Use of Interpolating Functions in Fast State
Estimation for Dynamic Systems with Missing
Observations 7
by KERiM DEMiRBASDcpurtnzent ~f’Elcctrica1 Engineering und Computer Scierzce (M/C 154).
Universit~~ of Illinois at Chicqo, Clricugo, IL 60680, U.S.A.
I. lntvoduction
For the last three decades, researchers have extensively treated recursive state
estimation of the classical dynamic models (or systems) (l-4). These classical
models must be linear functions of the disturbance noise and (additive) observation
noise. but they can be nonlinear functions of the state; and observations are
assumed to be available for all times within a considered time interval. Proposed
classical estimation schemes, such as the (extended) Kalman filter, have had many
applications in the areas of aerospace and electronic systems (5), and economics.
Recursive state estimation has been also considered for dynamic models which
are nonlinear functions of the state, disturbance noise and observation noise ; and
whose observations are assumed to be available for all times within a considered
time interval (6-8). The proposed schemes have been also applied to practical
systems (9). These schemes prevent the state estimate divergence caused by model
linearization errors which are introduced by the classical estimation schemes
such as the extended Kalman filter for state estimation of nonlinear dynamic
models (9).
Recursive state estimation has been recently considered for nonlinear dynamic
models with missing observations in a considered time interval. in which the Viterbi
decoding algorithm is used (IO). The implementation of the proposed scheme
requires an exponentially increasing memory with time, which makes state esti-
mation impractical for a long time interval.
In this paper, in order to overcome the obstacle that the implementation of the
scheme proposed in (10) requires an exponentially increasing memory with time.
a stack sequential decoding algorithm of Information Theory is used for state
TThis work was carried out while the author was visiting Bilkent University, Ankara, Turkey.
estimation
of
dynamic systems with missing observations. The proposed scheme isfaster and more practical than the scheme in (10). The performance of the proposed
scheme is also discussed. II. Problem Statement
Consider the closed time interval [0, L], where L is a positive integer. Let A be
the set of all discrete times in [O. L], and STOM be a subset of discrete times in the
open interval (0. L), that is STOM c A n (I, L). It is assumed that observations in
the set STOM are missing. In other words, STOM is the set of times at which the
observations are missing.
This paper deals with state estimation of nonlinear discrete dynamic models
which are defined by
.I-(k + I ) = ,f’(k. s(k). II(~)) the state model (1)
:(/i) = ,g(k. X(A). r(k)) the observation model. (2)
where li denotes the discrete time; n(k) is an II x I zero mean disturbance noise
vector at time /i with known statistics; .X(O) is an ~1 x I initial state vector with
known statistics; s(k), /i > 0, is an IIT x 1 state vector at time /i ; r(k) is an I’ x I
zero mean observation vector at time k with known statistics; z(k) is an .Y x I
observation at time h- ; ,f’(k, s(A-). II.(~)) and y(li, s(k), I) are given functions
which define the state at time li+ 1 and observation at time /i in terms of the state,
disturbance noise and observation noise at time k; and the initial state and all
samples of the disturbance noise and observation noise are independent.
The interest is to estimate the states from time zero to time L, denoted by
X’- A (.v(O), .Y( I). . , s(L)). by using the set of available observations, denoted by
Z 6 (:(I) : /E STOAi. where STOA is the set of times at which the observations
arc available, that is, STOA = A - STOM.
111. Estimation Scheme
The state model is first approximated by a time varying finite state machine (or
model). This finite state model is represented by a trellis diagram. Then the states
are estimated by using a stack sequential decoding algorithm.
The finite state model which approximates the state model is defined by
.r,,(k+ 1) = Q(f’(k,S<,(li), n,,,(k))). (3)
where I*,, is a discrete disturbance noise vector which approximates the dis-
turbance noise vector 11(/i), and the possible values of n‘,,(k) are denoted by ir,,,(/i). ~i;,~(k), . , and I,,,,,, ( w ere the first subscript h d stands for the discrete random
vector and the second subscript shows the label of the possible value of the discrete
random vector) (6), .x,(O) is an initial discrete random vector which approximates
the initial state vector X(O), and the possible values of X,,(O) are denoted by x,,(O)..
