Generalized Multiple-Model Adaptive Estimation

Using an Autocorrelation Approach

John L. Crassidis
Dept. of Mech. & Aero. Eng.
University at Buffalo
State University of New York
Amherst, NY 14260-4400 U.S.A.
[email protected]
Abstract - In this paper a generalized multiplemodel adaptive estimator is presented that can be
used to estimate the unknown noise statistics in filter designs. The assumed unknowns in the adaptive
estimator are the process noise covariance elements.
Parameter elements generated from a quasi-random
sequence are used to drive multiple-model parallel
filters for state estimation. The current approach
focuses on estimating the process noise covariance
by sequentially updating weights associated with the
quasi-random elements through the calculation of
the likelihood function of the measurement-minusestimate residuals, which also incorporates correlations between various measurement times. For linear Gaussian measurement processes the likelihood
function is easily determined. For nonlinear Gaussian measurement processes, it is assumed that the
linearized output sufficiently captures the statistics
of the likelihood function by making the small noise
assumption. Simulation results, involving a twodimensional target tracking problem using an extended Kalman filter, indicate that the new approach
is able to correctly estimate the noise statistics.

Keywords: Multiple-model adaptive estimation, filtering, extended Kalman filter, target tracking.

Introduction

Filtering algorithms, such as the extended Kalman filter (EKF) [1], the Unscented filter (UF) [2] and Particle filters (PFs) [3, 4], are commonly used to both
estimate unmeasurable states and filter noisy measurements. The EKF and UF assume that the process noise
and measurement noise are represented by zero-mean
Gaussian white-noise processes. Even if this is true,
both filters only provide approximate solutions when
the state and/or measurement models are nonlinear,
since the posterior density function is most often nonGaussian. The EKF typically works well only in the
region where the first-order Taylor-series linearization
adequately approximates the non-Gaussian probability density function (pdf). The Unscented filter works
on the premise that with a fixed number of parameters it should be easier to approximate a Gaussian

Yang Cheng
Dept. of Mech. & Aero. Eng.
University at Buffalo
State University of New York
Amherst, NY 14260-4400 U.S.A.
[email protected]
distribution than to approximate an arbitrary nonlinear function. This in essence can provide higher-order
moments for the computation of the posterior function
without the need to calculate jacobian matrices as required in the EKF. Still, the standard form of the EKF
has remained the most popular method for nonlinear
estimation to this day, and other designs are investigated only when the performance of this standard form
is not sufficient.
Like other approximate approaches to optimal filtering, the ultimate objective of a PF is to construct the
posterior pdf of the state vector, or the pdf of the state
vector conditioned on all the available measurements.
However, the approximation of a PF is vastly different
from that of conventional nonlinear filters. The central
idea of the PF approximation is to represent a continuous distribution of interest by a finite (but large) number of weighted random samples of the state vector, or
particles. Particle filters do not assume the posterior
distribution of the state vector to be a Gaussian distribution or any other distribution of known form. In
principle, they can estimate probability distributions
of arbitrary form and solve any nonlinear and/or nonGaussian system.
Even if the process noise and/or measurement noise
are Gaussian, all standard forms of the EKF, UF and
PFs require knowledge of their characteristics, such as
the mean and covariance for a Gaussian process. The
covariance and mean of the measurement noise can
be inferred from statistical inferences and calibration
procedures of the hardware sensing devices. The calibration procedures can also be used to determine the
nature of the measurement process distribution. The
kurtosis characterizes the relative compactness of the
distribution around the mean, relative to a Gaussian
distribution. A common kurtosis, called the Pearson kurtosis, divides the fourth moment by the second moment [5]. Positive kurtosis indicates a relatively
peaked distribution, while negative kurtosis indicates a
relatively flat distribution. However, the process noise
is extremely difficult to characterize because it is usually used to represent modeling errors. The covariance is usually determined by ad hoc or heuristic approaches, which leads to the classical tuning of the
filter problem. Fortunately, there are tools available

