5,

'

ELSEVIER

S:J TtiMS 6 ONTItOI. LI[TTIINS

Systems & Control Letters 26 (1995) 167-176

Model reference adaptive control for nonminimum phase systems

D a n i e l E. Miller*
Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Ont., Canada N2L 3G1

Received 8 September 1994; revised 20 February 1995

Abstract

In this paper we define a new model reference adaptive control problem which allows a large amount of plant uncertainty. Instead of the usual goal of asymptotic error regulation, here we require that the limit superior of the magnitude of the error be less than a predefined constant times the limit superior of the magnitude of the external reference input. This provides a natural way to incorporate uncertainty which includes nonminimum phase systems.
Keywords." Model reference adaptive control; Switching control; Nonminimum phase system; lj performance

1. I n t r o d u c t i o n

It has been shown that to carry out exact model reference adaptive control for an uncertain discrete-time plant, an upper bound on the relative degree must be known, and any plant zeros outside the open unit disk must lie in a finite set [8]; this information is required to design a compatible reference model, and is the reason why the problem is typically solved only for classes of minimum phase systems. Here we move away from the minimum phase assumption by modifying the problem slightly: we no longer require exact asymptotic model matching, but instead we require that the controlled plant asymptotically approximate the behaviour of the reference model. We approach this problem using ideas from the nonidentifier based approach to adaptive control (see [2] for an overview). This approach is based on adjusting the controller gains based on performance, and no parameter identification is carried out. The approach was initially used for stabilization, e.g. [15, 5, 10], and later for tracking, e.g. [6, 12, 1 I, 3]. However, in the latter case the results are restricted either to having the reference signal described by a finite dimensional unforced differential equation, e.g. [6, 12], or to that of allowing a much more general reference signal, but at the expense of requiring the plant to be minimum phase, e.g. [11, 3]. Here we will extend these ideas to the case of a nonminimum phase plant approximately tracking the output of a forced reference model. We will design an adaptive controller which switches between the elements of a set of LTI compensators; the difficult part is in determining when to switch compensators, and there is an extra twist not evident in earlier work in this area. This approach allows for a mismatch of relative degree and unmodeled dynamics. Here we will consider the discrete-time multivariable case. An earlier version of this work appeared in [7]. ~"This work was supported by the Natural Sciences and Engineering Research Council of Canada. * E-mail: [email protected]. 0167-6911/95/$9.50 (~) 1995 Elsevier Science B.V. All rights reserved SSDI 0167-6911(95)00012-7

168

D.E, Miller~Systems & Control Letters 26 (1995) 167-176

2. Notation
We use the Holder cxD-norm for vectors x E ~n: Ilxl[ = max Ixi[,
l~i~n

and the corresponding induced norm for matrices in ~,m. We let S(~ "m) denote the set of ~"m-valued sequences on Z +, For x = {x(i)} E S(R "m) we define

IIxll~ := sup IIx(i)ll.

iE~"

We let l~(~nxm) denote the subset o f x E S(~ "m) for which tlxll~ < ~ . With f E S(R"xm), we let f denote its z-transform, and we define the operator F : S ( R m) ~ S(~ ") by

Fu:=f*u.
In this paper we are interested in using the ~ - n o r m on our signals, so we define the norm of this operator by

IIF[I :=

sup

IIFull~

If we partition f = [ f T . . . f T ] T , i.e. so that

oo

fi E S(~lm),

then it is easy to see that

IlFit = m ~

S [[fgO)[I.

t ~ n .l=O

Last of all, to improve readability, we will drop the dimension of the difference spaces when there is no possibility of confusion, e.g. we write f E l ~ rather than f E l ~ ( R " ) when the dimension of f is already known from the context.

3. The problem
Suppose that the plant can be modeled by a finite-dimensional difference equation

x( k + 1) = Ax( k ) + Bu( k ),
y( k ) = Cx( k ),

x(0) =xo,

(1)

with k E Z +, x(k) E R n the state, u(k) E ~'~ the control input, and y(k) E ~r the plant output; we identify the model with the matrix triple (A,B,C). To minimize notation we use P to denote both this matrix triple and the corresponding operator u ~ y (with x0 = 0); its meaning will be clear from the context. We let ~p denote the set of all systems of the form (1) of arbitrary order n but with m inputs and r outputs, and for which (A,B) is stabilizable and (C,A) is detectable. Here our model is not known exactly: it lies in a subset of 2:p. We also have an asymptotically stable reference model described by

x,.(k + l) = A,.xm(k ) + B.,u,.(k ), ym(k ) = C,nxm(k ),

x,.(O ) = X,.o,
(2)

with k E Y+, xm(k) E R "~ the state, urn(k) E ~ " the exogenous input, and ym(k) E ~r the reference model output. The reference model embodies the desired behaviour of the controlled plant in response to the exogenous input urn, and is assumed to have a nonzero transfer function. We let Pm denote the operator um ~-~ Ym (with x,, 0 = 0).

