Next: Selection of Evolution Parameters
Up: Alvarez: Genetic Algorithm Inversion
Previous: Parameter Summary
When dealing with high dimensionality problems, it may be difficult or too time
consuming for all the model parameters to converge within a given margin of
error. In particular, as the number of model parameters increases, so does the
required population size. Recall that large population sizes imply large numbers
of cost-function evaluations.
An alternative is the use of micro-genetic algorithms Krishnakumar (1989),
which evolve very small populations that are very efficient in locating
promising areas of the search space. Obviously, the small populations are
unable to maintain diversity for many generations, but the population can be
restarted whenever diversity is lost, keeping only the very best fit
individuals (usually we keep just the best one, that is, elitism of one
individual). Restarting the population several times during the run of the
genetic algorithm has the added benefit of preventing premature convergence due
to the presence of a particularly fit individual, which poses the risk of
preventing further exploration of the search space and so may make the program
converge to a local minimum. Also, since we are not evolving
large populations, convergence can be achieved more quickly and less memory
is required to store the population.
Next: Selection of Evolution Parameters
Up: Alvarez: Genetic Algorithm Inversion
Previous: Parameter Summary
Stanford Exploration Project
11/11/2002