History Marginalization Improves Forecasting in Variational Recurrent Neural Networks

Summary: Deep state space models (SSMs) are actively researched in the deep learning community as a model class for temporal models with a close connection to classic SSMs. The flexibility of deep neural networks allows deep SSMs to describe a wide range of dynamics as a black-box identification model. The probabilistic nature of the model class enables the modeling of system uncertainty, and in this work, a deep SSM class and its parameter learning algorithm are explained to expand nonlinear identification methods with a deep learning approach. Six recent deep SSMs are evaluated together on nonlinear system identification benchmarks.