.u,,:(O). . , sy, ,,,, (0). which arc called the initial quantization levels or the quan
tization levels at time zero; .v,(k) is the quantized state at time k, and the quan-
Fast Stute Estimation
first subscript q denotes the quantized state and the second subscript shows the
label of the quantization level), and Q( .) is the quantizer defined in (6). This is a
function which divides the entire m-dimensional Euclidean space into non-
overlapping generalized rectangles (called gates) of the same size and which then
assigns to each rectangle its center. The size of each rectangle is referred to as the
gate size.
The finite state model is represented by a trellis diagram (Fig. I) in which the
quantization levels of x,(k) are denoted by nodes at the (k+ 1)th column of the
trellis diagram, and the transitions between quantization levels are denoted by
directed lines. The transition probability from a quantization level s,,(k- I) to a
quantization level .u,,(k) is denoted by T(+,(k - I ) + x,,(k)) and defined by
T(X,,(k- 1) +X4, (k)) 4 Prob jr,(k) = .u,,(k)Is,(k- I) = .~(,,(k- I))
= CProb {~(k- 1) = t~.,,~(k- 1)).
where the summation is taken over all Y such that
s,;(k) = Q(,f;(k- 1,X(/G I) = .v,,(k- I), IL.(k- I) = W,,J- 1))).
The following metrics are also assigned to each node, branch and path of the
trellis diagram. The metric of a node (or quantization level) .u,,(k) is denoted by
MN(.u,,(k)), and defined as zero except for an initial node, whose metric is defined
as the natural logarithm of its occurrence probability, that is
In {Prob (.u,(h-) =
+,(k)) j
ifk=OMM%,,(k)) 5% o
otherwise
where In denotes the natural logarithm. The metric of the branch connecting the
node .~,~(k - 1) to the node -y<,,(k) is denoted by MB(s,,(k - 1) + .u,,(k)) and defined by
M&,,(k- 1) -+.x,,,(k)) e In {T(x,;(k- 1) + .u,,(k))p(-(k)l.~(k) = s,,(k))),
where p(~(k)l.v(k) = r,/(k)) is the conditional density function of the observation
Time 0 Time 1 Time k-l Time k Time L
FIG. 1. Trellis diagram of state.
Vol 327. No. 3. pp. 491-501. ,990
K. Denliuhq
at time k given x(k) = x,,(k). If:(k) is a missing observation, then it is first estimated
by a function which interpolates available observations in the neighborhood of
STOM (11). The metric of a path is defined as the sum of all metrics of the nodes
and branches along the path.
The trellis diagram from time zero to time L shows all possible paths, the
quantization levels along which can be taken by the state with time. Then the state
estimation problem is to find a path through this trellis diagram so that the
quantization levels along this path become the estimates of the state from time zero
to time L. It can be shown (6) that the optimum rule which minimizes the overall
error probability is to choose the path with the greatest metric (if there is more
than one path with the same metric. choose any one of these at random). The
metric with the greatest metric can be chosen by the Viterbi decoding algorithm
(VDA) (6, 7, 10). But the implementation of the VDA requires an exponentially
increasing memory with time. To overcome this obstacle, in this paper a stack
sequential decoding algorithm (12-14) is used to choose a path so that the quan-
tization levels along this path become the estimates of the state from time zero to
time L. Stack sequential decoding algorithms estimate (not find) the path with the
greatest metric by searching only the paths which most likely contain the path with
the greatest metric. Hence, a stack sequential decoding scheme is suboptimum and
its implementation requires an increasing (not exponentially as in the Viterbi
decoding algorithm) memory with time. Hence, the estimation scheme using a
stack sequential decoding algorithm is more practical and faster than the estimation
scheme using the Viterbi decoding algorithm.
IV. Performance
Performance of the proposed scheme is quantified by an ensemble upper bound
of the Gallager type. When the stack sequential decoding algorithm presented in
(14) is used for estimation, it can be shown (6) that the ensemble overall error
probability for choosing the correct path is bounded by
where
where r’, is the ensemble overall error probability for choosing the correct path;
K is the number of possible paths through the trellis diagram; T$” and nTax are
the minimum and maximum values of the occurrence probabilities of the initial
Fast State Estimutim
quantization levels ; T$“” and 7~ “‘d’ are the minimum and maximum of the transition
probabilities from time k- 1 to time k, respectively : x’ is the set of all quantization
levels from time I to time L; and N’ is the number of elements in X’.
