young1980.pdf

download young1980.pdf

of 25

Transcript of young1980.pdf

  • 8/18/2019 young1980.pdf

    1/25

    This article was downloaded by: [Ohio State University Libraries]On: 19 June 2012, At: 05:18Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House37-41 Mortimer Street, London W1T 3JH, UK

    International Journal of ControlPublication details, including instructions for authors and subscription information:

    http://www.tandfonline.com/loi/tcon20

    Refined instrumental variable methods of recursivetime-series analysis Part III. ExtensionsPETER YOUNG

    a & ANTHONY JAKEMAN

    a

    a Centre for Resource and Environmental Studies, Australian National University, Canberra

    Australia

    Available online: 21 May 2007

    To cite this article: PETER YOUNG & ANTHONY JAKEMAN (1980): Refined instrumental variable methods of recursive time-

    series analysis Part III. Extensions, International Journal of Control, 31:4, 741-764

    To link to this article: http://dx.doi.org/10.1080/00207178008961080

    PLEASE SCROLL DOWN FOR ARTICLE

    Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

    This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form toanyone is expressly forbidden.

    The publisher does not give any warranty express or implied or make any representation that the contentswill be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses shouldbe independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly inconnection with or arising out of the use of this material.

    http://dx.doi.org/10.1080/00207178008961080http://www.tandfonline.com/page/terms-and-conditionshttp://dx.doi.org/10.1080/00207178008961080http://www.tandfonline.com/loi/tcon20

  • 8/18/2019 young1980.pdf

    2/25

    Refined instrumental variable methods of recursive

    time-series n lysis

    Part 111. Extensions

    PETER YOUNGtS and ANTHONY JAKEMANt

    This is th e final paper in a series of t hre e which have been concerned w i t h th e compre-

    hensive evalua tion of the refined instrumental variab le

    I V )

    method of recursive

    time-series analysis. The paper shows how the refined

    IV

    procedure can be extended

    in various important directions and how it can provide the basis for the synthesis of

    optim al generalized equat ion error GE E) algor ithms for a wide class of etochastic

    dynamic systems. The topics discussed include the estimation of param eters in

    continuous-time differential equation models from continuous o r discrete d at a

    the estimation of time-variable parameters in continuous o r discret e-tim e models of

    dynamic systems the design of stochestic state reconstruction Wiener-Kalmen)

    filters direct from da ta the estimation of parametere in multi-input, single ou tp ut

    MISO)

    transfer function models the design

    of

    simple stochastic approximation

    SA)

    implementatio ns of t he refined

    I V

    algorithms and the use of the recursive algorithms

    in self-adaptive self tuning) control.

    1 ntroduction

    I n the first two parts of this paper Young and Jake man

    1979

    a, Jakeman

    and Young 1979 a ) we have been concerned with th e description an d compre-

    hensive evaluation of th e refined instrumental variable

    I V )

    approach to

    time-series analysis for single input, single output SISO) and multivariable

    dynamic systems described by discrete-time series models. In

    this, the third

    and final part of th e paper, we consider how the refined I V method can be

    extended in various directions to handle continuous time-series models and

    discrete or continuous time-series models with time-variable parameters. We

    also discuss briefly other extensions including off-line and on-line adaptive

    methods of designing stat e reconstruction Kalm an) filters for stochastic

    systems

    ;

    th e development of

    IV

    estimation procedures for specific time-series

    models, such as the multiple input-single ou tput transfer function model an d

    finally the estimation of parameters in multivariable system models in those

    situations where the observation space is less tha n the dimension of th e model

    space. Fo r convenience, in a11 cases except the latt er , we shall consider refined

    I V estimation algorithms with non-symmetric matrix gains. Bearing in mind

    th e results of th e first two parts of the paper, however, it is clear tha t symmetr ic

    matrix gain alternatives could be implemented and i t is likely th at, a t least for

    reasonable sample size, they would perform in a similar manner.

    Received 10 August, 1979

    t Centre for Resource and Environmental Studies, Australian National University,

    Canberra, Australia.

    1Currently Visiting Professor, Control and Management Systems Division,

    Engineering Department, University

    of

    Cambridge

    002&7179/80/3104 0741 02.00 1980 Taylor Francls Ltd

  • 8/18/2019 young1980.pdf

    3/25

    742

    P Young and

    A

    akeman

    2

    Continuous time dynamic systems described by ordinary differential equations

    models

    The refined

    IV

    procedure can be applied to both SISO and multivariable

    continuous time-series models bu t, for simplicity of exposition, we will describe

    here only the SISO implementation. The extension of the SISO procedures

    to the multivariable situation is, however, quite obvious by analogy with the

    discrete-time case discussed in Part

    I1

    of the paper (Jakeman an d Young

    1979 a) . Using nomenclature similar to tha t used previously, the continuous-

    time SISO model is illustrated in block

    M

    (within dotted lines) of Fig. and

    can be written as

    (ii)

    (iii

    YV

    = x t )

    f

    1)

    where

    s

    is the differential operator,

    i

    .e. sx(t

    =

    dx t /dl (Ioosel

    y

    interpreted here

    as the Laplace operator)

    ; A B,

    C and

    D

    are polynomials in s of the following

    .

    form,

    and

    [ t)

    is a continuous-time white noise process.

    It is well known (e.g.

    AstrGm 1970, Jazwinski 1970) th at theoretical and analytical difficul ties

    arise because of the use of continuous white noise in mathematical models of

    dynamic systems, particularly transfer function formulations such as eqn.

    1 ) (ii). In the present context , this difficulty is manifested in the form of

    practical problems associated with the recursive estimation of t he parameters in

    the

    C

    and

    D

    polynomials characterizing the noise model. For the moment ,

    however, i t is convenient to assume a continuous-time model of the form

    1 (ii)

    although, as we shall see, i t is necessary in practice to evaluate t he noise-

    components of the model in discrete-time

    in

    order to circumvent estimation

    problems.

    2.1. Discrete and continuous time recursive

    algorithms

    It

    is clear that the model

    (1)

    is algebraically equivalent to the discrete-time

    SISO model discussed in previous parts of this paper.

    Let us consider,

    therefore, the situation whcre we wish to implement the estimation algorithm

    in discrete-time using sampled da ta from the continuous-time system we will

    refer to this as CD (continuous-discrete) analysis (Young 1979 a). Using an

    approach similar to th at used in previous parts of the paper, it is then possible

    to obtain estimates of t he parameters in the continuous-time model polynomials

  • 8/18/2019 young1980.pdf

    4/25

    Refined instrumental variable

    methods

    of recursive time-se ries an aly sis 743

    A

    and B by minimizing a least squares cost function of the form

    Here

    y

    =

    [yo yo

    yOTlTand

    u

    =

    [u

    uo2

    ,

    u O T I T

    where the first zero

    subscript on u and y indicates that the variables are, respectively, the basic

    input and out put variables i.e. the

    '

    zeroth

    '

    derivatives of u and y ), while the

    second subscript

    i

    = l

    2

    , T denotes the sampled values of the variables

    at

    time ti i.e. y t,) and

    u t,).

    Figure

    1.

    Refined IV

    algorithm for continuous-time systems.

    'noise

    M

    :

    w .

