Enhancing the Adaptive Dissortative Mating Genetic Algorithm in Fast Non-stationary Fitness Functions

摘要

The Adaptive Dissortative Mating Genetic Algorithm (ADMGA) is a variation of the standard GA in which a mating restriction based on the genotypic similarity of the individuals is introduced. The algorithm mimics a mating strategy often found in nature: dissimilar individuals mate more often than expected by chance and, as a result, genetic diversity throughout the run is maintained at a higher level. ADMGA has been previously applied to non-stationary fitness function, performing well when the changes hit the function at a medium and slow rate, while being less effective when the frequency is higher. Due to the premises under which the algorithm was tested, it has been argued that the replacement strategy that results from the implementation of the dissortative mating strategy may be harming the performance when solving high-frequency dynamic problems. This paper investigates alternative replacement strategies for ADMGA with the objective of improving its performance on this class of non-stationary problems. The strategies maintain the simplicity of the algorithm, i.e., the parameter set is not increased. The replacement schemes were tested in dynamic environments based on stationary functions with different frequency and severity, showing that it is possible to improve standard ADMGA's performance in fast dynamic problems by simple modifications of the replacement strategy.