Diagonal Crossover in Genetic Algorithms for Numerical Optimization

In this paper the results of a detailed investigation on a multi-parent recombination operator, diagonal crossover, are reported. Although earlier publications have indicated the high performance of diagonal crossover on a number of problems, so far it has not been investigated whether high performance is indeed a result of using a high number of parents. Here we formulate three hypotheses to explain why GA performance increases when more parents are used. Based on an extensive study on a test suite containing eight numerical optimization problems we are able to establish that the higher number of parents is indeed one of the sources of higher performance, if and when this occurs. By the diversity of the test functions (unimodal, multi-modal, quasi-random landscapes) we can also make observations on the relationship between tness landscapes and operator performance.