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

169

Ur n

Fig. 1. The Standard Feedback System. Here we consider controllers of the form

: S

--~ S,

(y, ym, Um) ~

U,

(3)

so that we end up with a feedback configuration given in Fig. 1. This form of K is very general; in particular, it includes every controller of the form

z(k + 1) = f ( z ( k ) , y(k), uz(k), k), u(k) = g(z(k), y(k), urn(k), k).

z(O) = zo,
(4)

[Of course, in order for K to be a well-defined map, zo must be fixed.] Our goal is to design a controller which forces the plant to act asymptotically like the reference model. Ideally, for each P E ~ , xo E R n, Xmo E R nm, and Um E loo, we would like u,x E l ~ and

e(k) := y ( k ) - ym(k)
to go to zero as k ~ co. However, as shown in [8] for the single-input single-output case, this goal is not achievable if ~ is too large; indeed, there is a bound on the asymptotic performance which can be achieved. Instead, we will be content with a weaker tracking requirement, to be made precise shortly. Consider a general LTI compensator of the form

z(k + 1 ) = Fz(k ) + Gy(k ) + Hym(k ) + Jura(k), u(k ) = Lz(k ) + M y ( k ) + Nym(k ) + Qum(k );

z ( O ) = zo,

(5)

we identify this compensator with the 8-tuple (F, G,H,J,L,M,N, Q). To minimize notation we use K to denote both this 8-tuple and the corresponding operator (y, Ym, Urn) H U (with z0 = 0); its meaning will be clear from the context. We let Sk denote the set of all (F, G,H,J,L,M,N, Q) with appropriate dimensions for which (F, G) is stabilizable and ( L , F ) is detectable. Suppose that (5) is applied to (1). With

.f:=

/} :=

A :=

x.,

LBmJ

GC o

F o

HCm Am

:=

(6)

the closed loop system is given by

,2(k + 1 ) = AY(k) + Bum(k),

e(k) = C(k).

Hence, we say that (5) stabilizes (1) (or the closed loop system is stable) if the corresponding matrix given in (6) is asymptotically stable. Denote the closed loop transfer function from Um to y by [(P,K); the transfer function from Um to e is simply i ( P , K ) - Pro, which is exactly C ( z l - j ) - l ~ . Now suppose that the closed loop system is stable. If xo = O, Xmo = 0, and z0 = 0, then the tracking error e(k) satisfies

[lell~ ~

[IT(P,K)

e,,,ll x Humlloo.

170

D.E. Miller~Systems & Control Letters 26 (1995) 16~176

In fact, we can also show that lim sup Ile(k)ll ~< IIT(P, K) - Pm II x lim sup Ilum(k)[I,
k~o k---* cx~

and that this bound is tight. The latter inequality is more useful here, since in adaptive control the transient response is often poor during the tuning phase, and we are often willing to settle for good asymptotic performance. This leads us to the following modified version of the model reference adaptive control problem. Definition 1. Given Pm, the controller (3) solves the Approximate Model Reference Adaptive Control Problem ( A M R A C P ) f o r ~ w.r.t. ~ if, for every P E ,~, x0 E ~n, Xmo E ~n,, and u,, E loo, we have x,u E l ~ and

lim sup Ile(k)ll ~

k----~ oo

lim sup
k---*cx~

Ilum(k)bl,

Implicit here is the requirement that the closed loop system be well-posed for every P E ~@: for every x0 E R n, Xmo E II~ and um E IN, we have unique u,y E l ~ which satisfy (1) and (3). nm, For the single-input single-output case, a conservative lower bound on ~ is given in [8]. Also, observe that the performance objective is similar to that desired in the /1-optimal control problem, e.g. see [14, 1]. While it has been shown recently that in some situations a nonlinear controller can provide better tracking (in the /l-sense) for a fixed plant than the best LTI controller [13], given that a typical adaptive controller works by tuning the gains of an LTI controller, the best ~ that we can reasonably hope for is
PC.~ K E Zk stabilizes P

sup

inf

IIT(P,K) - Pmll.

