Abstract:Generative adversarial networks (GAN) have been effective for learning generative models for real-world data. However, existing GANs (GAN and its variants) tend to suffer from training problems such as instability and mode collapse. In this paper, we propose a novel GAN framework called evolutionary generative adversarial networks (E-GAN) for stable GAN training and improved generative performance. Unlike existing GANs, which employ a pre-defined adversarial objective function alternately training a generator and a discriminator, we utilize different adversarial training objectives as mutation operations and evolve a population of generators to adapt to the environment (i.e., the discriminator). We also utilize an evaluation mechanism to measure the quality and diversity of generated samples, such that only well-performing generator(s) are preserved and used for further training. In this way, E-GAN overcomes the limitations of an individual adversarial training objective and always preserves the best offspring, contributing to progress in and the success of GANs. Experiments on several datasets demonstrate that E-GAN achieves convincing generative performance and reduces the training problems inherent in existing GANs.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the instability and mode collapse problems encountered in the training process of existing Generative Adversarial Networks (GANs). Specifically, traditional GANs and their variants are prone to gradient disappearance and mode collapse during the training process, and these problems limit the performance and application scope of GANs. To solve these problems, the author proposes a new GAN framework - Evolutionary Generative Adversarial Network (E - GAN), aiming to improve the training stability and generation performance of GANs by introducing evolutionary algorithms. ### Core contributions of the paper 1. **Introduction of evolutionary algorithms**: Different from traditional GANs, E - GAN uses different adversarial training objectives as mutation operations and optimizes the generator population through evolutionary algorithms to adapt to the environment (i.e., the discriminator). This evolutionary mechanism enables E - GAN to overcome the limitations of a single adversarial training objective and always retain the best offspring. 2. **Multi - objective optimization**: E - GAN uses multiple adversarial training objectives (such as min - max objective, heuristic objective, least - squares objective, etc.) to optimize the generator, thereby combining the advantages of different objectives and avoiding their respective disadvantages. 3. **Evaluation mechanism**: E - GAN introduces an evaluation mechanism to assess the quality and diversity of generated samples. This helps to select well - performing generators and further improve the generation performance. ### Specific methods - **Mutation**: Generators produce offspring through different mutation operations. These mutation operations correspond to different training objectives and aim to narrow the distance between the generation distribution and the data distribution from different perspectives. - **Minimax Mutation**: Corresponds to the min - max objective function of the original GAN and aims to minimize the Jensen - Shannon divergence. - **Heuristic Mutation**: Aims to maximize the probability that the discriminator is deceived and avoid gradient disappearance. - **Least - Squares Mutation**: Based on the least - squares objective, it avoids gradient saturation and partially avoids mode collapse. - **Evaluation**: The performance of generators is evaluated through the quality fitness score \(F_q\) and the diversity fitness score \(F_d\). - **Quality fitness score**: Measures the quality of generated samples through the output value of the discriminator. - **Diversity fitness score**: Measures the diversity of generated samples through the gradient value of the discriminator to suppress mode collapse. - **Selection**: Select well - performing generator offspring according to the fitness scores, retain them and use them for subsequent training. ### Experimental results - **Synthetic data set**: On the 2D Gaussian mixture distribution, E - GAN can effectively avoid mode collapse and accurately fit the target distribution. - **CIFAR - 10 data set**: E - GAN obtains a higher Inception score within fewer training steps and shows good stability when converging. - **LSUN bedroom data set**: E - GAN can generate reasonable results under different network architectures, demonstrating its advantages in architecture robustness. In conclusion, by introducing evolutionary algorithms and multi - objective optimization, E - GAN significantly improves the training stability and generation performance of GANs and solves the common problems in the training of traditional GANs.

Evolutionary Generative Adversarial Networks

GEGAN: Generative Adversarial Networks Training Based on Gradient Directed Evolutionary Strategy

Creative and Diverse Artwork Generation Using Adversarial Networks

Evolutionary Generative Adversarial Networks Based on New Fitness Function and Generic Crossover Operator.

Evolutionary Architectural Search for Generative Adversarial Networks

PEGANs: Phased Evolutionary Generative Adversarial Networks with Self-Attention Module

CDE-GAN: Cooperative Dual Evolution Based Generative Adversarial Network

IE-GAN: An Improved Evolutionary Generative Adversarial Network Using a New Fitness Function and a Generic Crossover Operator

Evolutionary Generative Adversarial Networks with Crossover Based Knowledge Distillation

The application of evolutionary computation in generative adversarial networks (GANs): a systematic literature survey

Evolutionary Multiobjective Optimization Driven by Generative Adversarial Networks (Gans)

Multi-branch evolutionary generative adversarial networks based on covariance crossover operators

Enhanced Evolutionary Generative Adversarial Networks

EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs

Online Kernel based Generative Adversarial Networks

A New Distributed Method for Training Generative Adversarial Networks

CAGAN: Consistent Adversarial Training Enhanced GANs.

Training Generative Adversarial Networks by Auxiliary Adversarial Example Regulator

Evaluation of GAN-Based Model for Adversarial Training

Host-Pathongen Co-evolution Inspired Algorithm Enables Robust GAN Training

Exploring generative adversarial networks and adversarial training