Self-adjusting the intensity of assortative mating in genetic algorithms

Abstract

Mate selection plays a crucial role in both natural and artificial systems. While traditional Evolutionary Algorithms (EA) usually engage in random mating strategies, that is, mating chance is independent of genotypic or phenotypic distance between individuals, in natural systems non-random mating is common, which means that somehow this mechanism has been favored during the evolutionary process. In non-random mating, the individuals mate according to their parenthood or likeness. Previous studies indicate that negative assortative mating (AM)--also known as dissortative mating--, which is a specific type of non-random mating, may improve EAs performance by maintaining the genetic diversity of the population at a higher level during the search process. In this paper we present the Variable Dissortative Mating Genetic Algorithm (VDMGA). The algorithm holds a mechanism that varies the GA's mating restrictions during the run by means of simple rule based on the number of chromosomes created in each generation and indirectly influenced by the genetic diversity of the population. We compare VDMGA not only with traditional Genetic Algorithms (GA) but also with two preceding non-random mating EAs: the CHC algorithm and the negative Assortative Mating Genetic Algorithm (nAMGA). We intend to study the effects of the different methods in the performance of GAs and verify the reliability of the proposed algorithm when facing an heterogeneous set of landscapes. In addition, we include the positive Assortative Mating Genetic Algorithm (pAMGA) in the experiments in order test both negative and positive AM mechanisms, and try to understand if and when negative AM (or DM) speeds up the search process or enables the GAs to escape local optima traps. For these purposes, an extensive set of optimization test problems was chosen to cover a variety of search landscapes with different characteristics. Our results confirm that negative AM is effective in leading EAs out of local optima traps, and show that the proposed VDMGA is at least as efficient as nAMGA when applied to the range of our problems, being more efficient in very hard functions were traditional GAs usually fail to escape local optima. Also, scalability tests have been made that show VDMGA ability to decrease optimal population size, thus reducing the amount of evaluations needed to attain global optima. We like to stress that only two parameters need to be hand-tuned in VDMGA, thus reducing the tuning effort present in traditional GAs and nAMGA.