[Ml-stat-talks] Fwd: David Wingate's talk - Wed 11/24 10:30am
sjgershm at princeton.edu
Tue Nov 16 12:16:32 EST 2010
Dave Wingate is visiting from MIT and he'll give an informal talk in the
psychology department (Green Hall, room 3-N-4) on Wed 11/24 at 10:30am.
Title and abstract below.
Probabilistic Programming for Bayesian Reinforcement Learning
In model-based reinforcement learning, an agent uses its
experience to construct a model of the world, which is in turn
used for planning. Typical modeling algorithms are based on
simple frequentist estimators of transition probabilities. How
we can extend this to deal with rich, highly structured
environments? We borrow tools from hierarchical Bayesian
modeling, which naturally allows us to share statistical strength
between model parameters and deal with richly structured
At the same time, probabilistic programming is a new paradigm
for specifying complex probabilistic models that are compositional,
recursive and structured, using the familiar tools of modern
programming languages. In this informal talk, I'll present some
work we've done on probabilistic programming,
and sketch how I think they could serve as a basis for model
building in reinforcement learning. Time permitting, I will also
discuss planning in a Bayesian reinforcement learning context.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Ml-stat-talks