Evolutionary generative adversarial networks (E-GAN) attempts to alleviate mode collapse and vanishing gradient that plague generative adversarial networks by introducing evolutionary computation. However, E-GAN lacks a reasonable evaluation mechanism, which limits its effect. Moreover, E-GAN only contains mutation operators in its evolutionary step, while ignoring crossover operators. The crossover operator generates more competitive individuals by combining the good traits of multiple individuals, thus it can complement the mutation operator. In this paper, we propose a novel evolutionary generative adversarial networks framework called improved evolutionary generative adversarial networks (IE-GAN), which introduces a new fitness function and generic crossover operator. A more efficient fitness function can measure the evolutionary degree of individuals more precisely. And with the help of knowledge distillation, crossover offspring can learn knowledge from multiple networks simultaneously. Experiments on various datasets demonstrate the effectiveness of IE-GAN, and show that our framework is competitive in terms of the quality of generated samples and time efficiency.
View on arXiv