A preliminary investigation into directed mutations in evolutionary algorithms

Abstract

The traditional mutation operator within evolution strategies and evolutionary programming relies on adding a multivariate zero mean Gaussian random vector to each parent solution. An alternative method is proposed that allows for optimizing the direction of such mutations. The notion of mutation in polar coordinates is adopted such that parents generate offspring in a selected direction with a random step size. Experiments on four functions suggest that the independent adjustment of direction of travel and step size can produce improvements in rate of convergence on some functions.