Multi-population GAN Training: Analyzing Co-Evolutionary Algorithms

Generative adversarial networks (GANs) are powerful generative models but remain challenging to train due to pathologies such as mode collapse and instability. Recent research has explored co-evolutionary approaches, in which populations of generators and discriminators are evolved, as a promising solution. This paper presents an empirical analysis of different coevolutionary GAN training strategies, focusing on the impact of selection and replacement mechanisms. We compare (μ, λ), (μ+λ) with elitism, and (μ+λ) with tournament selection coevolutionary schemes, along with a non-evolutionary population based multi-generator multi-discriminator GAN baseline, across both synthetic low-dimensional datasets (blob and gaussian mixtures) and an image-based benchmark (MNIST). Results show that full generational replacement, i.e., (μ, λ), consistently outperforms in terms of both sample quality and diversity, particularly when combined with larger offspring sizes. In contrast, elitist approaches tend to converge prematurely and suffer from reduced diversity. These findings highlight the importance of balancing exploration and exploitation dynamics in coevolutionary GAN training and provide guidance for designing more effective population-based generative models.