    I

    ~ s i

    ,

    ~ t )

    ,

    YO*

    t

    -

    r.5thte.--

    I

    1) REFIND

    reconstruct~oni

    I A s I

    o r

    1

    - - tate

    I

    fi lter

    Now, by direct analogy with the analysis in the DD discrete-discrete) case

    of Part I, the recursive estimate d of the unknown parameter vector

    a

    [a a ..

    ,

    a b b , , ..., b - IT can be obtained from th e following discrete-time

    algorithm,

    i a = a,-,

    P,-,ak*[sz

    +z , * ~ k - l ~ k* ] -l ~ k *T,-, Ok*)

    or

    ii) a

    =

    -P,lt,*{~,*~

    -

    Ok*)

    and

    iii) P, =

    P,-,- Pk-,g,*[e2 + z ~ * ~k-lltk*]-l~k*TPk-l

    u(t)

    where

    Here 6 is an estimate of the variance of

    e t ) ; i k

    s the outpu t of an adaptive

    ' auxiliary model ' as shown in Fig. 1 and t he st ar superscript indicates that

    z

    ariable

    ...

    I

    i

    ~ l t e r s

    * ,

    a c

    kit)

    ~ S I -

    recurswe

    or

    u*.dt)

    -

    a u x ~ lary

    model

    iteraiive update

    ,

    I

    v

    C sl

    - GoRITw

    [s)A s)

    a

    k i t )

  • 8/18/2019 young1980.pdf

    5/25

    744

    P. Young and

    A .

    Jakeman

    the Jariables are filtered by adaptive prefilters

    ' C B A

    again as shown in

    Fig. 1. The i =

    1, 2 ,

    n subscript on the variables within th e square brackets

    in

    5)

    denotes th e ith time-derivative of the variabIe, while the

    k

    subscript

    outside th e brackets indicates th at the enclosed variables are all sampled a t th e

    kth sampling instant.

    .

    This algorithm has close similarity with the

    I V

    algorithm suggested some

    years ago by Young 1969), h e only difference lies in the na ture of the prefilters ;

    in the previous algorithm, these were termed ' sta te variable filters an d were

    introduced mainly to avoid direct differentiation of noisy signals. I n th is

    sense, the function of the present filters is identical

    :

    their presence means that

    it is not the direct derivatives of the variables

    y l ) , i t )

    and

    u t )

    hat are required

    for estimation b ut t he derivatives of the filtered variables y* t), j . * t ) and u* t).

    And these filtered derivatives, unlike th e direct derivatives, are physically

    realizable s a product of th e filtering operation Young 1964, 1969). Of

    course the prefilters here do more than just avoid differentiat ion of noisy

    signals ; they also represent the mechanism for inducing asymptot ic statisti cal

    efficiency.

    In the present case, the ' optimal prefilters are defined in terms of estimates

    of t he a

    priori

    unknown polynomials

    A ,

    C and

    D. t

    is necessary, therefore,

    to define some adaptive procedure for synthesizing the prefilters as t he estima-

    tion proceeds. In the situation where C .0, i.e.

    t )

    is white noise, the

    adaptive synthesis of the prefilters 1 A is fairly straightforward

    :

    both the

    prefilter and auxiliary model parameters can be updated either recursively or

    itera tively as shown in Fig. 1, exactly as .in the discrete-time model case

    described in Part I of this paper. When th e noise

    [ t )

    is coloured i.e. C 1.0

    and/or

    D

    1.0), however, the situation is no t so straightforward : in contrast to

    the discrete-time model situation, i t is not easy to construct a similarly moti-

    vated recursive estimator for the continuous-time noise model parameters since

    the derivatives of the white noise e t ) do not exist in theory.

    While i t may be possible to solve this noise estimation problem by con-

    sidering either band-limited noise or purely autoregressive noise where

    derivatives of

    e t )

    do not occur), we feel that it may be better to consider a

    hybrid approach. Here, the noise is estimated in purely discrete-time DD)

    terms by the use of the

    A M L

    or refined

    A M L

    algorithms described previously.

    This does not create any implementation problems because the noise model is

    only required for adaptive prefiltering operations, which can easily be carried

    out in discrete-time when using D analysis. The general implementat ion

    in this case is shown in Fig. 2 a) and the detailed structure of the deriva tive

    generating filters

    l / A s )

    is illustrated in

    Fig.

    2 b).

    t

    should be noted here

    th at t he filter in Fig. 2 b ) s similar to the ' stat e variable filter suggested by

    Kohr see, e.g. Kohr and Hoberock 1966) : the only difference is that the

    coefficients ti i = 1, 2, n are not constant, as in the Kohr case, but are

    aduplively

    adjusted, either iteratively or recursively, as the estimation proceeds.

    U p to this point we have assumed that, while the algorithm

    4)

    is imple-

    mentcd in discrete-time, the signals y t),4 t)and

    u t )

    are available in continuous-

    time form so that they can be passed through the continuous-time prefilters

    prior to sampling. In practice, however, it could well be th at both in put an d

    outp ut signals are naturally in sampled data form. This difficulty can be

    circumvented, albeit in an approximate manner, by assuming th at t he signals

  • 8/18/2019 young1980.pdf

    6/25

    Refined instrumental variable methods

    of

    recursive time series analysis

    7 4 5

    remain constant over the sampling interval and passing them directly into the

    continuous-time filters. In other words, the sampled dat a ar e converted to a

    continuous time staircase form prior to filtering. I n this manner, the

    prefilters perform an additional, useful, interpolation role and provide

    estimates of the continuous-time filtered variables.

    and I

    and

    I hold I

    I hold i

    x

    J i. .Y J

    a1

    Figure

    2.

    Refined I V algorithm

    : D

    implementation.

    a) verall implementation

    X

    closed : continuous-time data available

    ;

    X operative

    :

    discrete data only

    available ; b ) state variable filter l / A s ) applied to u t) .

    r - - . -

    Of course th e estimates of the filtered variables emerging from the prefilters

    are in no sense optimal and th e efficacy of this approach is clearly dependent

    upon the sampling period T the approximation will be good for small T

    and will become progressively worse as T s lengthened. Fortunately the

    estimation results do not appear particula.rly sensitive to the choice of T nd

    acceptable performance can be obtained from qui te coarse sampling frequencies.

    I

    v F

    (see block

    above 1n a1)

    I

    u:(t)

    I

    uE-, t )

    I

    u:-l t)

    I :

    I

    : 4u; ( t )

    I

    u. ( t )

    I

    I

    I

    I

    m > I

    =-

    I

    I

    _

    - .

    (b

  • 8/18/2019 young1980.pdf

    7/25

    746

    P Young

    and

    A

    Jakeman

    Finally, i t is worth noting tha t, if continuous-time measurements of y(t ) an d

    u(t)are available, then it is possible to consider a continuous-time implementa-

    tion.of the estimation algorithm itself (i.e. using CC analysis). This is a logical

    development of early continuous-time gradient procedures for estimat ing

    dynamic system parameters (see, e.g. Young 1965 a, Levadi 1964, Ka ya an d

    Yamamura 1962, Young 1976). The most obvious impleinentation would be

    an estimation algorithm of the form

    (i)

    (ii)

    which is a continuous-time equivalent of the discrete-time recursive algorithm.

    Algorithms of thi s form are also discussed by Solo (1978).

