Generative adversarial networks (GANs) are pivotal for generative artificial intelligence, but their training is plagued by three intertwined core challenges: instability, gradient vanishing, and mode collapse. Existing methods primarily address these issues in isolation, lacking a unified theoretical framework to achieve synergistic optimization. To address this fundamental gap, we reinterpret GAN training as a closed-loop dynamical system and propose a control-theory-driven objective design paradigm based on state-feedback control. Specifically, we formulate the training dynamics in a state-space form and introduce a feedback control input u as a function of the training state. Under this unified closed-loop framework, we jointly optimize the above three problems. A closed-loop control (CLC) regularization term for the discriminators and a content loss term for the generator are designed to ensure the stability of the training dynamics. A weight decay strategy for the CLC term is proposed to maintain effective gradient flow and mitigate gradient vanishing. A dynamic weight-adjustment strategy for the content loss term is developed to perturb the equilibrium point of the closed-loop system, alleviating mode collapse from a dynamics perspective. We have validated the effectiveness and generalization of our method across two distinct generation tasks: image and molecular sequence generation.
Li et al. (Tue,) studied this question.