Using the uncertainty handling CMA-ES for finding robust optima

Algorithms that search for robust optima often evaluate the effective fitness (robust fitness) based on stochastic approximation schemes. In this setting, finding robust optima can be recast as an optimization problem with an/a uncertain/noisy objective function. This paper studies whether state-of-the-art uncertainty handling techniques, proposed in the context of optimizing noisy objective functions, can be applied for finding robust optima. In this paper, the UH-CMA-ES is modified to handle approximations of the effective fitness. This modified approach, named RO-UH-CMA-ES, is evaluated empirically and compared to other schemes that aim to find robust optima. The experiments on multiple benchmark problems show that the RO-UH-CMA-ES yields comparable results for multi-modal problems and it outperforms other schemes on unimodal problems.