    Note that it would be difficult to implement the estimation algorithm (6)

    for other than

    D

    1.0 because of the difficulty in estimating th e an d D

    polynomials (unless we once again consider some hybrid mechanization which

    would be rather impractical). Thus when f(t ) s not white noise, th e estimates

    produced by the algorithm will not have any optimal properties. They will,

    however,. be consistent, asymptotically unbiased and, on the basis of previous

    experience, they should be reasonably efficient (see Jakeman 1979). Note also

    th at we can reduce the computational complexity furthe r by replacing (t) in

    (6) by a simpler stochastic approximation

    SA)

    gain (e.g. Young 1976). Thi s

    would be a continuous-time equivalent of t he SA algorithms discussed in

    7

    2.2. Experimental results

    The CD approach to t he continuous-time model estimation discussed in t he

    previous section has been evaluated by Monte Ca;rlo simulation analysis applied

    to two systems described by second order differential equations.

    In

    the first

    case, the sys tem was of th e form

    with u (t ) chosen as a random binary signal with levels plus and minus 1.0.

    In the second, the system was modified to

    with K 0-781, w 1-6,

    0.5

    and u(t) chosen as the following combination

    of three sinusoidal signals,

    u(t ) =sin (0.5wdt) sin (w,t) +sin (l.5wdt)

    where w is the damped natura l frequency of the system i.e. w o,Z/(1-

    c2 .

    In both of these examples, the noise l ( t) was simulated white noise adju ste d

    to give several different signal/noise ratios S (defined as in Young and Ja keman

    1979 a) .

  • 8/18/2019 young1980.pdf

    8/25

    Refined instrumental variable methods. of recursive time-series analysis

    747

    Tables 1

    a )

    nd 1 b )are typical of the results obtained during the analysis.

    For each sample size, column represents the average parameter value over 10

    experiments, while columns and 3 represent standard deviation from th e true

    and average parameter value, respectively. I n Table 1 , th e sampling interval

    T,

    is chosen to be quite rapid a t

    0.1

    sec which represents

    1/31.4

    of th e Shannon

    maximum sampling period, P ,

    =

    n /w , In Table 1 b) , wo different sampling

    intervals are compared one fairly coarse

    P, /8) ,

    he other quite small

    P, /40)

    there seems to be some bias on a,

    in

    the coarse sampling situation.

    Number

    o

    samples

    Parameter True

    value

    100 500

    Table

    a). S=10, T,=0.1.

    Sampling rate

    Parameter True

    value PJ40 Pel8

    Table b ) .

    S= 10, 500 samples.

    Other simulations have tended to confirm that quite coarse sampling

    intervals can be tolerated but suggest that the degree of bias is, no t surprisingly,

    a function of the system dynamic characteristics and the value of S As a

    result, the algorithm should always be used with great care if th e sampling ra te

    is low and, as

    a

    rule of thumb, sampling intervals should always be chosen less

    than P,/10. But there is clearly a need for more research on this topic before

    the algorithm can be used with confidence with coarsely sampled data.

    The algorithm 4) has also been applied with, some success both to multi-

    variable systems Jakeman

    1979

    and to real data. Typical of the latt er are

    the results shown in Table c) and Fig.

    3.

    Table c)compares the estimates

    obtained when carrying out

    CD

    and DD time-series analysis on da ta obtained

    during fluorescence decay experiments on 1-naphthol Jakeman e t al 1978).

    I t is clear th a t the continuous-time and discrete-time models have virtually

    identical dynamic characteristics in this case, where the sampling period was

    short in relation to P , approximately P,/113).

  • 8/18/2019 young1980.pdf

    9/25

    P .

    Young nnd

    1. akernan

    6 dye

    tr cer

    data

    model a t ) = b , l ~ ( t - ~ l

    +

    b,stGs

    odel

    vtput

    .

    C t )

    time

    hours)

    Figure 3 .

    Results from model of dye tracer concentration in Murrumbidgee River,

    Australia.

    Dynamic characteristics

    Parameter

    Model estimates Time constant Steady state.

    (nsec) gain

    Discrete time T 0.212 nsec a

    =

    .9724 7.575 1.0

    bo= 0.0276

    Continuous time; time unit = 0- 212 a

    =

    35.9622 7.624 1-0

    nsec bo= 1.00

    Table c).

    Figure 3 shows the observed and estimated dye concentration in a river,

    where the estimated concentration is generated by

    a

    second order differential

    equation model estimated using algorithm 4. The data used in this exercise

    wcre collected during dye tracer experiments carried out on t he Murrumbidgee

    River system in Australia (Whitehead et al 1978). Here, as can be seen from

    Fig.

    3 ,

    i t was not possible to maintain a completely regular sampling interval

    but T s approximately

    half

    an hour P , / 3 0 ) . This demonstrates how data

    with irregular sampling intervals can be used, provided the longest sampling

    interval does not lead to serious interpolation errors and estimation bias.

    3

    Syst em s described by time variable parameter models

    The idea of modifying recursive algorithms to allow for the estim ation of

    time-variable model parameters has been exploited many times since th e

    publication of R. E. Kalman s seminal papers on st ate variable filter-estimation

    theory in the early nineteen sixties (Kalman

    1960,

    Kalman and Bucy

    1 9 6 1 ) .

    In the case of V algorithms, thi s particular extension has so far been heuristic

    (see Young 1 9 6 9 ) . Bu t now, with the advent of the refined V algorithm, i t is

    possible

    to

    pu t such modifications on a sounder theoretical base and t o const ruct

    algorithms which have greater practical potential.

    There are a num ber of ways in which th e time variable parameter modifica-

    tions can be introduced. The most straightforward is simply to take no te of

  • 8/18/2019 young1980.pdf

    10/25

    Refined instrumental variable nzetho s of recursive t ime series ana lysis

    749

    the relationship between the refined I V algorithm and the Kalman estimation

    algorithms and introduce additional a priori information in the form of a

    stochastic model for the parameter variations. The general form of this model

    is the following discrete-time, Gauss-Markov model,

    Here

    @

    and I are assumed

    known

    and possibly time variable matrices, while

    qk is a discrete white noise vector with zero mean .and covariance matrix Q

    which is independent of the observational white noise source e i.e.

    where S,, is th e Kronecker delta function.

    This device is now well known in the recursive parameter estimation litera-

    ture and it is straightforward to show (see, e.g. Young 1974) how the simple

    recursive linear least squares regression equation can be modified in the light

    of the additional a priori information inherent in (7) to include additional

    prediction equations which allow for the update between samples of both the

    parameter estima tes and the covariance matrix of th e parametric estimation

    errors,

    P .

    Unlike the basic

    I V

    algorithm, th e refined I V algorithm would appear to

    have certain optimal properties. In particular, the theoretical results of Pierce

    (1972), together with the stochastic simulation results reported in Pa rt s and

    I1

    of this paper, have shown th at the pk matrix generated

    by

    the refined

    algorithm provides a good empirical estimate of t he covariance matr ix

    P of

    the estimation errors, where

    P

    =

    E

    piT

    and

    B =

    a

    A It

    is possible, therefore, to employ the same approach used to

    modify the recursive linear regression equations to similarly modify the refined

    I V

    algorithms. The resulting algorithmt takes the following prediction-

    correction form (see, e.g. Young

    1974,

    p.

