Digital Signal Processing Reference
In-Depth Information
as a tracking filter. The NSPP-based detector threshold optimization for the PDAF
case appeared in the works [
5
,
20
,
22
,
48
], and [
2
] which are all given in Fig.
6.2
.In
this section, we define the underlying optimization problems of these approaches.
The presentation is given as in Fig.
6.2
, i.e., in two parts, static and dynamic opti-
mization schemes. The section ends with a comparison of these approaches via a
simulation scenario.
6.3.1 Static Threshold Optimization (STOP)
The NSPP-based static threshold optimization problem is [
2
] to determine the opti-
mal
P
FA
value such that
P
FA
f
S
[
P
NSPP
]
P
FA
=
arg min
(6.26)
subject to
P
D
=
f
ROC
(P
FA
,ζ)
and
0
≤
P
FA
≤
1
,
n
x
×
n
x
where
f
S
: R
is an appropriate scalar measure deduced from a matrix
(such as trace, determinant, or a matrix norm) and
P
NSPP
is the steady-state covari-
ance matrix obtained by propagating one of the NSPP recursions to its steady-state,
i.e.,
→ R
P
NSPP
P
NSPP
(k
|
k),
lim
k
→∞
(6.27)
where
P
NSPP
(k
k)
corresponds to the output of either the HYCA or the MRE algo-
rithm at time step
k
. The equality constraint of the optimization problem is nothing
but an ROC curve relation which links
P
D
to
P
FA
, or vice-versa, through current
SNR (
ζ
), and the inequality constraint ensures that the resultant operating false
alarm value is a valid probability.
|
Remark 6.1
Note that this optimization is performed
offline
. The MRE and
HYCA recursions are initialized as explained in the Sects.
6.2.3
and
6.2.4
.In
the practical implementation, one cannot iterate the recursion given in (
6.27
)
indefinitely. One should check whether the value of a suitably chosen norm of
the difference matrix between two consecutive covariance matrices is below
a chosen threshold to conclude that the recursion is converged, or whether a
maximum number of iterations is reached to conclude on its divergence.
The optimization problem given in (
6.26
) is a line search. Provided that the cost
function is
unimodal
, the global optimum point can be found directly applying well-
known numerical techniques, such as the Golden-Section or Fibonacci Search meth-
ods [
11
]. For each function evaluation at an arbitrary point
P
FA
, one needs to obtain
Search WWH ::
Custom Search