Inference and learning in models of complex stochastic systems
Daphne Koller
Stanford University
Consider a real-world dynamic system such as a complex physical device whose
state changes over time, or a robot interacting with a complex environment.
Such systems involve many relevant variables, exhibit unpredictable
dynamics, and involve variables that we can never observe. Dynamic Bayesian
networks (DBNs) can be used to provide compact models of such systems.
I will discuss how to perform inference in these models and how to learn
them from data. The main technical difficulty is that most algorithms
involve the use of a belief state --- a probability distribution over the
state of the process at a given point in time. Unfortunately, most
real-world systems are too complex to allow a belief state to be represented
exactly. In the talk, I will discuss an approximate inference algorithm
that exploits the hierarchical structure of real-world domains to allow
efficient inference even for large complex systems. I will present a
theoretical analysis that allows us to bound the error resulting from our
approximation. Empirical results show that our algorithm achieves orders of
magnitude faster inference with only a tiny degradation in accuracy. We can
extend these techniques to apply to hybrid systems -- ones involving both
discrete and continuous variables. I will present some results for the
hybrid algorithm on a complex diagnosis task. The inference algorithm also
forms the key subroutine within algorithms that learn models directly from
data, allowing dramatic speedups in learning. These techniques allow us to
apply probabilistic modeling to complex dynamic systems.
Joint work with: Xavier Boyen, Uri Lerner, and Ron Parr.