Comparative Analysis of Two Asynchronous Parallelization Variants for a Multi-Objective Coevolutionary Solver

摘要

We describe and compare two steady state asynchronous parallelization variants for DECMO2++, a recently proposed multi-objective coevolutionary solver that generally displays a robust run-time convergence behavior. The two asynchronous variants were designed as trade-offs that maintain only two of the three important synchronized interactions / constraints that underpin the (generation-based) DECMO2++ coevolutionary model. A thorough performance evaluation on a test set that aggregates 31 standard benchmark problems shows that while both parallelization options are able to generally preserve the competitive convergence behavior of the baseline coevolutionary solver, the better parallelization choice is to prioritize accurate run-time search adaptation decisions over the ability to perform equidistant fitness sharing.