Training Strategy Based on Game Theory for Generative AI

With the emergence of ChatGPT, a variety of cutting-edge applications and services of generative AI have been continuously developed. Among them, generative AI is primarily built upon techniques such as Generative Adversarial Networks (GANs) and reinforcement learning, which involve training generators and discriminators to provide generative AI services. While various models for image and language generation have been developed, effectively training these models remains a crucial research topic. Therefore, this study focuses on exploring game theory, particularly models such as the Cournot game, the Stackelberg game, and the repeated game, to design suitable training strategies for generative AI models. The paper will provide theoretical proofs and designs, along with empirical validation using open datasets.