EA stability visualization: perturbations, metrics and performance

It is well-known that Evolutionary Algorithms (EAs) are sensitive to changes in their control parameters, and it is generally agreed that too large a change may turn the EA from being successful to unsuccessful. This work reports on an experimental hybrid visualization scheme for the determination of EA stability according to perturbation of EA parameters. The scheme gives a visual representation of local neighborhoods of the parameter space according to a choice of two perturbation metrics, relating perturbations to EA performance as a variant of Kolmogorov distance. Through visualization and analysis of twelve thousand case study EA runs, we illustrate that we are able to distinguish between EA stability and instability depending upon perturbation and performance metrics. Finally we use what we have learned in the case study to provide a methodology for more general EAs.