From: Russell Power (rjpower_at_u.washington.edu)
Date: Mon Nov 24 2003 - 10:14:52 PST
Acting Optimally in Partially Observable Stochastic Domains - Anthony
Cassandra, Leslie Kaelbling, Michael Littman.
The authors provide an introduction to POMDPs and describe a new algorithm
for efficiently finding near-optimal policies.
The paper describes the 'Witness' algorithm, based on previous work by
Cheng, for efficiently finding policies in a partially observable
environment. Sadly, the details of how this is accomplished are not
expounded on by the authors. They also describe the usage of a policy graph
to represent final policies efficiently.
For a paper describing efficient methods for find policies, the results
section of the paper is quite a bit shorter than expected. The authors
don't provide any tables or data indicating the performance of their methods
in comparison to the methods in use before the paper appeared.
Also, it is difficult to tell what additional information this paper has
over the original Witness paper the authors cite. The graph representation
of policies is somewhat interesting, but certainly does not seem enough to
warrant an entire paper.
Future research: It would be interesting to see this approach extended and
applied to larger problem spaces - the problems mentioned in this paper seem
a bit too trivial to arise in real life. Also, some investigation into the
performance of the approximate policies generated by this and other methods
would be insightful - how much of a price are we paying for a speedy
solution?
This archive was generated by hypermail 2.1.6 : Mon Nov 24 2003 - 10:14:59 PST