Multiobjective coevolutionary training of Generative Adversarial Networks

Abstract

This article presents a multiobjective evolutionary approach for coevolutionary training of Generative Adversarial Networks. The proposal applies an explicit multiobjective optimization approach based on Pareto ranking and non-dominated sorting over the co-evolutionary search implemented by the Lipizzaner framework, to optimize the quality and diversity of the generated synthetic data. Two functions are studied for evaluating diversity. The main results obtained for the handwritten digits generation problem show that the proposed multiobjective search is able to compute accurate and diverse solutions, improving over the standard Lipizzaner implementation.