Switching Linear Dynamics for Variational Bayes Filtering

TL;DR

This paper introduces a method for system identification that combines switching linear dynamics with variational Bayes filtering, enabling better modeling of complex, nonlinear systems from high-dimensional observations.

Contribution

It proposes a novel approach integrating switching linear dynamics with variational Bayesian methods to improve system identification and state space learning.

Findings

Enhanced accuracy in learned dynamics on simulated tasks

Ability to encode constraints and collisions in state representations

Improved modeling of complex systems from partial observations

Abstract

System identification of complex and nonlinear systems is a central problem for model predictive control and model-based reinforcement learning. Despite their complexity, such systems can often be approximated well by a set of linear dynamical systems if broken into appropriate subsequences. This mechanism not only helps us find good approximations of dynamics, but also gives us deeper insight into the underlying system. Leveraging Bayesian inference, Variational Autoencoders and Concrete relaxations, we show how to learn a richer and more meaningful state space, e.g. encoding joint constraints and collisions with walls in a maze, from partial and high-dimensional observations. This representation translates into a gain of accuracy of learned dynamics showcased on various simulated tasks.