to aid the filter designer. For example, several tests can

be applied to check the consistency of the filter from
the desired characteristics of the measurement residuals. These include the normalized error square test,
the autocorrelation test and the normalized mean error test [6]. These tests can, at the very least, provide
mechanisms to show that a filter is not performing in
an optimal or desired fashion.
In practice the tuning of a filter can be arduous and
time consuming. A classic approach to overcome this
difficulty is to use adaptive filters. Adaptive filtering
can be divided into four general categories: Bayesian,
maximum likelihood, covariance matching, and correlation approaches [7]. Bayesian and maximum likelihood methods may be well suited to a multi-model approaches, but sometimes require large computational
loads. Covariance matching is the computation of the
covariances from the residuals of the state estimation
problem, but have been shown to give biased estimates
of the true covariances. A widely used correlationbased approach for a linear Kalman filter with stationary/Gaussian process and measurement noise is based
on residual whitening [8]. In particular, the autocorrelation matrix, which can be computed from the
measurement-minus-estimate residuals, is used with
the system state matrices to provide a least-squares
estimate of the Kalman filter error covariance times
the measurement output matrix. If the number of unknowns in the process noise covariance is equal to or
less than the number of states times the number of
outputs, then the error-covariance/output-matrix estimate can be used to find an estimate of the process
noise covariance by solving for a set of linear equations.
These equations are not linearly independent and one
has to choose a linearly independent subset of these
equations [8].
Adaptive filtering for nonlinear systems has recently
gained attention. Parlos et al. shows a neural net
to constructively approximate the state equations [9].
The proposed algorithms in their paper make minimal assumptions regarding the underlying nonlinear
dynamics and their noise statistics. Nonadaptive and
adaptive state filtering algorithms are presented with
both off-line and on-line learning stages. Good performance is shown for a number of test cases. Lho
and Painter show an adaptive filter using fuzzy membership functions, where the fuzzy processing is driven
by an inaccurate online estimate of signal-to-noise ratio for the signal being tracked [10]. Good results are
shown for a simple tracking problem. Lee and Alfriend
show an adaptive scheme that can be used to estimate
the process noise covariance for both the UF and the
first-order divided difference filter [11]. The standard
update approach requires proper selection of a window
size to control the level of the variance update. The
innovation of their paper is a procedure that automatically calculates the window size using a derivative-free
numerical optimization technique. Good results are
shown for satellite orbit determination applications.
In this paper a new approach is derived for adaptive
filtering based on generalizing the standard multiple-

model adaptive estimation (MMAE) algorithm [12]. A

MMAE algorithm uses a parallel bank of filters to provide multiple estimates, where each filter corresponds
with a dependence on some unknowns, which can be
the process noise covariance elements if desired. The
state estimate is provided through a sum of each filters
estimate weighted by the likelihood of the unknown elements conditioned on the measurement sequence. The
likelihood function gives the associated hypothesis that
each filter is the correct one. Standard MMAE algorithms use only the current time measurement-minusestimate residual to test the hypothesis. Our approach
is a generalization of Ref. [13], which uses the time correlation of the filter residuals to assign the likelihood
for each of the modeled hypotheses. In particular, the
spectral content of the residuals is used and only scalar
measurements are assumed in Ref. [13]. The authors
also state that if multiple measurements are available,
then a diagonal matrix can be used with elements given
by the spectral content of each measurement residual,
but this assumes that the cross-correlation terms are
negligible. Also, the focus of their paper is on the detection of actuator failures with known control-input
frequency content.
Our approach is based on calculating the timedomain autocorrelation function, which is used to form
the covariance of a generalized residual involving any
number of backward time steps. This covariance matrix also includes the time correlated terms, thus providing a more rigorous approach. The unknown elements in our design are the parameters of the process
noise covariance. Process noise covariance elements
can be drawn from any sample distribution as long as
the resulting covariance matrix remains positive semidefinite. A Hammersley quasi-random sequence [14]
is chosen due to its well distributed pattern. The covariance elements are estimated using a weighted sum
of the quasi-random elements, similar to the approach
used for state estimation in PFs. An expression for
the error-covariance of the estimates is also provided,
which gives a bound on the process noise parameter
estimates.
The organization of the remainder of this paper proceeds as follows. First, the standard EKF equations
are summarized, since this filter will be used in the
simulations. Then, a review of the standard MMAE
algorithm is given. Next, the new adaptive approach
is shown, including the assumptions used for a method
that can incorporate nonlinear measurement models in
the adaptive approach. Finally, simulation results involving a two-dimensional target tracking problem are
shown.

