Systems & Control Letters 42 (2001) 379393

www.elsevier.com/locate/sysconle
Adaptive l
1
robust control for SISO systems

V.F. Sokolov
a, b,
a
Department of Mathematics, Syktyvkar State University, 55, Oktyabrsky prospekt, Syktyvkar, 167001, Russia
b
Department of Mathematics, Ural Division, RAS, Russia
Received 12 September 1996; received in revised form 12 September 2000; accepted 26 October 2000
Abstract
This paper addresses the problem of nonconservative adaptive robust regulating a SISO system, allowing for model
perturbations in the coprime factor representation of the system as well as external additive bounded disturbance signals.
The perturbations and disturbances are assumed to belong to a particular parameterized class and are norm bounded. The
norm bounds are not available for control design, but the class knowledge is available. The performance measure is the
worst-possible steady-state absolute value of the controlled output, over the class of allowable disturbances and model
perturbations. An identication-based adaptive control method is proposed where the identication criterion coincides with
the control objective. The proposed adaptive scheme achieves near-optimal performance. Simulations comparing the new
scheme with an adaptive controller based on classical recursive least-squares identication are presented. c 2001 Elsevier
Science B.V. All rights reserved.
Keywords: Adaptive control; Robust control; l
1
control
1. Introduction
1.1. System description
We consider the problem of nonconservative adaptive regulating, a SISO system, subjected to both norm
bounded perturbations and external bounded disturbance. The system is represented by
a(q
1
),(t) =q
d
b(q
1
)u(t) + t(t, ,, u), t N:={0, 1, . . .}, (1)
where ,(t), u(t) and t(t, ,, u) denote, respectively, the output, control, and total disturbance, q
1
is the back-
ward shift operator, d is the delay in control,
a(q
1
) =1 + a
1
q
1
+ + a
n
q
n
, b(q
1
) =b
0
+ b
1
q
1
+ + b
m
q
m
.
The total disturbance t is the sum of disturbances of dierent nature,
t(t, ,, u) =w
a
(t) + w
,
(t, ,) + w
u
(t, u),

This work was supported by the Russian Fund of Basic Research under grant 96-01-00459.

Correspondence address: Department of Mathematics, Syktyvkar State University, 55, Oktyabrsky prospekt, Syktyvkar, 167001, Russia.
E-mail address: [email protected] (V.F. Sokolov).
0167-6911/01/$ - see front matter c 2001 Elsevier Science B.V. All rights reserved.
PII: S0167- 6911( 00) 00110- 9
380 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
where
w
a
=o
w
w, w

:= sup
t
|w(t)| 61, w
,
=o
,
A
1
,, w
u
=o
u
A
2
u. (2)
We will call w
a
an external additive disturbance, w
,
and w
u
perturbations in output and control, respectively.
The operator A=diag{A
1
, A
2
} will be called the operator of perturbations and we will consider two dierent
classes of operators A. The rst one is the class D
j
B
(2) of bounded memory perturbations and in this case,
|A
1
,(t)| 6m
,
(t), m
,
(t) := max
0k 6j
|,(t k)|,
|A
2
u(t)| 6m
u
(t), m
u
(t) := max
0k 6j
|u(t k)|, t N.
(3)
The class D
j
B
(2) is characterized by the parameter j 1, the memory of perturbations. Without loss of
generality, we assume j max{n, d + m} (see Section 6 for details). The second class is the class D
z
ED
(2),
z =(z
,
, z
u
), 0 6z
,
, z
u
1, of perturbations with exponentially decreasing impulse responses and in this case,
|A
1
,(t)| 6m
,
(t), m
,
(t) :=z
,
m
,
(t 1) + (1 z
,
)|,(t 1)|, m
,
(0) :=0,
|A
2
u(t)| 6m
u
(t), m
u
(t) :=z
u
m
u
(t 1) + (1 z
u
)|u(t 1)|, m
u
(0) :=0, t N.
(4)
Here the multipliers 1 z
,
and 1 z
u
are introduced to ensure the inequalities A
i
61, i =1, 2, for the
induced norms A
i
of the operators A
i
: l

. The class D
z
ED
(2) is characterized by the set z of decay
rates (see [14] for details).
Since A
i
61, i =1, 2, for both classes of perturbations, the (nonnegative) numbers o
,
and o
u
are the
norms of perturbations in output and control, respectively. In terms of the robust control theory, independent
perturbations in output and control represent a structured uncertainty in the system (see [5] for details). The
nonnegative number o
w
is the

norm of the external additive disturbance w

a
.
Introduce the notations
0
c
:=(a
1
, . . . , a
n
, b
0
, . . . , b
m
)
T
, 0 :=(0
T
c
, o
,
, o
u
, o
w
)
T
R
n
0
, n
0
:=n + m + 4.
The vector 0
c
will be called the coecient vector and 0 the parameter vector of the system (1).
Assumption 1. The set of structural parameters o :=(n, m, d) and the class of perturbations, D
j
B
(2) or D
z
ED
(2),
are assumed to be known. The parameter vector 0 is assumed to be unknown but a compact set O
c
R
n+m+1
of allowable values of the coecient vector 0
c
is known: 0
c
O
c
. It is assumed also that polynomials a(, 0
c
)
and b(, 0
c
) associated with the coecient vector 0
c
O
c
do not have common zeros in the closed unit disk
of the complex plane for all 0
c
O
c
.
A priori information on system (1) can be summarized in the set
I
pr
:={D, o, O
c
}
with D denoting the class D
j
B
(2) or D
z
ED
(2).
An additional a priori information on allowable values of the norms of perturbations o
,
and o
u
will be
claried later on and will be dierent in two adaptive control problems discussed below.
1.2. Controller, robust stability and robust performance measure
It will be assumed that for any 0
c
O
c
, a (stabilizing) controller
:(q
1
, 0
c
)u(t) =[(q
1
, 0
c
),(t), :(0, 0
c
) =1 (5)
is chosen such that the characteristic polynomial of the closed-loop system (1) and (5),
,(:, 0
c
) :=a(:, 0
c
):(:, 0
c
) :
d
b(:, 0
c
)[(:, 0
c
),
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 381
is stable (i.e., all zeros of , lie outside of the closed unit disk of the complex plane). The existence of a
stabilizing controller is guaranteed by the condition on polynomials a and b in Assumption 1.
Robust performance of the closed-loop system (1) and (5) will be characterized by the worst-case steady-state
robust performance measure
I (0) := sup
AD
sup
w

