[ieee 1982 21st ieee conference on decision and control - orlando, fl, usa (1982.12.8-1982.12.10)]...

4
FAI-11:15 EXPONENTIAL CONVERGENCE OF RECURSIVE LEAST SQUARES WITH EXPONENTIAL FORGETTING FACTOR* Richard M. Johnstonel, C. Richard Johnson Jr.2, Robert R. Bitmead3 and Brian D.O. Anderson1 'DepartmentofSystemsEngineering,Instituteof Advanced Studies, 2School of Electrical Engineering, Cornell University, Ithaca, NY 14853, USA 3DepartmentofElectricalandElectronicEngineering, James Cook Australian National University, Canberra, ACT 2600, Australia University, Townsville, Qld 4811, Australia *The work of Authors1 was supported by the Australian Research Grants Committee,whilethatofAuthor2 was supportedby NSF grant ECS-8120998. ABSTRACT Thispaperdemonstratesthat,providedthesystem input is persistently exciting, the recursive least squaresestimationalgorithmwithexponentialfor- getting factor is exponentially convergent. Further, it is shown that the incorporation of the exponential forgetting factor is necessary to attain this conver- gence and that the persistence of excitation is vir- tually necessary. The resultholdsforstablefinite- dimensional, linear, time-invariant systems but has its chief implications to the robustness of the parameter estimator when these conditions fail. 1. IhTRODUCTION This paper deals with the exponential stability of a popular adaptive estimation scheme - the recursive least squares (RLS) algorithm. In practice, the RLS algorithm is modified to incorporate an exponential forgettingfactor (X) between zero and one. The reason for including X has most often been justified by noting thatsuchaninclusiondiscountspastmeasurementsin favourof more recentones[l]and so offers some capacity to identify time-varying plants. Without exponentialforgetting (X = l), and with rich inputs, t h e RLS a l g o r i t h m effectively turns itself off, after some time. This, of course, means t h a tt h e RLS a l g o - rithm p r o g r e s s i v e l y l o s e s t h e a b i l i t y t o i d e n t i f y the time-varying plant parameters. Conversely, with A less than 1, one maintains tracking capabilities at the cost of non-zero limiting parameter covariance. Our main result is that, if the system input {ukl is p e r s i s t e n t l y e x c i t i n g the RLS algorithm with expo- nential forgetting factor is exponentiallyconvergent, in that the parameter error vector (the difference betweentheactualplantandtheestimation scheme parameters) approaches zero exponentially fast. Further it is shown t h a t without this exponential for- getting factor (ie A = 1) exponential convergence is lost. It is this property of exponential convergence which is necessary to ensure robust performance in the presence of measurement noise, time-variations of plant parameters, and under-modelling of plant order. The easewithwhichanexponentialoverbound on the para- meter error is found contrasts sharply with the rela- tive difficulty [2,9] in establishing such a conver- gence rate for steepest-descent type algorithms, This reflects the Newton-Raphson-like nature of RLS. In view of the recent results of [2] the need for persistent excitation of {uk} shouldnot be surprising. Indeed, if we do n o t have persistent excitation or' {ukl,then it is trivial to show that we cannot en- sure convergence of the parameter error to zero, at all. We also show how anestimation scheme due to Landau [3] can be modified so that an exponential forgetting factor is incorporated as in [4]. An out- line of the proof of exponential convergence of the modifiedalgorithm is given. The approach of the present note is l i m i t e d to linear, time-invariant, finite-dimensional, stable plants which aresingle-input,single-output (SISO) and driven by deterministicinputs. The SISO r e - striction can doubtless be removed with an appropriate parameterisation scheme [SI. Extensions to stochastic inputsalongthelinesof[6]also seem readily possible, although technically involved. Nonlinear, time-varying or infinite-dimensional plants can only be accommodated to the extent that the results given here are robust in the presence of some nonlinearity, some parameter time-variation and some modes which a r e neglected. See [7-91 for discussion of the effects of departure from ideality in adaptive system problems. The significance of this paper is that it demonstrates that the convergence rate advantages of RLS o v e r steepest-descent algorithms need not be abandoned in order to gain adaptability, 2. THE RLS ALGORITHM WITH EXPONENTIAL FORGETTING The RLS a l g o r i t h m can be described by the following eauations and r or (2.3) where X E [O,l], ?-l is a positive definite matrix, and 8 is arbitrary. The parameter estimate vector e can be shown [ l ] to converge to the value e which - 0