Extended Kalman Filter

A summary of the continuous-discrete EKF is given in

Table 1, where x(t) is the n 1 state vector, u(t) is the
known control input, G(t) is the process noise distribution matrix, w(t) is the process noise vector which
is assumed to be a zero-mean Gaussian noise process
k is the discrete-time meawith spectral density Q(t), y

Table 1: Continuous-Discrete Extended Kalman Filter

Model

x(t)
= f (x(t), u(t), t) + G(t) w(t), w(t) N (0, Q(t))
k = h(xk ) + vk , vk N (0, Rk )
y
(t0 ) = x
0
x


(t0 ) x
T (t0 )
P0 = E x

Initialize

T
1
Kk = Pk HkT (
x
x
x
k )[Hk (
k )Pk Hk (
k ) + Rk ]

h 
Hk (
x
k)
x 

Gain

k
x

x
yk h(
x
k = x
k + Kk [
k )]

Update

Pk+ = [I Kk Hk (
x
k )]Pk

(t) = f (
x
x(t), u(t), t)
Propagation

P (t) = F (
x(t), t) P (t) + P (t) F T (
x(t), t) + G(t) Q(t) GT (t)

f 
F (
x(t), t)
x 
(t)
x

surement, vk is the measurement noise vector which is

assumed to be a zero-mean Gaussian noise process with

+
covariance Rk , x
k and x
k are the propagated and updated state estimates, respectively, and Pk and Pk+ are
the propagated and updated covariances, respectively.
Oftentimes, if the sampling interval is below Nyquists
limit, a discrete-time propagation of the covariance is
used:

Pk+1
= k Pk+ Tk + Qk
(1)

Note that in most cases Q(t) is a diagonal matrix, while

Qk contains cross-correlation terms due to sampling.
These cross-correlations can be ignored if Eq. (6) is a
good approximation for Qk .

where k is the discrete-time state transition matrix

of F (
x(t), t) and Qk is the discrete-time process noise
covariance matrix. These matrices can be numerically
computed for a constant sampling interval using an
algorithm given by van Loan [15]. First, the following
2n 2n matrix is formed:

F (
x(t), t)
G(t) Q(t) GT (t)
t
A=
(2)
0
F T (
x(t), t)

where t is the constant sampling interval. Then, the

matrix exponential of Eq. (2) is computed:

B11
B12
B11
1
k Qk
=
(3)
B = eA
T
0
B22
0
k

where B11 is not needed in the computation of k and

Qk . The state transition matrix is then given by
T
k = B22

(4)

Also, the discrete-time process noise covariance is given

by
Qk = k B12
(5)
If the sampling interval is small enough, then a good
approximation for the solution of Qk is given by
Qk t G(t) Q(t) GT (t)

(6)

Multiple-Model Adaptive Estimation

In this section a review of MMAE is shown. More details can be found in Refs. [16, 17]. Multiple-model
adaptive estimation is a recursive estimator that uses
a bank of filters that depend on some unknown parameters. In our case these parameters are the process
noise covariance, denoted by the vector p, which is assumed to be constant (at least throughout the interval
of adaptation). Note that we do not necessarily need to
make the stationary assumption for the state and/or
output processes though, i.e. time varying state and
output matrices can be used. A set of distributed elements is generated from some known pdf of p, denoted
by p (p), to give {p() ; = 1, . . . , M }. The goal of the
estimation process is to determine the conditional pdf
of the th element p() given the current-time measurek . Application of Bayes rule yields
ment y

k) =
p (p() |Y

k |p() ) p (p() )
p (Y
M
X
k |p(j) ) p (p(j) )
p (Y
j=1

(7)

 k denotes the sequence {

1 , . . . , y
k }. The
where Y
y0 , y
a posteriori probabilities can be computed through [18]
k) =
p (p() |Y
=

k1 )
p (
yk , p() |Y
k1 )
p (
yk |Y
()
k1 )
p (
yk |
xk ) p (p() |Y
M h
i
X
(j)
k |
k1 )
p (Y
xk ) p (p(j) |Y

