Information Technology Reference
In-Depth Information
0
3
6
9
12
15
18
21
i
Fig. 4.
Example of an incremental swarming slope where
g
describes the best fitness of the actual
evaluation step
i
Min
describes the relationship of the minimum and the corresponding increment. There
are two subtypes for this feature.
μ
IP.Min
is divided by the slope of the model's straight
line by the spread of the first increment whereas the second subtype
μ
IP.MinQ
divides
the slope by the interquartile range of the first increment. The features
Incremental
Max
and
Incremental Range
are handled accordingly. Three additional features
are created by separately looking at the fitness values of the individual increments. The
fitness values of each increment is sorted in ascending order and normalized into the
interval
[0
,
1]
. This results in a sequence
x
k
=
x
1
,...,x
k
and we calculate a measure
of linearity by
x
i
−
2
k
i
−
1
μ
IP.Fit
=
(3)
k
−
1
i
=1
where
∀
i<j
:
x
i
<x
j
.
3.3
Incremental Swarming
The features of
Incremental Swarming
use the dynamic behavior of PSO to extract fea-
tures of the objective function. Therefore, we construct a small swarm of two particles
and initiate an optimization run. The particles are initialized with a defined distance
to each other. We use a inertia PSO with parameter
θ
=(0
.
6221
,
0
.
5902
,
0
.
5902)
and
record the best solution found in the first
t
=20
iterations. To get the parameter set
θ
we evaluated a few parameter sets empirically to find good values which lead to a fast
convergence of the small swarm. The spread of the global best fitness is the difference
between the first and the last fitness value. The development of the global best fitness
