used not only to generate and analyse a wide class of robust adaptive laws for MIMO
plants, but also to check whether a given adaptive law is robust. In the following lemmas we
present and analyse several robust adaptive laws for different choices of G I ,G2.

Lemma 5.2 (dead-zone)

Assume that d ; ~an , upper bound for I di(t)l, is known, i = 1, ..., N , where d ; ( t )is the ith
component of d ( t ) . Let G I ,G2 be chosen as
-slT
G T1 - - G2 = gtT

gi= i -

where g; is the ith component of g(t), i = 1, ..., N , and gio > 0, i = 1, ..., N, are design
parameters. There exists a g* > 0 such that if gio E (0, g * ) then the existence of a p * > 0 is
guaranteed, so that for p E [ 0 ,I*) all signals in the closed-loop plant (25), (36) (resp. (41)), (SO)
and (52) are uniformly bounded for any bounded initial conditions and the tracking error
el E &where

ydER+ and go=maxi[gio].

Proof. In view of (AS) and (AS'), we choose the positive definite function
+
V ( t )= tr[@(t)KiKTQT(r)] tr[J/T(t)r-lJ/(t)] (54)
Then

(55)

From Lemma 111.1, q/m is bounded, and from ( 5 2 ) and ( 5 5 ) it follows that
 MODEL REFERENCE ADAPTIVE CONTROL 233

where
0 if I el; I < g;om + d;o
s;= (57)
1 otherwise
and V;O is the upper bound for 1 7;I/m, i = 1, ..., N.
Setting

p? = min (g;o/m)
i

it follows that for p E [0, I?), < 0, V t 2 0, i.e. condition (i) is satisfied.
(ii) Since 6, E L , implies that t/mE L,, we have

II GI II + II G2 II < I/ s Il(llt I1 + II t II) < 41/m+ 8 0 4 2

where ql, 4 2 E R+.Choosing In. = go@, condition (ii) is satisfied.
(iii) Using Lemma 111.2, we have

Hence G I , & E L,. This implies that condition (iii) is satisfied. Therefore using Theorem 5.1
and setting g* = j i * / q 2 , the proof is complete. rn

Remark 5.1
In order to implement (52), the upper bounds d;o for the unknown disturbance
1 d;(t)l,i = 1, ..., Nhave to be known. Furthermore, the design parameters g;o have to be
sufficiently small in order to satisfy the inequality 0 < gio < g*, since g* = p*/q2 which depends
on many unknown terms and cannot be calculated.
Our analysis shows that making g;o smaller results in a small p * , since p * < min;(g;o/v;o), and
therefore in a smaller stability margin. Another drawback of the dead-zone is that in the
absence of unmodelled dynamics (i.e. p = 0) and disturbances (i.e. w ( f )= 0) the tracking error
cannot be shown to go to zero unless the design parameter g;o = d;o = 0 , i = 1, ..., N.

Lemma 5.3 (fixed-a modifrcation)

Assume that w0, an upper bound for 11 w(t)ll, is known. Let G I , G2 be chosen as

where a > 0 is a design parameter. There exists a a* > 0 such that if a E (0, a*) then the existence
of a p* > 0 is guaranteed, so that for p E [0, p * ) all signals in the closed-loop plant (25), (36)
(resp. (41)), (50) and (60) are uniformly bounded for any initial conditions and the tracking
error e l 6 S , where

y o €R + .
234 G . TAO AND P. A. IOANNOU

Proof. Choose the positive definite function

V ( t )= tr[+(t)K;KT+'(t)] + tr [$T(t)I'-l$(t)l (62)
Then

p < - ( y ) 2 - u ( 2 ( t r [ @ K K , @ T ] + tr[KK,+'B*] +tr[$'P-'$] +tr[r-'$$:] '""')

-- U

):(
-
2
[2(tr[+KK,+'I + tr[KK,iPTO*l + tr[$'r-'$] +tr[r-l$$~l)-fll (63)

where f i s the upper bound for 11 7 Il/m,fi E R + .

Since there exist m3, c f R', cij E R such that

tr[KK,3T8*] =
i
c c@;j(t) < CZZI
j i j
@ij(t)l

and there exist @3, C E R', Cij E R such that

for some c0,q3,q4ER+.The boundedness of e(t),$ ~ ( t implies ) that q3,q4are finite, so

condition (ii) is satisfied by setting ,G = uq4, u* = ,G*/q4.
(iii) From Lemma 111.2 we have that GL, G 2 exist and are bounded and therefore condition
(iii) is satisfied. Therefore using Theorem 5.1, the proof is complete.

