A dynamical system is a mathematical model of a physical process.
A state is a complete description of a dynamical system at an instant in time. Structure refers to properties of a dynamical system that make it possible to compactly represent its states and the associated temporal relations involving those states.
This talk investigates the issues involved in representing and learning dynamical systems using graphical models that consist of a set of state variables that aggregate states and functions that relate these state variables. This representation allows us to make explicit the structure in a process and suggests learning
algorithms to exploit this structure computationally. We link previous work in learning dynamical systems to current theory and practice in learning Bayesian networks. We establish the connection between techniques in computer science concerned with inferring finite automata and those in physics, speech recognition and time series analysis concerned with recovering the dynamics of nonlinear dynamical systems.