Modeling Dynamical Systems with Structured Predictive State Representations

Predictive state representations (PSRs) are a class of models that
represent the state of a dynamical system (e.g. an agent's environment)
as a set of predictions about future events. PSRs are capable of
representing partially observable, stochastic dynamical systems,
including any system that can be modeled by a finite partially
observable Markov decision process (POMDP). There is evidence that
predictive state is useful for generalization and helps to learn
accurate models.

This talk will focus upon two classes of PSR models, factored PSRs and
multi-modal PSRs, which exploit different types of structure in a
dynamical system in order to scale up PSR models to large systems. The
factored PSR exploits conditional independence, allowing a trade-off
between model compactness and accuracy. The multi-modal PSR is designed
for systems that switch between different modes of operation; the model
makes specialized predictions for each mode. The model also maintains
predictions about the current mode of the system, because the current
mode is only observable after some delay. Both the factored PSR and the
multi-modal PSR were evaluated on the task of predicting highway traffic
on a six-lane portion of Interstate 80. The learned PSR models compare
favorably with other prediction techniques, achieving an average error
as low as one car length when predicting the distance a car will travel
over five seconds.