61
limsup
t
|,(t)|. (6)
Exact computing of the performance index I (0) is an open and dicult mathematical problem and we will
use an upper bound for I (0) presented in [14]. The stable transfer functions connecting the disturbance t with
the output , and control u are of the form
G
,t
(:, 0
c
) :=
:(:, 0
c
)
,(:, 0
c
)
, G
ut
(:, 0
c
) :=
:
d
[(:, 0
c
)
,(:, 0
c
)
.
Let G(:)
1
denote the l
1
norm of the impulse response of a stable transfer function G(:). The closed-loop
system (1) and (5) is robustly stable in any one of the classes D
j
B
(2) and D
z
ED
(2), if
o
,
G
,t
(, 0
c
)
1
+ o
u
G
ut
(, 0
c
)
1
1 (7)
and for robustly stable system
I (0) 6J(0) :=
G
,t
(, 0
c
)
1
1 o
,
G
,t
(, 0
c
)
1
o
u
G
ut
(, 0
c
)
1
o
w
. (8)
The robust stability condition (7) and the upper bound J(0) for the worst-case steady-state robust performance
measure I (0) are nonconservative in the case of bounded memory perturbations (see [14] for details).
1.3. Problem formulation
We consider two dierent problems of adaptive regulation of system (1).
The problem of adaptive robust stabilization is in the synthesis of adaptive controller which ensures the
boundedness of all signals in the closed-loop system.
For any 0
c
O
c
, dene the set
S(0
c
) :={(o
1
, o
2
)
T
| o
1
G
,t
(, 0
c
)
1
+ o
2
G
ut
(, 0
c
)
1
1, o
1
0, o
2
0}
which will be called the robust stability region associated with the coecient vector 0
c
. The set
S
o
(O
c
) :=

0
c
O
c
S(0
c
) (9)
will be called the maximal 0
c
independent robust stability region.
In the problem of adaptive robust stabilization, our main purpose will be to make the set of pairs (o
,
, o
u
)
T
that can be tolerated under adaptive control to be arbitrarily close to the set S
o
(O
c
) possible.
The problem of nonconservative adaptive robust control is in the synthesis of adaptive controller which
not only stabilizes system (1) with the unknown parameter vector 0, but also ensures the same upper bound
J(0) for the output of the closed-loop adaptive system.
In fact, a solution to a slightly weaker problem will be proposed below and adaptive controller will be able
to provide the upper bound J(0) with a prespecied accuracy.
In the problem of nonconservative adaptive control, we will assume that the set of allowable vectors
(0
T
c
, o
,
, o
u
)
T
will be nearer to the set
S
NA
(O
c
) :={(0
T
c
, o
,
, o
u
)
T
| 0
c
O
c
, (o
,
, o
u
)
T
S(0
c
)}. (10)
For a given prior set O
c
, the set S
NA
(O
c
) is the maximal set of vectors (0
T
c
, o
,
, o
u
)
T
which are allowable by the
robust stability condition (7) for systems with known parameters and the subscript NA refers to nonadaptive
case. Thus, the nonconservative adaptive robust control must be nonconservative with respect to both robust
performance and robust stability.
382 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
Since any S(0
c
) is a triangle in the (o
,
, o
u
)-plane, the set S
o
(O
c
) is convex as an intersection of convex
sets and is generally a strictly proper subset of the projection of the set S
NA
(O
c
) onto the (o
,
, o
u
)-plane (it
is not dicult to see that this projection is nonconvex).
Solution to both of the stated problems is based on the certainty equivalence principle according to which
an adaptive controller is of the form (5)
:(q
1
, 0
c
(t))u(t) =[(q
1
, 0
c
(t)),(t), (11)
where an estimate 0
c
(t) substitutes for the unknown vector 0
c
. The problem of synthesis of adaptive control
law is thus reduced to the synthesis of appropriate estimation algorithm.
Problems of stabilization are well studied in classical robust adaptive control where the term robust
refers to the fact that boundedness of all signals in the closed-loop system is guaranteed in the presence of
some perturbations. The norm of perturbations that can be tolerated is usually described by one real parameter
and is determined by the particular estimation algorithm. The problem of performance evaluation is less well
understood in classical robust adaptive control and the most advanced results (see, e.g., [810,15]) provide
fairly conservative upper bounds for robust performance of adaptive systems.
In classical adaptive control, the synthesis of estimation algorithm is usually based on a widely accepted
separation principle according to which it is desirable to consider the tasks of estimation and controller
synthesis separately, so that each can be optimally tuned to enhance the overall performance [3]. In this
paper, we use quite a dierent approach to the synthesis of estimation algorithms for robust adaptive control.
This approach proceeds from the question how to apply the results, obtained in the framework of a certain
robust control theory, to the control of a system when the parameters of the nominal system (i.e., the system
without disturbances and perturbations) and the norms of perturbations and disturbances are assumed to be
unknown. In particular, we are interested in the application of the results of robust control theory in
1
setting.
Basic results of this theory were presented in [46] and extended for the purposes of adaptive control in [14].
In the case of SISO system under coprime factor perturbations, these results take the form of the robust
stability condition (7) and of the upper bound (8) for the worst-case steady-state robust performance measure
(6).
The assumptions on perturbations (2)(4) come to the concept of a model which is compatible with (or
equivalently, unfalsied by) data and a priori information on a certain time interval. Finding a model which
is asymptotically compatible with data and a priori information on innite time interval, provides the solution
to the problem of adaptive robust stabilization. This idea is used in Section 3 for the synthesis of a projection
estimation algorithm that ensures the convergence of estimates in a nite time and approximate compatibility
of the limit estimate on innite time interval. A similar idea was used in [18] for robust adaptive stabilization
of SISO systems with unstructured uncertainty of the form
|t(t, ,, u)| 6

o
w
+

o

max
0k 6j
max{|,(t k)|, |u(t k)|}, t N (12)
or of the form
|t(t, ,, u)| 6

o
w
+

o

(t), (t) :=z(t 1) + max{|,(t 1)|, |u(t 1)|}, 0 6z1, (13)

where the memory of perturbations j (or the decay rate z), the upper bound

o
w
for the norm of additive
disturbance and the upper bound

o

for the norm of unstructured uncertainty were assumed to be known. In

adaptive control literature, the description of perturbations in the form of unstructured uncertainty which is
characterized by one real parameter, the upper bound

o

for the norm of perturbations, is the most widespread.