    214)

    i )

    klkvl

    (ii) Pklk-

    ~ f i - @ ~

    r Q F T

    ii

    a k ~ ~ ~ ~ - ~ - p ~ ~ ~ - ~ ~ ~ * [ 6 ~z ~ ~klk-l~k*] -l

    correction

    on receipt { ~ k * ~k~k-l-~k*)

    ( i ~ ) i PkIk-l- kIk-I~k*[@ bkIk-I~k*]-l

    sample

    ~ k * ~

    klk-l

    Equations 8) (i) to (iv) constitute the refined I V algorithm for estimating

    stochastically variable parameters in a discrete time-series model of

    a

    S SO

    t t will be noted tha t the derivation of this algorithm is made a little more

    obvious if the symmetric gain matrix form of the refined IV algorithm is utilized.

    ? from this algorithm is a somewhat closer approximation to P than

    f ,

    n the non-

    ,

    symmetric gain

    c se

    (see Young and Jakeman 1979 a).

  • 8/18/2019 young1980.pdf

    11/25

    750 P

    Young and

    A

    Jakeman

    system.

    At first sight, this algorithm appears somewhat restrictive since i t

    requires

    a priori

    knowledge of th e matrices nd

    I

    in the stochastic model of

    th e parameter variations. Bu t pas t experience with similar algori thms e.g.

    Young 1969, Norton 1975) has indicated th at t he assumption is no t as limiting

    as i t might appear. First , it is possible to consider a class of simple random

    walk models which represent special cases of 7) with very simple and

    I

    matrices and which seem to offer some considerable practical potential.

    Second,

    priori

    information on the natur e of parameter variations can some-

    times be utilized to arrive

    at

    simple Gauss-Markov models.

    In the first case, the three random walk models that have proven most

    useful in practice are as follows :

    The pure random walk

    RW)

    The smoothed random walk SRW)

    where a is a constant scalar with 0

    < a <

    1.0.

    The

    integrated random walk

    IRW)

    Th e R W model 9) was first used in the early nineteen sixties see, e.g. Kop p

    and Orford

    1963,

    Lee

    1964).

    The

    IRW

    model 11) was suggested in t he

    parameter estimation con text by Norton 1975) who has used it successfully

    in a

    number of practical applications. The SRW model 10) is of more rec ent

    origin Young and .Kaldor

    1978

    and seems to provide .a good compromise

    between models 9) and

    l l ) ,

    lthough it requires the specification of one addi-

    tional parameter, th e smoothing constant u

    =

    1/ ~ , , here i- is the approximate

    exponential smoothing constant in sampling intervalst.

    All of the models 9) to

    1

    1) are non-stationary in a statistical sense and so

    they allow for wide variation in the parameters. Their different characteris tics

    are described fully by Norton 1975) and Young and Kaldor 1978). P u t

    simply, the progression from model 9) through 10) to 11)allows for grea ter

    overall variation in the estimated parameters for any specified covariance

    matrix Q accompanied by greater smoothing of the short -term variations.

    Jn th e case where more general and

    I

    matrices are considered it may

    often be possible to a ss l~m e ha t, for physical reasons, the variations in th e

    parameter are correlated with the variations in other

    measured

    variables

    affecting th e system. For example, the parameters in an aerospace vehicle are

    known t o be functions of variables such as dynamic pressure, Mach number,

    altitude, etc. Young 1979 b) . Or again, th e numerator coefficients in a

    t

    Strictly, the time constant, T = -T,/log,

    1

    -a time units, where

    T,

    s the

    sampling interval in time units.

  • 8/18/2019 young1980.pdf

    12/25

    Refine d in strumen tal variable methods of recursive time -ser ies an aly sis 751

    transfer function model between rainfall and runoff flow in hydrological systems

    are known to be functions of soil moisture and evapo-transpiration (Young 1975,

    Whitehead and Young i975).

    I n such examples, t is often possible to define a in th e following form

    where

    T k

    s a matrix (often diagonal) of t he relevant mea suredvari able s ;

    a,*

    is a vector of residual parameter variations which,

    if

    the T ransformation is

    effective, will be only slowly variable and can be described, for example,. by

    one of th e random walk models. I n the case of the RW (9), i.e.

    ak*= akPl*

    Tk -1

    (13)

    we see th at , upon subs titution from 13) into

    12),

    the variations

    in

    ak are given

    by a Gauss-Markov m odel such as (7 ) with

    @

    =0

    TkTk-,-I

    and r

    = I?,= Tk .

    t is clearly possible, therefore, t o utilize the refined I V algorithm

    (8)

    with

    0

    and

    Q

    in the prediction eqns. (8) (i) and (ii) defined accordingly.

    Such an

    approach has been used previously with other recursive algorithms by Young

    (1969, 1979 b).

    The implementation of the algorithm defined by eqns. 8 (i) t o (iv) offers

    several problems. I n particular, the equations imply th e parallel implementa-

    tion of t he refined AML algorithm and its interactive use with th e refined I V

    algorithm, as described by Young an d Jakeman (1979 a) . This introduces

    considerable complexity and, for the present paper, we have once again

    implemented only the special case where

    El

    is white noise, i.e. C(z-1) = D z-1) =

    1.0. Here, the full refined AML is not required and the prefilters (nominally

    e/Ab are defined

    s 1/A^

    This simpler algorithm works very well and

    seems to give good results even if 5 is coloured noise. Moreover, the algori thm

    in this form

    has

    also been modified furthe r to allow for an off-line smoothing

    solution in which the recursive estimate a t any sampling ins tan t k is a condi-

    tional estimate ti based on t he whole da ta se t of

    N

    sampIes. Th e smoothing

    algori thm is an extension of Norton s work (Norton 1975) within a n I V context

    and it requires both forward and backward recursive processing of t he dat a

    (Young and Kaldor 1978, Kaldor 1978, Gelb 1974).

    Finally, it should be remarked that the above approach to time variable

    parameter estimation can be extended straightforwardly both to the multi-

    variable and continuous-time situations. Such extensions are fairly obvious

    and so they are not considered in detail in the present paper.

    3.1. Experimental results

    The IV algorithm (8) has been applied to the follpwing second order discrete

    time system

    where 5 is white noise chosen

    to

    make S = 20 ; while a, = 0.5, Vk

     

    and both

    b and

    a

    are time-variable with

    0.3 ; k= 1,

    ..., 30

    0.5 ; k=31, ... 60

    0.4 ; k=61,

    ..., 100

  • 8/18/2019 young1980.pdf

    13/25

    752

    P oung nd A . Jake-

    and

    -0.35

    ;

    k

    1 ...,

    60

    a l k

    =

    -0.6 ; k=61, ... 100

    Figures

    4

    (a),

    b )

    and c) show the estimation results obtained when an

    IRW

    model (11) is assumed, with

    Q

    selected as a diagonal matrix with elements

    0.001, 0, and 0-001 respectively. Both t he recursive filtering an d smoothing

    estimates are shown in all cases (for the constant parameter a,, the smoothed

    estimate is, of course, itself a constant). I t is interesting to note th a t very

    similar results to these were obtained using the SRW model with a = 0.9 nd

    the diagonal elements of Q set

    to

    0.05, 0, and 0.05, respectively.

    t is clear th at in thi s example where step changes in parameters occur, th e

    smoothing algorithm is not pa,rticularly appropriate, since it attempts to

    provide a smooth transition where abrupt changes are actually being en-

    countered. I n practice, however, it is quite likely th at smoother changes in

    parameters will often occur nd t is here th at the smoothing algorithm will have

    maximum potential. Bu t it should be emphasized th at t he smoothing

    algorithm used.here is

    a n

    off-line procedure and is computationally expensive

    in comparison to

    the

    filtering algorithm (8). On line, fixed lag smoothing

    algorithms (Gelb 1974) could be developed, however, if

    c i r cu m s t an ~ s o

    demanded.

