CDE-GAN: Cooperative Dual Evolution-Based Generative Adversarial Network

Generative adversarial networks (GANs) have been a popular deep generative model for real-world applications. Despite many recent efforts on GANs that contributed, mode collapse and instability of are still open problems caused by their optimization difficulties. In this article, motivated the cooperative co-evolutionary algorithm, we propose dual evolution-based GAN (CDE-GAN) to circumvent these drawbacks. essence, CDE-GAN incorporates evolution with respect generator(s) discriminators into unified evolutionary framework conduct effective multiobjective optimization. Thus, it exploits complementary properties injects mutation diversity training, steadily diversify estimated density in capturing multimodes improve performance. Specifically, decomposes complex problem two subproblems (generation discrimination), each subproblem is solved separated subpopulation (E-Generators E-Discriminators), evolved its own algorithm. Additionally, further Soft Mechanism balance tradeoff between E-Generators E-Discriminators steady training CDE-GAN. Extensive experiments one synthetic dataset three benchmark image datasets demonstrate proposed achieves competitive superior performance generating good quality diverse samples over baselines. The code more generated results available at our project homepage https://shiming-chen.github.io/CDE-GAN-website/CDE-GAN.html.