Graph Neural Networks for Learning System Dynamics

Learning multi-agent system dynamics is important yet challenging. The complex interplay among agents is often explained with a one-step discrete model in existing work, which predicts the next observations for all agents given the current observations and their interaction graph. In reality, the interaction graph among agents may evolve over time and not always be observable. Moreover, many real-world systems are continuous in nature and using discretized modeling may significantly impair the precision of nonlinear dynamics. In this talk, I will present our recent work on modeling dynamic interactions using graph neural networks. Our models can jointly learn the latent representations of agent trajectories and the interaction graph in an unsupervised manner over time. Experiments show that they can accurately predict system dynamics especially in the long range and generalize well to low-resource systems that have only few training samples.