In the present paper, we use less conservative description of perturbations in the form of structured uncertainty
parameterized by the norms o
,
and o
u
. The main contribution of Section 3 of the present paper compared
with the results in [18] is the rejection of the assumption on the prior knowledge of some upper bound

o
w
for the norm o
w
of the additive disturbance and in considering the structured uncertainty.
Compatibility of a model on innite time interval is not sucient for solving the problem of nonconservative
adaptive robust control. The key idea for solving of this problem is in choosing the best model among the
elements of the set of unfalsied models. Here the best model means the model that promises the best
robust performance. This idea was proposed in [11] for the synthesis of nonconservative adaptive control of
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 383
systems with bounded disturbances and used in [13] for the synthesis of adaptive robust control. The notions
of asymptotic compatibility and the best compatible model are similar to the notions of asymptotic eciency
and preferential identication introduced and discussed in [7] in the framework of L
2, o
-stability.
2. Estimation problem
Denition. Given data ,
t
0
:=(,(0), . . . , ,(t)), u
t1
0
:=(u(0), . . . , u(t 1)) for some t N, a vector

0 =
(

0
T
c
,

o
,
,

o
u
,

o
w
)
T
will be said to be compatible with the data ,
t
0
, u
t1
0
and a priori information I
pr
, if

0
c
O
c
and there exist perturbations w
,
and w
u
and a disturbance w
a
satisfying restrictions (2) on the interval [0, t]
such that the data ,
t
0
, u
t1
0
satisfy Eq. (1) associated with the coecient vector

0
c
.
Lemma 1. For any data ,
t
0
, u
t1
0
, any perturbation class D, any vector

0
c
O
c
, and any nonnegative

o
,
and

o
u
, the vector

0 =(

0
T
c
,

o
,
,

o
u
,

o
w
)
T
is compatible with the data ,
t
0
, u
t1
0
and a priori information I
pr
for all
suciently large

o
w
.
Proof. Indeed, for any data ,
t
0
, u
t1
0
, any perturbation class D, any vector

0
c
O
c
and any allowable pertur-
bations w
,
(0, ,), . . . , w
,
(t, ,) and w
u
(0, u), . . . , w
u
(t, u), we dene an additive disturbance w
a
as follows:
w
a
(s) = a(q
1
),(s) q
d

b(q
1
)u(s) w
,
(s, ,) w
u
(s, u), s {0, 1, . . . , t},
where a and

b are the polynomials associated with the vector

0
c
. Then the vector

0 is compatible with the
data ,
t
0
, u
t1
0
and a priori information I
pr
if

o
w
sup
s 6t
| w
a
(s)|.
Lemma 1 is almost trivial but very important for understanding the stated adaptive problems. Because of
the possibility to attribute any model mismatch, on any nite time interval, to an appropriate external additive
disturbance, neither the unknown coecient vector 0
c
of the true system (1) nor the true norms o
,
and o
u
of perturbations can be identied under a priori information available. Moreover, arbitrary

0
c
,

o
,
, and

o
u
can
be considered as the part of a parameter vector

0 which is compatible with the data and a priori information.
The situation remains the same if additional restrictions on the norms of perturbations are imposed in the
form of allowable sets (9) or (10).
The following lemma gives a criterion of the compatibility of estimates for the classes of perturbations
D
j
B
(2) and D
z
ED
(2).
Lemma 2. A vector

0 =(

0
T
c
,

o
,
,

o
u
,

o
w
)
T
is compatible with the data ,
t
0
, u
t1
0
and a priori information I
pr
if
and only if

0
c

O
c
and
| a(q
1
),(s) q
d

b(q
1
)u(s)| 6

o
,
m
,
(s) +

o
u
m
u
(s) +

o
w
for s 6t, (14)
where a and

b are the polynomials associated with the vector

0
c
and the sequences m
,
and m
u
are dened
in (3) and (4).
Proof. Lemma 2 follows from Eq. (1), restrictions (2), and the description of the classes D
j
B
(2) and D
z
ED
(2).
Dene the unfalsied parameter set O(t) as the set of all parameter vectors that are compatible with
the data ,
t
0
, u
t1
0
and a priori information I
pr
. The set O(t) presents complete information on the unknown
parameter vector of system (1) at the time instant t: 0 O(t). In view of Lemma 2, the set O(t) is determined
by the inclusion

0
c
O
c
and by the family of 2(t +1) linear, with respect to the components of

0 inequalities
generated by inequalities (14). Since the number of inequalities grows innitely as t , a complete
information on the unknown parameter vector 0 cannot be stored for large t.
384 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
The discussion presented in this section claries the hardness of the problem of nonconservative adaptive
robust control where the desired bound for robust performance of adaptive control depends on the true
nonidentiable parameter vector 0.
3. Adaptive robust stabilization
In this section, we show that a simple projection estimation algorithm provides a solution to the problem of
adaptive robust stabilization such that a priori set of allowable (o
,
, o
,
)
T
pairs that are tolerable under adaptive
control can be as close to the maximal 0
c
independent robust stability region S
o
(O
c
) as desired.
The synthesis of stabilizing adaptive control will be based on the method of recursive goal inequalities
proposed by Yakubovich in 1960s [1,2]. According to this method, estimates 0(t) should be designed in such
a way as to ensure, rst, their convergence in a nite time and, second, to satisfy, approximately, inequalities
(14) for all suciently large t. We will use the following assumptions.
Assumption 2. For any allowable parameter vector 0 :=(0
T
c
, o
,
, o
u
, o
w
)
T
of system (1),
0
o
:=(o
,
, o
u
)
T