Remark 5.2
The fixed-a modification scheme has the same drawbacks as the dead-zone scheme. T o
implement (60), an upper bound for the output disturbance is needed and the design parameter
a should satisfy 0 < u < u* where a* cannot be calculated unless the plant is known exactly.
Making a sufficiently small reduces the stability margin expressed by u* > 0 since p* < p: = ,u.
Furthermore, the tracking error may not go t o zero at t co when w ( t ) = 0, p = 0, unless
+

u, w o are set to zero in (60).

 MODEL REFERENCE ADAPTIVE CONTROL 235

Lemma 5 . 4 (El modiBcation)

Assume that the upper bound wo for /I w ( f ) / / is known and G I ,G2 are chosen as

where XO > 0 is a design parameter. There exists a A,* > 0 such that if XOE (0, A$) then the
existence of a p* > 0 is guaranteed, so that for p E [0, p * ) all signals in the closed-loop plant
(25), (36) (resp. (41)), (50) and (68) are uniformly bounded for any initial conditions and the
tracking error el E SElwhere

Proof. Define the positive definite function

V ( r )= tr [9(t)KiKTQrT(t)]+ t r [ $ T ( t ) r - l $ ( t ) ]
Then

Xo[2(tr[cPKK,9T] + tr[KK,QrTB*]
m

+ tr[$Tr-l$l +tr[r-l$$Or])] -pf+""

m
[2(tr[QrKK,9Tl +tr[KK,cpT8"]

+tr[$Tr-l$] +tr[r-'$$b])-m4]) (71)

for some m4 E R + . Following (64), (65), we have that for any finite c3 > max [ m4,fl there exists
c4 > 0 such that
2(tr [ KK,QrT8* + Q ~ K K+, r~- ~l $ ~ $ T + 2 $ T r - '$1 > c3 (72)
whenever /I Qr 11 + 11 $11 > c4 and p E [ O , XO]. It follows that V(t) is bounded. Hence condition
(i) is satisfied by setting p? = XO.
(ii) From Lemma 111.2 it follows that El/m E L,. Therefore
11 G I I\ + 11 G2 11 < 4 5 x 0 + 46/m
where 45,46 > 0 depend on the bounds for 11 f I Il/m, 11 8 (I, 11 $0 11 and on WO.Setting G, = 4 5 x 0 ,
condition (ii) is satisfied.
(iii) From Lemma 111.2 it follows that l/m, 11 E l ll/m, 0, GO are Lipschitz functions of time.
Since .cl/m, l/m E L,, the Lipschitz property of G I , G2 follows directly.
Therefore using Theorem 5.1 and setting Xo*=,l*/q5, the proof is complete.

Remark 5.3
The el modification scheme suffers from the same drawbacks as the dead-zone and the fixed-a
modification schemes. The design parameter XO in (69) should satisfy 0 < XO < A,* = ji*/qs for
some unknown constant X,* which depends on the plant. Choosing XO sufficiently small may
meet the upper bound A$; however, this may reduce the stability margin since p * < XO. When
236 G . TAO A N D P . A . IOANNOU

w(t) = 0, p = 0, the tracking error may not be reduced to zero as t -+ co unless XO, wg are set to
zero in (68).

Lemma 5.5 (switching-a modification)

Let G I ,G Lbe chosen as

where

d , = min [ X,(KK,)] d2 = m i n [ X , ( r - ' ) I

I I

and uO> 0 is an arbitrary design parameter.

There exists a p* > 0 such that for p 6 [ O , p * ) all signals in the closed-loop plant (25), (36)
(resp. (41)), (50) and (73) are uniformly bounded for any bounded initial conditions and the
tracking error el E So,, where

Proof. Choose the positive definite function

V ( f )= tr [+(t)K,fKT+T(t)] + tr [ $ 7(t)F-'$(f)l (75)
Then
 MODEL REFERENCE ADAPTIVE CONTROL 237

from (76) we have

tr[2uIKK,iPT€3 + 2d-'J/J/:] G - e+ + dnl/m (78)
Owing to the choice of 61,u2, we have
ultr [ KK,iPT€3] 2 0 uztr [ ~ - I J / ~ 3J o/ ~ ] (79)
Therefore for some m5, m6 E R+ we have
/I G I I/ G m s a 1 t r [ K K , 9 ~ 0 1 /I ~2 I/ G m 6 ~ z t r [ r - ' $ ~ h 1 (80)
Hence
I/ G I (1 + /I Gz /I G m7[ - e+( P V O ) +~ dln/ml (81)
for some m7 > 0.
Setting
2
g = m7 [ - ri + (p110) I k, = m7dnl = ffl7(pvO) (82)
condition (ii) is satisfied.
(iii) From Lemma 111.2 we have that cl/m, f / m ,E/m, 6,$0 E L,. Therefore €3, $0 are
Lipschitz. Hence G I ,Gz are Lipschitz. Hence Lemma 5.5 follows from Theorem 5.1.

