Mirrored variants of the (1, 2)-CMA-ES compared on the noisy BBOB-2010 testbed

Derandomization by means of mirrored samples has been recently introduced to enhance the performances of (1, λ) and (1 + 2) Evolution-Strategies (ESs) with the aim of designing fast local search stochastic algorithms. In this paper, we investigate the impact of mirrored samples for noisy optimization. Since elitist selection is detrimental for noisy optimization, we investigate non-elitist ESs only here. We compare on the BBOB-2010 noisy benchmark testbed two variants of the (1,2)-CMA-ES where mirrored samples are implemented with the baseline (1,2)-CMA-ES. Each algorithm implements a restart mechanism. A total budget of 104 D function evaluations per trial has been used, where D is the dimension of the search space. The experiments clearly show a ranking among the three algorithms: both mirroring variants have lower expected running times than the (1,2)-CMA-ES by at least 50% on 5 functions and they solve three additional functions in 20D that the (1,2)-CMA-ES cannot solve (or only ...

• Type: Conference
• Year: 2010
• Where: GECCO
• Authors: Anne Auger, Dimo Brockhoff, Nikolaus Hansen
• Keywords: GECCO 2010, Mirrored Samples, Mirroring Variants, Noisy Optimization, Optimization