S
o
(O
c
), 01, (15)
where is a known constant and

S denotes the closure of the set S.
Assumption 3. Constants

G
,
and

G
u
are assumed to be known such that
G
,
:= sup
0
c
O
c
G
,t
(, 0
c
)
1
6

G
,
, G
u
:= sup
0
c
O
c
G
ut
(, 0
c
)
1
6

G
u
.
Assumption 4. The prior set O
c
is convex.
Now we describe the projection estimation algorithm. Let c0 be a positive design parameter of estimation
algorithm. Let the initial estimate of the unknown vector 0 be of the form 0(0) =(0
c
(0)
T
, 0, 0, 0)
T
with arbitrary
0
c
(0) O
c
and let 0(t 1) be the estimate of the vector 0 at the time instant t 1. Using the standard regression
vector
(t 1) :=(,(t 1), . . . , ,(t n), u(t d), . . . , u(t d m))
T
,
dene an extended regression vector
[(t 1) :=(p(t) (t 1)
T
, m
,
(t), m
u
(t), 1)
T
, (16)
where
p(t) :=sign(,(t) 0
c
(t 1)
T
(t 1)) (17)
and the sequences m
,
and m
u
are dened in (3) and (4). Dene a preliminary estimate

0(t) :=
_

_
0(t 1) if 0(t 1)
T
[(t 1) p(t),(t) c|[(t 1)|,
0(t 1) +
,(t)p(t) 0(t 1)
T
[(t 1)
|[(t 1)|
2
[(t 1) otherwise,
(18)
where |[(t 1)| denotes the Euclidean norm of the vector [(t 1). The estimate 0(t) is dened now by
projecting

0
c
(t) onto the set O
c
:
0
c
(t) :=P
O
c

0
c
(t)
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 385
(P
O
c
denotes the operator of orthogonal projecting onto O
c
), and by subsequent projecting

0
o
(t) :=(

o
,
(t),

o
u
(t))
T
onto the set

S(0
c
(t))
0

o
(t) :=

0
o
(t) + (t)


0
o
(t)
T
G(0
c
(t))
|G(0
c
(t))|
2
G(0
c
(t)),
0
o
(t) :=
_

_
0

o
(t) if 0

o
(t) 0,
_

G
,t
(0
c
(t))
1
, 0
_
T
if o

u
(t)0,
_
0,

G
ut
(0
c
(t))
1
_
T
if o

,
(t)0,
(19)
where the inequality 0

o
(t) 0 is taken componentwise and
G(0
c
) :=(G
,t
(, 0
c
)
1
, G
ut
(, 0
c
)
1
)
T
,
(t) :=
_
0 if

0
o
(t)
T
G(0
c
(t)) 6,
1 otherwise.
Finally, o
w
(t) :=

o
w
(t).
Geometric interpretation: A geometric interpretation of the projection estimation algorithm is simple. The
inequality
0(t 1)
T
[(t 1) p(t),(t)
is equivalent to the inequality
|a(q
1
, 0(t 1)),(t) q
d
b(q
1
, 0(t 1))u(t)| 6o
,
(t 1)m
,
(t) + o
u
(t 1)m
u
(t) + o
w
(t 1)
which is inequality (14) with respect to 0(t 1) for s =t. The preliminary vector

0(t) is the projection of
0(t 1) onto the half-space
O(t) :={t|t
T
[(t 1) ,(t)p(t)}. (20)
Subsequent projecting was explained in the formulation of the algorithm. Introducing the term c|[(t 1)|
in the rst line of (18) implies that the estimate 0(t 1) is not updated, if the distance of 0(t 1) to the
half-space O(t) is not greater than c. Such a dead zone provides the convergence of the sequence {0(t)}
to its limit in a nite time and simplies the analysis of closed-loop system. The properties of the algorithm
are presented in Lemma 3.
Lemma 3. Let Assumptions 1, 2, and 4 hold. Let the estimation algorithm (18) and (19) be applied to the
system (1) with some control sequence u. Then,
(1) there exists a time instant t

N such that
0(t) =0

=(0
T
c,
, o
,,
, o
u,
, o
w,
)
T
:= lim
t
0(t) for all t t

,
0
c,
O
c
, (o
,,
, o
u,
)

S(0
c,
);
(2) the number of corrections of the estimates 0(t) is not greater than |0 0(0)|
2
}c
2
;
(3) t t

|,(t) 0
T
c,
(t 1)| 6(o
,,
+ C
1
c)m
,
(t) + (o
u,
+ C
2
c)m
u
(t) + o
w,
+ c, (21)
where C
1
=

n + 1, C
2
=

m + 2 in the case D=D

j
B
(2) and C
1
=
1
z
n1
,
(1z
,
)
+ 1, C
2
=
1
z
m+d1
u
(1z
u
)
+ 1 in
the case D=D
z
ED
(2).
386 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
Proof. It follows from (18) that
|

0(t) 0|
2
6|0(t 1) 0|
2
c
2
,
in the case of updating the estimate 0(t 1). In view of the convexity of the sets O
c
(Assumption 4) and

S(0
c
(t)), and in view of the inclusions 0
c
O
c
(Assumption 1) and 0
o

S(0
c
(t)) (Assumption 2) we have
|0
c
(t) 0
c
| 6|

0
c
(t) 0
c
|, |0
o
(t) 0
o
| 6|

0
o
(t) 0
o
|.
Taking into account the equality o
w
(t) =

o
w
(t) we get the inequality
|0(t) 0|
2
6|0(t 1) 0|
2
c
2
which implies statements (1) and (2) of Lemma 3. Statement (3) follows now from the inequalities 0
T

[(t
1) p(t),(t) c|[(t 1)| for t t

, where |[(t 1)| is bounded in the case D=D

j
B
(2) as follows:
|[(t 1)| 6

_
n

k =1
,(t k)
2
+ m
2
,
(t) +

_
m+d

k =d
u(t k)
2
+ m
2
u
(t) + 1
6

n + 1m
,
(t) +

m + 2m
u
(t) + 1
and in the case D=D
z
ED
(2) as follows:
|[(t 1)| 6
n

k =1
|,(t k)| +
m+d

k =d
|u(t k)| + m
,
(t) + m
u
(t) + 1
6
n

k =1
z
1k
,
|,(t k)| +
m+d

k =d
z
1k
u
|u(t k)| + m
,
(t) + m
u
(t) + 1
6
_
1
z
n1
,
(1 z
,
)
+ 1
_
m
,
(t) +
_
1
z
m+d1
u
(1 z
u
)
+ 1
_
m
u
(t) + 1.
Lemma 3 is proven.
Theorem 1. Let Assumptions 14 hold. Consider the closed-loop system consistent of system (1), the adap-
tive controller (11), and the estimation algorithm (18) and (19). Let the design parameter c of the algorithm
be chosen suciently small such that 0c(1 )}(C
1

