High-Probability Mutation in Basic Genetic Algorithms

Abstract

Customarily, Genetic Algorithms use lowprobability mutation operators. In an effort to increase their performance, this paper presents a study of Genetic Algorithms with very high mutation rates (≈ 95%) . A comparison is drawn, relative to the low-probability (≈ 1%) mutation GA, on two large classes of problems: numerical functions (well-known test functions such as Rosenbrock's, Six-Hump Camel Back) and bit-block functions (e.g. Royal Road, Trap Functions). A large number of experimental runs combined with parameter variation provide statistical significance for the comparison. The high-probability mutation is found to perform well on most tested functions, outperforming low-probability mutation on some of them. These results are then explained in terms of dynamic dual encoding and selection pressure reduction, and placed in the context of the No Free Lunch theorem.