(8)

4.1

Autocorrelation for Time-Varying

Systems

In this section the autocorrelation matrix for timevarying systems is derived, which is an extension to
the approach shown in Ref. [8]. Here we assume that
the model is linear with
xk+1 = k xk + k uk + k wk
k = Hk xk + vk
y

j=1

()
k1 , p() ) is given by p (
since p (
yk , |Y
yk |
xk ) in the
Kalman recursion. Note that the denominator of
Eq. (8) is just a normalizing factor to ensure that
k ) is a pdf. The recursion formula can now
p (p() |Y
()
be cast into a set of defined weights k , so that
()

()

()

k = k1 p (
yk |
xk
k
M
X
(j)
k

(9)

()

x
k =

M
X

(j) (j)

k
k x

(10)

j=1

Also, the error covariance of the state estimate can be

computed using


(j)
k
x

x
k



(j)
k
x

x
k

T

(11)

k,
The specific estimate for p at time tk , denoted by p
and error covariance, denoted by Pk , are given by
k =
p

M
X

(j)

k p(j)

(12a)

j=1

M
X
j=1



T
(j)
k p(j) p
k
k p(j) p



where Ck, i E ek eTki and E{} denotes expecta
tion. The propagation of x
k is given by

x
k = k1 (I Kk1 Hk1 ) x
k1

(16)

+ k1 Kk1 vk1 k1 wk1

Carrying Eq. (16) i steps back leads to

i
Y

x
kj (I Kkj Hkj ) x
k =
ki
+

i
X

"j1
Y

j=2

=1

j=2

k (I Kk Hk ) kj Kkj vkj

=1

"j1
i
X
Y

k (I Kk Hk ) kj wkj

+ k1 Kk1 vk1 k1 wk1

(17)

(12b)

Equation (12b) can be used to define 3 bounds on the

k.
estimate p

where x
k x
k xk . The following autocorrelation
function matrix can be computed:

Hk Pk HkT + Rk
i=0

 T T
Ck, i =
(15)

k x
ki Hki
Hk E x

 T

k vki
Hk E x
i>0

j=1

(j)
k

j=1

Pk =

(14)

= Hk x
k + vk

where k p (p() |
yk ). The weights at time t0
()
are initialized to 0 = 1/M for = 1, 2, . . . , M .
The convergence properties of MMAE are shown in
Ref. [18], which assumes ergodicity in the proof. The
ergodicity assumptions can be relaxed to asymptotic
stationarity and other assumptions are even possible
for non-stationary situations [19].
The conditional mean estimate is the weighted sum
of the parallel filter estimates:

k Hk x

ek y
k

j=1

M
X

where k is the discrete-time process noise distribution

matrix. Consider the following discrete-time residual
equation:

()

()

Pk

(13a)
(13b)

Adaptive Law Based on Autocorrelation

In this section the adaptive law, based on an autocorrelation approach, for the process noise covariance
matrix is shown. First, the autocorrelation for timevarying systems is derived, followed by the associated
likelihood functions for the defined measurement residuals.

where

i
Y

Zkj Zk1 Zk2 Zki

(18)

j=1

Performing the expectations in the definition

leads to

Hk Pk HkT + Rk

Hk k1 (P H T Kk1 Ck1, 0 )
k1 k1
Ck, i =

hQ
i

i1

Hk
kj (I Kkj Hkj )

j=1

T
ki (Pki Hki Kki Cki, 0 )

of Ck, i
i=0
i=1

i>1
(19)

()

where

T
Cki, 0 Hki Pki
Hki
+ Rki

(20)

Note that storage of the state model and covariance

matrices to the k i point is required to compute Ck, i
in general.

