Beruflich Dokumente
Kultur Dokumente
Abstract
Computable Lz and L, performance bounds are derived for a recently
proposed class of adaptive systems which show that, in addition to
global stability and asymptotic tradcin a systematic improvement of
transient performance can be achieved. %he underlying linear nonadaptive controller is shown to possess a parametric robustness property, but
for large parameter uncertautty it requires high gain. A comparison between the ada tive and the nonadaptive performance bounds demonstrates that alaptation improves the overall performance without the
undesirable effects of high gain.
We make the following assumptions about the plant and the reference
signal:
Assumption 2.1 The plant is minimum phase, i.e., the polynomial
b1.9
bo is Huururitt, and the plant order (n),
relative degree ( p = n - m), and high-frequency gain b, are known.
Assumption 2.2 The reference signal y,(t) and its first p derivatives
ape known and bounded, and, in addition, d"(t) is piecewise wntinuow. In particular, %(t) may be the output of a reference model of
relative degree h 2 p with a piecewise continuous input r(t).
in = -Qy+bou
y =
4 3 )
b,S"
+ . . . + b l s + bo u(s).
3" + a,-ls"-'
+ . ..t a13 + a0
21.
To design the estimate of the state 2 we construct the two filters, one
at the output and the other a t the input of the plant (2.1):
where & =
[ 1:.
i
1::;
0
+.
Let
(2.4)
..,-@,bm-l,.. .,bo],
(2.5)
y(s) = -u(s)
B(3)
+ +
B ( s ) = b,sm t
To simplify our presentation we will consider the case where the highfrequency gain is known, b, = 1.
Introduction
In the absence of disturbances and unmodeled dynamics, the tracking error of most adaptive control schemes converges to zero, that is,
they satisfy the asymptotic performance requirement. In applications,
however, the system's traniiient performance is often more important.
Numerous simulations indicate that the transient response of adaptive
systems may be unacceptable due to large initial swings. An example was presented in [12], where an extremely poor transient behavior
occurs together with ideal asymptotic performance. It is therefore necessary that in the performance analysis of adaptive systems both transient and asymptotic behavior be adressed. Recently, such an analysis
of transient performance hals led to its improvement, suggested in [ll]
and further developed in [Z]. The proposed modifications of adaptive
controllers render the tracking error arbitrarily small in terms of both
mean-square and L, bounds. This important advance is conceptual,
because the bounds derivecl in [2] depend on the normalizing signal
and therefore are not a priori verifiable. Other efforts for estimating or
improving transient performance are presented in [7, 9, 131.
In this paper we undertitke a performance analysis for a new class
of adaptive controllers [5, 61 which in simulations exhibited a transient
behavior far superior to other adaptive schemes. Here we demonstrate
that the observed improvement of the transient performance is systematic, and due to the nonlinear adaptive design and the structure of the
underlying nonadaptive linear controller. We prove this by deriving L2
and Is, bounds which show that all the error states of the adaptive_system can be made arbitrarily small, except for the parameter error e(t),
which is bounded by a constant proportional to e'(0). The performance
bounds are computable and informative: they are explicit functions of
initial conditions and des& parameters. Another remarkable property
of the new class of adaptive controllers is that for a known bound on
the uncertain parameters the stability can be guaranteed even without
adaptation [6]. When the adaptation is switched off, the underlying
linear nonadaptive controller satisfies L, bounds which can be used for
a comparison with the adaptive controller. The adaptation, although
not necessary for stabilizatilon when the bounds on the parametric uncertainties are known, results in a smaller L- bound and achieves the
asymptotic tracking not possible with nonadaptive controllers. It is
ala0 important that the new adaptive Controllers avoid the use of high
gain with which the nonadaptive controllers counteract large parameter
uncertainties.
Ioannis Kaneflakopoulos
Dept. of Electrical Engineering
University of California LOSAngeles, CA 90024
(2.1)
'This work wm supported in part by the Air Force Office of Scientific Research
nnder Grant F-49620-925-0004, and in part by the National Science Foundation
under Grant ECS-9203491.
