ODE-RSSM: Learning Stochastic Recurrent State Space Model from Irregularly Sampled Data
Published in In the proceedings of Conference on Artificial Intelligence, AAAI2023, 2023
This work proposes the Ordinary Differential Equation Recurrent State Space Model (ODE-RSSM), a continuous-time deep state space framework designed for nonlinear and stochastic input-output systems with irregularly sampled observations. By integrating an ODE-based network to model latent state evolution between time points, the approach overcomes the limitations of conventional discrete-time SSMs. An efficient reparameterization method is introduced to enable parallel training on batched ODEs with non-uniform time spans, significantly improving computational efficiency. Extensive experiments on multiple datasets, including a private industrial system exhibiting long-term delay and stochasticity, demonstrate superior open-loop prediction performance, highlighting ODE-RSSM’s effectiveness for representation, forecasting, and decision-making in real-world dynamic environments.
Recommended citation: Zhaolin Yuan, Xiaojuan Ban, Zixuan Zhang, Xiaorui Li, Hong{-}Ning Dai, "ODE-RSSM: Learning Stochastic Recurrent State Space Model from Irregularly Sampled Data." In the proceedings of Conference on Artificial Intelligence, AAAI2023, 2023.
Download Paper