Mitchell, T. J. (2010) An exploration of evolutionary computation ...

object parameters, within what is termed phenotype space. In contrast, traditional GAs
employ a binary representation, in which object parameters are encoded into discrete,
usually fixed length, bitstrings. These algorithms are said to work in genotype space, and
require functions that map individuals between genotype and phenotype space.
Before the reproductive cycle may begin, it is necessary to initialise the system‘s
population by generating separate random numbers for each individual. Thereafter,
genetic material from the parent population is blended via recombination to generate
offspring, which are subsequently varied by means of mutation. Mutation is implemented
with the random perturbation of offspring, to introduce chance positive stochastic
variation. Each offspring is then evaluated as a solution to the objective function
assigned a fitness quotient in proportion to its performance. New parents are selected based
on their relative fitness, ensuring that high-performing individuals are then chosen to take
part in the next round of variation more frequently than low-performing individuals.
A widely accepted viewpoint of the evolutionary process considers selection to encourage
the exploitation of high-fitness regions of the solution space, while recombination and
mutation facilitate the exploration of new regions which are not currently represented by
population. This interplay of exploitation and exploration directs the evolving population
towards higher levels of fitness, and thus, evolutionary computation has several advantages
over more traditional optimisation methods. For example, enumerative and random-based
optimisation techniques are only capable of exploration; consequently the process of
optimisation is costly. Hill-climbing-based techniques only exploit and are therefore
susceptible to becoming trapped within local optima. The implementation of both search
tactics within EAs offers a heuristic optimisation method, which is both robust and
efficient.
However, EAs do not provide a universal solution to all optimisation problems; there are
certain problem characteristics for which evolutionary algorithms are not well suited. In his
study of epistasis, Davidor (1991) identifies two environmental extremes for which EAs
have no advantage over more traditional optimisation methods. At one extreme the
problem is so well structured and easy to solve that an EA would be unable to perform
better than a hill-climber. At the other extreme, the problem domain is so complex and
unstructured that an EA would be unable to perform better than a random strategy. Davidor
concludes that application domains with characteristics between these two extremes are the
types of problem for which optimisation by EA might be advantageous.
and
13