195
Y-VI
(trackkg error)
-%
z;
= eTArX-ai-1 ( y , ,...,
~ yp-*),q,X,i) , i = 2 ,...,p(2.6)
- '
- e:&'
e;
,T
'
e i h
W =
constant, the nonlinear terms a;jI'w vanish from (2.9) and, as explained
in [SI,the nonlinear terms
reduce to known constants, so that the
whole closed-loop system becomes linear. In Section 4 we analyze the
CM and mean-square performance of this unerlying linear system. Finally, in Section 5, we use the derived bounds to qualitatively evaluate
the advantages of adaptation.
m-1
2-40
'
(2.8)
+elY.
e:&
-
3
-
...
024rw
...
azprw
~~
-CI
51
-l
22
i3
24
- dl
-1
1
-c2 - d2
(t)'1 + aZ3rw
- d3 (g)' 1 + c34rw
-1 - usrw
-cq - d4 (%)'
-1 - a23rw -c3
-uz,rw
...
. ..
21
z2
u3prw
23
u4prw
- 2, -
ZP
v, 5 -c0Izl2
1 tb J
The closed-loop adaptive system consists of the plant, the input and
output filters, and the parameter update law. In our analysis, the
states of the closed-loop system are %pressed in the error coordinates
E E W,fj E EL",( E R",t E RP,B E Et"+*. The filter states are
represented by E , and the error f j = 7 % of the output filter state ?j
with respect to its ideal tracking reference %(t). Since ~ ( tis)not an
implemented signal and is used only for analysis purposes, without loss
of generality, its initial value q ( 0 ) can be chosen such that fj(0) = 0.
The error state ?j is governed by:
- 4d01&13
2 5 -colz12.
(3-4)
6 = hi t e,z1
i(0) =0 .
(2.11)
=A d
bbzi
t(0) = 0
A
where = C Cr,and c(t) is the ideal tracking reference. Likewise, the
( = ( 5 1 - Ab)-'bbZI = wi(")ZI,
(3.8)
initial reference d u e cr(O) is chosen such that {(O) = 0.
Global uniform stability of the origin of the adaptive system (2.4)- which yields
(2.10), and the convergence of the states t o the manifold M =
1
{ z = 0, E = 0, ?j = 0, = 0} were proven in 151 using the positive def(3-9)
I I ~ I5I IIw~ll-llzlllz
~
I
inite function
1
1
1-2 ,
(2.13) and proves (3.3).
U
v, = -I.#-I&~
zieir-l
2
do
It should be pointed out that although the initial states
zz(O), ...,zp(0) may depend on c;,di, r (see [5]), they are at the dewhere l = p 1 and 121; = zTPz.'
signer's disposal. As explained in [5],by appropriately intializing the
do j = 1 dJ.
In Section 3 we analyze the L2 and L, performance of the adaptive reference model or the filter states, z ( 0 ) can be set to zero, while
system (2.4)-(2.10). The dosed-loop adaptive system is nonlinear, and &(O),j(O) and the initial conditions of the reference model and the illters remain independent of c,, d i , l'. Therefore, by setting z(0) = 0, we
'Notation: The C
, h and truncated L nonas for signals d be denoted by achieve that
11 . llmt I(. 112 and 11 . Ilz,t, respectively. The H' , norms for trsnsfer functions will be
(3.10)
denoted by 11. ]Im, and the CI norms of their impnlse responses by 11 . Ill.
~h2311~~11w
196
IIiIIz :S -I%O)IWiIIm
(3'12)
m
1 --l~(O)lllw&.
m
IIAlz :s
d
dt (IzI2eZqt) i
Iz(t)12 5 Iz(0)(2e-2at+L
be arbitrarily small.
Another advantage of the derived bounds is that they are computable.
The bound for llzllz is explick, while the bound for 11fi112 involves IlW,&
which is known from (3.6). Only the factor l[Willmin the bound for the
zero dynamics ll(112 dependii on the unknown parameters bo,. ,bm-l.
When these parameters belong to known intervals, IIWJlm can be computed using [l].
For a more complete characterization of the achieved performance, we
proceed to derive C, norm bounds for the error states of the adaptive
system (2.4)-(2.10). These bounds are also useful for a comparison with
nonadaptive systems.
lz(t)I2
5 Iz(0)12e-2'at +
-&lE(O)lP0
(3.16)
I Iz(0)12e-2af
(3.18)
d r , (3.25)
e-2*(f-')dr
SUP
te[o.m)
l P ~ (+
t )&z(t)l*
+~ z l l t ,
(3.26)
gives
O)
(3.27)
+ IIYrIIm) + we-"*
IlbWlll
(I@%%+ YI J), +
llhwlll (ll~1Iloo
(3.29)
{ jRT+GG[tlhwll1 (m+
I I ~ I I ~I,
tlvrllm)
Itwill1 3
+m
I+,.
