Multi-parent recombination to overcome premature convergence in genetic algorithms

Recent research shows that enlarging the arity of recombination operators in a Genetic Algorithm lowers the probability of premature convergence. This results in more robust genetic function optimizers. In this paper we try to give an explanation why these multi-parent operators are better. In particular, we discuss two operators: the uniform scanning crossover and the diagonal crossover operator. First we show that these operators are better than the standard ones by testing them on an extensive test-suite of function optimization problems. Second, we explain the empirical results by first looking at the influence that the operators have on the evolution of populations, and then by using a new kind of description we are able to explain the convergence curves and rates of success as obtained in the experiments.