Predictive State Representations

(A new class of models for discrete-time dynamical systems)

Professor Satinder Singh

Computer Science and Engineering Division

University of Michigan, Ann Arbor

Modeling dynamical systems, either for control purposes or to make predictions about their behavior, is ubiquitous in science and engineering. I will present predictive state representations (PSRs), a new class of models for discrete-time dynamical systems. The key idea in PSRs is to represent the state of the system as a set of predictions of observable outcomes of tests or experiments one can do in the system. This key idea sets PSRs apart from history-based models

such as n-th order Markov models as well as from hidden-state-based models such as HMMs or POMDPs. I will show how PSRs can be derived simply from a system-dynamics matrix, a conceptual construct that I will also use to formally show that PSRs are more general than both n-th order Markov models and HMMs/POMDPs. Finally, I will point to some recent work on PSR models for continuous observation systems.

Friday, October 27, 2006

4:00 – 5:30 p.m.

1500 EECS