Digital Signal Processing Reference
In-Depth Information
Tabl e 1
Comparison of ordering and operating point selection algorithms
Ordering and Operating
System
Utility
Message
Speed of
Adapt- Control
Point Selection Algorithm compliance achieved
exchange convergence ability
SQP-Greedy
Low
Bound; Local Opt. Heavy
Medium
Little
∅
Safe Experimentation
High
Local Opt.
∅
Medium
∅∅
Partial Search
Complete
Local Opt.
Light
Rapid
Total
Yes
Tabl e 2
Utilities and computational time achieved for different ordering algorithms
Algorithm
Order obtained
Utility
Comp. time
Centralized
Optimal
[
C
6
C
2
C
1
C
4
C
3
C
7
C
5
]
100
>
5min
Greedy
[
C
4
C
1
C
6
C
7
C
2
C
3
C
5
]
95
0.002s
Decentralized
Safe Experimentation
[
C
6
C
2
C
1
C
4
C
3
C
7
C
5
]
100
2.09
Partial Search
[
C
6
C
2
C
1
C
4
C
3
C
7
C
5
]
100
1.2s
4.5.2
Comparison of Ordering and Operating Point Selection Algorithms
Our preliminary results in Table
1
compare the performance of several joint
ordering and operating point selection algorithms based on important criteria in the
considered stream mining system.
4.5.3
Order Selected by Various Classifiers for Different Ordering
Algorithms
The performance of the different ordering algorithms are shown in Table
2
for 7
classifiers with fixed operating points per classifier. The classifier's characteristics
(
p
F
p
D
(i.e. the
resource requirements) were generated randomly. The misclassification cost
c
M
,
)
(i.e. the DET curve),
ψ
(i.e. the
ex-post
selectivities), and
α
=
10, false alarm cost
c
F
=
1, and
λ
=
0
.
1. The input data rate
t
0
=
g
0
was selected to
normalize the optimal utility to 100.
As expected, the globally optimal centralized solution requires too much com-
putation time, while the centralized greedy algorithm does not lead to the optimal
order, but results in very little computational time. The Parametric Partial Search
Algorithm (here with
p
=
0
,
1,
T
=
1and
β
=
0) converges quicker than Safe
Experimentation (here with
1), to the optimal order. Decentralized algorithms
converge to the optimal order, given that they ultimately probe all the possible
orders, but they require longer computational time than the centralized greedy
only a few iterations.
ε
=
0
,