Corollary 5.2
In the absence of plant uncertainties (i.e. p = 0, w(t) = 0 ) the switching-u modification
guarantees that the tracking error el(() reduces to zero asymptotically with time.

~ k, = 0
Proof. It follows directly from Corollary 5.1 by noting that f ( p ) = m - / ( p 7 0 )and
when w(t) = 0, as in (82). m

Remark 5 . 4
The switching-u modification scheme guarantees that the tracking error goes to zero as t 05 -+

in the absence of unmodelled dynamics and bounded disturbances. It requires no information

about the bound for w ( t ) in the implementation of the adaptive law. The design parameters
M I , Mz are much easier to choose than go, a and XO in the dead-zone, fixed-u modification and
C I modification schemes respectively.

6. DOMINANTLY RICH SIGNALS AND IMPROVEMENT OF PERFORMANCE

The robust MRAC schemes developed in Section 5 guarantee that the closed-loop plant is
globally stable and the tracking error is small, i.e. of the order of the modelling error
characterized by the scalar p and the bound for the disturbance, in the mean for any bounded
piecewise continuous reference input signal r(t). Smallness in the mean, however, does not
guarantee small bounds at steady state. That is, even though the tracking error is guaranteed
to be bounded for all time, it may assume values within short intervals of time which are larger
than the order of the modelling error. Such a behaviour may give rise to what are called burst
and/or bifurcation phenomena, recently discussed in References 18-21. In this section we show
that if the reference signal r ( t ) belongs to a certain class, i.e. the class of dominantly rich
238 G . TAO AND P . A. IOANNOU

signals, l 4 then the convergence rate and tracking performance of the robust MRAC scheme are
improved considerably.
The notion of the dominant richness of input signals was used in Reference 22 for the case
of identification, and then used in Reference 14 for the case of robust adaptive control for SISO
plants in the presence of unmodelled dynamics. In this section we extend the results of
Reference 14 to the multivariable case where both unmodelled dynamics and bounded
disturbances are allowed to be present. For clarity we consider the case when K, is known, and
without loss of generality we assume that K, = K = 1. For the sake of completeness we present
the following definitions which can also be found in Reference 14.

Defrnition 6.1 l4
A bounded function x : [0, m) R" is said to be persistently exciting (PE) with level of
+

excitation p if there exist 6 > 0, /3 > 0 such that

.\ t6
x(t)x7 (t) dt 2 PSI vs 20 (83)
5

Defrnition 6.223
A bounded function x: [ O , o o ) -+ R" is said to have a spectral line at frequency w of amplitude
X(w) iff

lim
T-w T
s $+

9
7
x(t)exp(-jut) d t = X(W) vs 2 0 (84)

When X ( o ) # 0 we say that x(t) has a spectral line at w .

Defrnition 6.3
The reference input signal r(t) is dominantly rich for the closed-loop plant if the vector signal
{(t) is PE with level of excitation /3 > O ( p ) + kowo where ko t R * .

Remark 6.1
Note that owing to the presence of disturbances Definition 6.3 for dominantly rich signals
differs from that of Reference 14 where only unmodelled dynamics was allowed to be present
in the plant. The presence of bounded disturbances requires the level of excitation to be larger
than some constant24which depends on the bound for the disturbance plus the O ( p ) term due
to the unmodelled dynamics. In References 25 and 26 classes of PE signals were introduced for
parameter identification of systems without any unmodelled dynamics and disturbances, where
the level of excitation of a PE signal, which may guarantee the parameter error to go to zero
exponentially fast, can be very small. But for dominantly rich signals the level of persistency
of excitation should be larger than the order of modelling errors. In the case of no modelling
errors,23925*26if the reference input r ( t ) to the adaptive system has as many frequencies as the
estimated parameters, then the signal vector r ( t ) in the closed-loop system is PE. If some
frequencies are too large or some magnitudes are too small, then the level of PE is small. A
small level of PE results in a slow rate of convergence in the case of no modelling errors.
 MODEL REFERENCE ADAPTIVE CONTROL 239

However, as shown in Reference 14, in the presence of modelling errors reference inputs r ( t )
with arbitrarily large frequencies or small magnitudes may not guarantee { ( t ) to be PE.
The following theorem specifies the class of dominantly rich signals and shows that such
reference signals guarantee exponential convergence and tight bounds for the parameter and
tracking error.