()

k Hk x
k .
where ek y
The new adaptive law, which we call the generalized
multiple-model adaptive estimation (GMMAE) algorithm, is based on carrying Eq. (8) i steps back to give
the new update law:
()

()

()

k = k1 Li

4.2

Likelihood Function

In this section the likelihood function for the measurement residual is shown. First, the following residual is
defined:

ek
ek1

i .
(21)
..
eki
The likelihood function associated with i is given by


1
1 T 1
exp i Ci i
(22)
Li =
2
[det(2 Ci )]1/2
E{i Ti }

where Ci

Ck, 0
T
Ck,
T1

Ci = Ck, 2
.
..
T
Ck,
i

is given by

Ck, 1
Ck1, 0
T
Ck1,
2
..
.
T
Ck1,
i1

Ck, 2
Ck1, 2
Ck2, 0
..
.
T
Ck2,
i2

..
.

Ck, i

Ck1, i1

Ck2, i2

..

.
Cki, 0
(23)
When i = 0 the likelihood function reduces down to
1
L0 = 
1/2
T
det[2 (Hk Pk Hk + Rk )]
(24)


1 T
T
1
exp ek (Hk Pk Hk + Rk ) ek
2
This likelihood is widely used in MMAE algorithms
[12, 16], but ignores correlations between different measurement times. However, it is simpler to evaluate than
the general likelihood function in Eq. (22) since no storage of data or system matrices is required.

4.3

New Adaptive Law

In this section the new adaptive law based on the autocorrelation is shown. In the traditional MMAE approach only the current the measurement information
is used in the update law given by Eq. (9). Therefore,
the update law is given by
()
k

()
()
= k1 L0
()

()
k M k
X (j)
k
j=1
()

since p (
yk |
xk

(25)

()

o1/2
+ Rk )]

1 ()T
()
()
exp ek (Hk Pk HkT + Rk )1 ek
2



1 ()T  () 1 ()
1
()

exp

Li = h

i1/2
i
i
2 i
()
det 2 Ci

(28)

where

()
i

is defined as

()
i

()T
[ek

()T
ek1

()T
eki ]T .

()
Ci
th

The matrix
is given by Eqs. (19) and (23) evaluated at the covariance and the optimal Kalman
gain. Unfortunately, the optimal gain is a function of
the actual covariance Qk , which is not known. Specifically, if Kk from Table 1 is substituted into Eq. (19),
then for i 1 the correlated terms Ck, i will always
be zero. One way to overcome this problem is to esti()
mate the Ck, i terms using the residuals, which is the
approach taken in Ref. [8]. But, this requires a stationary process and a sufficiently large set of measurements
over time, which would not work properly for timevarying system matrices and/or a sequential updating
scheme. A different approach is taken here, which is
also expanded for nonlinear systems. This assumes
that the measurement noise is small compared to the
signal so that the Gaussian nature of the measurement
()
residual is maintained. Estimates for Ck, i are given by

()
Ck, i =

()
()
()

xk )Pk HkT (
xk ) + Rk

Hk (

()
()

Hk (
xk )k1 (
xk1 )

()
()
T

k1 C () ]

xk1 ) K
[Pk1 Hk1 (
k1, 0

() Qi1
()

xk ){ j=1 kj (
xkj )

Hk (

()
kj Hkj (

[I K
xkj )]}

()
()
()

ki (
xki ) [Pki Hki
(
xki )

K
ki C () ]
ki, 0

()

()

()

()

()

is computed using

i=0

i=1

i>1
(29)

T
(
xki ) + Rki (30)
Cki, 0 Hki (
xki ) Pki Hki

()

 (26)

(27)

with

The covariance matrix Pk

) = L0 , which is defined by

()
det[2 (Hk Pk HkT

k
M
X
(j)
k
j=1

where

()

L0 = n

()

()

()

+()

()