    I recursive es t im te

    I S

    0 5 0

    1 O

    number of

    sarqdes

    Figure

    4.

    ime variable parameter estimation for second order, stochastic system.

  • 8/18/2019 young1980.pdf

    14/25

    Refined instrumental variable methods of recursive time-series analysis

    75

    Algorithms such as 8) with.parameter variation model 12) and 13) have

    been used successfully for the estimation of the rapi d.1~ hanging parameters

    of a simulated missile system e.g. Young 1979 b). I n this example, additional

    flexibility was required to estimate the particularly rapid changes in parameters

    th at occurred over the rocket boost period and this was introduced by making

    the covariance matrix

    Q

    also a function of

    k

    The algorithm 8) has also been applied to many other sets of s imulated

    and real data . These include time-series obtained from a large econometric

    model of the Australian economy Young and Jakernan 19.79 b) real da ta

    from the United States economy Young 1978) ; an d various sets of environ-

    mental dat a Young 1975, Whitehead and Young 1975, Young 1978). In the

    latter examples, the time-variable estimation was utilized specifically for the

    identification of non-linearities in the model structure, a procedure for which it

    seems singularly well suited.

    4. Sta te rCconstruction

    filter

    design

    Recently Young 1979 c) has shown how the estimates of th e st at e of

    a

    stochastic, discrete-time, SISO system can be obtained as a linear function

    of the output s of t he adaptive prefilters used in the refined IV algorithm when

    i t is applied t o t he following

    ARMAX

    model

    In particular, i t can be shown tha t the sta te estimate jCk, is generated theoreti-

    cally from a relationship of th e following form,

    'jCk='k~

    16)

    where

    z k = [ N l < l k :Nn

  • 8/18/2019 young1980.pdf

    15/25

    764

    P

    oung nd

    A J a k e m n

    The s ta te reconstruction filter (16) can be implemented in practice by replacing

    by its estimate obtained from a recursive IVAML algorithm and using t he

    outputs of t he prefilters in the same algorithm to define

    2 .

    f th e fully recursive I V solution is utilized in the above fashion then t he

    overall procedure constitutes an optimal adaptive Kalman filter, in the sense

    th at th e asymptotically optimal state estimates are obtained as a by-product of

    an asymptotically efficient parameter estimation scheme.

    In the off-line

    recursive-iterative implementation, use of th e algorithm in this fashion

    represents a Kalman filter design procedure in which the asymptotic gain

    Kalm an (or Wiener) filter represented by eq n. ,( l6 ) s synthesized direct ly from

    th e system data?. This should be particularly useful in practice because i t

    obviates the need for specifying noise statistics and solving the covariance

    matrix R,iccati equation, as required in normal Kalman filter design. Note also

    that this approach to state estimation can also be applied in the purely

    stochastic situation where no input variables

    u

    are present : here the refined

    AML algorithm would provide the source of parameter estimates and prefiltered

    variables.

    An extension of t he above approach to continuous-time systems with

    deterministic inputs is fairly obvious but involves some technical problems.

    The expression for the estimate of th e continuous-t ime sta te -vec tor .jC(t) is of

    the same basic form as

    16)

    (see Fig. I , i.e.

    where, theoretica.lly, Z(t) is t he continuous equivalent of Zk. But the noise

    model in th e continuous-time equivalent of eqn. (15) has equal order numerato r

    and denominator polynomials which introduces estimation problems (see, e g

    Phillips 1959, Phadke and Wu 1974).

    We

    will not discuss these problems in this present paper but will merely

    note th at , in this situation, a straightforward ye t clearly suboptimal approach

    is to use a more arbitrary state variable filter ( l / B ) . For example, th e choice

    of

    D

    could be based on the heuristic notion t ha t its passband should encompass

    the passband of the system under study (e.g. b = d ) . In this manner, the

    filter will pass frequency components of interest in the estimation bu t at te nu at e

    .high frequency noise. The resultant IV algorithm is then identical to that

    suggested by Young (1969), and the s tate estimate obtained from (22) with

    replaced by

    p

    although not optimal in any minimum variance sense, will be

    asymptotically unbiased and consistent, i.e. jC(t)+x(t) for 1-co where is

    the true stat e vector.

    The I V algorithm in this latter case can be considered as

    a

    stochastic

    equivalent of the adaptive observer suggested by Kreisselmeier 1 977 which

    is based on th e deterministic equivalent of eqn. (22) and uses

    a

    continuous-time

    deterministic estimation algorithm with .constant gains. Bu t unlike th e

    Kreisselmeier observer, this stochastic sta te reconstruction filter will function

    satisfactorily, albeit non-optimally, in the presence of even high levels of noise.

    t

    This would seem to satisfy Kalman s requirement (1960), hat the two problems

    (parameter nd state estimation) should be optimized jointly if possible .

  • 8/18/2019 young1980.pdf

    16/25

    Refined instrumental variable

    metho s

    of

    recursive

    time series analys is

    7

    Final ly, with the utilization of suitable multivariable canonical forms, it is

    possible to extend the arguments in this section to multivariable systems

    Jakeman and Young 1979 c . Such extensions will, however, suffer from the

    disadvantages of complexity associated with all multivariable black box

    methods see Jakernan and Young

    1979

    a) and they will need

    to

    be considered

    in

    detail before their practical potential can be evaluated.

    4.1 . Experimental

    results

    Figures 5 a) nd b ) show the results obtained when the off-line Kalman

    filter design procedure is applied to the following model

    which is simply the nnovations or Kalman filter description

    of

    model 5,

    considered

    in

    Par t of the paper. These results were obtkined using Monte

    Carlo analysis with ten random simulations and the figures show the ensemble

    averages of the two sta te variable estimates compared with the true st ate

    variables generated by the model. The variance associated with the ensemble

    averages was quite small as shown by the standard error bounds marked on t he

    la1

    ru

    value

    l o \

    st imt

    0 5 100

    number of samples

    Figure

    5.

    Joint parameter state estimation : output

    of

    adaptive state reconstruction

    Kalman) filter for second order, stochastic system.

  • 8/18/2019 young1980.pdf

    17/25

    756

    P . Young

    and

    A . Jakernan

    plots.

    t

    is interesting to observe th at t he estimate of the first element of

    gk s the optimally filtered output of the system ik hich corresponds with t he

    optimal one step ahead prediction of the output.

    5.

    The general refined

    IV

    approach to estimation and

    ts

    application to some

    special model forms

    So far in this paper we have talked mainly in terms of specific mathematical

    model forms. One attraction of th e refined IV method, however, is t ha t it

    suggests a general

    approach

    to stochastic model estimation that has some

    similarities with the alternative prediction error

    (PE)

    minimization approach,

    but which leads to algorithms with subtle bu t important practical differences of

    mechanization. In this section, we will outline this approach and show tha t

    it can be considered

    to

    arise from a conceptual basis which we will term

    generalized equation-error (GE E) minimizationt. We will also demons trate

    the efficacy of the approach by showing how i t can be applied to two specific

    model forms, namely the multi-input, single output (MISO) transfer function

    model ; and the tanks in series representation, as used in chemical engineering

    and water resources modelling work.