Upload: brain

Post on 27-Jan-2017

214 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: [IEEE 1982 21st IEEE Conference on Decision and Control - Orlando, FL, USA (1982.12.8-1982.12.10)] 1982 21st IEEE Conference on Decision and Control - Exponential convergence of recursive

FAI-11:15

EXPONENTIAL CONVERGENCE OF RECURSIVE LEAST SQUARES WITH EXPONENTIAL FORGETTING FACTOR*

Richard M. Johns tone l , C . Richard Johnson J r . 2 , Robert R . Bitmead3 and Brian D . O . Anderson1

'Department of Systems Engineer ing, Inst i tute of Advanced S t u d i e s ,

2Schoo l o f E lec t r i ca l Eng inee r ing , Corne l l Un ive r s i ty , I t haca , NY 14853, USA

3Depar tment o f E lec t r ica l and Elec t ronic Engineer ing , James Cook

Aus t r a l i an Na t iona l Un ive r s i ty , Canbe r ra , ACT 2600, Aus t ra l ia

University, Townsville, Qld 4811, Australia

*The work of Authors1 was suppor ted by t h e A u s t r a l i a n R e s e a r c h G r a n t s Commit tee , whi le that of Author2 was supported by NSF g r a n t ECS-8120998.

ABSTRACT

This paper demonst ra tes tha t , p rovided the sys tem i n p u t i s p e r s i s t e n t l y e x c i t i n g , t h e r e c u r s i v e l e a s t squa res e s t ima t ion a lgo r i thm wi th exponen t i a l fo r - g e t t i n g f a c t o r i s exponen t i a l ly conve rgen t . Fu r the r , i t i s shown t h a t t h e i n c o r p o r a t i o n o f t h e e x p o n e n t i a l f o r g e t t i n g f a c t o r is n e c e s s a r y t o a t t a i n t h i s c o n v e r - g e n c e a n d t h a t t h e p e r s i s t e n c e o f e x c i t a t i o n i s v i r - t u a l l y n e c e s s a r y . The r e s u l t h o l d s f o r s t a b l e f i n i t e - d imens iona l , l i nea r , t ime- inva r i an t sys t ems bu t has i ts c h i e f i m p l i c a t i o n s t o t h e r o b u s t n e s s o f t h e p a r a m e t e r e s t i m a t o r when t h e s e c o n d i t i o n s f a i l .

1. IhTRODUCTION

T h i s p a p e r d e a l s w i t h t h e e x p o n e n t i a l s t a b i l i t y o f a popu la r adap t ive e s t ima t ion scheme - t h e r e c u r s i v e least squa res (RLS) a l g o r i t h m . I n p r a c t i c e , t h e RLS a lgo r i thm i s m o d i f i e d t o i n c o r p o r a t e an exponen t i a l f o r g e t t i n g f a c t o r ( X ) between zero and one. The reason f o r i n c l u d i n g X has mos t o f ten been jus t i f ied by no t ing tha t such an inc lus ion d i scounts pas t measurements in f avour o f more r ecen t ones [ l ] and so o f f e r s some c a p a c i t y t o i d e n t i f y t i m e - v a r y i n g p l a n t s . W i t h o u t e x p o n e n t i a l f o r g e t t i n g ( X = l), and w i th r i ch i npu t s , t h e RLS a l g o r i t h m e f f e c t i v e l y t u r n s i t s e l f o f f , a f t e r some time. Th i s , o f cou r se , means t h a t t h e RLS a lgo - rithm p r o g r e s s i v e l y l o s e s t h e a b i l i t y t o i d e n t i f y t h e t ime-vary ing p lan t parameters . Converse ly , wi th A less than 1, o n e m a i n t a i n s t r a c k i n g c a p a b i l i t i e s a t t h e c o s t o f non-ze ro l imi t ing pa rame te r cova r i ance .