In the next section we design an adaptive controller to solve the AMRACP based on switching between a countable set JU of LTI compensators; the choice of ~ will depend on jug, but it is never smaller than the above quantity.

4. The main result At this (i) A (ii) A (iii) A point we assume that we have set of LTI models ~ C Sp. reference model Pm of the form (2). countable set of LTI compensators
~ } C ~'k,

= {Ki = (Fi, G i , H i , J i , L i , M i , N i , Q i ) : i E

for which there is a uniform upper bound on the order. (iv) A constant 7 E ~ satisfying sup
PE~

inf
K C.~Y stabilizes P

I[T(P,K)

- Pmll < ~.

In principle, a suitable ~ can be obtained from Pro, ~, and ~ , although it may be difficult to compute in general; see the next section for some special cases. Without loss of generality, we assume that the controller gains do not blow up too fast, namely that there exists a constant c > 0 so that IIFi[I + Ilai[I + IIn~[I + IIJ~ll + tlZ~bl + IIMil[ + IIN~II+ IIa~ll ~ cil/3, i E ~, (7)

and that every element in the sequence of compensators {Ki}i~l appears infinitely often: {Ki:i>~n}=~, ne~.

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

171

Finally, because of Assumption (iii), we can assume that the elements of )f" are all of the same order, since we can augment our state-space with an unobservable part if necessary. The following result allows us to obtain a bound on the magnitude of the plant state by filtering u and y. L e m m a 1. Suppose that (C,A) is detectable; then there ex&t constants 2 E (0, 1) and 7 > 0 so that the solution o f (1) satisfies

[[x(k)ll ~<~, Ilxoll + ~ ,~k-~-~([lu(i)ll + Ily(i)ll) ,

i=0

k~>

l.

I f (C,A) is observable, then for every 2 E (0, 1) there exists a 7 > 0 so that the above equation holds.
Proof. The first part is proved in [9] under the additional (but unused) hypothesis that (A,B) is stabilizable; the second part follows from a straight-forward modification of that proof. E2 While we know that there exists a Ki E ~ for which [ [ T ( P , K ~ ) - emil < c~, the impulse response of T(P, K i ) - Pm may go to zero very slowly, so that e(k) may depend strongly on values of urn(i) far into the past. Using the following result we shall construct a filter of Um whose output tracks um immediately when it is increasing, but decays slowly if Um suddenly becomes small; it plays a crucial role in our controller. L e m m a 2. Suppose that p E [0, 1) and ~b(k + 1) = max{p~(k),llum(k)ll}, (a) Then lim sup q~(k) = lim sup I[um(k)l I.
k---~oo k---~ oo

~b(O)~>O.

(b) Further suppose that h E S ( ~ rm) satisfies h(O) = O, there exist 7 > 0 and 2 E (0, 1) so that

IIh(k)ll ~ ) ? ,

k>~O,

and e := h * u~. Then for every 6 > 0 there exists a ~ E (0, 1 ) so that if p E (fi, 1 ), then for every u~ E S we have

I[e(k)ll<~([lnl] +'~)l'P(k)l,
Proof. See the Appendix.

k~>0.

Before presenting the controller, we define the switching function f : E ~ f(x)= 1 x~<l i xE(i!,(i+l)!], i E N.

~ by

With )o E (0, 1 ), we define the causal controller K). : S S S --~ S as follows: with y, Ym, Um E S we define u E S by

r(k + 1) = r(k) + max{(2 - 1)r(k) + [ly(k)[[ + Ilu(/)ll + IIz(k + 1)ll, Ile(k)ll - ~ffm(k),0}, ffm(k + 1) : max{[1 - f(r(k))-l]ffm(k), I[um(k)ll},
if,,(0) : 0, z(0) = 0,

r(0) = 1,

z(k + 1 ) = Ff(r(k))z(k ) + Gf(r(k))y(k ) + Hy(r(k))ym(k ) + Jf(r(k))um(k ),

R e m a r k 1. The idea is that we generate the auxiliary signal r(k), and use Ki in the feedback loop while r(k) E (i!,(i + 1)[].