Theorem 6.1
If r ( f ) has spectral lines at frequencies W I ...,
, WN,,N r 2 2Nu with (w;,R(w;)) E Q where
11 R(w)/l > O ( p ) + kswo I
0 = [ w E R , R(w ) E C N: I w I < O(1 / p ) , (85)
I W; - wj I > O ( p ) ,for i # j , and some ks E R + ,then for some p$ > 0, i.O* > 0 and v p E [ 0 , p:),
,!Ei (0,,Go*), r ( t ) is dominantly rich for the plant (25). Furthermore, all signals in the closed-loop
plant (25), (50) are uniformly bounded for any initial conditions and the tracking and
parameter error converge exponentially to the residual set
D,= ( e l E R N ,i p E R ( 2 Y N ) xIllNel I1 + /I Q, I1 < k l ( P + wo+ g1)I (86)
where gl = supI,,, /I G ~ ( t ) l for
l some tl > 0, kl E R t

Proof. The proof is similar to that in Reference 14 for SISO plants and is outlined as
follows. Define r m ( t )to be the vector signal corresponding to { ( t ) when €3 = 8*and w ( t ) = 0,
p = 0; then = Q 1(s)r where Q I (s) is a strictly proper transfer matrix with stable poles. It can
be shown that the GRAL guarantees that { ( t ) - f m ( t ) is 7 3 ( p + wo + i + &)-small in the mean,
where 7 3 E R for any general reference input signal r ( t ) . If r ( t ) has Nf2 2mu spectral lines
+

which belong to the set Q given by ( 8 5 ) , then it can be shown that r m ( t ) is PE with level
Prn> O ( p ) + k9wo for some k9 E R'. This property of t m ( t ) together with the fact that { * ( t )
is y 3 ( p + w o + i + €)-close to { ( t ) most of the time guarantees that { ( t ) is PE with level
/3 > O ( p ) + kowo for some ko E R + and V p E [0, PO*), I*. E (0,I*.$) and some p: > 0, i$> 0. From
(47) and (50) we have that

where q l ( t )E L,, 11 wl(t)ll < bswo, bs E R + . Since { ( t ) is PE with level /3 > O ( p ) + kowo, the
homogenous part of (87) is exponentially stable. Hence there exist constants a2, a3 E R + such
that the transition matrix +(t,t o ) of (87) satisfies
// + ( t , to)\/ < wexp [ - a2(t - toll vt 2 t o (88)
Using (87) and (88) and the fact that
el = Q z ( ~ ) + ~ ( t ) x (+t )pq2 + w2 (89)
where x(t) E L,, q z ( t )E L,, (1 wz(t)ll < b 6 W & b6 E R ', Q2(s)is some transfer matrix with stable
poles, the residual set (86) can be derived.

Remark 6.2
As in Reference 14, it can be shown that if r(t) is not dominantly rich, then the convergence
of the signals in the adaptive loop may be slow and the bound for the tracking and parameter
error at steady state may be large.
240 C. TAO AND P . A. IOANNOU

Remark 6.3
The bound for the tracking and parameter error in (86) depends on the constant gl in
addition to 1 and wg terms; this arises because of the design matrix G , ( t )in the adaptive law.
I t can be shown that if G I( I ) is chosen as in (73), then gl = 0 and therefore the residual tracking
error depends only on the bound for the unmodelled plant uncertainties. Hence the
performance of the robust MRAC is qualitatively the same as that achieved in Section 4 for
the known parameter case.

Remark 6.4
If the desired reference input signal r d ( t ) to be tracked is not dominantly rich, then the
properties of Theorem 6.1 can be achieved by adding a signal r o ( t ) to r d ( t ) so that
r(t) = r d ( f ) + ro(t) is dominantly rich. In this case the convergence of the signals will be
exponential but the bound for the tracking error will be given by

lim
/-m
SUP
120
(I el (t)\l < kl ( p + w0 + gl + 7 0 ) (90)

where kl E A+ and FO is the bound for ro(t). The bound (90) indicates that there is trade-off
between convergence rate and tracking performance. For exponential convergence ro has to be
larger than O ( p ) + kswo, as indicated by (85). Large Fo, however, increases the bound for
tracking error in (90). Therefore when the desired reference input signal is not dominantly rich,
an additional test signal can be added to it to make it dominantly rich. Such a signal, however,
has to be switched on and off by monitoring the adaptation process so that the convergence
rate and tracking error are both acceptable.

