Information Technology Reference
In-Depth Information
number of generations, also see figure 3.5. The step sizes have to be adapted in
order to speed up the optimization process. In the approach of the 1/5th-rule
the whole populations makes use of the same
σ
for all individuals. It states that
the ratio of successful mutations should be 1/5th. If the ratio is greater than
1/5th the step size should be increased, if it is less than 1/5th the step size
should be decreased. This rule is applied every
g
generations. Figure 3.3 shows
the pseudocode of Rechenberg's adaptation method. It must hold 0
<λ<
1. The
1. Perform the (1+1)-ES for a number g of generations:
•
Keep
σ
constant during this period,
•
Count the number
s
of successful mutations during this period
2. Estimate success rate
p
s
by
p
s
:=
s/g
(3.2)
3. Change
σ
according to
⎨
⎩
σ/λ,
if
p
s
>
1
/
5
σ
:=
σ
·
λ,
if
p
s
<
1
/
5
(3.3)
σ,
if
p
s
=1
/
5
4. Go to step 1.
Fig. 3.3.
Pseudocode of the 1/5th rule of Rechenberg
aim of this approach is to stay in the so called
evolution window
guaranteeing
nearly optimal progress. Of course, the 1/5th rule is not the proper heuristic for
every problem and is outperformed by self-adaptive step size approaches. The
optimal value for the factor
a
depends on several factors such as the number of
generations g and the dimension of the problem
N
.
Self-Adaptive
The concept of self-adaptation is a step into the direction of parameter indepen-
dent EC. Self-adaptation is usually associated with the step size control of ES.
It was originally introduced by Rechenberg and Schwefel [132], later by Fogel
for EP. Self-adaptive algorithms are characterized by the integration of strategy
parameters which influence the genetic operators into the individuals' chromo-
somes. These parameters undergo genetic variation and are therefore involved
into the evolutionary optimization process
Hybridization
The hybridization of different parameter control techniques is a source for many
interesting variants. An example is the DSES, a heuristic for constraint handling,
see chapter 7.4. The self-adaptive mutation strength control of ES fails in the
Search WWH ::
Custom Search