Introduction
Generative Adversarial Networks (GANs) have revolutionized generative modeling but remain notoriously difficult to stabilize. Two primary challenges hinder their performance: mode collapse, where the generator produces a limited variety of outputs, and vanishing gradients, where a superior discriminator fails to provide useful feedback. While Evolutionary Algorithms (EA) offer a potential solution through broad parameter space exploration, they are often used in isolation from gradient-based methods. This paper proposes a hybrid framework that bridges this gap by running backpropagation and neuroevolution in parallel within a multi-generator architecture.
Hybrid Training Framework
The proposed method utilizes a population of generators ($G_i$) competing against a single shared discriminator ($D$). The training cycle alternates between two distinct optimization routines to balance "exploitation" and "exploration."
- Adversarial Optimization ($O_A$): Standard GAN training where all generators and the discriminator are updated via backpropagation. This phase focuses on "exploitation," refining the models toward the target data distribution.
- Evolutionary Optimization ($O_E$): This phase introduces "exploration." Generators are evaluated for fitness based on their loss. The "least fit" half of the population undergoes hypermutation, where random Gaussian noise is added to the weights of specific layers—most effectively the last convolutional layer.
By forcing generators to both converge (via $O_A$) and diverge (via $O_E$), the framework prevents the model from becoming trapped in local optima or collapsing into a single mode.
Experimental Results
The framework was evaluated on the MNIST dataset across several variants (e.g., AllGPCGAN, AllGLCGAN). A key finding was the efficiency of the hybrid approach: while a standard Deep Convolutional GAN (DCGAN) required 25 epochs of adversarial training, the hybrid models achieved comparable image quality with only 5 epochs of $O_A$ per generation.
Qualitative analysis shows that "All-Generator" variants—where the entire population is evolved—exhibit greater robustness than "1-Generator" variants. Furthermore, the research highlighted that mutating only the final convolutional layer provides the best balance between stability and diversity, mitigating the risk of repetitive content.
Conclusion
This hybrid neuroevolutionary approach demonstrates that combining gradient-based fine-tuning with evolutionary exploration can significantly stabilize GAN training and improve convergence speed. While currently tested on the relatively simple MNIST dataset, the results suggest that layer-specific mutation is a powerful tool for maintaining diversity. Future work will focus on scaling this architecture to more complex, high-dimensional datasets and refining the computational efficiency of the dual-optimization process.
Summary updated