Our main r e s u l t i s t h a t , i f t h e s y s t e m i n p u t { u k l

i s p e r s i s t e n t l y e x c i t i n g t h e RLS algori thm with expo- n e n t i a l f o r g e t t i n g f a c t o r i s exponen t i a l ly conve rgen t , i n t h a t t h e p a r a m e t e r e r r o r v e c t o r ( t h e d i f f e r e n c e be tween the ac tua l p lan t and the es t imat ion scheme parameters ) approaches zero exponent ia l ly fas t . Fur the r i t i s shown t h a t w i t h o u t t h i s e x p o n e n t i a l f o r - g e t t i n g f a c t o r ( i e A = 1) exponent ia l convergence is l o s t . I t i s th i s p rope r ty o f exponen t i a l conve rgence which i s n e c e s s a r y t o e n s u r e r o b u s t p e r f o r m a n c e i n t h e presence o f measurement no ise , t ime-var ia t ions o f p lan t parameters , and under -model l ing o f p lan t o rder . The ease with which an exponential overbound on t h e p a r a - m e t e r e r r o r i s f o u n d c o n t r a s t s s h a r p l y w i t h t h e r e l a - t i v e d i f f i c u l t y [ 2 , 9 ] i n e s t a b l i s h i n g s u c h a conver- g e n c e r a t e f o r s t e e p e s t - d e s c e n t t y p e a l g o r i t h m s , T h i s r e f l e c t s t h e Newton-Raphson-like nature of RLS.

I n v i e w o f t h e r e c e n t r e s u l t s o f [ 2 ] t h e n e e d f o r p e r s i s t e n t e x c i t a t i o n o f { u k } s h o u l d n o t b e s u r p r i s i n g .

Indeed, i f we do n o t h a v e p e r s i s t e n t e x c i t a t i o n or' { u k l , t h e n it is t r i v i a l t o show t h a t we cannot en-

s u r e c o n v e r g e n c e o f t h e p a r a m e t e r e r r o r t o z e r o , a t a l l .

We a l s o show how an es t imat ion scheme due t o Landau [3] can be modified so t h a t an exponent ia l f o r g e t t i n g f a c t o r i s inco rpora t ed as i n [ 4 ] . An o u t - l i n e o f t h e p r o o f o f e x p o n e n t i a l c o n v e r g e n c e o f t h e modi f ied a lgor i thm i s g iven .

The approach o f t he p re sen t no te is l i m i t e d t o l i n e a r , t i m e - i n v a r i a n t , f i n i t e - d i m e n s i o n a l , s t a b l e p l a n t s which a r e s i n g l e - i n p u t , s i n g l e - o u t p u t (SISO) and dr iven by d e t e r m i n i s t i c i n p u t s . The SISO r e - s t r i c t i o n c a n d o u b t l e s s b e removed wi th an appropr i a t e pa rame te r i sa t ion scheme [SI. E x t e n s i o n s t o s t o c h a s t i c i n p u t s a l o n g t h e l i n e s o f [ 6 ] a l s o seem r e a d i l y p o s s i b l e , a l t h o u g h t e c h n i c a l l y i n v o l v e d . N o n l i n e a r , t ime-vary ing o r in f in i t e -d imens iona l p l an t s can on ly be accommodated t o t h e e x t e n t t h a t t h e r e s u l t s g i v e n h e r e a r e r o b u s t i n t h e p r e s e n c e o f some n o n l i n e a r i t y , some pa rame te r t ime-va r i a t ion and some modes which a r e neglected. See [7-91 f o r d i s c u s s i o n o f t h e e f f e c t s o f d e p a r t u r e f r o m i d e a l i t y i n a d a p t i v e s y s t e m p r o b l e m s . The s i g n i f i c a n c e o f t h i s p a p e r i s t h a t it demonst ra tes t h a t t h e c o n v e r g e n c e r a t e a d v a n t a g e s o f RLS over s t eepes t -descen t a lgo r i thms need no t be abandoned i n o r d e r t o g a i n a d a p t a b i l i t y ,

