Generative Adversarial Networks Based on Cooperative Games

Generative adversarial networks (GANs) have become a hot research topic in recent years, representing an unsupervised learning method based on zero-sum games. Due to the high complexity of real samples, GANs still face challenges in training stability and the quality of generated samples. Mode collapse is a common problem in GANs. To overcome the drawbacks of mode collapse, this paper introduces a novel approach called Cooperative-GAN (Coop-GAN) by incorporating a beating discriminator system instead of the original discriminator model. Cooperative games, which differ from zero-sum games, are employed in Coop-GAN. Unlike traditional GANs’ zero-sum games, Coop-GAN incorporates the concept of cooperative games, which involve non-adversarial interactions, distinguishing it from traditional GANs’ adversarial nature. In cooperative games, the overall benefits of the game system increase, and all parties can benefit from cooperation, achieving a win-win or mutually beneficial outcome.This collaborative framework forms the foundation of Coop-GAN. In Coop-GAN, the discriminator model learns the real data distribution exclusively by assigning high scores to real samples and refraining from assigning low scores to generated samples. This cooperative mode enables the discriminator model to better guide the generator, enhancing the discriminator’s capabilities without increasing the generator’s loss. Through mutual cooperation, Coop-GAN achieves an overall gain in the cooperative game, contributing to improved diversity and authenticity of the generated samples. The beating discriminator system in Coop-GAN reduces the frequency of rejecting generated samples, introducing a new cooperative mode where the beating discriminator system and the generator model mutually collaborate and learn together. Through multiple experiments on the Fashion-MNIST dataset, the results demonstrate that Coop-GAN satisfies superadditivity in cooperative games. As the batch size increases, the proposed Coop-GAN model exhibits greater stability compared to traditional GANs. Particularly, when the batch size is 512, Coop-GAN exhibits outstanding performance, reducing the FID value by 10% and increasing the IS score by 4% compared to the multi-discriminator model GMAN.In conclusion, Coop-GAN generates samples with lower FID values and fewer mode collapse phenomena compared to various other GAN models. © 2024, Taiwan Ubiquitous Information CO LTD. All rights reserved.