Section: New Results

Benchmarking of CMA-ES Variants for Numerical Blackbox Optimization

Participant from DOLPHIN: Dimo Brockhoff; External Participants: Anne Auger and Nikolaus Hansen (Inria Saclay - Ile-de-France)

The covariance matrix adaptation evolution strategy (CMA-ES) is one of the state-of-the-art optimization algorithms for numerical single-objective blackbox optimization [81] , [80] [67] . Previously, we proposed to use so-called mirrored mutations to generate new candidate solutions in evolution strategies which turned out to increase the convergence rate for certain variants [71] , [65] , [66] . Another recent approach to speed up the CMA-ES is to perform an active (i.e. negative) covariance matrix update [60] . In [32] , [35] , [34] , [33] , we tested empirically how the combination of mirrored mutations and active CMA-ES perform on the COCO framework [77] , [78] . It turned out that both concepts complement each other well without a significant decrease in performance on any of the 24 test functions. Moreover, the main improvement over the standard CMA-ES could be shown to come from the active covariance matrix adaptation while the addition of mirrored mutations only slightly improves the algorithm.