7. CONCLUSIONS
In this paper we have presented a complete procedure for designing and analysing robust
MRAC schemes for MIMO plants. We use the notion of the MLI and MRI matrix to
characterize the integral structure of the modelled part of the plant and develop convenient
plant parametrizations. These parametrizations are used to design MRC structures which are
robust with respect to plant uncertainties such as unmodelled dynamics and bounded
disturbances. A general robust adaptive law (GRAL) for MIMO plants is developed and shown
to guarantee global stability for the closed-loop plant when combined with the MRC structures,
despite the presence of plant uncertainties. As in the SISO case, l 3 the CRAL involves a general
non-linear design function which can be used to generate a wide class of robust adaptive laws
for MIMO plants. We show that robust adaptive laws for SISO plants which employ a dead-
zone, fixed-cr, E I and switching-a modifications can be extended to the MIMO case and shown
to be special cases of the GRAL. It is also shown in this paper that the convergence rate and
tracking performance of globally stable robust adaptive controllers can be considerably
improved if the reference input signal is chosen to be dominantly rich. Loss of the dominant
richness of the reference input signal may lead to a slow Convergence rate and large tracking
erros at steady state. Since the robust adaptive controller is globally stable independent of the
richness of the reference input signal, dominantly rich signals can be switched on and off
depending on the performance requirements of the particular adaptive control system without
losing stability and robustness.
 MODEL REFERENCE ADAPTIVE CONTROL 24 1

ACKNOWLEDGEMENTS

This work was supported by the National Science Foundation under Grant DMC-8452002 with
industrial matching funds from General Motors Foundation, Inc.

APPENDIX I: PROOF OF THEOREM 4.1

MRI matrix case

From (34) and (37) it follows that

y = W,,,(s)r + p [ (I +f,:'(s)K/i*-'Fi(~)) Al,(s) +f,-'(s)K'.*-'(I - F;(s)) AL(s)l

' ' '
[ I - pT; (s)f,: (s)KC;*- 'F$(s)(A; + T,.(s) A;)] - (T; (s)W,,,(s)r '
'
+ T; ' ( s ) f , (s)K$-'Fi(s)w(t)) + c f y ' (s)K(,*-'F$(s) + I)w(t) (91)
where
F'i (s) = C, (s)N, '(s) F$(s) = D,.(s)N,'(s)

with N,(s), C,.(S),D,- defined in Section 4.1.

(91) is stable if
A i ( s ) + Ah(s)))-'
(I - pTl'(s)KC;*-'F$(~)T,.(s)f,-'(s)(T;'(s)
is stable.
From the choice of fr(s), W,,(s) and the stability of T;'(s), A i ( s ) , A h ( s ) we have that
f,:'(s)(T;'(s)A:(s) + Ain(s)), T;'(s)W,(s), T;'(s)f,-'(s) are stable and strictly proper.
From the model matching condition (37) we have

'
T; (s)K(," 'Fi (s)T, (s) = T; ' (s)K6* * (I - F \ (5)) - f , (s)l
i.e. T , ' ( ~ ) K ; ~ * - ' F $ ( ~ ) T ,is
( s )stable, and from the choice of J ( s ) we have

s-m
'
Iim T,- (s)K(,*- (s)K6* -'Fi(s)T, (s) = lim T;
I-m
' @If,- ' (s)KA*- 'F'z@If, (SIT,(s)

= lim F $ ( s ) K ~ * - '
s -10

which means that T; I (s)Kh*-'Fi(s)T, (s) is proper. Hence

'
T; (s)K6*- 'F$(s)T,(s)J (s) (T; ' (s) Ab (s) + Ah (3))

is stable and strictly proper. Therefore there exists c E R f such that

1
>C
11 T,'(s)Ko'*-'Fi(s)T,(s)f,-'(S)(T;'(s) Ai(s) + Ah(s))II$ = J W
for voE [ - m , + . o ] .
Using the small-gain theorem and setting p * = c, we have that for V p E [O, p*) the closed-loop
plant is stable, and the bound for el([) can be easily obtained from (91).
242 G . TAO AND P. A . IOANNOU

MLI matrix case

Lemma 1.1
If y ( t ) = g - ' ( s ) H ( s ) x ( t ) ,where H(s) is a polynomial matrix, g(s) is a polynomial and
g-'(s)H(s) is proper and stably invertible, and x ( t ) is bounded for any finite time and
(1 x ( t ) ( ( 6 CI (1 x ( 7 ) ( 1 for some C I E R', then

for some C ~ RE + .

Proof. Let g ( s ) = fi(s)fi(s) such that f~(s)H-'(s)is proper and define x ~ ( t=) f i - ' ( S ) x ( t ) ;
then y ( t ) = f i ' ( s ) H ( s ) x l( t ) and we have that 11 x l ( t ) l l < c3 supl 2 11 y(7)11 for some c3 > 0
since f; * (s)H(s) has a stable inverse by assumption. The remainder of the proof is similar to
that for the SISO case27928 and is omitted.
From (39) and (42) it follows that
'
y f = fr- (s)EP (s)Wm(s)r + p (I + A- (sW? - F i (s)) A', (s) '
+fi-'(s)K?-'(I- F:(s)) Ak(s)l [ I - kTI'(~)fi-'(s)K~*-'F:(s)(A',(s)
+ TI(s) A k (s))l-' (TT ' (s1.K '47 (s) W (s)r + T I ' (s )A- ' (s)K:* - 'A- ' (s)EP (s)w ( t1)
+ w- ( s ) K ~ IF: ( s ) + I )A- ( ~ ) f i(S)W
- " (t ) (93)
where
F: (s) = C/(s)NI (s) ' F:(s) = D/(s)N['(s)

with Nl(s), C/(s), Dc defined in Section 4.2.

