Summary: | 碩士 === 國立交通大學 === 資訊工程研究所 === 83 === Genetic algorithms are adaptive search techniques that are
inspired from natural selection, i.e. survival of the fittest.
A genetic algorithm determines how fit an individual is by its
evaluation value of the objective function, i.e. a fitness
function of objective values. With reproduction through
generation after generation, the adaptiveness is achieved by
means of the change in the proportions of individuals in a
population. But when a population is subjected to long
convergence in a mature run, the population will consist of
similarly good individuals. Then the adaptiveness of a genetic
algorithm will be lost due to the problem of declining
selective pressure. We proposed a translation invariant and
nonmonotonic fitness function to try to solve the problem of
declining selective pressure. This new fitness function uses
improvement of objective values in replace of using objective
values themselves as in traditional genetic algorithms. The
improvement is obtained by computing the difference of an
individual's new objective value and its old objective value.
Because every offspring is produced by two parents, we use the
average of the two parents' objective values as an individual's
old objective value. Then we translate the improvement by
adding the negative value of the minimum improvement to
guanrantee the positiveness of the fitness function. The
empirical results showed that it succeeded in fighting the
problem of declining selective pressure. In addition, the
nonmonotonically way in which the new fitness function attacks
an objective function provides different clues to complement
the traditional genetic algorithms.
|