Multi-Optimiser Training for GANs Based on Evolutionary Computation

Generative adversarial networks (GANs) are widely recognized for their impressive ability to generate realistic data. Despite the popularity of GANs, training them poses challenges such as mode collapse and instability. To address these issues, many variants enhance GAN performance through improvements in network architecture, modifications to loss functions, and the inclusion of regularization techniques. However, a limited number of methods focuses on optimising GAN performance from the optimiser's perspective, despite the distinct roles different optimisers play in training. Existing GANs typically employ a single optimiser throughout, with Adam being the default choice for GAN training. Approaches using multiple optimisers have shown improved performance but are often task-specific. This paper introduces a novel, portable approach that lever-ages the strengths of multiple training methods, with an evolutionary supervisor coordinating the training of different sections of the network. Initially, real-number-encoded vectors representing the optimiser for each sub-parameter layer undergo pro-gressive enhancement using a standard evolutionary algorithm (SGA). Through the evolutionary process, optimal optimiser combinations are retained based on the performance of the trained GAN. The proposed method, SGA-GANs, is validated on the CIFAR10 dataset by integrating the steps into five benchmark GAN models: GAN, DCGAN, BEGAN, WGAN, and WGAN-GP. Experimental results demonstrate that SGA-GANs outperforms single optimiser training methods, achieving superior evaluation results and generating higher-quality images. Source code can be found at https://github.com/lizzhang-spec/SGA-GANs.