Pk+1 = k (
xk )Pk Tk (
xk ) + Q()
(31a)
h
i
+()
()
()
()
Pk
= I Kk Hk (
xk ) Pk
(31b)
h
i1
()
()
()
()
()
Kk = Pk HkT Hk (
xk )Pk HkT (
xk ) + Rk
(31c)

where Q() is computed using p() . The estimate of the

optimal gain is computed using

h
i1

k = P H T (
K
x
)
H
(
x
)
P
H
(
x
)
+
R
(32)
k
k
k
k
k
k
k
k
k

0.8

with

k is computed using p
k.
where Q
k , along with the
Using the current measurement, y
th element, p() , 1 M , a bank of filters are ex()
k , and
ecuted. For each filter the state estimates, x
()
measurements are used to form the residual, i , go()
ing back i steps. The filter error covariance, Pk , and
()
()
state matrices, k
and Hk , evaluated at the current estimates are used to update the estimate of the
()
autocorrelation, denoted by Ci . Note that at each
()
new measurement time, all elements of Ci need to
k is provided,
be recalculated since a new estimate p
which is used to compute an estimate of the optimal
gain. Unfortunately, this can significantly increase the
computational costs. The diagonal elements do not
need to be recomputed though, since they are not a
function of the optimal gain. The residuals and autocorrelations are then used to evaluate the likelihood
()
functions Li . These functions are used to update the
k using Eq. (12a).
weights, which gives the estimate p
There are many possibilities for the chosen distribution of the process noise covariance parameters. A simple approach is to a assume a uniform distribution. We
instead choose a Hammersley quasi-random sequence
[14] due to its well distributed pattern. A comparison
between the uniform distribution and the Hammersley quasi-random sequence for 500 elements is shown
in Figure 1. Clearly, the Hammersley quasi-random
sequence provides a better spread of elements than
the uniform distribution. In low dimensions, the multidimensional Hammersley sequence quickly fills up
the space in a well-distributed pattern. However, for
very high dimensions, the initial elements of the Hammersley sequence can be very poorly distributed. Only
when the number of sequence elements is large enough
relative to the spatial dimension is the sequence properly behaved. This isnt much of a concern for the
process noise covariance adaption problem since the
dimension of the elements will be much larger than
the dimension of the unknown process noise parameters. Remedies for this problem are given in Ref. [20]
if needed.

p2
0.4

(33b)

Numerical Simulations

In this section a numerical simulation result is given.

A two-dimensional target tracking problem is used
to compare the performance of the MMAE and the
GMMAE. The state vector is chosen as the position
and velocity in x and y directions, given by x =

0.2

0
0

0.2

0.4

()

0.6

0.8

0.8

p1

(a) Uniform Distribution

1

0.8

0.6
()

(33a)

p2

+ T x ) + Q
k
Pk+1
= k (
x
k )Pk k (
k
h
i
k Hk (
P + = I K
x ) P

()

0.6

0.4

0.2

0
0

0.2

0.4

()

0.6

p1

(b) Hammersley Quasi-Random Sequence

Figure 1: Uniform Distribution and Hammersley

Quasi-Random Sequence Comparison

T
x x y y . The target motion model is given by
a linear dynamical system:
xk+1 = xk + wk
with

1
0
=
0
0

t
1
0
0

0 0
0 0

1 t
0 1

(34)

(35)

The sampling period t = 0.01. The 4-by-4 process

noise covariance matrix is parameterized by qx and qy ,
given by
 3


t /3 t2 /2
022
qx t2 /2

t
 3

Qk =
2

t /3 t /2
022
qy
2
t /2
t
(36)
The true values of qx and qy are chosen as qx = qy = 10.
In the MMAE and GMMAE, the elements of qx and qy
for the individual Kalman filters are generated using a
two-dimensional Hammersley sequence under the assumption that qx and qy are independently uniformly

