System identification in Gaussian process dynamical systems (original) (raw)

This paper introduces the GPIL algorithm for system identification within Gaussian process dynamical systems, specifically focused on nonlinear dynamic systems where the transition and measurement functions are defined by Gaussian processes. The method learns the Gaussian process models for these functions without the need for any observed latent states, relying instead on pseudo training sets to effectively parameterize distributions over nonlinear functions. Through the use of the Expectation Maximization algorithm, the approach models relationships between latent states and observations, demonstrating promising results against established methods in both synthetic and real-world datasets.