G
,
+ C
2

G
u
), where the constants C
1
and C
2
are
dened in Lemma 3 and the constants

G
,
and

G
u
in Assumption 3. Then,
(1) for all suciently large t
0(t) =0

=(0
T
c,
, o
,,
, o
u,
, o
w,
)
T
, o
w,
6o
w
+|0
c
0
c
(0)|;
(2) the control u and output y are bounded;
(3)
limsup
t+
|,(t)| 6
G
,t
(, 0
c,
)
1
1 (o
,,
+ C
1
c)G
,t
(, 0
c,
)
1
(o
u,
+ C
2
c)G
ut
(, 0
c,
)
1
(o
w,
+ c). (22)
Proof. Theorem 1 follows from Assumptions 14 and Lemmas 2 and 3. The proof of Theorem 1 is similar
to the proof of Theorem 2 in [13].
Main merits of the projection estimation algorithm are, rst, its simplicity and, second, the nearness of
the adaptive robust stability region (15) to set (9) for close to 1. It is worth pointing out that no prior
knowledge of set (15) is required for the implementation of the projection estimation algorithm, except for the
known that species a degree of robust stability of closed-loop adaptive system. Note also that the upper
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 387
bounds

G
,
and

G
u
are used for the choice of suciently small design parameter c only, and can therefore be
arbitrarily conservative. However, the second merit of the algorithm is not a merit at all if to take the point
of view that an adaptive controller should guarantee the same robust stability and robust performance as the
nonadaptive controller (5) for system with known parameters are taken into consideration. In this connection
it should be added that the upper bound (22) guaranteed by the projection estimation algorithm is close, for
small c, to the upper bound J(0

) associated with the nal estimate 0

but a nearness of J(0

) to J(0)
cannot be guaranteed.
4. Nonconservative adaptive robust control
The projection estimation algorithm is oriented towards getting convergent estimates 0(t) which are asymp-
totically and approximately compatible with the data and a priori information. However, the compatibility of
estimates is not sucient for the synthesis of nonconservative adaptive robust control. The main idea of the
synthesis is in choosing such an estimate, among the elements of the unfalsied parameter set O(t), which
promises the best robust performance. A motivation for this idea was presented in [11,13]. Since the unfalsied
parameter sets O(t) cannot be employed in closed-loop, we will use set valued estimates E(t) of restricted
complexity to get an approximate solution to the problem of nonconservative adaptive robust control.
4.1. Set estimation algorithm for nonconservative adaptive robust control
Fix a positive c, a design parameter of the set estimation algorithm, and set
E(0) :=S
NA
(O
c
) [0, ), 0(0) :=(0
T
c
, 0, 0, 0)
T
(23)
with arbitrary 0
c
O
c
. Let E(t) be a set estimate and 0(t) be a vector estimate of the unknown parameter
vector 0 at the time instant t. Having measured the output ,(t + 1), set
E(t + 1) :=
_
E(t) if 0(t)
T
[(t) ,(t + 1)p(t + 1) c|[(t)|,
E(t) O(t + 1) otherwise,
(24)
0(t + 1) :=
_
_
_
0(t) if E(t + 1) =E(t),
arg min
t E(t+1)
J(t) otherwise,
(25)
where [(t), p(t), and O(t) are dened by (16), (17), and (20), respectively.
For the sake of simplicity, we assume that the minimum in (25) is achievable. In fact, Theorem 2, the
main result of this paper, can be reformulated for the situation when the minimization in (25) is performed
approximately.
Geometric interpretation: A geometric interpretation of the set estimation algorithm is simple. Every set
E(t) is determined by a subset of the family of 2(t + 1) linear inequalities produced by inequalities (14).
The new inequality

0
T
[(t) ,(t + 1)p(t + 1) is included in the description of the set E(t + 1) only if
the distance of the current estimate 0(t) to the half-space O(t + 1) is greater than c. Note that some of the
inequalities describing E(t) can become redundant after including the new inequality and therefore the number
of inequalities involved in the description of the set estimate E(t + 1) can decrease.
Lemma 4. Let Assumptions 1 and 2 hold. Let the estimation algorithm (23)(25) be applied to system (1)
with some control sequence u. Then,
(1) there exists t

N such that the set estimates E(t) and the vector estimates 0(t) are constant for t t

;
(2) the nal estimate 0

:=lim
t
0(t) satises statement (3) of Lemma 3;
(3) J(0

) 6J(0).
388 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
Proof. The proof of statement (1) is analogous to the proof of Lemma 2 in [13]. Statement (2) follows from
the condition in the rst line in (24) and from the nite convergence of estimates. Since the true parameter
vector 0 is compatible with data and a priori information for all t, we have 0 E(t) for all t. Now statement
(3) follows directly from the second line in (25).
The main merit of the set estimation algorithm is in ensuring the inequality J(0

) 6J(0) for the nal

estimate 0

. Note that the right-hand side of this inequality, J(0), is associated with the true unknown
nonidentiable parameter vector 0. This inequality is crucial in solving the problem of nonconservative adaptive
robust control.
The nite convergence of the set estimates E(t) follows from the fact that updating E(t) implies the
exclusion of at least a ball of radius c}2 from an initial set of the form S
NA
(O
c
)[0, C
w
) (C
w
is a proportional
to o
w
constant). Then the number of possible corrections and, consequently, the number of linear inequalities
involved in the description of the sets E(t), has an evident upper bound of the form C(dim(S
NA
(O
c
))o
w
2}c)
n
0
,
where dim(S
NA
(O
c
)) denotes the diameter of the set S
NA
(O
c
) and C is a constant. This upper bound is too
large from practical point of view and is not included in the formulation of Lemma 4. However, this upper
bound is too conservative by several reasons. First, after each updating of the set E(t), much more region than
the above-mentioned ball is excluded from E(t) by the new half-space O(t + 1) so that a better theoretical
upper bound can be derived (this problem is out of our consideration). Second, the latter upper bound can
be achieved under the worst-case sequence of the half-spaces (20). Since these half-spaces are determined by
the extended regression vectors [(t) and outputs ,(t + 1), they can hardly generate the worst-case sequence
of half-spaces. Unfortunately, it is dicult to express this second reason of conservatism of theoretical upper
bound in strict mathematical terms.
Instead of Assumption 2, we use now more general Assumption 5.
Assumption 5. For any allowable parameter vector 0 :=(0
T
c
, o
,
, o
u
, o
w
)
T
of system (1),
(0
T
c
, o
,
, o
u
)
T