60
GMMAE
MMAE

50

40

qx

It is assumed that the noise in range and azimuth

measurements is uncorrelated, so the measurement
noise covariance matrix is diagonal, given by Rk =
diag[r rA ]. We choose r = 0.01 and rA = 0.000001
in the simulations. Since the measurement model is
nonlinear in the state vector, EKFs are used in the
MMAE and GMMAE with the sensitivity matrix Hk
evaluated at the current estimate. The EKFs are provided with good initial estimates of the state vector, so
they do not suffer any divergence problems. An alternative approach to working with the nonlinear measurement model is to first convert the original range
and azimuth measurements to the effective position
measurements on x and y and then apply the linear
Kalman measurement update. In the latter approach,
the covariance matrix for the effective measurement
does not take as simple a form as the shown Rk and
becomes data-dependent. The resulting residual sequence is not stationary or ergodic in either case.
The two adaptive estimators are run 50 seconds to
process the same measurement data. As discussed in
the previous section, the GMMAE goes back i time
steps in order to form the residual. We choose i = 4 for
the GMMAE. The size of the corresponding residual
()
i is 10 by 1. The results are given in Figures 2 and
3. It can be seen that both estimators converge within
50 seconds, but the MMAE takes more than twice as
much time as the GMMAE to converge. Both estimators converge with a relative small set of elements in
the above-mentioned example. A closer examination of
the automatically-generated element set shows that of
the 250 elements there is a element that is close to the
true values of qx and qy . If all of the 250 elements are
far away from the true values, then the MMAE and
GMMAE with 250 static elements may fail to yield
satisfactory performance, since the estimates of qx and
qy are given by the average of the elements. Increasing
the element size is always a solution, but as the dimensionality of the problem increases, the computational
complexity involved may quickly become prohibitive.
Meanwhile, it should be noted that most of the elements have almost zero weights at the end of the simulation and their contributions to the final estimates
of qx and qy are negligible.
Resampling and Markov-Chain Monte Carlo or regularization techniques used in particle filtering can
be used in order to prune elements with very small
weights and dynamically generate new elements with
high weights in the neighborhood of existing elements,
while maintaining a constant number of elements during the estimation process. However, since a long data
sequence is usually required (and sometimes maneuvers are also required) to estimate the process noise
covariance and to discriminate between good elements and bad elements, straightforward application

of resampling and Markov-Chain Monte Carlo or regularization based on just a few consecutive data points
does not yield a satisfactory result. It is found that
good elements could be pruned in the early stage and
promising elements could end up with small weights.
Further investigation of the problem is under way.

30

20

10

10

15

20

25

Time (s)

30

35

40

45

50

Figure 2: Estimates of qx

80
GMMAE
MMAE
70

60

50

qy

distributed in [0, 100]. The number of elements in the

MMAE or the GMMAE is 250. The measurement
model is given by
 p

x2 + y 2
k =
+ vk
(37)
y
arctan(y/x)

40

30

20

10

10

15

20

25

30

35

40

45

50

Time (s)

Figure 3: Estimates of qy

Conclusions

In this paper a new approach for multiple-model

adaptive estimation was derived, which can be used
for time-varying and nonlinear systems. This approach is based on using the autocorrelation of the
measurement-minus-estimate residual. A formulation
was developed for the extended Kalman filter, however
computation of the off-diagonal elements of the autocorrelation matrix is intractable in practice since these
matrix elements are a function of the true parameters.
This difficulty was overcome by using the estimated
values in the Kalman gain. Another approach involves
using the estimated values in the state matrices and us-

ing the elements in the Kalman gain computation. Further studies are under way to assess the differences between these approaches. Simulation results indicated
that the new multiple-model adaptive estimation approach can provide better convergence properties than
the standard approach.

Acknowledgment
This work was supported through an Independent Research and Development (IRAD) award by L-3 Communications Integrated Systems, Greenville, TX, under the supervision of Dr. Gerald L. Fudge. The authors greatly appreciate the support.

[10] Y. H. Lho and J. H. Painter, A Fuzzy-Tuned

Adaptive Kalman Filter, Proceedings of the
Third International IEEE Conference on Industrial Fuzzy Control and Intelligent Systems, Houston, TX, Dec. 1993, pp. 144148.
[11] D.-J. Lee and K. T. Alfriend, Adaptive Sigma
Point Filtering for State and Parameter Estimation, AIAA/AAS Astrodynamics Specialist Conference and Exhibit , Providence, RI, Aug. 2004,
AIAA-2004-5101.
[12] D. T. Magill, Optimal Adaptive Estimation of
Sampled Stochastic Processes, IEEE Transactions on Automatic Control , Vol. AC-10, No. 4,
Oct. 1965, pp. 434439.

