2019/09/30 by Zhengdao Chen, Jianyu Zhang, Martin Arjovsky +1 · 2 citations
Computer Science · Mathematics · #cs.LG #stat.ML
published as 8th International Conference on Learning Representations (ICLR 2020) · Added link to GitHub repository
arxiv created 2020/04/25 · arxiv updated 2020/04/28
We propose Symplectic Recurrent Neural Networks (SRNNs) as learning algorithms that capture the dynamics of physical systems from observed trajectories. An SRNN models the Hamiltonian function of the system by a neural network and furthermore leverages symplectic integration, multiple-step training and initial state optimization to address the challenging numerical issues associated with Hamiltonian systems. We show that SRNNs succeed reliably on complex and noisy Hamiltonian systems. We also show how to augment the SRNN integration scheme in order to handle stiff dynamical systems such as bouncing billiards.