2 . THE RLS ALGORITHM WITH EXPONENTIAL FORGETTING

The RLS a lgor i thm can be descr ibed by the fo l lowing eaua t ions

and r

o r

(2.3)

where X E [ O , l ] , ?-l i s a p o s i t i v e d e f i n i t e m a t r i x ,

and 8 is a r b i t r a r y . The p a r a m e t e r e s t i m a t e v e c t o r

e can be shown [ l ] t o c o n v e r g e t o t h e v a l u e e which - 0

Page 2: [IEEE 1982 21st IEEE Conference on Decision and Control - Orlando, FL, USA (1982.12.8-1982.12.10)] 1982 21st IEEE Conference on Decision and Control - Exponential convergence of recursive

m i n i m i z e s t h e q u a d r a t i c c r i t e r i o n

where y i s t h e o u t p u t o f t h e p l a n t b e i n g i d e n t i f i e d and + i s the measurable information vector f rom which y i s presumed t o b e f o r m e d ,

I f we now d e f i n e a p a r a m e t e r e r r o r v e c t o r a s

e k = e - 'k ( 2 . 5 )

and r e s t r i c t o u r a t t e n t i o n t o t h e "homogeneous" case [6], where yk i s g iven exac t ly by & 0 , then (2 .3)

takes on the form

which i s a homogeneous, t ime-varying, l inear d i f f e r e n c e e q u a t i o n .

With A = 1 i n t h e above equat ions we have t he s t anda rd RLS a lgor i thm, whi le choos ing A < 1 can be s e e n v i a ( 2 . 4 ) t o b e e f f e c t i n g a n e x p o n e n t i a l w e i g h i n g on t h e p a s t d a t a . S i m i l a r l y , (2 .3) i s an asymptot i - c a l l y s t a b l e r e c u r s i o n o n l y f o r A c 1 s o t h a t i f $ i s p e r s i s t e n t l y e x c i t i n g P-' will remain bounded.

The q u e s t i o n o f t h e o r i g i n o f t h e i measurement vec to r will f o r t h e moment b e l e f t i n a b e y a n c e , t o be a d d r e s s e d l a t e r i n t h e d i s c u s s i o n o f t h e p e r s i s t e n c e o f e x c i t a t i o n c o n d i t i o n a n d i t s i n t e r p r e t a t i o n i n , for example, A R M model l ing .

3. MAIN RESULTS

As a p r e l i m i n a r y t o t h e main r e s u l t we d e f i n e p e r s i s t e n c e o f e x c i t a t i o n a n d t h e n p r o v e t h a t i f i+ 1 i s p e r s i s t e n t l y e x c i t i n g t h e n P-' i s uni formly

bounded above and un i fo rmly pos i t i ve de f in i t e ( i e bounded away f r o m z e r o ) . T h i s l a t t e r p r o p e r t y c a n b e seen to fo l low f rom (2 .3) because , wi th X < 1 , P i 1

i s t h e o u t p u t o f a n e x p o n e n t i a l l y s t a b l e s y s t e m d r i v e n by a p e r s i s t e n t i n p u t ,

D e f i n i t i o n

k

k -1

We s a y t h a t t h e v e c t o r s e q u e n c e I$ j i s p e r s i s -

t e n t l y e x c i t i n g i f f o r some c o n s t a n t i n t e g e r S and a l l j t h e r e e x i s t p o s i t i v e c o n s t a n t s a and E s u c h t h a t

k

j + S O < a I , < c ' j i d i , < B 1 < r n (3.1)

i= j

Lemma 1 : Suppose t h a t I $ 1 i s p e r s i s t e n t l y e x -

c i t i n g t h e n f o r a l l k > S k

e -1 a(T; - 1) 1

s+l I + L A 1 5 'k-1 2 k

s+l I > 0 (3 .2) 1 -A

Proof:

From (2.3) and (3.1) we o b t a i n t h a t

-1

-

Also, from (2.3) i t i s obv ious t ha t

P& 2

SO t h a t f o r j 3 0,

( - + - + . . . + 1) P: AS XS-1 1 1 1

a n d c o n s e q u e n t l y f o r a l l k 2

.l 5 a1 +s-1

: s

The upper bound follows from the solution of

Using (3.1) we conc lude t ha t

( 3 . 4 )

(3.5)

(3 .7)

+ AJ-s[P;J1 + . . . + p - 5 -1

We can now s t a t e our main r e s u l t .

