Controlled Predictive State Models for Dynamical Systems with Continuous Observations

Predictive state representations are an interesting class of models in which the state of the system consists of predictions about the future, in contrast to distributions over latent variables as in many traditional models. I will present the controlled Predictive Linear-Gaussian (PLG) model, a predictive state model for controlled, discrete-time dynamical systems with continuous-valued observations. The PLG has equivalent reprentational power to the Linear Dynamical System / Kalman filter model. I will also outline an algorithm that can be used to obtain consistent estimates of the model's parameters from data, and present experimental results comparing it to the Expectation Maximization algorithm for Linear Dynamical Systems.