Generative Adversarial Network (GAN) is a generative model proposed to imitate real data distributions. The original GAN algorithm has been found to be able to achieve excellent results for the image generation task, but it suffers from problems such as instability and mode collapse. To tackle these problems, many variants of the original model have been proposed; one of them is the Evolutionary GAN (EGAN), where a population of generators is evolved. Inspired by EGAN, we propose here a new algorithm, called Multi-Objective Evolutionary Generative Adversarial Network (MOEGAN), which reformulates the problem of training GANs as a multi-objective optimization problem. Thus, Pareto dominance is used to select the best solutions, evaluated using diversity and quality fitness functions. Preliminary experimental results on synthetic datasets show how the proposed approach can achieve better results than EGAN.
Multi-objective evolutionary GAN
Baioletti M.;DI Bari G.;Poggioni V.
2020
Abstract
Generative Adversarial Network (GAN) is a generative model proposed to imitate real data distributions. The original GAN algorithm has been found to be able to achieve excellent results for the image generation task, but it suffers from problems such as instability and mode collapse. To tackle these problems, many variants of the original model have been proposed; one of them is the Evolutionary GAN (EGAN), where a population of generators is evolved. Inspired by EGAN, we propose here a new algorithm, called Multi-Objective Evolutionary Generative Adversarial Network (MOEGAN), which reformulates the problem of training GANs as a multi-objective optimization problem. Thus, Pareto dominance is used to select the best solutions, evaluated using diversity and quality fitness functions. Preliminary experimental results on synthetic datasets show how the proposed approach can achieve better results than EGAN.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