Theorem 1: I f t h e measurement vector sequence I d 1 k i s p e r s i s t e n t l y e x c i t i n g , t h e n t h e RLS es t ima t ion scheme w i th exponen t i a l fo rge t t i ng f ac to r is exponen- t i a l l y s t a b l e .

We choose a Lyapunov func t ion cand ida te as Proof: Consider equat ions ( 2 . 2 ) , (2 .3) and ( 2 . 5 ) .

- T -1 - v k = e P k k-1 'k (3.10)

Then s t r a igh t fo rward a lgeb ra u s ing (2 .3 ) and (2 .1 ) y i e l d s

'k - "k-1 = 'kPk-1"k - 'k-1 k-2 k-1 -T -1 ; -T p-l 6

(3.11)

and t h e r e f o r e

F ina l ly , (2 .3 ) , ( 3 .10 ) and Lemma 1 ( p a r t i c u l a r l y the lower bound o f ( 3 . 3 ) ) i m p l i e s f o r k s

, s+l

A Q . E . D .

The main r e s u l t of Theorem 1 d e p e n d s c r i t i c a l l y o n t h e p e r s i s t e n t e x c i t a t i o n o f ill. 1 . Lemma 1 showed k *

995

Page 3: [IEEE 1982 21st IEEE Conference on Decision and Control - Orlando, FL, USA (1982.12.8-1982.12.10)] 1982 21st IEEE Conference on Decision and Control - Exponential convergence of recursive

t h a t Pk and Pi1 were bounded wi th pers i s ten t ly

e x c i t i n g I9 1 and i f P is unbounded i t i s not poss ib le

t o c o n c l u d e f r o m t h e f a c t t h a t BkPk-lSk +. 0 t h a t B k + 0 . k k - T -1

One c a n a l s o e a s i l y show t h a t if Jik i s bounded,

p e r s i s t e n c y b f e x c i t a t i o n of Jik i s necessa ry fo r P t o k - be bounded. Suppose that 0 < A < 1 1 Pk I / 2 B < m f o r

a l l k. Choose s s o t h a t A B<T. Then from ( 2 . 3 )

-1 S A

Applying the bounds shows that

o r

( 3 . 1 5 )

so t h a t qk i s p e r s i s t e n t l y e x c i t i n g .

ARMA Modelling

Given t h a t Theorem 1 guaran tees exponen t i a l convergence of RLS wi th A < 1 s u b j e c t t o s u f f i c i e n t exc i ta t ion o f the measurement vec tor Jik, we n o t e t h a t t h i s p rope r ty o f $k can be ach ieved i n t he impor t an t ARMA m o d e l l i n g s i t u a t i o n by r e q u i r i n g s t a b i l i t y of t h e p lan t , min imal i ty o f the parameter iza t ion , and pers i s tence o f exc i ta t ion o f the input sequence {ukl o n l y . I n p a r t i c u l a r , i f t h e p l a n t s a t i s f i e s

yk+alyk-l+...+ anyk-n = 8d~k-d+.. .+B m u k-m ( 3 . 1 6 )

and i s s t a b l e , and f u r t h e r

a ( z ) = z +alz +...+a n n-1 ( 3 . 1 7 )

8(z) = Bdz +8d+l~ +.. .+B z ( 3 . 1 8 ) n-d n-d-1 n-m

m

are r e l a t i v e l y p r i m e , t h e n Theorem 2 . 2 of [ 2 ] ensu res pe r s i s t ence o f exc i t a t ion o f $k = u ~ - ~ . . . u l T i f [uk-d.. .u 1 i s p e r s i s t e n t l y e x c i t i n g .

