 |
|
Roulette Wheel selection is the first used and most popular operator. A selection probability proportional to its fitness is assigned to each individual in the population. If no scaling, i.e. adimensionalisation of fitness in the [min, max] range, is applied premature convergence occur that can be avoided only with large population size. The operator is robust but computationally expensive and is not used in modeFRONTIER.TM
Tournament Selection overcome the problem of fitness scaling with direct comparison of fitness value: on a "tsize" group of individuals the best is selected. This type of selection is generally considered more efficient and more robust than roulette wheel.
Local Geographic Selection [12], elsewhere named as step-stone island model, is a particular case of Tournament Selection. The "n-size" individuals participating to the tournament are not selected randomly in the population but through a local random walk in the neighborhoods of a given individual being the population distributed in a N dimensional grid.