Consider the discrete models defined by
s(k + 1) = ,f’(k. x(k), iv(k)) the state model
z(k) = g(k, x(k)) +r(k) the observation model,
where x(0) and r(k) are Gaussian noises with means nz(, and 0; and covariances
A,, and A,,(k), respectively. Substituting p(z(k)lx(k) = s), which is Gaussian, into
the expressions for Rk and S, above, we can obtain
U(.ul,s2)
A
exp- [s(k,s,)-y(k,.uZ)]‘A, ‘(k)[~(k,s,)--g(k..u,)l6
where the superscript .s is the dimension of the observation r(k), the superscript T
indicates the transpose, and det stands for the determinant. The bound in (4) is
used as the performance measure of the proposed scheme.
V. Simulations
State estimation of many examples with Gaussian noise and missing observations
was simulated on the IBM 308 1 K mainframe computer. In simulations : the initial
state and disturbance noise were approximated by the discrete random variables
presented in (6), and the stack sequential decoding algorithm presented in (14) was
used. Simulation results of three examples are presented in Figs 2(a)-4(c), where
the observations z(4) and ~(5) were assumed to be missing. In Figs 2(a)4(c) : the
simulated state and observation models are stated in the first and second rows at
the left upper corner; ACTUAL, SSDS and KALMAN show the actual values.
estimated values by the proposed scheme, and estimated values by the (extended)
Kalman filter of the states ; AAEOP and AAEK indicate averaged absolute errors
for estimates obtained by the proposed scheme and Kalman filter; E(Y) and
VAR(Y) denote the mean value and variance of the random variable Y : BOUND
and ER.COV. yield the bound in (4) and the error variances for the Kalman
estimates, respectively ; NUM. OF DISC. FOR Y denotes the number of possible
values of the discrete random variable which approximates the random variable
Y; and GATE SIZE shows the gate size used in (3).
Figures 2(a*) present simulation results of a linear example, whereas Figs 3(a)-
4(c) present simulation results of two nonlinear examples. The missing observations
z(4) and z(5) were first estimated by using a first-order polynomial which interp-
Vol 327, No. 3. pp. 491-501, 1990
K. Drmirhrr~
10 1'.60 3'.20 4'.80 6.40 s'.oo TIME
XlK+11=0.79XiKl+WlK1 ZIK1:7X[Kb+VIKI NUil.
OF
OiSC-
FOR
XlOI=l
VRRIXl011=0.001 ElXIOi1=30.000 NUN.
OF OISC. FOR WC.)=3 “ARIW, .,lk4.000 VRQlVI .11=3.000 GATE SILE=0.250 LEGENO A: KRLnRN +: 550s fiREK=O.B45866EO AAEOP:O.l62780El FIG. 2(c). Absolute and time averaged absolute errors for estimates of states.
XIK~II~XIKili+O.5SS~NIX!K~11+W~~? ZIK1=6X!<l-VIKi NUti.
OF OiSC. FOR XICIkI 8 VRRlX1011k1.500 4 EIXIOi1=3.000 _, NuM. 0' CISC. FOR kc.;=3 / IRR~WI.il=2.COO _EGENO 0: RCTdAL A: 7RLflRN 7: ssos lo.00 1.60 3.20 4.30 6.40 8.00 TIME FIG. 3(a). Actual and estimated values of states
Fat State Estimution
olates the available observations ~(3) and ~(6) for the linear and nonlinear examples
in Figs 2 and 3, but by using a second-order polynomial which interpolates the
available observations z(3), z(6) and ~(7) for the nonlinear example in Fig. 4. In
Fig. 2 : the Kalman estimates are better than the estimates obtained by the proposed
scheme since the Kalman estimates are optimum for linear models with white
Gaussian noise, whereas the proposed scheme is suboptimum. Both the proposed
scheme and extended Kalman filter are suboptimum for nonlinear models. The
Kalman estimates are better than the estimates obtained by the proposed scheme
for the models in Fig. 3, whereas the estimates obtained by the proposed scheme
are better than the Kalman estimates for the models in Fig. 4. One must keep in
mind : (i) A stack sequential decoding algorithm estimates the path with the greatest
metric. Therefore it may also pick up as the path with the greatest metric a path
which does not have the greatest metric. This may cause a state estimate divergence
from the actual state values, as in the state estimate divergence caused by model
linearization errors which are introduced by the extended Kalman filter. (ii) The
bound of (4) may sometimes become a number greater than one (i.e. useless) for
some dynamic models since it is derived by using some inequalities, and this bound
is an ensemble bound for the performance of the proposed scheme. Hence, it does
not exactly determine the performance of the proposed scheme for a given dynamic
system (6).