-
4z1
( - m+ I ~
~ ( o ) I ~ - * 't1wi111
)
+ tz(011e-a')
+ E,(,)>'
t IJ(0)l.
llillm I (;-M
(3.24)
1
am
parameters.
%CO
eZat
e-2a(t-T) ( p w ( r )
/*
(3.17)
In this way, Ili((,is explicitly related with initial conditions and design
tliillm I
+w
llullm I
that
I'
26 0
1
-1IeTw
4cod0
..
Since
The error states t,ij, < , E are guaranteed to have good La performance.
2&
(3'13)
&
1 M
+ Iz(O)le-a'.
(3.30)
(3.19)
From (2.11) we have
(3.20)
(3.21)
llijllm 5 llWill1ll~1llw5
(&1
+ lz(0)Ie-m')
IlWilll,
(3.31)
where
5 12 { &&74
(\/2ypo+ 11YrIl~)
[llhwlll
1
+Kw]
+ -&m140)lPo}
IItllm 5 ~
(3.22)
~ 5 ~(=M
~ ~
(3.32)
0
The expression (3.10) for V,(O) and (3.22) show that M is a decreasing
function of 4 independent of Q, whenever z(0) = 0 due to the filter or
reference model initializaticn.
proof. Differentiating ilz12 = i z T t along the solutions of (2.9) we get
Thus, the entire state z, 0, E , fi, 1is guaranteed to have good Lm per2
P
This is an improvement over [2] where only the tracking error
(1
' z1)'
=Ckzi dkZi
- Zkaak-l ( p , &Z) formance.
of a modified MRAC scheme wan proven to be arbitrarily small.
dt 2
k=i
k=1
k=i
Since M in (3.30) depends on Ilhwlli,the bounds (3.19)-(3.21)require
1
computation of Ilhwlll,llwilli and llw~lli. Although llhwlliand llwilli
= - Ckzi dk *Z[k
- (pU &2
depend on uncertain parameters we can employ the procedure of [l]
k=l
k=l
2dk
to compute their H
' , norms and then apply the well known inequality
llgllr 5 (2n l)llGll,, where G(s) is a stable transfer function, n is its
+
p w Ez)2
Mchfillan degree, and g ( t ) is its impulse response.
=1
Let us now give a special but more revealing form of the above L,
1
2
I--colz12 4& ( P w a,) .
(3.23) bounds.
and llw.jll1, IIwilli, llhwlll,K cm independent of CO,
do.
f:
{<
2 {c
[' )
(9)
+
+
+- +
197
K,
(3.33)
1
2&
5 -JtezW'
The assumption t ( 0 ) = 0, ~ ( 0=) 0, K, = 0 is satisfied in the particular case where the initial conditions of the plant and the filter states
are zero and the system is driven by r(t).
This form clarifies the dependence of the ,
C performance on the
parameter uncertainty lO(0)l and the design parameters Q, & and 7.
Any increase in those parameters results in an improvement of the L,
performance. It is of interest to observe that 4, present in the L,
bounds (3.33)-(3.35), is absent from the Lz bounds (3.11)-(3.13). This
is consistent with the "peak-shaving" ability of the nonlinear damping
terms observed in [5].
2co
(4.11)
(4.12)
On the other hand, since Hw(s) is stable and proper, then the truncated
norms of w and z are related as
132
II~IIz,~
I II~wIlmllYllz,t i I l z w l l m (II4l2,t + I l ~ 4 z , t ) *
(4-13)
From (4.12) and (4.13),by the small gain theorem, C2 stability is guar> 18111Hwllm,and asymptotic stability is argued aa
anteed for 2&&
in the proof of Theorem 6.1 of [6].Substituting (4.13)into (4.12) and
solving for
2 1
( p w ( s ) ) -e-2co'ds
11z112,t=
It is of interest to evaluate the performance achievable with the underlying linear controller resulting from setting r = 0 in the adaptive
C performance and the
system (2.4)-(2.10). Here we investigate the ,
mean square performance of this nonadaptive system.
