A theoretical and empirical study on unbiased boundary-extended crossover for real-valued representation

We present a new crossover operator for real-coded genetic algorithms employing a novel methodology to remove the inherent bias of pre-existing crossover operators. This is done by transforming the topology of the hyper-rectangular real space by gluing opposite boundaries and designing a boundary extension method for making the fitness function smooth at the glued boundary. We show the advantages of the proposed crossover by comparing its performance with those of existing ones on test functions that are commonly used in the literature, and a nonlinear regression on a real-world dataset.