S
NA
(0
c
), 01, (26)
where is a known constant.
Theorem 2. Let Assumptions 1, 3, and 5 hold. Consider the closed-loop system consistent of the system (1),
the adaptive controller (11), and the estimation algorithm (23)(25). Let the design parameter c of the
algorithm be chosen as in the formulation of Theorem 1. Then,
(1) for all suciently large t,
E(t) =E

:= lim
t
E(t) and 0(t) =0

:= lim
t
0(t)

S
NA
(0
c
) [0, +);
(2) the control u and output y in the closed-loop system (1), (2), (11) and (23)(25) are bounded;
(3) the output y satises the inequality (22), and J(0

) 6J(0).
Proof. Theorem 2 follows from Assumptions 1, 3 and 5, Lemmas 2 and 4.
5. Implementation
Comment on the choice of the design parameter c of the set estimation algorithm: In contrast to the
projection estimation algorithm, introducing the positive design parameter c into the set estimation algorithm
is crucial for getting set estimates E(t) of bounded complexity. The smaller is c, the closer is the upper bound
(22) to the desired upper bound (8) and the closer is the adaptive prior set (26) to the nonadaptive prior
set (10). On the other hand, the smaller is c, the larger is the number of possible corrections the set estimates
E(t) and, consequently, the more is computing power required for the implementation of the set estimation
algorithm. In practice, a compromise in the choice of the parameter c can be found by comparing the achieved
upper bound (22) with the value J(0

) associated with c =0.

Comment on computing the set estimates E(t): Under additional assumption that the prior set S
NA
(O
c
) is
approximated by a polytope, the set estimates E(t) become polyhedra and their updating requires computing
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 389
the intersection of the polyhedron E(t) and the half-space O(t +1). Eective algorithms and software packages
to handle polyhedra are now available [16,17].
Comment on minimizing the control criterion J(t): Computing the vector estimate 0(t) requires the mini-
mization of the control criterion J(0) over the polyhedron E(t). Any appropriate minimization algorithm can
be used for this purpose. Note that only several steps of chosen minimization algorithm can be made at a
current time instant, depending on the computer power available. Further steps, if necessary, can be continued
at subsequent time instants without any eect on the steady-state robust performance of the adaptive system.
Minimization over the set of vertices of the polyhedron E(t) seems to be an acceptable simplication and
works well in simulations.
Comment on minimizing the 0
c
independent upper bound for control criterion: Considerably less complexity
of the set estimation algorithm can be achieved at the cost of weakening the control objective. For example,
consider the following 0
c
independent upper bound for the control criterion J(0):

J(0) :=

G
,
1

G
,
o
,


G
u
o
u
o
w
J(0),
where

G
,
and

G
u
are dened in Assumption 3. In this case, the cone set estimation algorithm using only n
0
linear inequalities at any time instant is applicable (see [12] for details).
Comment on possible heuristic simplications: A direct application of the set estimation algorithm (23)
(25) is possible for the systems with moderate orders n and m or for problems with prior sets O
c
of moderate
size. Applications to more general cases are possible on heuristic grounds with the use of limited complexity
algorithms. The following heuristic idea of a set estimation algorithm of restricted complexity seems to be
promising. Instead of storing all the linear inequalities describing the updated set E(t + 1), one can store a
restricted family of inequalities. Various heuristic ideas can be involved for such simplications. For example,
one can store only such inequalities that generate the vertex minimizing the function J or the adjacent vertices.
More complex modications are also obvious. Such heuristic modications seem to be more justied for the
synthesis of adaptive robust control in l
1
setting than adaptive control schemes based on estimation algorithms
of gradient type or recursive least-squares because such modications allow to store and utilize as detailed
information on unknown parameters as desired.
6. What model of perturbations is preferable?
Two dierent models of perturbations were used above in the system description, namely, the class D
j
B
(2)
of bounded memory perturbations and the class D
z
ED
(2) of perturbations with exponentially decreasing impulse
responses. In this section, the merits and defects of both models will be discussed.
The main merit of these models of perturbations is that they are veriable on-line in closed-loop. The
compatibility of a parameter vector

0 with a priori information and data on some time interval validates
simultaneously the specied model of perturbations, D
j
B
(2) or D
z
ED
(2), on this time interval. Inequalities (21)
thus ensure, both asymptotic approximate compatibility of the limit estimate 0

and the validation of the

specied model of perturbations. Quite a dierent situation prevail in stochastic adaptive control where a priori
assumptions on stochastic properties of disturbances remain in fact without verication. Of course, there is
an implicit verication of those assumptions through practical implementation of stochastic adaptive control
laws.
At rst sight the class D
z
ED
(2) seems to be preferable for implementations because previous values of output
and control have vanishing inuence on current values of output in this class, whereas in the class D
j
B
(2),
the previous values, ,(t k) (u(t k)), k {1, . . . , j}, can inuence the current value of output ,(t) equally
well. In the rest of the section we discuss problems associated with making use of each of the classes and
an additional conservatism introduced by the class D
z
ED
(2).
Dealing with the class of perturbations D
j
B
(2), we need the knowledge of the memory of perturbations
j. If j is unknown a priori, we can use some upper bound j without any eect on the upper bound (22)
guaranteed by Theorems 1 and 2 for steady-state robust performance of adaptive control. It follows from the
inclusion D
j
B
(2) D
j
B
(2) and inequality (8). The only complication of both estimation algorithms amounts to
390 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
storing a longer piece of data ,
t1
t j
and u
t1
t j
for computing m
,
(t) and m
u
(t) according to (3). One can also
see that there is no conservatism in making use of the same memory for the description of perturbations in
output and control.
Now we discuss the problems associated with making use of the class D
z
ED
(2) and show that the situation
is considerably worse for this class. Dealing with this class of perturbations, we need the knowledge of the
vector z =(z
,
, z
u
)
T
of decay rates for computing m
,
(t) and m
u
(t) according to (4). If z is unknown a priori,
some upper bounds