(93) is stable if
(I - k T l ' ( s ) K ~ - ' F : ( s ) T / ( s ) ~ - ' ( s ) ( T i ' ( s )A/,@) + A ~ ( s ) ) ) - '
is stable.
Following a similar analysis to the case of the MRI matrix, we can show that
TI'(s)K~-'F:(s)Ti(s)fi-'(s)(Ti '(s) A',@) + AkI(s))
is stable and strictly proper, and choosing

p* = inf
[-m, + W J
1
(11 TI'(s)K6*-'F:(s)Ti(s)f;-'(s)(Ti1(s)Ah@) +Ah(s))lls=/w 1
(93) is stable for all k c [ O , p * ) . Since y f =f;-'(s)ET"(s)y,where f i - ' f ? ( s ) is proper and stably
invertible, using Lemma I . 1 we have the that the closed-loop plant is stable; and a similar bound
for elf(t) = yf(t) - y m f ( t )where
, ymf = f;'(s)4i"(s)ym,can be easily obtained from (93), and the
bound for el ( t )= y ( t ) - y m ( t )follows from the fact that elf(t) = fi-'E,!"(s)el( t )and Lemma I. 1.

APPENDIX 11: PROOF OF LEMMA 5.1

We will give a unified proof for Lemma 5.1 based on both the MRI matrix and the MLI matrix,
so we omit the subscripts and superscripts 'r'and 'I' of the notation in the following analysis.
From the model matching condition we have
f-'(s)K;-'(I- FI(s)- Fz(s)T(s))=T(s) (94)
 MODEL REFERENCE ADAPTIVE CONTROL 243

Hence it follows that

f - (s)T- ' ( s ) - f- ' (s)FI (s)T- ' ( s ) - f- (s)F2(s) = K$
f- ' (s)T-' (s)Y - f - ' (s)F I (s)T- ' (s)Y - f - ' (s)F2 (s)? = KtY (95)
Since
= T (s)U + yT ( s ) An,(s)U + pAa (s)U + W
where A,,, = AL(s)(resp. A',@)) and A, = AL(s) (resp. A:(S)) when the MRI (resp. MLI) is
used, (95) can be written as
f I (s)F I (s)U + f ' ( s ) F ~(s)? + K,"ym= f - (s)U - Ko*@ - ym) + P H I(s)U + H2 (s)W (96)
where
Hl(s) = f-'(s)(I- Fl(s))(Ams) + T-'(s)Aa(s))
H ~ ( s =) f - ' ( s ) ( I - FI(s))T-'(.s) (97)
with H I(s) being stable and strictly proper, H ~ ( s )being stable and proper.
Letting

the proof is complete.

APPENDIX 111: PROOF OF THEOREM 5.1 AND COROLLARY 5.1

For the proof of Theorem 5.1 and Corollary 5.1, we need the following lemmas whose proofs
follow directly from those in Reference 12 for SISO plants and are omitted.

Lemma III.1
Let Y = W (s)u, where W (s) is a strictly proper transfer matrix, the poles of W (s - q ) are
located in Re[s] < 0 for some q > 0, and u ( t ) is continuous, and let m ( t ) be a scalar function
of time which satisfies the equation

r i l ( f )= -60m(t) + 6l(ll Vo(t)ll + 1) m(0) > 61/60 (99)

where 61 > 0, h0 + 62 < q for some 62 > 0, V o ( t ) is continuous and 60,61 > 0; define

Y / mE L,
for any bounded initial conditions.
244 G.TAO AND P. A . IOANNOU

Lemma 111.2
Assume that the estimated parameters 0 ( t ) , $ o ( t ) are uniformly bounded. Then
d'yldt'
u -_
_ _ _ _d'{ldt'
_ _ _ _d'elldt'
_ _ _ _d'[/dt' in
, - € L,
m ' in ' rn m m in
for i = O , l ; j = O , l , ...,n * ; n " is the degree of f(s).

