Self-Adaptation - Self-Adaptive Heuristics for Evolutionary Computation - page 26

3 Self-Adaptation
The adjustment of parameters and adaptive operator features is of crucial impor-
tance for reliable results to given problems and the eciency of the evolutionary
heuristics. But furthermore, proper parameter settings are important for the
comparison of different algorithms on given problems. How can evolutionary pa-
rameters be controlled, how can they be tuned? This chapter gives an insight
into parameter adaptation techniques. Beginning with the history of adaptation
in EA, an extended taxonomy of parameter setting techniques is presented. Typ-
ically adapted components of EAs are presented. Afterwards, the chapter defines
self-adaptation in a classic way and from a point of view of estimation of distri-
bution algorithms, i.e. it gives insight into generalized views on self-adaptation
and its limitations.
3.1 History of Parameter Adaptation
Historically, the development of adaptation mechanisms began in 1967, when
Reed, Toombs and Baricelli [116] learned playing poker with an EA. The genome
contained strategy parameters determining probabilities for mutation and
crossover with other strategies. In the same year, Rosenberg [119] proposed to
adapt the probability for applying crossover.
With the 1/5th rule Rechenberg [114] introduced an adaptation mechanism
for step size control of ES, see section 3.2.2. One of the earliest realization of
self-adaptation can be found in the work of Bagley [4], who was the first inte-
grating the control parameters into the representation of the individuals. To-
day, the term self-adaptation is commonly associated with the self-adaptation
of mutative step sizes for ES and was introduced by Schwefel [132] in 1974.
For numerical representations the self-adaptation of mutation parameters seems
to be an omnipresent feature. After ES, Fogel introduced self-adaptation to EP
[43]. However, for binary-coded EAs self-adaptation has not grown to a standard
method. Nevertheless, there are several approaches to incorporate it into binary
representations, e.g. by Back [6], [7], Smith and Fogarty [144], Smith [142] and
Stone and Smith [148]. Schaffer and Morishima [128] let the number and location