z
,
and

z
u
for z
,
and z
u
can be used again. However, the inclusion D
z
ED
(2) D

z
ED
(2) is
wrong now and more sophisticated analysis than for the class D
j
B
(2) is necessary. In order to get expansion of
the class of perturbations with the use of upper bounds

z
,
and

z
u
, we consider a description of perturbations
w
,
and w
u
in output and control in equivalent form:
|w
,
(t, ,)| 6o
,

,
(t), |w
u
(t, u)| 6o
u

u
(t), (27)
where the sequences
,
and
u
are computed as follows:

,
(t) =z
,

,
(t 1) +|,(t 1)| =
1
1 z
,
m
,
(t),

u
(t) =z
u

u
(t 1) +|u(t 1)| =
1
1 z
u
m
u
(t).
(28)
We continue further discussion considering, for simplicity, the perturbation w
,
in output only. In terms of
description (2), the norm of perturbation w
,
becomes now equal to o
,
}(1 z
,
). If we use an upper bound

z
,
for the unknown value of z
,
, then we have to use in the estimation algorithm, the sequence
,
associated
with

z
,
:

,
(t) =

z
,

,
(t 1) +|,(t 1)|.
The problem is reduced to nding the least upper bound for w
,
in terms of the signal
,
, that is, to nding
the minimal constant C such that
o
,

,
(t) 6C
,
(t), t N.
In view of the inequality |
,
(t)| 6|
,
(t)|, t N, we can take C 6o
,
. On the other hand,
,
(1) =
,
(1) =
|,(0)| and, consequently, C o
,
. Thus, C =o
,
and the norm of perturbation w
,
in terms of modied signal
,
becomes equal to o
,
}(1

z
,
). One can see now, that the upper bound (22) guaranteed by the control-oriented
set estimation algorithm will correspond not to the norm o
,
}(1 z
,
) but to a worse value o
,
}(1

z
,
). So,
the use, e.g., of upper bound

z
,
=0.9 instead of true value z
,
=0.5 makes the norm of perturbation w
,
5
times more that leads both to considerable reducing of the regions of robust stability and to increasing of the
guaranteed steady-state upper bound (22). So, the choice of the vector z in (4) remains an open problem.
Another conservatism associated with making use of the model of perturbations with exponentially decreas-
ing impulse responses is in inequality (8). It was shown in [14] that the upper bound J(0) for the worst-case
steady-state robust performance measure I (0) is nonconservative in the case of the class D
j
B
(2) as j
and is generally conservative in the case of the class D
z
ED
(2).
In practice, the choice of more appropriate model of perturbations among the classes D
j
B
(2) and D
z
ED
(2)
can be made through the use of both models in parallel and producing control actions based on the model
promising better steady-state robust performance.
7. Simulation example
In this section, we show how the modication of the set estimation algorithm based on minimizing the
control criterion over the set of vertices of the polyhedron E(t) performs in simulations.
Let the model of the system under control be of the form
,(t) + a
1
,(t 1) + a
2
,(t 2) =b
0
u(t 1) + t(t, ,, u), t N,
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 391
Fig. 1.
so that n=2, m=0, d=1, and the number of parameters to be estimated n
0
=6. Let controller (5) be of the form
b
0
u(t) =a
1
,(t) + a
2
,(t 1).
Then the transfer functions from the disturbance t to the output , and control u are G
,t
(:) =1 and
G
ut
(:) =(a
1
+a
2
:)}b
0
and the l
1
norms of their impulse responses are G
,t
(:)
1
=1 and G
ut
(:)
1
=(|a
1
| +
|a
2
|)}|b
0
|. The upper bound J(0) takes the form
J(0) =
o
w
1 o
,
o
u
(|a
1
| +|a
2
|)}|b
0
|
.
In the simulations below, the minimization of the function J was carried out over the set of vertices of E(t).
A priori information about unknown parameters was obtained in the form
|a
1
| 63, |a
2
| 63, 0.4 6b
0
63, o
,
+ o
u
60.8.
Note that not all of the systems with parameters belonging to the chosen set are robustly stabilizable by
corresponding controllers. In simulations, the minimization in (25) was realized over the family of vertices
associated with robustly stabilizable systems.
Figs. 1 and 2 show simulation results for two dierent systems. The graphs on the left of the gures
correspond to the adaptive control based on the set estimation algorithm with the class of perturbation D
10
B
(2)
(j =10). The design parameter c of the set estimation algorithm was taken c =0.02. The graphs on right
of the gures correspond to the adaptive control based on the classical recursive least-squares algorithm
for estimating the coecient vector 0
c
=(a
1
, a
2
, b
0
). The upper graphs show the total disturbance t in the
actual system. The additive disturbances under both types of adaptive laws were identical and the norm of
additive disturbance was o
w
=0.1. The perturbations w
,
and w
u
under dierent estimation algorithms had the
same coecients of proportion as m
,
(t) and m
u
(t), respectively. The middle graphs present the output of the
392 V.F. Sokolov / Systems & Control Letters 42 (2001) 379393
Fig. 2.
systems, while the bottom graphs are the same in another scale to make out the dierence in outputs for two
types of estimation algorithms. The initial values of outputs were random from the interval [ o
w
, o
w
].
Comments to Fig. 1: The system under control was described by the equation
,(t) ,(t 1) + 1.5,(t 2) + 0.4,(t 3) =0.8u(t 1) + 0.1u(t 2) + w
a
(t).
Since n =2 and m=0, the terms 0.4,(t 3) and 0.1u(t 2) present the unmodeled dynamics in the system, and
can be considered as perturbations from the class D
3
B
(2) with the norms o
,
=0.4 and o
u
=0.1, respectively.
The additive disturbance w
a
was a sample from the uniform distribution on the interval [ 0.1, 0.1]. The poles
of the nominal system,
,(t) ,(t 1) + 1.5,(t 2) =0.8u(t 1),
are (1 + i