Without 1088 of generality we assume that all initial conditions of the
linear nonadaptive system are zero.
Using Theorem 6.1 of [6],the following result is immediate.
T h e o r e m 4.1 (Stability and L, performance) The nonadaptive
(NA) system (2.4)-(2.9) is asymptotically stable for
The 13,
s,'
Theorems 4.1 and 4.2 provide two different stability conditions (4.1)
and (4.5), of which (4.5)is directly computable [l]and less conservative
5 llhwlll.
because llHwllm
Another usual way of expressing performance properties of a h e a r
system is to examine the difference between the actual and the desired
closed-loop transfer function. In the case of tracking the desired transfer
= 1. The actual closed-loop transfer function of the
function is
nonadaptive system was derived in [6] as
> l~lllHwllm.
&
2
The menn-square values o f t , f j ,
In this expression,
is the transfer function from $w t o q in (2.9)
when r = 0. This transfer function is stable, relative degree one, and
with deg a, = p. Its poles can be arbitrarily placed by using the design
parameters c;,di.
T h e o r e m 4.3 (Tracking performance) In the nonadaptive system
(4.15), the design parameters c;,d;, 15 i 5 p can be chosen to satisfy,
for any 6, > 0 , the following tmcking performance specification:
IGc(jw)- 11 < 6,,
(4.5)
By setting t =
of
00
Vu E
R.
(4.16)
&,
2 -
which implies
5 *,
Vu E
have
Proof. For
from (3.25):
(4.9)
Now, integrating (4.9)over
a
As expected, the tracking condition (4.18) is more stringent than the
corresponding stability condition (4.5). The required value of 2&
is increased by the factor 1
and tends to infinity as 15, 4 0. In this
sense the underlying linear controller is a "high-gain" controller which
achieves a good tradring performance at the expense of an increase of
the bandwidth of the closed-loop system.
+&
198
performance is systematic. It is crucial that with adaptation this performance improvement, in the presence of large parameter uncertainty,
can be achieved without large bandwidth required by the nonadaptive
linear controller.
The performance bounds derived in this paper are for the ideal w e ,
that is, in the absence of disturbances and unmodeled dynamics. The
problem of robustness with respect to Such modeling errors is yet to be
addressed, and robust update laws for the new class of adaptive systems
will have to be developed. As in [2, 3, 4, 81, it is of interest to study
performance of the new adaptive controllers in presence of unmodeled
dynamics.
References
[l] H. Chapdat, M. Dahleh and S. P. Bhattacharyya, "Robust st%
b i t y under structured and unstructured perturbations", IEEE
Zhneactiona on Automatic Control, vol. 35, pp. 1100-1108,1990.
Corollary 5.1 Let the initial conditions of the ermr states I,E , ij, t be
zero, and w(0) = 0. Then with adaptation gain
- 18111hwlll
m ELc, - ( 2 J m - 18111hyI11)
F 5&2
72 2
l$lllh&,
1'-
''I2 ,
IlVJt
(5.2)
[S] S. M. Naik, P. R. Kumar and B. E. Ydstie, "Robust continuoustime adaptive control by parameter projection", IEEE %nsactions on Automatic Control, vol. 37, pp. 182-197,1992.
GAdtA.
GAGA
[9] R. Ortega, "On Morse's new adaptive controller: parameter convergence and transient performance" ,to appear in ZEEE 2hm"tions on Automatic Control.
[ll] J. Sun, "A modified model reference adaptive control scheme for
improved transient performance", Pmceedings of the 1991 American Control Conference, Boston, MA, pp. 150-155.
[12] Z. Zang and R.R, Bitmead, "Transient bounds for adaptive control
systems", Proceedings of the 29th IEEE Conference on Decision
and Control, Honolulu, HI,1990, pp. 2724-2729.
[13] B. E. Ydstie, "'Bansient performance and robustness of direct
adaptive control", IEEE hnsactions on Automatic Control,
vol. 37, pp. 1091-1106,1992.
Conclusions
Our L2 and L, bounds show that the performance of the new class
of adaptive controllers proposed in [5, 61 can be made a8 good as d e
sired. The use of design parameters for improvement of the transient
199