Outperforming Mutation Operator with Random Building Block Operator in Genetic Algorithms

Abstract

The refinement process in genetic algorithms is carried out mainly by crossover and mutation operators. In their classical forms these operators need to be tuned through parameters and they are not efficient enough. Moreover, lack of sufficient variation in the population causes genetic algorithms to stagnate at local optima. In this work a new dynamic mutation operator called random building block operator with variable mutation rate proportionate to the number of variables in the problem is proposed. This operator does not require any pre-fixed parameter. At runtime it dynamically takes into account the length of the binary presentation of the individual and the number of variables in the problem and replaces a randomly selected section of the individual by a randomly generated bit string of the same size. Experimentation with 33 test functions, 231 test cases and 11550 test runs proved the superiority of the proposed dynamic mutation operator over single-point mutation operator with 1%, 5% and 8% mutation rates and the multipoint mutation operator with 5%, 8% and 15% mutation rates. Based on the experimentation results the random building block operator can be proposed as a better substitution of single-point and multipoint mutation operators.