Nonexponential Convergence of RLS wi th X = 1

k-m+l k-m-n

To d e m o n s t r a t e t h a t RLS wi th X = 1 is n o t e x p o n e n t i a l l y c o n v e r g e n t i n s p i t e o f s u f f i c i e n t l y e x c i t i n g $ we cons ider the fo l lowing sca la r example .

Let e and 8 b e s c a l a r s a n d t h e p e r s i s t e n t l y e x c i t i n g scalar Jik = 1 f o r a l l k . Then choosing P =

c we have from ( 2 . 3 )

k A

-1

1 c+k+l Pk = -

and using (Z.l),

Consequently 8 decay a s - which cannot be over 1 K bounded by an exponential and s o does no t converge

e x p o n e n t i a l l y f a s t .

These problems with slow convergence ra te and non- p e r s i s t e n t l y e x c i t i n g i n p u t s are a l s o known t o o c c u r i n Kalman f i l t e r i n g w i t h o u t e x p o n e n t i a l d a t a w e i g h t i n g . A d i s c u s s i o n a l o n g t h e s e l i n e s c o n c e r n e d w i t h d a t a s a t u r a t i o n and f i l t e r d i v e r g e n c e is g i v e n i n [lo].

4 . THE "CLASS B" ALGORITHM OF LANDAU WITH EXPONENTIAL WEIGHTING

I n [ 3 ] Landau d e t a i l s a "Class B" algorithm which i s very much l i k e t h e RLS a lgo r i thm o f t h i s pape r . The equa t ions wh ich desc r ibe t h i s a lgo r i thm a re :

1 p k-1 JI k-1 J I T P k-1 k-1 P k = P - - k-1 X 1 T '+?k-lpk-1'k-l

and

T A n

i= 1 vo = k yk - "-lek-, + d iEk- i ( 4 . 3 )

'k = Yk - $k-Iek T A

( 4 . 4 )

where

= [Zk-l"zk-n Uk-l*''Uk-m] ( 4 . 5 )

and Zk i s t h e o u t p u t of t h e " p a r a l l e l " e s t i m a t i o n model ( s e e [ 3 ] , pp. 164-167 and [ 2 ] , S e c t i o n 3 ) .

Prov ided t ha t

1 + diz-'

1 + z aiz-i

h (z 1 = -1 i=l 1

2A - - ( 4 - 6 )

i= 1

is a s t r i c t l y p o s i t i v e r e a l t r a n s f e r f u n c t i o n , t h e a lo r i thm conve rges i n t he s ense t ha t t he pa rame te r e r r o r g o e s t o z e r o . The extra complexity of t h e a lgo r i thm i s o f f s e t by a tendency for measurement noise t o c a u s e less damage t h a n f o r a n o r d i n a r y least squa res a lgo r i thm.

The s i m i l a r i t y of ( 4 . 1 ) - ( 4 . 5 ) t o ( 2 . 1 ) - ( 2 . 3 ) is o b v i o u s . I n f a c t t h e o n l y s t r i k i n g d i f f e r e n c e s are t h a t t h e Landau a lgo r i thm assumes un i ty t ime-de lay i n t h e p l a n t ( i . e . d = 1); t h e r e i s a n a d d i t i o n a l t e r m a p p e a r i n g i n ( 4 . 3 ) which does not appear in ( 2 . 1 ) ; "pa ra l l e l " e s t ima t ion mode l ou tpu t s , z ~ - ~ , a p p e a r i n

( 4 . 3 ) whereas ARMA p l a n t o u t p u t s y a p p e a r i n t h e

d e f i n i t i o n o f $ i n S e c t i o n 3 of th i s pape r and ,

f i n a l l y , ( 4 . 2 ) does no t i nco rpora t e an exponen t i a l f o r g e t t i n g f a c t o r .

