Online convergence detection for evolutionary multi-objective algorithms revisited

The design and application of termination criteria has become an important aspect in evolutionary multi-objective optimization. Online convergence detection (OCD) determines when further generations are no longer promising based on statistical tests on a set of performance indicators. The behavior of OCD mainly depends on two parameters, the number of preceding generations considered in the statistical tests and the desired variance limit. In this paper, guidelines for selecting appropriate combinations of these parameters are empirically derived based on design-of-experiment methods. Furthermore, a variant of OCD is introduced which directly operates on the hypervolume indicator - the internal measure of the SMS-EMOA. This allows a separated analysis of the variance criterion and reduces the complexity of OCD. Based on the experimental design, a systematic comparison with the classical OCD approach is performed and differences between the appropriate parameterizations of both variants are highlighted.