Proof of Theorem 5 . I
Since
+ tr[$(t)$T(f)r-'l
V ( 1 )= tr[@(t)KIKT@T(t)]
is positive definite from the positive definiteness of r and KK,, (assumption (A5), (A5')), from
condition (i) we have that
p<0 for I/ +(t)ll + I/ $(t)ll 2 No
Therefore V ( t ) ,@ ( t ) $, ( t ) E L , for any given bounded initial conditions @(O), $(O).
Using (46) and (50), we have that

From Lemma 111.1 we have that (I/ q 11/m)' is bounded. Therefore

we have
K,,@'c+$[=STZ 40 = KP 'STf(s)Z
Let T(s) have a left matrix faction representation A-l(s)B(s), where B - ' ( s ) is stable by
assumption (A7); it can be shown from (25), (94) that
f(s)Y = f + qTf(s)Z + f(s)(pflll+ d,,,)
B (s)f(s)ii= A (s)T + A ( s ) S ~ ~ ( S+ ) Z +
&ll dlll (1 06)
where
f, = (K;-'(I +
- F I ( s ) ) A ~ , ( s ) ~ - ~ ( . (f-'(s)K;-'Fz(~)
s) + I)A~(s))U
 M O D E L R E F E R E N C E ADAPTIVE C O N T R O L 245

where f = &"'(s)W,,(s)r (resp. f = f(.s)W,ll(.s)r),W = W F (resp. W = w ) if the MLI (resp. MRI)

matrix is used.
From (103) we have

Next we will show that there exists p * > 0 such that m ( t ) is bounded for p E [0, p * ) by
contradiction. I f m ( t ) grows unbounded with time, then for any given E E (0, 1 ) there exists a
t , ( c ) 2 0 such that m ( t ) 3 l / , c , vt E [ t , , t, + T,] where T, = ( 1 / 2 6 0 ) l n ( l / ~ ) .
Using condition (ii) and (108), we have

,< (2kzji + k l p 2 + d k + 2kzkg,~)T+k3 TE [ t s ,t, + T,]

for some k3 > 0.
(109) can be rewritten as

with T E [ f S , f ,+ T r l , where

Following the same steps as in Appendix C of Reference 12 and using (110) and condition
(iii), we can show that

with T E[t,, t, + T r ] ,where

k,, C I , c3, c4 E R+ and EO E (0, 1 1 is arbitrary.

Using the stability of f- '(s), B - '(s) and boundedness of f,,,/m, f,,,/rn, stability and
properness of T- '(s)Ko*- 1F2(s)T(s)(Appendix I) and (106), we have that for t E [t,, I , + T,]

11 Y /I < a l h l ( t ) * (11 *T(flf(~)Z(f)ll + b l p / n ( t ) )+ cs

11 u I] < u2h2(t)* (11 * T ( t ) f ( S ) Z ( f ) l+I b2pm(f))+ c6 (1 13)

where U I ,61, a2, b2, C S , c6 E R + , h l ( t ) ,h 2 ( f ) are some positive exponentially stable impulse
functions, and * denotes the convolution operator.
From (49) and ( 1 1 3 ) we have

for some u3, u4, b3, b4 > 0.

246 G. TAO AND P. A. IOANNOU

Let

then from (114), (115) if follows that m ( t ) < m(t),r 2 t,.

Using (1 15), we have

Owing to ( l l l ) , the fundamental solution of (116) is

for some k4 > 0.

G,
If a4 > f s ( p , E , E O ) + b3p, then the homogeneous part of (1 16) is exponentially stable and
E is bounded. For stability we should have

Setting
1
&o = 0" p = 4 x=2 I,*+

7c4

p* = min r*=- p2"a4 (119)

14kzkc
it follows that for G, E 10, p*), p E [ 0, p*)(118) is satisfied and therefore m ( t ) is bounded. This
is a contradiction. Hence m ( t ) ,y ( t ) , ii(t) and all the signals in the system are bounded.
Using (106), we get
el =f + pfm + d m
- *(s)STf(s)Z ( 120)
Since m ( t )E L,, it can be shown by using (103), (109) and (110) that

for some k s ,k6 > 0, to 2 0 and an arbitrarily small number & > 0. From (120), (121), the
stability of f- '(s) and Lemma 1.1 it follows that

for some y > 0 and exponentially decaying term E ~ Letting

. T-t 00, Theorem 5.1 follows. B

Proof of Corollary 5.1

Since p = ji = (1 w(t)ll = k, = 0, from (109) we have
 MODEL REFERENCE ADAPTIVE CONTROL 247

Using the same steps as in Appendix C of Reference 12, we can show that

dt m
is bounded and therefore

is uniformly continuous. From condition (iii) we have that 11 G I 11 + 11 G2 I/ is also uniformly

continuous, hence

which implies that

From (120) and Lemma 1.1 it follows that el([) -P 0 as t + a.

