DevoGAN: A Method for Enhancing Quality and Diversity in Sequence Generation using Generative Adversarial Networks with Dual Evolutionary Algorithms

Sequence generation research based on Generative Adversarial Networks is often limited by the specified adversarial optimization strategy, which seriously affects the optimization performance during the training process. How to improve the quality of sequence data is the main concern, especially for the category sequence data, which not only needs to improve the quality of the generated data but also to ensure the diversity of the data. Previous research is mainly committed to the introduction of additional dimensional information to achieve the purpose of increasing the diversity of the generation, but these methods often lead to the model overly relying on the specific label information, which results in the generation of sequence data diversity and quality can not be guaranteed. In this paper, inspired by evolutionary algorithms, we propose an algorithm based on the dual evolution of the discriminator and the generator, which improves the diversity and quality of the generated category sequence data by using complementary evolutionary strategies and collaborative fitness functions. In order to evaluate the performance of the model proposed in this study, experiments were conducted on the MR dataset using Negative Log Likelihood (NLL) and Bilingual Evaluation Understudy (BLEU) as evaluation metrics. After comparing the results with other algorithms, it is shown that the method in this paper has advantages in improving the quality and diversity of sequence data generation.