This paper introduces GAGAN, a hybrid model that integrates Generative Adversarial Networks (GANs) with Genetic Algorithms (GA) to improve GAN performance in image generation. Traditional GANs often face challenges such as mode collapse and unstable training, where the generator and discriminator struggle to consistently improve. To address these issues, GAGAN employs a hybrid approach: the discriminator’s weights are optimized using GA, while the generator is trained through standard gradient-based backpropagation. The GA evolves the discriminator’s weights, enhancing its ability to distinguish real from fake images, providing more robust feedback to the generator. This hybrid method combines the exploratory nature of evolutionary algorithms with the efficiency of gradient-based optimization. The model was trained on 2,000 images from the CelebA dataset,
generating images at a resolution of 128x128. The results demonstrate that GAGAN outperforms traditional GANs, leading to higher-quality images and more stable convergence. This novel approach enhances adversarial training by leveraging the strengths of both GA and backpropagation techniques.
.jpg){kind=logo}
University of West Attica