172

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

Theorem 1. For every 2 E (0, 1), the controller K~ solves the A M R A C P for ~ w.r.t. ~.
Proof. Let the plant model belong to ~ , x0 E ~", xmo E ~,m, and u,, E l ~ ; define J? as in (6). For now we assume that (C,A) is observable. Now for a brief outline of the proof. First, we use Lemma 1 to show that r(k) provides a bound on IIJ?(k)[I, and then use Lemma 2(a) to show that if r is bounded, then we achieve our control objective. We then assume that r is unbounded and derive a contradiction. We first show that for large i we will not jump over the controller Ki. We then use Lemma 2(b) to show that (k) and r(k) are well-behaved while a sub-sequence of these Ki's are being used, thereby obtaining a contradiction. From Lemma 1 there exists a constant c~ > 0 so that

IIx(k)ll<~el IIx01l+ ~ 2k-J-z([ly(J)l[ + [[u(j)ll)

j=0

E
k-I j-0

k>~l.

Using the difference equation defining r, we see that

r ( k + l ) ~ > 2 r ( k ) + Ily(k)ll + Ilu(k)ll,

k>~0,

r(k)>~ ~ '~-J-~(llY(J')[[ + Ilu(j)ll), k>~ 1,

Ilx(k)l [~<c~[llxotl+r(k)], k>~l.
Since r ( k ) ~ 1 provides a bound on IIz(k)[I and Am is asymptotically stable, there exists a constant c2 > 0 so that II~(k)ll <<,c2r(k),

k >~O.

(8) cxz, and

First, let us suppose that r E l ~ . Since r(k) is monotonically increasing, it converges as k ~ u E l ~ , and that lim sup Ile(k)ll ~<~ lim sup fire(k),
k ~:x~ k---*~

f ( r ( k ) ) converges in finite time. From (8) we see that x E l ~ , and from the definition of r we see that

so by invoking Lemma 2(a) with p = l i m k _ ~ ( 1 - f ( r ( k ) ) - t ) , lim sup

k---~oo

we see that

Ile(k)ll ~<= lim sup Ilum(k)[[,

k---*~

as required. Now suppose that r is unbounded; we will obtain a contradiction. Since Pm is stable,

c3 := Ilymll~ + Ilu,,ll~ < ~ .

By assumption, there exists a constant ca so that
liE, I[ + [[GII + [[nil[ + [IJ~ll + Ilt, ll + IIm~ll + I[g,[[ + [IQi[l~c4i 1/3,

(9)

i E N.

(10)

For each i E N, let ki denote the smallest k E Z + such that r(k) > i!. Making use of (8), (9), and (10), we see that there exists a constant cs so that

r(ki) ~ r(k~ - 1 ) + [[y(ki - 1)ll + I[u(k, - 1)[I + Ilz(kg)[I + lie(k, - 1)[[ <~ (i!) + c2[[cll(i! ) + 2c4 il/3 <~ csil/3(i!),
Choose il so
that

[(i]) -+- e2llfll(i!) + c3] --I--[c3 -~- c2llCIl(i!)] (11)

i E ~.
1;

csil/3 < i I -t-

then ki+j > ki for i>~ij. Hence, every Ki, i>~il, is used for awhile.

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

173

Choose/~ E 3 ( so that IIT(P,f;)-P~II < ~; let (A,B, C) denote the corresponding state-space representation given by (6). Choose 21 E (0, 1 ) and c6 > 0 so that [I/i*ll ~<c62ff, k~>0. (12)

With h(0) := 0, h(k) := ~ j k - l / } , k E ~, and 6 E (0,7 - [IHII), choose 15 as in Lemma 2(b). Choose i2 > il so that (1 - 1/i2) > 3, and let I denote the infinite set I := { i E ~ :Ki = K , i>~i2}. Note that u(k) is being generated b y / for

k E {j E ~ " f ( r ( j ) ) E I} =: [;
indeed x ( k + l) =~L~(k)-[-Bum(k),

e(k) = C;(k),

k EL

(13)

Let i E I. Then {ki, ki + 1 . . . . . ki+l - 1} C [ , so using the definition of h we have

k-1 j=k~ k--I

e(k) = CAk-k~Y(ki)+ ~ h ( k - j ) u m ( j ) ,
j=ki

ki < k<~ki+l.