    5 . 1 . Generalized equution-error G E E )minimizat ion

    The general refined I V approach

    to

    time-series analysis consists of th e

    following three steps.

    1 ) Formula te the stochastic, dynamic model so th at t he stochastic charac-

    teristics are defined in relation t o a source term consisting of a white noise

    innovations process

    ek

    (or k n the multivariable case), and then obtain an

    expression for e, in terms of all the other model variables.

    2) By defining appropriate prefilters, manipulate the expression for

    e,

    until i t is a linear relationship (or set of relationships) in the unknown para-

    meters of the basic

    deterministic

    part

    of

    the

    model

    (i.e. the an d polynomial

    coefficients in all examples considered so far). Because of their similarity to

    the equation-error relationships used in ordinary IV analysis, these linear

    expressions can be considered as generalized equation-error (GEE) functions.

    3 )

    Ap ply the recursive or recursive-iterative I V algorithms to the estima-

    tion of th e parameters in the GEE model(s) obtained in st ep (2 ) , with t he I V s

    generated in the usual manner as functions of the ou tpu t of a n adap tive

    auxil iary model (in th e form of the deterministic par t of the system model).

    f

    prefilters are required in the definition of GEE , then they will also need to be

    made adaptive, if necessary by reference to a noise model parameter estimation

    algorithm (e.g. th e refined

    A m

    lgorithm) utilized

    in

    parallel a nd co-ordinated

    with the IV algorithm.

    This decomposition of t he estimation problem into parallel but co-ordinated

    system and noise model estimation, as outlined in step

    3 ) ,

    is central to the

    concept of

    GEE

    minimization

    ;

    and it contributes

    t

    the robustness of t he

    resultant algorithms in comparison with equivalent prediction error. (PE)

    minimization algorithms (Ljung

    1976,

    Young and Jakeman 1979 c).

    f

    GEE

    has been used previously

    t

    denote any EE function defined in terms of

    prefiltered variables ; here we use i t more specifically to mean an optimal

    GEE

    function with prefilters chosen to induce asymptotic statistical efficiency.

  • 8/18/2019 young1980.pdf

    18/25

    Refined instrumental variable metho s of recursive time series analysis 7 7

    The robustness is enhanced further by the IV mechanization, which ensures

    th at the algorithm is not susceptible to contravention of theoretical assumptions

    about the nature of the noise process. In particular, the supposition th at the

    noise arises from a white noise source, usually via some dynamic model (e.g. a

    rational transfer function, as in the

    ARMA

    model) is not restrictive

    :

    provided

    that the system input signals (u, or uk)are independent of the noise, then t he

    refined IV algorithms will yield estimates which are asymptotically unbiased

    and consistent

    even if the noise ssumplions are imrrect.

    This remains true

    even if t he noise is highly structured, e.g. a periodic signal or d.c. bias. On th e

    other hand, if the assumptions

    are

    valid, then the resulting estimates will, as

    we have seen in this series of papers, have the additional desirable property of

    asymptotic statistical efficiency.

    5 . 2 .

    The M I S transfer function model

    To exemplify the

    GEE

    minimization approach, let us consider the MIS0

    transfer function model. Most time-series research in the M I S 0 case has been

    directed towards the so-called

    ARMAX

    representation, where the transfer

    functions relating each separate input to the output have common denominator

    polynomials. An alternative and potentially more useful model form is the

    following MISO transfer function model where them individual input-output

    transfer functions are defined independently (see, e.g. Box and Jenk ins 1970),

    This model can be considered as the dynamic equivalent of regression

    analysis, with the regression coefficients replaced by transfer functions. I n

    this sense, such model has wide potential for application.

    Considering the two input case for convenience, we note from (24) th at the

    white noise source e is defined

    as

    I t is now straightforward to show that (25) can be written

    in

    two GEE forms.

    First,

    i

    a single star superscript is utilized to denote prefiltering by CIDA,,

    then

    Here lk is defined as

    t l k =Yk- k

    where gZk s the output of the auxiliary model between the second input

    u

    and the output, i.e. i t is th at part of the output explained by the second input

    alone. Similarly,

    e

    can be defined in terms of

    t2

    where

    f2k =Yk

    28)

    In this case,

    ek

    =

    A,(2k** B2u2,**

    P9)

    where the double star superscript denotes prefiltering by CIDA,.

  • 8/18/2019 young1980.pdf

    19/25

    758

    P. Young

    and

    A. Jakeman

    By decomposing the problem into the two expressions (26) and (29), we have

    been able to define two separate GEE S which are linear

    in

    t2ie unknown moae?

    parameters for each transfer function in turn.

    Now let us define,

    where aif,j

    = 1 2, ,

    n

    ; and bij, j = 0,

    1,

    ,

    n

    are th e j t h coefficients of

    A

    and Bi respectively. It is now possible to obtain estimates 8 and 6 of a

    and a, from the following refined I V algorithm

    where Qi and Kik* are defined as follows

    -

    *

    Qik*= [gi,k-l

    ,

    ..,xi,k-n

    uik*, . . . I Ui,k-n*]T

    I

    (ik*=[fi,k-L .., ' i,k-n*r ~ i k * ,.., Ui,k-n*IT

    Algorithm (30) is used twice for

    =

    1 2 but when i = 2 the single star super-

    scripts are replaced by double sta r superscripts. The adaptive prefiltering is

    then executed in the same manner as for the SISO case, with th e refined AML

    algorithm providing estimates of the and polynomial coefficients. The

    extension to th e general case of inputs is obvious. There is also a symmetric

    gain version of (30)with gik*Tbeing replaced by gik* everywhere except within

    th e braces.

    5.3. The tanks-in-series model

    In chemical engineering and water resources research i t is quite often useful

    to describe a dynamic system by means of a serial connection of identica l

    tanks with each tank described by a first order differential equa tion with

    transfer function b/(s+a) ; in other words the input u t ) and the output y t )

    are related by

    where m is the number of tanks in series and t (t ) s a noise term. Using GEE

    approsch,it is possible to obtain refined I V estimates of a and b for different

    and so identify and estimate the tanks-in-series model. We will not discuss

    this in detail here since i t is done elsewhere (Jakeman and Young 1979 b), b u t

    an example of it s use will be described

    in

    the next section.

    5.4. Experimental results

    Jakeman et a2 (1979) have evaluated the M I S 0 transfer function model

    estimation procedure using both simulated and real da ta.

    Figure 6 compares

    the deterministic out put of a

    MIS0

    air pollution model obtained in thi s manne r

    with the measured data. Here the d i t a are in the form of atmospheric ozone

    measurements a t a downstream location in the San Joaquin Valley of Cali-

    fornia. These are modelled in terms of two inpu t ozone measurements a t

  • 8/18/2019 young1980.pdf

    20/25

    Refined instrumental variable methods of recursive time-series analysis

    759

    upstream in relation to th e prevailing wind) locations.

    This analysis proved

    particularly useful for interpreting across gaps in downstream da ta .

    Table shows the results obtained when th e tanks-in-series estimation

    procedure was applied t 100 samples of simulated dat a obtained f rom a second

    order system

    m

    =

    2). The results again include averages an d sta nda rd errors

    over ten experiments.

    mo el output i t

    5

    100

    num er of mrrQles-

    Figure 6. Results from model of ozone concentration in San Joaquin Valley,

    California.

