Adaptive Group Mutation for Tackling Deception in Genetic Search

In order to study the efficacy of genetic algorithms (GAs), a number of fitness landscapes have been designed and used as test functions. Among these functions a family of deceptive functions have been developed as difficult test functions for comparing different implementations of GAs. In this paper an adaptive group mutation (AGM), which can be combined with traditional bit mutation in GAs, is proposed to tackle the deception problem in genetic searching. Within the AGM, those genes that have converged to certain threshold degree are adaptively grouped together and subject to mutation together with a given probability. To test the performance of the AGM, experiments were carried out to compare GAs that combine the AGM and GAs that use only traditional bit mutation with a number of suggested “standard” fixed mutation rates over a set of deceptive functions as well as non-deceptive functions. The results demonstrate that GAs with the AGM perform better than GAs with only traditional bit mutation over deceptive functions and as well as GAs with only traditional bit mutation over non-deceptive functions. The results show that the AGM is a good choice for GAs since most problems may involve some degree of deception and deceptive functions are difficult for GAs. Key-Words: Genetic algorithm, adaptive group mutation, bit mutation, deceptive functions, building blocks.