Biomedical Engineering Reference
In-Depth Information
called generation procedures were later termed—in
the context of supervisory programs, that is, pro-
grams that allow a system to adapt to different
environmental conditions— mutation . After intro-
ducing differential selection as a key driving force
for adaptation, “Adaptation, then, is based upon
differential selection of supervisory programs.
That is, the more 'successful' a supervisory pro-
gram, in terms of the ability of its problem-solving
programs to produce solutions, the more predomi-
nant it is to become (in numbers) in a population
of supervisory programs. [...] Operation of the
selection principle depends upon continued gen-
eration of new varieties of supervisory programs.
There exist several interesting possibilities for pro-
ducing this variation” [9, pp. 300-301] . “The pro-
cedure described [...] requires that the supervisory
program duplicate, with some probability of vari-
ation or mutation [...]” [9, p. 309] .
Thus, as early as 1962, a sketch of evolutionary
algorithms was in place that could only become
more pronounced over the years [10] . The ran-
domness of Turing's paper, later reflected in
Friedberg's work [11] , gave way to a variation-
selection loop with accumulation of beneficial
variations in a population and the regular infor-
mation exchange between individuals.
Other paradigmatic developments in evolu-
tionary algorithms at the time include evolution-
ary programming [12] and evolutionary strategies
[13] . For a more thorough review of early work
in evolutionary computing, the reader is pointed
to Ref. 14 , discussing a selection of papers from
the fossil record of evolutionary computing.
All evolutionary algorithms follow the
Darwinian principle of differential natural selec-
tion. This principle states that the following pre-
conditions must be fulfilled for evolution to
occur via (natural) selection:
3. In the course of reproduction there is
variety that affects the likelihood of survival
and, therefore, the reproducibility of
individuals. This variety is produced by
stochastic effects (such as random mutation
and recombination) as well as by systematic
effects (mating of like with like, etc.).
4. There are finite resources that cause the
individuals to compete. Due to overrepro-
duction of individuals, not all can survive
the struggle for existence. Differential
natural selection is a result of this competi-
tion exerting a continuous pressure toward
adapted or improved individuals relative to
the demands of their environment.
In evolutionary algorithms, there is a popula-
tion of individual solutions. Usually this popula-
tion is initialized as a random population,
i.e., from random elements determined to be
potentially useful in this environment. Next is a
determination of fitness in the process of evalua-
tion. This could take any form of a measurement
or calculation to determine the relative strength of
an individual solution. The outcome of this meas-
urement or calculation is then used in the selec-
tion step to determine which individual solutions
are to survive the competition for resources and
Initialization
?
Evaluation
Termination
Variation
Selection
1. There are entities called individuals that
form a population. These entities can
reproduce or can be reproduced.
2. There is heredity in reproduction; that is to
say that individuals of this population
produce similar offspring.
Reproduction
FIGURE 17.1 The general process of an evolutionary
algorithm is a cycle of evaluation, selection, reproduction,
and variation that accumulates beneficial changes. Variation
operators could be mutation, duplication, or crossover/
recombination.
Search WWH ::




Custom Search