[Ml-stat-talks] Fwd: RL talk
blei at CS.Princeton.EDU
Fri Jan 21 16:07:24 EST 2011
from warren powell... an interesting talk on reinforcement learning.
---------- Forwarded message ----------
Seminar: Reinforcement learning and stochastic optimization
Monday, January 24, 4pm
On the construction of scenario-tree approximations to multistage stochastic
University of Liege
Multistage stochastic programming is a promising framework for formulating
sequential decision making problems under uncertainty with high-dimensional
decision spaces. However, multistage stochastic programs are notoriously
difficult to solve, even approximately. Existing theoretical results I am
aware of on scenario-tree approximations (the mainstream method for
obtaining a first-stage approximate solution) are discouraging, in the sense
that the rate of convergence of finite-dimensional approximate problems to a
"true" multistage problem seems too slow in practice.
In this talk, after a brief presentation of the multistage stochastic
framework, I will present a pragmatic approach for building scenario trees.
It consists in generating several scenario-tree approximations to a true
problem, and then selecting a good approximation by a fast Monte Carlo
validation technique. The validation is based on the inference of decision
policies by a mix of supervised learning and feasibility restoration
techniques (repair procedures). Results obtained with this approach will be
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Ml-stat-talks