Learning latent variable and predictive models of dynamical systems

摘要: A variety of learning problems in robotics, computer vision and other areas artificial intelligence can be construed as statistical models for dynamical systems from sequential observations. Good system allow us to represent predict observations these systems, which turn enables applications such classification, planning, control, simulation, anomaly detection forecasting. One class assumes the existence an underlying hidden random variable that evolves over time emits we see. Past are summarized into belief distribution this variable, represents state system. This assumption leads ‘latent models’ used heavily practice. However, algorithms still face a issues model selection, local optima instability. The representational ability also differs significantly based on whether latent is assumed discrete Hidden Markov Models (HMMs), or real-valued Linear Dynamical Systems (LDSs). Another recently introduced set predictions about future rather than summarizing past. These ‘predictive models’, Predictive State Representations (PSRs), provably more powerful hold promise allowing accurate, efficient since no quantities involved. has not been realized. In thesis propose novel address minima instability models. We show certain ‘predictive’ methods bridge gap between predictive model, Reduced-Rank HMM (RR-HMM), combines desirable properties latent-variable reparameterizing RR-HMMs yields subset PSRs, asymptotically unbiased algorithm PSRs along with finite-sample error bounds RR-HMM case. In terms efficiency accuracy, our outperform alternatives dynamic texture videos, mobile robot visual sensing data, domains.