Home

About

Advanced Search

Browse by Discipline

Scientific Societies

E-print Alerts

Add E-prints

E-print Network
FAQHELPSITE MAPCONTACT US


  Advanced Search  

 
The Center for Control, Dynamical Systems, and Computation University of California at Santa Barbara
 

Summary: The Center for Control, Dynamical Systems, and Computation
University of California at Santa Barbara
Spring 2009 Seminar Series
Presents
Multi-Agent Machine Learning
Geoff Gordon
Carnegie Mellon University
Friday, June 5, 2009 3:00 - 4:00 PM WEBB 1100
Abstract: Intelligent agents constantly face uncertainty, and are therefore a rich source of machine learning
problems. Unfortunately, many well-known ML tools can behave unpredictably on these problems. For exam-
ple, if we use a learned classifier to help decide how to act, we have created a feedback loop: our actions affect
both our environment and the behavior of other agents, which can change our observations, which can change
our learned hypothesis, which can change our actions. (Imagine predicting which restaurants will be busy on
Friday night: "nobody goes there any more, it's too crowded.") This sort of feedback loop "voids the warranty" of
many standard ML algorithms, and can lead to instability and poor performance, both in theory and in practice.
To design learners which are robust to this sort of multi-agent uncertainty, a promising strategy is to incorporate
ideas from game theory into our ML tools and algorithms. Unfortunately, doing so is not always straightforward:
many game-theoretic tools were not designed to provide the kinds of specific action recommendations that
agents need, and many do not scale to the sizes of problems that ML algorithms regularly handle.
In this talk, I'll describe some recent progress in this direction. In particular, I'll describe no-Phi-regret learners

  

Source: Akhmedov, Azer - Department of Mathematics, University of California at Santa Barbara

 

Collections: Mathematics