REIERENCES

I . Morse, A. S., ‘Parametrizations for multivariable adaptive control’, h o e . IEEE CDC, San Diego, California,
1981.
2. Singh, R . P . and K . S. Narendra, ‘A globally stable adaptive controller for multivariable systems’ Report N0.8111,
Electrical Engineering, Yale University, 1982.
3. Wolovich, W. A. and P. L. Falb, ‘Invariants and canonical forms under dynamic compensation’, S I A M J . Control
Optirn., 14 (6), (1976).
4. Elliott, H . and W. A. Wolovich, ‘A parameter adaptive control structure for linear multivariable systems’, IEEE
Trans. Automutic Control, AC-27, (1982).
5. Elliott, H . and W. A . Wolovich, ‘Parametrization issues in rnultivariable adaptive control’, Auton?atica, 20 ( 5 ) .
(1984).
6. Singh, R. P. and K . S. Narendra, ‘Priori information in the design of multivariable adaptive controllers’, IEEE
Trans. Autoinutic Control, AC-29, (1984).
7 . Peterson, B. B. and K . S. Narendra, ‘Bounded error adaptive control’, IEEE Truns. Auroinatic Control, AC-27
(6), (1982).
8. Kreisselmeier, G. and B. D. 0. Anderson, ‘Robust model reference adaptive control’, IEEE Trans. Automatic
Control, AC-31 (2), (1986).
9. loannou, P. A. and P. V. Kokotovic, Aduprive Systems with Reduced Models, Springer-Verlag. Berlin, 1983.
10. Narendra, K . S. and A. M. Annaswamy, ‘A new adaptive law for robust adaptation without persistent excitation’,
Report No. 8510, Electrical Engineering, Yale University, 1985.
I I . loannou, P. A., ‘Robust adaptive controller with zero residual tracking errors’, IEEE Trans. Autoinuric Control,
AC-31 (S), (1986).
12. loannou, P. A. and K. Tsakalis, ‘A robust direct adaptive controller’, IEEE Truns. Automatic Control, AC-31
( I I ) , 1133-1143 (1986).
13. loannou, P. and J . Sun, ‘Theory and design of robust direct and indirect adaptive control schemes’, I n / . J .
Control, 47 (3). (1986).
14. loannou, P. A. and G . Tao, ‘Dominant richness and improvement of performance of robust adaptive control’,
Report 04-01-87, EE-Systems, USC, 1987.
15. Tsiligiannis, C. A. and S. A. Svoronos, ‘Multivariable self-tuning control via right interactor matrix’, IEEE Trans.
Autornaiic Control, AC-31, (1986).
16. loannou, P . A . and K . Tsakalis, ‘The class of unmodeled dynamics in robust adaptive control’, Proc. ACC,
Atlanta, Georgia, June 1988.
17. Praly, L., ‘Robust model reference adaptive controllers, part 1: stability analysis’, Proc. 23rd IEEE CDC, 1984.
18. Hsu, L and R. R. Costa ‘Bursting phenomena in continuous-time adaptive systems with a a-modification’, IEEE
Trans. Automatic Control, AC32(1), (1987).
248 G. TAO AND P. A. IOANNOU

19. Praly, L. and M . Espana, ‘An example of oscillations in adaptive linear control’, Proc. 3rd Yale Workshop on
Applications of Adaptive Systems Theory, May 1987.
20. Anderson, B. D. O., (1985). ‘Adaptive systems, lack of persistency of excitation and bursting phenomena’,
Autornafica, 21 (3), (1985).
21. Mareels, I . M. Y. and R. R . Bitmead, ‘Nonlinear dynamics in adaptive control: chaotic and periodic stabilization’,
Technical Report, Australian National University, August 1985.
22. loannou, P. A. and P. V. Kokotovic, ‘An asymptotic crror analysis of identifiers and adaptive observers in the
presence of parasitics’, IEEE Trans. Auto~naticControl, AC-27 (4), (1982).
2 3 . Boyd, S. and S. Sastry, ‘On parameter convergence in adaptive control’, Syst. Control Lett., 3 (6), (1983).
24. Narendra, K. S. and A. M . Annaswaniy, ‘Robust adaptive control in the presence of bounded disturbances’, IEEE
Trans. Auronintir Control, AC-31 (4), ( I 986).
25. Yuan, J . Y. and W. M. Wonharn, ‘Probing signals for model reference identification’, IEEE Trans. Autoniatic
Control, AC-22 (4),(1977).
26. Anderson, B. D. 0.. ‘Exponential stability of linear equations arising in adaptive identification’, IEEE Trans.
Automatic Conlrol, AC-22, (1977).
27. Narendra, K . S., Y. H. Lin and L. S. Valavani, ‘Stable adaptive controller design part 11: proof of stability’, IEEE
Trans. Autonlalic Control, AC-25, (1980).
28. Sastry, S. S . , ‘Model reference adaptive control: rtability, parameter convergence and robustness’, IMA J . Control
Info. 1, (1984).