Wed 10 Apr 1996, 12:00, WeH 7220
==================================================================
Making Decisions with a "quasi"-Bayesian Model
Fabio Cozman
There has been great progress in finding best decisions when
probabilities are known precisely (e.g. Michael Littman's talk). So
it seems that an agent should strive to "learn" the probabilities of
interest and then pick the best decision.
In this talk I'll present a class of Markov decisions where we *can*
find admissible decisions very fast if we work with a set of
probability distributions, but we *can't* if we constrain ourselves to
a single probability. I'm currently applying this problem to
estimation of material properties with a robotic probe.
So I'd like to discuss some points: Do we really need to "learn" a
single probability distribution? Wouldn't it be more reasonable to say
that, when we learn, we are actually "constructing" the probability?
And how would we look for best decisions while our probabilities are
being learned --- i.e., constructed? I'll show how these questions
can be formalized in the framework of Quasi-Bayesian theory, which has
been used in Statistics for robustness studies. Hopefully we will have
time to discuss how this fresh look at old problems can reduce the
complexity of our learning problems.