k-i '

k-d

Landau h a s p r o v e d t h e a s y m p t o t i c s t a b i l i t y o f t h e c l a s s B a l g o r i t h m u s i n g h y p e r s t a b i l i t y a r g u m e n t s . U n f o r t u n a t e l y o n e c a n n o t p r o v e e x p o n e n t i a l s t a b i l i t y o f t h e a l g o r i t h m a s i t s t ands . Th i s is e s s e n t i a l l y b e c a u s e the matr ix sequence {Pi ; l} cannot be guaranteed to be bounded. I n p a r t i c u l a r i f {$k-l} i s p e r s i s t e n t l y

e x c i t i n g t h e n use o f t h e m a t r i x I n v e r s i o n Lemma and ( l r . 2 ) shows tha t P- l+- . This p roblem may be c i rcum-

v e n t e d i f w e u s e a s l i g h t m o d i f i c a t i o n o f t h e class B a lgo r i thm [ 4 ] , which amounts to introduct ion of e x p o n e n t i a l f o r g e t t i n g .

k

Use m

P k = - 1 [I - - 1 pk-l'k-l'i-l X 1 T 'k-1

a?k-lPk-lJik-l < a <

( 4 . 7 )

996

Page 4: [IEEE 1982 21st IEEE Conference on Decision and Control - Orlando, FL, USA (1982.12.8-1982.12.10)] 1982 21st IEEE Conference on Decision and Control - Exponential convergence of recursive

i n p l a c e of (4 .2 ) . By u s i n g s imi l a r a rgumen t s t o ' those i n [2] ( see Sec t ion 3 of [2] ) i t i s then possible t o p r o v e t h e f o l l o w i n g r e s u l t . Due t o s p a c e l i m i t a t - ions on ly an ou t l ine o f the p roof is g iven .

Theorem 2 : Suppose that

( a ) t h e p l a n t is desc r ibed by (3.16) (with d = 1) and a(z-1) i s s t a b l e ,

( b ) a(2- l ) and B(2- l ) (wi th d = l), of (3 .17) , ( 3 . 1 8 ) , a r e r e l a t i v e l y p r i m e ,

( c ) {uk l i s a p e r s i s t e n t l y e x c i t i n g i n p u t sequence

-1 (d) h(z ) i n ( 4 . 6 ) i s s t r i c t l y p o s i t i v e r e a l .

t hen t he mod i f i ed c l a s s B a lgor i thm of ( 4 . 1 ) , (4 .7) and (4 .3)-(4.5) is e x p o n e n t i a l l y s t a b l e .

( a ) A Lyapunov f u n c t i o n c a n d i d a t e i s chosen as

where

(b) Fol lowing [2] we s e t - u p s t a t e v a r i a b l e equa t ions equ iva len t t o <4.1), (4 .7) and (4 .3 ) - (4 .4 ) .

( c ) Making use of t h e Kalman-Yakubovic l e m a equa t ions [ Z ] we o b t a i n an e x p r e s s i o n f o r V[ek+l ' 'k+11 - 'Ie k , 8,]. Th i s quan t i ty i s monotone dec reas ing un le s s ek = 0 ( and , i n ca se P i s bounded, -1

k-1 3k = 0 ) .

(d) We u s e a r g u m e n t s l i k e t h o s e i n t h e p r o o f of Lema 1 t o show t h a t i f { u 1 i s p e r s i s t e n t l y e x c i t i n g k then j : k-1 = [yk-l ' * ' yk-nuk-n * " ] i s p e r s i s t -

e n t l y e x c i t i n g .

T k-m

(e) From t h e a s y m p t o t i c s t a b i l i t y r e s u l t s of [4] we conc lude t ha t zk + yk. Lemma 3.2 of [2] then allows

u s t o conc lude t ha t {$k-l] is a l s o p e r s i s t e n t l y e x c i t i n g .

( f ) The Xa t r ix Inve r s ion Lema and (4 .7 ) g ives

lie conc lude , t hen , u s ing a rgumen t s a s i n t he proof of Lemmal tha t i f ;$,-,3 i s p e r s i s t e n t l y e x c i t i n g t h e n

{Pk I i s a bounded matrix sequence.

u s t o c l a im tha t t he mod i f i ed c l a s s B a lgor i thm of Landau i s e x p o n e n t i a l l y s t a b l e .