[1] R. E. Kalman and R. S. Bucy, New Results in

Linear Filtering and Prediction Theory, Journal
of Basic Engineering, March 1961, pp. 95108.

[13] P. D. Hanlon and P. S. Maybeck, Multiple-Model

Adaptive Estimation Using a Residual Correlation Kalman Filter Bank, IEEE Transactions on
Aerospace and Electronic Systems, Vol. AES-36,
No. 2, April 2000, pp. 393406.

[2] E. Wan and R. van der Merwe, The Unscented

Kalman Filter, Kalman Filtering and Neural
Networks, edited by S. Haykin, chap. 7, John Wiley & Sons, New York, NY, 2001.

[14] J. M. Hammersley, Monte Carlo Methods for

Solving Multivariable Problems, Annals of the
New York Academy of Sciences, Vol. 86, 1960,
pp. 844874.

[3] N. J. Gordon, D. J. Salmond, and A. F. M. Smith,

Novel Approach to Nonlinear/Non-Gaussian
Bayesian State Estimation, IEE Proceedings,
Part F - Communications, Radar and Signal Processing, Vol. 140, No. 2, Seattle, WA, April 1993,
pp. 107113.

[15] C. F. van Loan, Computing Integrals Involving

the Matrix Exponential, IEEE Transactions on
Automatic Control , Vol. AC-23, No. 3, June 1978,
pp. 396404.

References

[4] M. S. Arulampalam, S. Maskell, N. Gordon, and

T. Clapp, A Tutorial on Particle Filters for Online Nonlinear/Non-Gaussian Bayesian Tracking,
IEEE Transactions on Signal Processing, Vol. 50,
No. 2, Feb. 2002, pp. 174185.
[5] K. Pearson, Das Fehlergesetz und seine Verallgemeinerungen durch Fechner und Pearson. A Rejoinder, Biometrika, Vol. 4, 1905, pp. 169212.
[6] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navigation, John Wiley & Sons, New York, NY, 2001,
pp. 232245.
[7] B. J. Odelson, M. R. Rajamani, and J. B.
Rawlings, A New Autocovariance Least-Squares
Method for Estimating Noise Covariances, Tech.
Rep. 2003-04, Texas-Wisconsin Modeling and
Control Consortium, Sept. 2003.
[8] R. K. Mehra, On the Identification of Variances
and Adaptive Kalman Filtering, IEEE Transactions on Automatic Control , Vol. AC-15, No. 2,
April 1970, pp. 175184.
[9] A. G. Parlos, S. K. Menon, and A. F. Atiya, An
Algorithmic Approach to Adaptive State Filtering
Using Recurrent Neural Networks, IEEE Transactions on Neural Networks, Vol. 12, No. 6, Nov.
2001, pp. 14111432.

[16] R. G. Brown and P. Y. C. Hwang, Introduction

to Random Signals and Applied Kalman Filtering,
John Wiley & Sons, New York, NY, 3rd ed., 1997,
pp. 353361.
[17] R. F. Stengel, Optimal Control and Estimation,
Dover Publications, New York, NY, 1994, pp.
402407.
[18] B. D. O. Anderson and J. B. Moore, Optimal Filtering, chap. 10.1, Dover Publications, Mineola,
NY, 2005.
[19] B. D. O. Anderson, J. B. Moore, and R. M.
Hawkes, Model Approximations via Prediction
Error Identification, Automatica, Vol. 14, No. 6,
Nov. 1978, pp. 615622.
[20] L. Kocis and W. J. Whiten, Computational
Investigations of Low-Discrepancy Sequences,
ACM Transactions on Mathematical Software,
Vol. 23, No. 2, June 1997, pp. 266294.