    Signal to noise ratio,

    Parameter True

    value 10 30

    Table 2 T,=0.2 sec P,/15-7).

    6. ultivariable systems

    with

    limited dimension observation space

    Another example of t he many possibilities which are opened up by th e

    refined

    V

    approach

    to

    time-series analysis is the case where the number

    of

    observed variables in a multivariable system is less than th e number of out pu t

    variables in the model. In theory, the symmetric matr ix gain refined

    V

    algorithm for multivariable systems described by Ja kem an an d Young 1979 a)

    can be modified to allow for such a situation. This will bnly be possible, .of

    course, provided the complete model is identifiable from the limited observa-

    tions. The conditions for identifiability in these situations are not t he subject

    of the present paper but, if we assume th at the model is identifiable i.e. unique

    estimates of

    all

    the model parameters can be obtained -from the available

    observat ions) then the modifications to t he symmetric gain refined

    V

    algorithm

    are fairly straightforward.

  • 8/18/2019 young1980.pdf

    21/25

  • 8/18/2019 young1980.pdf

    22/25

    Refined instrumental variable methods of recursive time-series analysis 761

    7. Stochastic approximation algorithms

    t is well known that the recursive least squares and related algorithms,

    such as those discussed in th is series of papers, can be interpre ted as special

    examples of matr ix-gain , multi-dimensional, stochastic approximation (SA)

    procedures (Tsypkin 1971, Young 1976). t is clearly possible, therefore, to

    modify any of t he refined IV algorithms to form simpler, scalar gain al ter-

    natives. While such SA procedures are computationally efficient, they will

    not usually possess the rapid convergence characteristics and low sample

    statistical efficiency of their matrix gain equivalents. They may prove

    advantageous, however, where da ta are plentiful bu t computational load must

    be kept t o a minimum.

    I n the basic SA algorithms, the matrix gain

    p is replaced by a scalar gain

    which obeys the conditions of Dvoretzky (see, e.g. Tsypkin 1971). I n the

    discrete-time case, the best known gain sequence of thi s type is y, = y/k when y

    is a constant scalar

    :

    in other words, the gain sequence is made a monotonically

    decreasing function of the recursive step number , k. I n the continuous-time

    case the best known example is simply the continuous-time equivalent of y,.

    Such SA algorithms can also be modified (normally heuristically) to allow

    for variation in the estimated parameters

    :

    this is achieved by restricting the

    monotonic decrease in gain in some manner, usually by making y, or y(t)

    approach a constant yo exponentially as k or approaches infinity. This

    modification is based on a partial analogjr with the behaviour of th e f , matrix

    in algorithm

    (8)

    when a W model (9) for parameter variations is used to define

    and I (Young 1979 d) .

    The simple SA versions of the refined IV algorithms cannot be recom-

    mended for general application since their rates of convergence can be intoler-

    ably low (see, e.g. Ho and Blaydon 1966). Bu t i t is possible to consider

    somewhat more complicated algorithms which represent a compromise between

    the simplicity of basic SA and the complexity of the fully recursive matrix-gain

    algorithms. simple example would be the following modification t o the

    refined IV algorithm given by eqn. (4) (i) in Pa rt

    I

    of the paper (Young and

    Jakeman 1979 a, p. 4)

    ak

    YkfjkD ik*{zk*T -yk*)

    36) .

    Here f , ~s a 2n 1, diagonal matrix with elements ( k - i*) -2 , i = 1 2 n,

    and u , - ~ * ) - ~ , 0 1,

    .

    .

    ,

    n ; while y, is a SA sequence, say y/k.

    In other

    words, the gain matrix pk s replaced by a purely diagonal approximation

    y,P,D, so tha t the computational burden is proportional to n rather than n2

    for the full refined IV algorithm.

    Algorithms such as (36) seem to work reasonably well (see, e.g. Kuma r and

    Moore 1979). As might be expected, their performance seems to

    fall some-

    where between th at of th e full algorithm and the scalar gain equivalent. I n

    general, however, the simpler algorithms should only be used when this is

    . necessitated by computational limitations, as for example in on-line applica-

    tions using special low storage capacity microprocessors.

    8.

    Self adaptive control

    Perhaps the prime motivation for th e development of recursive estimation

    algorithms during the early nineteen sixties was their potential use in

  • 8/18/2019 young1980.pdf

    23/25

    76

    P

    Young

    and A Jakeman

    self-adaptive control.

    Of late, this, early stimulus has been revived with th e

    development of the self-tuning regulator (STR) based on recursive least

    squares (R LS) estimation (e.g. Astrom and Wittenmark 1973, Clarke an d

    Gawthrop 1975):. In the STR th e effect of t he noise induced asymptotic bias

    on the RLS parametric estimates is cleverly neutralized by embedding the

    algorithm within a closed adaptive loop which automatically adjusts the

    estimates and th e control law t o yield either minimum variance regulation or

    some other objective, such as closed loop pole assignment (Wellstead

    et a2

    1979).

    The concept of the STR can be contrasted with the earlier concept of self

    adaptive control by identification and synthesis (SAIS),where the objective

    is to explicitly obtain unbiased parameter estimates and then to separately

    synthesize the control law using these estimates (e.g. Kalman 1958, Young

    1965 b, Young 1979 b) .

    While the STR seems to possess good practical potential, there a re certa in

    situations where the alternative of SAIS will have some advantages. Fo r

    example, the stability of the adapt ive loop in the STR is not easy to ensure

    a

    pr or

    because of the close integration between the recursive algorithm and t he

    control law, and the highly non-linear nature of the resulting closed loop system.

    On the other han d, the separation of estimation .and synthesis in t he S AI S

    system means t ha t the question of convergence and stability is largely one of

    ensuring th e identifiability of t he system under closed loop control. Thi s will

    always be possible provided an external command input is present which is both

    sufficiently exciting and statistically independent of the noise in the closed

    loop. The requirement for such an inp ut can be problematical, however, in th e

    pure regulatory situation, where the STR clearly comes into its own.

    In cases where the SAIS procedure seems advantageous, the refined

    V

    algorith r provides the best, currently available recursive estimation str ategy

    :

    it is robust, can be applied in continuous or discrete-time and its results can

    be used for either deterministic or stochastic control system design. Th e

    efficiency of such an SAIS strategy is demonstrated in the self adapti ve

    autostabilization system described by Young (1979 b) : here the recursive

    estimation is used to synthesize a deterministically designed control system

    based on closed loop pole assignment using sta te variable feedback. This

    system achieved tight control of the simulated missile over the

    whoIe of the

    mission, which included a difficult boost phase where parameters changed

    rapidly by factors ,of up to 30 in sec.

    I n the case of adaptive, stochastic control by SAIS, the present paper has

    shown th at the refined V approach provides an added bonus : the single IV

    AML

    algorithm yields not only the estimates of the model parameters b ut also

    estimates of the sta te variables, which can then be used in st ate variable feed-

    back control. And in the discrete-time, linear case, such an SAIS system could

    be considered optima lly adaptive, since the st ate variable estimates would

    then, as we have seen, be optimal in a Kalman sense.

    9

    Conclusions

    This is th e third of three papers which have described and comprehensively

    evaluated the refined V approach to time-series analysis.

