Comparison of Steady-State and Generational Evolution Strategies for Parallel Architectures

Steady-state and generational selection methods with evolution strategies were compared on several test functions with respect to their performance and efficiency. The evaluation was carried out for a parallel computing environment with a particular focus on heterogeneous calculation times for the assessment of the individual fitness. This set-up was motivated by typical tasks in design optimization. Our results show that steady-state methods outperform classical generational selection for highly variable evaluation time or for small degrees of parallelism. The 2D turbine blade optimization results did not allow a clear conclusion about the advantage of steady-state selection, however this is coherent with the above findings.