Self-Adaptation - Self-Adaptive Heuristics for Evolutionary Computation - page 28

parameters
- exogenous
- endogenous
- population level
- individual level
- component level

parameter setting

parameter tuning parameter control

by hand DOE metaevolutionary deterministic adaptive self-adaptive

human
experience

sequential parameter
optimization
response surface
modeling

metaevolutionary
angle control ES
coevolutionary
approaches

dynamic penalty
function
annealing penalty
function

Rechenberg's
1/5th rule
covariance matrix
adaptation (CMS/CSA)

step size control
of ES
SA-PMX

methods

hybridization

Fig. 3.1. An extended taxonomy of parameter setting techniques based on the taxon-
omy of Eiben [37], complemented on the parameter tuning branch. For each parameter
setting class a couple of methods are presented exemplarily.

3.2.2 Parameter Tuning

Many parameters for EAs are static, which means that once defined they are not
changed during the optimization process. Typical examples for static parame-
ters are population sizes or initial strategy parameter values. Examples for static
parameters of heuristic extensions are static penalty functions. In the approach
of Homaifar, Lai and Qi [60], the search domain is divided into a fixed num-
ber of areas assigned with static penalties representing the constraint violation.
The disadvantage of setting the parameters statically is the lack of flexibility
concerning useful adaptations during the search process.

By Hand

In most cases static parameters are set by hand. Hence, their influence on the
behavior of the EA depends on human experience. Even though the user defined
settings might not be the optimal ones.

Design of Experiments, Relevance Estimation and Value Calibration

Design of experiments (DOE) offers the practitioner a way of determining opti-
mal parameter settings. It starts with the determination of the objectives of an
experiment and the selection of the parameters (factors) for the study. The qual-
ity of the experiment (response) guides the search to find appropriate settings.
Recently Bartz-Beielstein et al. [5] developed a parameter tuning method for
stochastically disturbed algorithm output, the sequential parameter optimiza-
tion (SPO). It combines classical regression methods and statistical approaches
for deterministic algorithms like design and analysis of computer experiments.