5)}2 and (1 i

5)}2, and the nominal system is unstable. The polyhedral set estimate E(2 0 0)
was described by 32 inequalities and had 415 vertices. The values of the robust performance function were
J(0) =0.348 and J(0(2 0 0)) =0.103.
Comments to Fig. 2: The system under control was described by the equation
,(t) ,(t 1) 2,(t 2) =1.5u(t 1) + t(t, ,, u).
The poles of the nominal system are 2 and 1, and the nominal system is unstable. The additive disturbance
and the perturbations were chosen as
w
a
(t) =0.1w(t), w
,
(t, ,) =0.4w(t)m
,
(t), w
u
(t, u) =0.2w(t)m
u
(t),
w(t) =0.3 sin(vt) + 0.3 sin(2vt) + 0.4 sin(4vt), v =10.
Thus, the norms of perturbations were o
,
=0.4 and o
u
=0.2 and the norm of additive disturbance is o
w
=0.1.
The polyhedral set estimate E(2 0 0) was described by 51 inequalities and had 899 vertices. The values of
V.F. Sokolov / Systems & Control Letters 42 (2001) 379393 393
the robust performance function were J(0) =0.348 and J(0(2 0 0)) =0, the latter was determined by the nal
estimate 0(2 0 0) =(0.025, 3, 2.207, 0.8, 0, 0)
T
.
In simulations, no eorts were made for feeding the controlled systems by the worst-case additive distur-
bances and perturbations. In most simulations, the set estimation algorithm provided a slightly better robust
performance as compared with the recursive least-squares algorithm. The simulations presented were selected
among those where an advantage of the set estimation algorithm is appreciable. It must be emphasized that
comparing the adaptive control based on the set estimation algorithm with the adaptive control based on the
recursive least-squares is incorrect from theoretical point of view. While the set estimation algorithm ensures
the best worst-case steady-state robust performance, according to the strict mathematical result of the Theorem
2, there are no similar results for the recursive least-squares algorithm.
8. Conclusion
Two dierent problems of adaptive robust regulation of SISO systems, allowing for model perturbations in
the coprime factor representation of the system as well as external additive bounded disturbance signals, have
been considered. Controllers for systems with known parameters were assumed to be given, and the problem
was in on-line control-oriented estimating both the unknown parameters of nominal system and the norms
of perturbations and additive disturbance. The solution of the problem of adaptive robust stabilization was
based on a simple projection estimation algorithm ensuring asymptotic approximate compatibility of the limit
estimate with data and a priori information. The solution to the problem of nonconservative adaptive robust
control was based on a set valued estimation and selecting the best, with respect to the control criterion, model
from current set valued estimate. Merits and defects of the models of perturbations with bounded memory
and with exponentially decreasing impulse responses have been discussed.
References
[1] V.A. Bondarko, V.A. Yakubovich, The method of recursive aim inequalities in adaptive control, Int. J. Adaptive Control Signal
Process. 6 (1992) 141160.
[2] V.N. Fomin, A.L. Fradkov, V.A. Yakubovich, Adaptive Control for Dynamic Systems, Nauka, Moscow, 1981 (in Russian).
[3] G.C. Goodwin, B. Ninness, P. Cockerell, M. Salgado, Illustration of an integrated approach to adaptive control, Int. J. Adaptive
Control Signal Process. 4 (1990) 149162.
[4] M.H. Khammash, Robust steady-state tracking, IEEE Trans. Automat. Control 40 (1995) 18721880.
[5] M. Khammash, J.B. Pearson, Performance robustness of discrete-time systems with structured uncertainty, IEEE Trans. Automat.
Control 36 (1991) 398412.
[6] M. Khammash, J.B. Pearson, Analysis and design for robust performance with structured uncertainty, Systems Control Lett. 20
(1993) 179187.
[7] J.M. Krause, G. Stein, P.P. Khargonekar, Sucient conditions for robust performance of adaptive controllers with general uncertainty
structure, Automatica 28 (1992) 277288.
[8] S.M. Naik, P.R. Kumar, Robust indirect adaptive control of time-varying plants with unmodeled dynamics and disturbances, SIAM
J. Control Optim. 32 (1994) 16961725.
[9] L. Praly, S.-F. Lin, P.R. Kumar, A robust adaptive minimum variance controller, SIAM J. Control Optim. 27 (1989) 235266.
[10] M.S. Radenkovic, A.N. Michel, Robust adaptive systems and self stabilization, IEEE Trans. Automat. Control 37 (1992) 13551369.
[11] V.F. Sokolov, Adaptive suboptimal control of a linear system with bounded disturbances, Systems Control Lett. 6 (1985) 9398.
[12] V.F. Sokolov, Cone algorithm of identication for adaptive robust control, Prepr. 10th IFAC}IFORS Symposium on System
Identication, Copenhagen Vol. 2, 1994, pp. 273278.
[13] V.F. Sokolov, Closed-loop identication for the best asymptotic performance of adaptive robust control, Automatica 32 (1996)
911928.
[14] V.F. Sokolov, l
1
robust performance of discrete-time systems with structured uncertainty, Systems Control Lett. 42 (2001) 363377,
this issue.
[15] J. Sun, P. Ioannou, Robust adaptive LQ control schemes, IEEE Trans. Automat. Control 37 (1992) 100106.
[16] S.M. Veres, D.S. Wall, S. Hermsmeyer, I. Valyi, A. Kuntsevich, S. Sheng, Geometric bounding toolbox for MATLAB, Version 5.2,
The MATLAB}SIMULINK Connection Catalogue, Math Works Inc., 1996.
[17] E. Walter, H. Piet-Lahanier, Exact recursive polyhedral description of the feasible parameter set for bounded error, IEEE Trans.
Automat. Control 34 (1989) 911915.
[18] E. Weyer, I.M.Y. Mareels, J.W. Polderman, Limitations of robust adaptive pole placement control, IEEE Trans. Automat. Control
39 (1994) 16651671.