Using recurrent selection to improve GA performance

Abstract

Genetic algorithms (GA's) are based on the idea that solutions to otherwise intractable problems can be derived by mimicking natural evolution. With a few exceptions, however, GA's are limited to haploid implementations with random breeding among a single population, failing to exploit a number of strategies that are found in nature. Plant breeders recognize that too-strict selection among the progeny in a given generation results in reduced diversity, which can, in turn, cause inbreeding depression, a decline in the average fitness of the population. Recurrent selection is a multistep breeding method designed to solve this problem, improving a plant population's fitness without excessive loss of diversity. This paper explored the use of recurrent selection for Genetic Algorithms. We see major advantages in this approach in oscillating environments which cycle among several different states as well as in stationary environments. The main advantage of recurrent selection is the maintenance of genetic diversity in even relatively small populations. We demonstrated the benefits of using recurrent selection for the 0/1 knapsack problem with changing weights and with the Schwefel function. Further research on harder dynamic and stationary problems is called clearly called for, as is further application of models based on a horticultural model.