    In the present

    paper, we have seen how this approach can be extended in various important

  • 8/18/2019 young1980.pdf

    24/25

    Refined instrumental variable methods of recursive t ime-se ries an aly sis 763

    directions a nd can also provide a conceptual basis for t he synthesis of refined

    c

    I V

    algorithms for a wide class of s tochas tic dyn amic systems.

    This conceptual base, which we have termed generalized equation-error

    GE E) minimization, has some similarities with th e alte rnat ive prediction error

    PE ) minimization concept but tends to yield algorithms which a re more ro bust

    both in a computational sense and to mis-specification of t he noise charac-

    teristics.

    We

    feel that this robustness, which arises primarily because of the

    errors-in-variables formulation Johnst on 1963) and conse quent IV mechaniza-

    tion, is an impo rta nt featur e of the refined I V algorithms and more detailed

    discussion on this topic will appear in a forthcoming paper Young an d Jak em an

    1979 c).

    This paper was completed while the authors were visitors in the Control

    an d Management Systems Division of the Engineering Dep art ment , University

    of Cambridge.

    REFERENCES

    ASTROM, . J. , 1970,

    Introduetwn to Stochastic Control Theory

    New York Academic

    Press).

    ASTROM,

    .

    J. , and WITTENMARK,. 1973,

    Automutica, 9,

    185.

    BOX,G. E. P., and JENKLNS. M. 1970, Time Series Analysis San Francisco:

    Holden Day) .

    CLARKE,D. W., and GAWTHROP,. J., 1975, Proc. Instn elect. Engrs, 122, 929.

    GELB,A. ed.), 1974, Applied Optimal Estimutwn Boston

    M I T

    Press).

    Ho,

    Y.

    C., and BLAYDON,., 1966, Proceedings of the

    N .

    E . C . Conference, U.S.A.

    JAKEMAN,. J. 1979,

    Proc. I P A C S ym p. on Identification and System Parameter

    E s t i m t w n , Darmstadt, Federal Republic of Germany.

    JAKEMAN, . J. STEELE,L. P., and YOUNG, . C., 1978, Rep. No. AS/R26, Centre

    for Resource and Environmental Studies, Australian National University ;

    1979, Rep. No. AS/R35.

    JAKEMAN

    :

    J.

    and YOUNG, . C., 1979 a,

    In t .

    J.

    Control,

    29,621

    ;

    1979

    b

    Rep. No.

    AS/R36, Centre for Resource and Environmental Studies, Australian National

    University 1979

    c

    Rep. No. AS/R37 submitted to Electron. Lett. .

    JAZWINSKI, . H., 1970, Stochastic Processes and Filtering Theory New York

    Academic Press).

    JOHNSTON., 1963,

    Econometric Methods

    New York McGraw-Eill).

    KALDOR, . , 1978, M.A. Thesis, Centre for Resource and Environmental Studies,

    Australian National University.

    KALMAN, .

    E.,

    1958, Tra ns. Am . Soc. mech. Engrs, 80-D, 468 ; 1960, T r a m . A m .

    Soc. mech. Engrs, 82-D, 35.

    KALMAN,. E., and BUCY

    .

    S., 1961, Tra ns. Am . Soc. mech. Engrs, 83-D, 95.

    KAYA,Y., and YAMAMURA,

    .,

    1962, A . I . E . E . T r a n s. A p p . I n d . , 80

    11

    378.

    KOHR,R.

    H., and HOBEROCK,. L., 1966,

    P ro c. J . A . C . C . , p.

    616.

    KOPP,

    R.

    E. nd ORFORD, . J . , 1963, A I A A

    J.

    I, 2300.

    KREISSELMEIER,., 1977,I E E E Trans . au tom. Contro l,

    22 ,

    2.

    KUMAR, ., and MOORE,

    .

    B.,

    1979,

    Automatics

    to appear).

    LEE, R . C. K., 1964, Optimal Estimation, Identification and Control Boston MIT

    Press).

    LEVADI,V. S., 1964, International Conference on M icrowaves, Circuit Theo ry and

    In.formation Theory, Tokyo.

    LJUNO .,

    1976, System Identif ication: Advances and

    Case

    Studies, edited by R K.

    Mehra and D.

    G.

    Lainiotis New York Academic Press).

    NORTON,. 1975,

    Proc. l ns tn elect.. Engrs,

    122, 663.

  • 8/18/2019 young1980.pdf

    25/25

    764

    Refined instrumental variable metho of recursive time -serie s an aly sis

    PHADKE,

    .

    S., and Wu, S. M., 1974,

    J . Am. statist. Ass.,

    69,

    325.

    PHILLIPS, A. W., 1959,

    Biometriku, 46

    67.

    PIERCE,D. A. 1972, Biometrika, 59, 73.

    SOLO,V.

    1978, Rep. No. AS/R20, Centre for Resource and Environmental Studies,

    Australian National University.

    T S Y P K I N ,

    A.

    Z . ,

    1971,

    Adaption and Learning

    in

    Adomatic Sys tems

    New York

    Academic Press).

    WELLSTEAD,.

    E.,

    EDMUNDS,. M.,.PRAoER,

    D.

    and ZANKER ., 1979,

    In t .

    J .

    Control, 30

    1.

    WHITEHEAD

    .

    G .

    and YOUNG,

    .

    C., 1975,

    Com pde r Simulution of Water Resource

    Systems,

    edited by

    G .

    C. Vansteenkiste Amsterdam North Holland).

    WHITEHEAD,.

    G.

    YOUNG,

    .

    C., and IV~ICHELL, ., 1978,

    Proc. I . E . Hydrology Sy mp. ,

    Canberra, p.

    1

    YOUNG,. C.

    964,

    I .E .E. E. Trans . Aerosp .,

    2, 1106 1965

    a,

    Rad.

    and Elect. Eng.

    J . E R E , 29,345 1965 b, The ory of Self Aahp tive Conlrol Sy ste ms , edited by

    P. H. Hammond New York Plenum Press)

    ;

    1969,

    Proce edings of the Wo rld

    Congress,

    Warsaw see also

    A d o m a t i m ,

    6 271) ; 1974,

    Bull.

    Inst. Math.

    Appl . ,

    10

    209

    ; 1975,

    J1 R . statist. Soc.

    B

    7, 149

    ;

    1976,

    Oplimisation

    in

    Action,

    edited

    b y L.

    C. W. Dixon New York

    Academic Press), 517

    ;

    1978,

    The Modeling of Enviro nmen tal Sy ste ms , edited by G . C. Vansteenkiste

    Amsterdam North Holland)

    ;

    1979 a,

    Proceedings of

    the

    I P A C S y m p. o n

    Identification

    nd

    System Parameter Estimation,

    Darmstadt, Federal Republic

    of Germany ; 1979 b, Ib id .

    ;

    1979 c,

    Electron. Lett.,

    15,358 1979d

    Recursive

    Estimation

    New York Springer).

    YOUNO,

    .

    C. nd

    JAKEMAN. J .

    1979 a,

    In t . J .Control,

    29, l

    ;

    1979 b,

    Proceedings

    01

    the I P A C Symposium on Computer

    ded

    Design of Control Sy ste ms ,

    Zurich

    ;

    1979 c, Rep. No. AS/R27, Centre for Resource and Environmental Studies,

    Australian National University.

    Y o u ~ a ,

    . C., and KALDOR.

    .

    1978, Rep. No. AS/R14, Centre for Resource and

    Environmental Studies, Australian National University.