Self-Adaptation - Self-Adaptive Heuristics for Evolutionary Computation - page 29

We conducted an experimental analysis of a particle swarm optimizer and ES
using design of experiments [73].
Nannen and Eiben [97], [98], [31] proposed the relevance estimation and value
calibration (REVAC) method to estimate the sensitivity of parameters and the
choice of their values. It is based on information theory, i.e. REVAC estimates
the expected performance when parameter values are chosen from a probability
density distribution

C

with maximized Shannon entropy. Hence, REVAC is an
estimation of distribution algorithm. It iteratively refines a joint distribution C
over possible parameter vectors beginning with a uniform distribution giving
an increasing probability to enlarge the expected performance of the underlying
EA. From a vertical perspective new distributions for each parameter are built
on estimates of the response surface, i.e. the fitness landscape. From a horizontal
point of view in the first step the parameter vectors are evaluated according to
the EA's performance, in the second step new parameter vectors are generated
with superior response.

Metaevolution

In metaevolutionary algorithms or nested EAs the evolutionary optimization
process takes place on two different levels [115]. The outer EA optimizes pa-
rameters of the inner EA. When other adaptation methods such as adaptive or
self-adaptive parameter control fail, nested EAs can be used to overcome the
parameter control problem. Furthermore, nested EAs can be used for the exper-
imental analysis of appropriate parameter settings. The PhD thesis of Kursawe
[80] gives an overview of methods using nested or related approaches. He an-
alyzed appropriate ES parameter settings, e.g. for the mutation parameter τ ,
with a nested ES. Coello [24] makes use of nested ES for adapting the factors of
a penalty function for constrained problems.
Figure 3.2 shows the pseudocode of a nested [ μ /ρ + ,λ ( μ/ρ + ,λ ) γ ]-ES. In
this algorithm μ parents represent the settings for the inner ( μ/ρ + ,λ )-ES and
reproduce λ offspring solutions. Each ( μ/ρ + ,λ )-ES runs for γ generations,
denoted as isolation time .Thebest μ settings form the parents for the next
generation on the outer level.

3.2.3 Parameter Control

Parameter control means the change of evolutionary parameters during the run.
Typical is the control of strategy parameters, but also classical exogenous pa-
rameters like population sizes can be controlled during the run.

Deterministic

Deterministic parameter control means that the parameters are adjusted ac-
cording to a fixed time scheme. Usually, the number of generations controls the
deterministic parameter control. E.g. it can be useful to reduce the mutation
strength during the evolutionary search in order to enable convergence of the