If we let p = 1 - 1/i > fi and apply Lemma 2(b), we see that

k-1 - j)u,.(l') j ~ h(k

~<([[H[I + O)tT~(k) ~< Sam(k),

kg < k<<.k~+j;

hence, if we take the norm of both sides of the previous two equations as well as (13) (to obtain the case of k = ki) and use (8), (9), (11), and (12), it follows that there exist constants c7 and c8 so that
1/3 'I II;(k)[) <<-C2C5C6l (l.) -~ ~ .1/3 .1 flBII ~<c7, (~.),

(14)

Ile(k)ll <~c2c5c61fcI1il/3(i!)2~ -k' + ~ffm(k ) <~c8il/3(i! )2~ -k' + ~tim(k), Using (9), (10) and (14) we see that there exists a constant c9 so that

ki <<.k<~ki+l.

(15)

Ily(k)ll + [[u(k)ll + Ilz(k+ 1)11<~c9i2/3(i!),

and by combining this with (15) we get

ki<~k < ki+l,

r(k + 1) <~r(k) + max{(2 -- 1 )r(k) + c9i2/3(i!), ~ :1/3, ~k-k, , 0}, c8l ~t.)~
We claim that

ki<~k < ki+l.

(16)

r(ki+l )~c5il/3(i!) + c8 [2 ~--j 21 ] il/3(i!) + c9 [

To prove this, first note that if
r ( k i + l ) ~ c 5 i l / 3 ( i ! ) + c9 ~ i2/3(i!),

i2/3(i!).

(17)

(18)

174

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

then this is clearly the case. If (18) does not hold, let/~ denote the largest k < ki+l for which

r(k)<~csil/3(i!)+c9

[']

~ - A i2/3(i!);

from (11) we see that fc>~ki. It follows from (16) that

r(lc 4-1)<~r(lc)+ c9i2/3(i!)4-c8i1"3(i') <~cst' (t.)4-c8il/3(i!)4-c9 [2 -- 2] i2/3(i!). .1:3 "1 ~

Furthermore, it follows from (16) and the definition of/~ that

(19)

r(k 4-1)<...r(k) 4- csil/3(i!)2~ -k~,

k 4-1<~k < ki+l.

Solving this iteratively with an initial condition satisfying (19) yields (17). Clearly the RHS of (17) is less than (i + 1 )! for i E I sufficiently large, which contradicts the definition of

ki+ I.
The extension of the proof to the case in which (C,A) is only detectable is easy - we apply the above argument to the observable part. E3 Remark 2. We can allow a more general switching function f : ~ ~ ~. Indeed, if we choose a strictly increasing sequence "rr"ti= satisfying l i m i n f i ~ ( r i + l / i r i ) > O, and redefine t /Ji=l f ( r ) := 1 if r < r I i if r E [ri, ri+l), i E N,

then the controller still works.

5. A special case

Perhaps the most interesting situation is one in which aug is finite. Suppose that ~ = UP=I every P E ~ i , and ~i := sup IIT(P,K~)-PmN < oo.
P E ,Y~i

' ~ i , Ki

stabilizes

If we choose ~ > max{~i: l ~ i ~ < p } and define

Ki = K1

+ (imodulop),

i > p,

then the corresponding adaptive controller presented in the previous section guarantees that all signals remain bounded and that limsup lie(OH ~<~ limsup Ilum(t)jl . Example 1. The simplest example of the above situation is the one in which .~ = {P1, P2} and there is no LTI compensator which stabilizes both systems. While computing the optimal LTI stabilizing controller Ki for Pi is difficult, we can use [1] to compute a sub-optimal LTI stabilizing Ki which achieves a performance level as close to the optimal one as we like. Furthermore, by examining the proof of Lemma 2 it is easy to see how we can choose p E (0, 1 ) so that the update equation for tTm can be simplified to ~Tm(k+ 1) = max {pg~m(k),Hum(k)[[},

tim(O) = O.

Example 2, Now suppose that ~@i has the form often considered in the robust control literature, e.g. [4]: with IV,. a stable LTI weighting function, we have

~i = {Pi(I + W i V ) : 6 is real rational and satisfies Ilvll ~< 1},

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

175

Assuming that a stabilizing controller for ~; exists, in the scalar case we can use [4] to design a suboptimal controller of the form Ki = [KiI - K i I 0] which achieves a performance level as close to the optimal one (of this form) as we like.