-1.

( 9 ) F i n a l l y , ( c ) , ( f ) and the form of ( 4 . 8 ) a l lows

5 . ~~o~cLus Ioss The main r e s u l t d e m o n s t r a t e s t h a t t h e RLS es t im-

a t i o n scheme w i t h e x p o n e n t i a l f o r g e t t i n g f a c t o r i s e x p o n e n t i a l l y s t a b l e p r o v i d e d t h a t + k i s p e r s i s t e n t l y e x c i t i n g . Wi th bounded input s igna ls the converse i s a l s o t r u e . T h i s a l g o r i t h m when used with fiK4 models i s sometimes called a " s e r i e s - p a r a l l e l ' ' a l g o r i t h m [ 3 ] ; we h a v e a l s o o u t l i n e d a n e x p o n e n t i a l s t a b i l i t y r e s u l t f o r t h e c o r r e s p o n d i n g " p a r a l l e l " RLS scheme, with e x p o n e n t i a l f o r g e t t i n g .

These r e s u l t s a r e p o t e n t i a l l y v e r y i m p o r t a n t i n the cases of t ime-varying systems or under-modelling [7-91 , and imply robus tness o f the a lgor i thms in the presence of measurement noise. A l s o , t h e r e s u l t s

complement those known fo r t he s tochas t i c app rox ima t io* a lgo r i thm [ Z ] ,

..

REFERENCES

G . C . Goodwin and R.L . Payne, Dynamic System Ident i f icat ion: Experiment Design and Analysis , Academic P res s , 1977 .

B . D . O . Anderson and C . R . Johnson J r . , Exponent ia l conve rgence o f adap t ive i den t i f i ca t ion and c o n t r o l a lgo r i thms , Automctica (1981) to appea r . A l so IFAC 8 t h T r i e n n i a l World Congress, Kyoto, Japan, August 1981.

I . D . Landau, .&?upzitre Control (Dekker, New York, 1979) .

i . D . Landau and H.M. S i l v e i r a , A s t a b i l i t y theorem w i t h a p p l i c a t i o n s t o a d a p t i v e c o n t r o l , -?roc. i75h ---- L A f i S CDC COR:, (1979) 1074-1079.

R . Schumann, I d e n t i f i c a t i o n and adap t ive con t ro l of m u l t i v a r i a b l e s t o c h a s t i c l i n e a r s y s t e m s , 5 t h i E A C S z r p c s i w sn Ide.<ti.=<:Catisr, a ~ d &rme$er : s ; ~ ~ t i ~ n (1979).

R . R . Bitmead and B.D.O. Anderson, Lyapunov t e c h n i q u e s f o r t h e e x p o n e n t i a l s t a b i l i t y of l i n e a r d i f f e r e n c e e q u a t i o n s w i t h random c o - e f f i c i e n t s , -is: :?an~. cn Auto. Ccntr. AC-25 (1980) 782-787. B . D . O . Anderson and R.X. Johnstone, Robust Lyapunov s t a b i l i t y r e s u l t s and adap t ive sys t ems , X E F C3C -mJ+., (1981) to appear .

B . D . O . Anderson and C . R . Johnson, J r . , On reduced o r d e r a d a p t i v e o u t p u t e r r o r i d e n t i f i c a t i o n and a d a p t i v e I I R f i l t e r i n g , IEEE l?anS. on A U Z O . COntZ', AC-27 (19821, t o a p p e a r .

A . Weiss and D . Y i t r a , D i g i t a l a d a p t i v e f i l t e r s : Cond i t ions fo r conve rgence , r a t e s of convergence, e f f e c t s of n o i s e and e r r o r s a r i s i n g f r o m t h e implementation, IEZE Trans. OK Iv..%. Ynecrd

- .

----

n -

IT-25, (1979) 637-652.

B . D . O . Anderson and J . B . Moore, Opt imal Fi l ter ing, Prent ice-Hal l , 1979.

997