From: Sandra B Fan (sbfan_at_cs.washington.edu)
Date: Wed Nov 19 2003 - 09:14:52 PST
Title:
"Symbolic Heuristic Search for Factored Markov Decision Processes"
Authors:
Zhengzhu Feng, Eric A. Hansen
One-line summary:
The authors expand on the work of Hoey et al. on SPUDD
by exploring the combination of factored representations of MDPs with
limiting computation to states reachable from the starting state.
Main points:
First, the authors show how to combine doing factored MDPs with computing
only reachable states using the LAO* algorithm. They show that symbolic
model-checking really is useful for these types of planning problems.
Secondly, they provide a comparison of the performance of their algorithm
with that of SPUDD. I thought their comments on how the factory examples
used by SPUDD were inadequate for testing their algorithm and their
subsequently having to come up with artificial examples was interesting
because it showed something about the nature of SPUDD and the examples it
had been run on. Symbolic LAO* seems to kick SPUDD's butt.
Flaws:
Expanding on what I just said above, the authors had to construct
artificial examples in order to show off what their method could do. If
that's the case then, are there actually any natural examples under which
their algorithm would be useful? It would have been nice if they could
have discussed that.
Open Questions:
The authors demonstrated success with the combination of different
strategies--how many more strategies can we combine to get an even better
result? And the other question is back to the whole, what kinds of
natural examples can we test this on and will our performance be the same?
This archive was generated by hypermail 2.1.6 : Wed Nov 19 2003 - 09:14:53 PST