The Distribution Genetic Algorithm: Evolving a Population of Distributions

We propose an alternative to the traditional representation used for real coded genetic algorithms (GA): here chromosomes consist of a vector of distributions instead of values. Two systems have been devised: one using a version of blended crossover along with uniform mutation, the second using binary crossover with a "directed" mutation-like change in the distributions through a weighted aggregation of samples from the population. These systems are merged using a novel two-population approach. Our experimental results show that the proposed system improves a GA\´s performance on most of the 17 functions used in our test suite.