The rapid advancement of generative models has demonstrated remarkable success across a spectrum of applications, including data augmentation, data interpolation, and image generation, among others. However, it’s important to note that while most existing works in generative modeling focus on learning the underlying data distribution through a single generative model, the utilization of generative models to learn the data generation process involving informative interactions is still in its early stages. For example, in a multi-agent system such as transportation system, different agents exhibit different kinds of behavior. A single generative model can capture the average behavior within such a system but struggles to model the varied behaviors of different agents. In order to model the varied behaviors, one must also model the interactions between agents.
Changyu’s study seeks to bridge this gap in existing research by emphasizing the crucial role of interaction in generative modeling. He refers to such models as Interactive Generative Models. His prior works have demonstrated the advantages of integrating interactions into generative modeling and shown its promising potential in two areas, simulation and reinforcement learning.