6. Concluding remarks
In this paper we looked at the problem of solving a model reference adaptive control problem for a possibly nonminimum phase plant. We argued that there is a natural formulation of such a problem, and presented a switching controller to solve it. It would be desirable to obtain a stronger form of tracking, namely that of recovering the best tracking which is possible when the plant is known. Furthermore, the noise properties of the proposed controller have not been investigated. These remain topics of further research.

Appendix Proof of Lemma 2. First observe that

limk__.o~sup q~(k) = limk__.oosup max{pq~(k), I]um(k)ll} = max {p limk~o~sup ~b(k), limk~supIlum(k)l[ } = lim sup Ilum(k)ll.
k---~o~

Now suppose that h E S(It~rxm) satisfies the hypothesis. Notice that

4~(k)>~l[u,,(k-1)ll,
so by iteration we see that

k>~l,

and

dp(k)>~pcb(k-1),

k>~l,

qb(k)>~pJ-J~(k-j + l)>~pJ-l[lu,,(k-j)l[,
Since
k

k >~j>~l.

e(k) = ~ h(j)um(k - j),

j=l

if we partition h . . . [ht . T

h T T,then f o r k / > l we have r]

lei(k)[ ~ ~ Ilhi(j)ll" Ilum(k -J)[I ~<

j=l =1

[[hi(J')l[p '-j qS(k),

1 <~i<~r.

For p E (2, 1) we have

k j=l ~o j=l j=l j=0 oo j=l

F, [Ihi(J)l[P1-j <~ ~ Ilhi(/)ll + ~ [rh,(j)ll(P 1-j - 1) ~< ~ Ilh;(J)[] + ~ T2J(p 1-j - 1)

--< IIHII + ~-2 (1 _ 1 - p - ,t)' ,t)(p so the result follows. []

l~i<.Nr,

References
[1] M.A. Dahleh and J.B. Pearson, Optimal rejection o f persistent disturbances, robust stability, and mixed sensitivity minimization, IEEE Trans. Automat. Control AC-33 (1988) 722-731.

176

D.E. Miller~Systems & Control Letters 26 (1995) 167-176

[2] A. llchmann, Non-identifier-based adaptive control of dynamical systems: A survey, IMA J. Math. Control Inform. 8 (1991) 321-366. [3] A. llchmann and E.P. Ryan, Adaptive tracking for nonlinear systems in the presence of noise, preprint. [4] M. Khammash and J.B. Pearson, Robust disturbance rejection in /I optimal control systems, Systems Control Letters 14 (1990) 93-101. [5] B. Mhrtensson, The order of any stabilizing regulator is sufficient a priori information for adaptive stabilization, Systems Control Lett. 6 (1985) 87-91. [6] B. M~rtensson, Adaptive Stabilization, Ph.D. Thesis, Dept. Automatic Control, Lurid Institute of Technology, Sweden (1986). [7] D.E. Miller, Model reference adaptive control for nonminimum phase systems, Proe. 1993 European Control Conf. (1993) 190-194. [8] D.E. Miller, On necessary assumptions in discrete-time adaptive control, Proc. First Asian Control Conf. 3 (1994) 65-68. [9] D.E. Miller, Adaptive stabilization using a nonlinear time-varying controller, IEEE Trans. Automat. Control AC-39 (1994) 1347-1359. [10] D.E. Miller and E.J. Davison, An adaptive controller which can stabilize any stabilizable and detectable LTI system, in: C.I. Byrnes, C.F. Martin and R.E. Saeks, eds., Analysis and Control of Nonlinear Systems (North-Holland, New York, 1988) 51-58. [11] D.E. Miller and E.J. Davison, An adaptive controller which provides an arbitrarily good transient and steady-state response, IEEE Trans. Automat. Control AC-36 (1991) 68-81. [12] D.E. Miller and E..I. Davison, An adaptive tracking problem, Internat. J. Adaptive Control and Signal Processing (1992) 45-63. [13] A.A. Stoorvogel, Nonlinear Li optimal control for linear systems, IEEE Trans. Automat. Control AC-40 (1995) 694-696. [14] M. Vidyasagar, Optimal rejection of persistent bounded disturbances, IEEE Trans. Automat. Control AC-31 (1986) 527-534. [15] J.C. Willems and C.I. Byrnes, Global adaptive stabilization in the absence of information on the sign of the high frequency gain, in: 6th Int. Conf. on the Analysis and Opt. of Systems, Lecture Notes in Control and Inform. Sci. No. 62 (Springer, New York, 1984) 49-57.