Evolutionary Computation - Computational Intelligence in Time Series Forecasting - page 210

promotes the process of reproduction by mating two given solutions and producing
a new one. It is expected that in this way good offspring are generated.
To estimate how far the generated offspring are good, the selected fitness
criteria are used. This process is repeated, producing better and better offspring by
mating and mutation operations. Although the evolutionary strategies are valuable
search concepts, they still have their limitations and drawbacks: in practical
applications, many decisions have to be made in the selection of an initial solution
set, the application of appropriate genetic operators at each evolutionary step, the
definition of an adequate fitness function, etc .
5.5 Evolutionary Programming
L.J. Fogel (Fogel et al. , 1966), in his search for a new evolutionary method for
developing artificial intelligence, elaborated a stochastic optimization methodology
relying on genetic principles that was later formulated by D. Fogel (1994) as
evolutionary programming . The new methodology differs substantially both from
genetic algorithms and genetic programming in that it evolves behavioural models
rather than genetic models . Hence, the objective of evolutionary programming is to
find a set of best behavioural models from a space of possible behavioural models.
Like other evolutionary methods, evolutionary programming also relies on
some repeated operational steps that are interrupted (before the next step
commences) by the evaluation of the results achieved using a fitness function. But
still, evolutionary programming is different from other genetic methods in that it
uses a population of parents , each of them producing a single offspring through
mutation, because in evolutionary programming no crossover operator is
implemented.
The algorithm of evolutionary programming can be outlined as follows:
x Generate randomly the initial generation as a set of initial problem
solutions and calculate the fitness value of each individual of the
population.
x For each individual (problem solution)
x generate a new solution set by copying the set and changing it
genetically
x calculate the fitness of each individual/new solution
x Store the new solution and fitness.
x From the new generation select the solution with the largest fitness and
delete the rest.
x If the best or nearly the best solution is found, stop the evolutionary
process; otherwise continue.
In practice, before the above search for the optimal solution runs, the population
size and the number of iterations ( i.e. number of generations) have to be fixed.
Also, the mutation operator to be used for generation of the next solution is to be
determined. This can be extended by integrating a randomly selected maturation