Genetic algorithms, schemata construction and statistics

Abstract

The paper performs a comparison between two types of binary Genetic Algorithms (GA): Statistical GA vs. Messy GA. They have the same challenge — solving “deceptive” problems — and, up to a point, they are designed on the same paradigm: improving the GA by directing the search using some statistical derived schemata. The Statistical GA that we propose does not use a larger population, but only a real-valued string with the average numbers of “ones” produced on each position of the chromosome during the GA's evolution. Assuming the stagnation of the GA in sub-optimal points (which is usually the case of deceptive problems), we extract — by imposing a threshold on the real-valued string — the schema responsible for stagnation, derive its complementary schema and resume the GA's evolution imposing that schema to all the new chromosomes.