Variations in GA Breeding
Traditionally, all of the parent individuals in one generation are
replaced by their children in the next generation. This is known as
Generational Replacement. Many researchers, however, now employ a
model known as Steady State Replacement in which only a small
number of new individuals are introduced in each generation, thus
requiring only a small number to be eliminated (often, the least fit are
most likely to face death). This potentially allows an individual to mate
with members of his parents' generation (or his offspring's generation),
while under Generational Replacement an individual can only mate with
members of his own generation.
The fitness function in a genetic algorithm assigns some sort of numeric
rating to each of the individuals it judges. Tournament Selection and
Fitness Remapping use data supplied by the fitness function to determine
which individuals actually reproduce.
- Fitness Remapping, of which there are several variations,
assigns a
probabilty to each individual based on its judged fitness, with fit
individuals earning a high probability of being selected for reproduction
and unfit ones getting a low probability.
- Tournament Selection omits the need for a second function to
determine probabilties and a third to actually run the probabilities to
choose the breeders. In a tournament, a certain number of individuals are
picked from the initial population and their fitnesses compared. The best
is then copied (not moved) into the group selected for mating.
All individuals are then replaced back into the initial generation.
Tournaments continue in this way until enough breeders have been
selected.