Adaptive Genetic Algorithms - Computational Intelligence in Time Series Forecasting - page 313

association with the concepts of adaptive control of dynamic systems, the idea was
widely accepted to work on adaptive GAs by incorporating the parameter
adaptation mechanism in conventional GAs. Different researchers have
concentrated their efforts on implementing genetic algorithms with different
parameters tuned. Herrera and Lozano (1996) later classified the proposed adaptive
GA systems as systems with
x adaptive parameter setting
x adaptive genetic operators
x adaptive operator selection
x adaptive representation
x adaptive fitness function.
In practical realizations of adaptive GA approaches it should first be decided at
what algorithm level the adaptation should work (Smith and Fogarty, 1997), i.e .
should it be at the
x population level , at which the global GA parameters of all individuals of
the population are on-line adapted
x individual level , at which the strategy parameters, usually mutation and
crossover, are adapted only in some elected population individuals in order
to effectuate only elected individuals
x component level , at which the strategic parameters of some components or
of some genes of population individuals are locally varied?
9.2 Genetic Algorithms Parameters to Be Adapted
Adaptive versions of genetic algorithms are particularly needed because, in the
process of the evolutionary search, the algorithm should converge to the global
optimum with a high speed of convergence, so that the global optimum value is
found in the minimum number of steps, i.e. it should be finished after a minimum
number of generations treated. This is usually achievable by on-line adapting of the
control parameters of the algorithm, such as the probability of crossover,
mutation, or of reproduction. Several empirical and theoretical studies devoted to
identifying the optimal mode of parameter settings for genetic algorithms (DeJong,
1985; Grefenstette, 1986; Hesser and Manner, 1990) have resulted in the following
general assessments:
x Crossover . This parameter controls the rate at which the solutions are
subjected to crossover effects. When its value is increased, new solutions
are more rapidly introduced into the population. Through this, the search
process can become so fast that it can be disrupted.
x Mutation. This parameter restores the genetic material and transforms the
GA - when its value is increased too much - into a purely random search
algorithm, whereas, a small value of the mutation parameter is required to
prevent the premature convergence of the GA to a suboptimal solution.