Technion, IEM faculty - Statistics Seminar

Speaker: Elad Hazan
Title: Interpolating worst-case and statistical learning
Date: 06/03/2011
Time: 14:30
Place: Bloomfield-527
Abstract:  <>
See also here:
The sequential decision problem is a fundamental pillar of 
statistics which has been studied since the
seminal work of Hannan and Robins in the fifties, with recent 
invigorated interest in the field of machine
learning. This framework is more general, but usually provides 
weaker convergence guarantees than those
attained in statistical learning theory. In 2005 Cesa-Bianchi, 
Mansour and Stoltz conjectured that the regret
of sequential decision making algorithms can be bounded by the 
variation in the observed data, thereby
providing a bridge between the two fields. We describe a 
solution to this conjecture in the fundamental
setting of discrete decision making (the so called "experts 
problem"), and proceed to more general results
along the same lines applied to decisions with partial 
information and to portfolio selection.
Based on joint work with Satyen Kale, published in COLT09, 
SODA09, NIPS09, and Machine Learning
Technion Math. Net (TECHMATH)
Editor: Michael Cwikel   <> 
Announcement from:  <>