ODE-RSSM: Learning Stochastic Recurrent State Space Model from Irregularly Sampled Data

For the complicated input-output systems with nonlinearity and stochasticity, Deep State Space Models (SSMs) are effective for identifying in latent state space, which of great significance representation, forecasting, planning online scenarios. However, most SSMs designed discrete-time sequences inapplicable when observations irregular time. To solve problem, we propose a novel continuous-time SSM named Ordinary Differential Equation Recurrent Model (ODE-RSSM). ODE-RSSM incorporates an ordinary differential equation (ODE) network (ODE-Net) to model evolution states between adjacent time points. Inspired from equivalent linear transformation on integration limits, efficient reparameterization method solving batched ODEs non-uniform spans parallel efficiently training irregularly sampled sequences. We also conduct extensive experiments evaluate proposed baselines three datasets, one is rollout private industrial dataset strong long-term delay stochasticity. The results demonstrate that achieves better performance than other open loop prediction even if predicted points uneven distribution length changeable. Code availiable at https://github.com/yuanzhaolin/ODE-RSSM.