XIK+1I=1.ZXIKI*WIK1 ZIK1=0.9XfKl*VIKI NUti. CF DISC. FOR XlOl=l B VRRLXlO11~0.800
A _- NUtI. OF DISC. FOR Wc.1~3 EIXlO11=1.800 VRRIWI.11=1.900 m VRR~VI.,~:,.OOO wo GRTE SlLE=O.ZSO +9 E?? Lo LEGEND 0: RCTURL A: KRLnRN +: S5D.s ml 1o.oo 1.60 3.20 4.80 6.40 8.00 TIME
FIG. 4(a). Actual and estimated values of states.
Vol.327.No.3,pp.491-SOI,
500
VI. Conclusions
A fast estimation scheme is presented for nonlinear discrete dynamic systems (or
models) with missing observations. These models can be nonlinear functions of the
state, disturbance noise and observation noise. The missing observations are first
estimated by interpolating functions. Hence, the accuracy of the proposed scheme
depends upon the estimation accuracy of missing observations, which is determined
by interpolating functions used. The proposed estimation scheme requires a
memory which increases less than exponentially with time, whereas the estimation
scheme using the Viterbi decoding algorithm requires an exponentially increasing
memory with time for the implementation.
References
(I) A. P. Sage and J. L. Melsa, “Estimation Theory with Applications to Communications
and Control”, McGraw-Hill, New York. 1971.
(2) T. Kailath. “A view of three decades in linear filtering theory”, IEEE Twn.s. h/b.
T/wor_r, Vol. IT-20, No. 2, pp. 146-I 81. 1974.
(3) J. Makhoul, “Linear prediction : a tutorial rcvicw”, Ptw. IEEE. Vol. 63. pp. 561-580,
1975.
(4) J. S. Mcdich, “A survey of data smoothing for linear and nonlinear dynamic systems”.
Autonwticu. Vol. 9, pp. I5 l-162, 1973.
(5) C. E. Hutchinson, “The Kalman Filter applied to aerospace and electronic systems”,
IEEE Truns. Awspuc~~ El~ctrorz. Swms, Vol. AES-20, No. 4, pp. 500-504, 1984.
(6) K. Dcmirba?, “New smoothing algorithms for dynamic systems with or without
interfercncc”, T/w N A TO A GA R Dog~.qd~ A thnc~cs in the Tmhicps cd T~c~hnolog~~
of’ Applicutiom of’ Nonlinewr Filtrvs md Kdn~rrn Filtcw, No. 756. AGARD, pp. l9- 11’66, Mar. 1982.
(7) K. Demirba? and C. T. Lcondes, “Optimum decoding based smoothing algorithm for
dynamic systems”, Int. J. .Sj.stc~nz.s Sci., Vol. 16. No. 8, pp. 95l~-966. Aug. 1985.
(8) K. Demirba? and C. T. Leondes, “A stack scqucntial decoding based smoothing
algorithm for dynamic systems”, /of. J. Srsterns SC;.. Vol. 17. No. 2. pp. 269-280,
Feb. 1986.
(9) K. Demirbav, “Maneuvering target tracking with hypothesis testing”. lEEE Trtms.
Arro.spocc Elrctrotz. .~~stm.s, Vol. AES-23, No. 6, pp. 757-766, Nov. 1987.
(IQ) K. Dcmirba$. “State estimation for nonlinear discrete dynamic systems with missing
observations”, J. Franklin ht., Vol. 321, No. I. pp. 49-59, 1990.
(11) S. D. Conte Carl dc Boor. “Elementary Numerical Analysis”, McGraw-Hill, New York, 1972.
(12) G. D. Forney Jr, “Convolutional codes 111. sequential decoding”, Icfh. Control, Vol.
25. pp. 267-297, 1974.
(13) F. Jelinek. “A fast scqucntial decoding algorithm utilizing a stack”, IBM J. Rcs. Dcr.,
Vol. 13, pp. 675-685, 1969.
(14) A. J. Viterbi and J. K. Omura, “Principles of Digital Communication and Coding”,
M&raw-Hill. New York. 1979.
Vol. 127. No. 3. pp. 491-501. 1990