In machine learning, generative models are used to create data samples that mimic the characteristics of the training data. Generative adversarial networks (GANs) are neural-network based generator models that have shown their capacity to produce realistic samples in different domains. In this paper we propose a neuro-evolutionary approach for evolving deep GAN architectures together with the loss function and generator-discriminator synchronization parameters. We also propose the problem of Pareto set (PS) approximation as a suitable benchmark to evaluate the quality of neural-network based generators in terms of the accuracy of the solutions they generate. The covering of the Pareto front (PF) by the generated solutions is used as an indicator of the mode-collapsing behavior of GANs. We show that it is possible to evolve GANs that generate good PS approximations. Our method scales to up to 784 variables and that it is capable to create architecture transferable across dimensions and functions.

Basque Government: IT-609-13, Spanish Ministry of Economy, Industry and Competitiveness: TIN2016-78365-R, University of the Basque Country: PIF16/238