Multi-Child DE - a Massively Parallel Differential Evolution Algorithm

To optimize objectives that require extensive simulations, engineers need a powerful algorithm that can run in parallel to minimize wall-clock time. Evolutionary algorithms excel in both regards: they are powerful optimizers that can be implemented in parallel by assigning one processor to each population member. This paper, however, addresses the question: How can we minimize a simulation's wall-clock time if we have many more processors than there are population members? Our solution is the multi-child differential evolution (MCDE) algorithm in which each parent competes against more than one child. We benchmarked the scaling performance of the MCDE version of classic DE with a test bed of ten functions. We also compared MCDE to a parallel CMA-ES variant with three of our test bed's functions. Results from both experiments show that MCDE achieves a significant speedup that scales well with the number of processors.