From: Parag (parag_at_cs.washington.edu)
Date: Tue Apr 29 2003 - 11:03:09 PDT
Paper Title: Probabilistic Planning in the Graphplan Framework
Authors: Avrim L. Blum and John C. Langford
The paper talks about how to extend the graphplan technique to the
case of probabilistic domains - in particular, the domains where
state of world is known completely but actions are probabilistic.
First, the idea of how to extend the representation of the planning graph
in case of probabilistic domains, though intuitive, is quite important
in the formulation of the problem.
Second, the authors do a good job of separating out two cases - one, where you
can find the optimal policy but the performance is not so good (PGraphplan).
The other case, where performance is comparable to the deterministic case
but only the optimal trajectory is output rather than the complete policy
(TGraphplan).
In case of PGraphplan, the way the information is propagated through
the graph seems a bit ad-hoc. There is no clear justification as
to how exactly the extra time spent during the backward sweep
would give benefits for the forward search for the goal, how much
extra information to be collected etc.
I had some problems with the experiments section. First, as usual,
representing the results in graphical form could have been more
useful. Second, I would have liked to get more insight into
where exactly TGraphplan performs almost optimally and where all
it does not perform well.
As the authors themselves point out, finding out better ways of propagating
information through the backward sweep might help in enhancing the performance
of PGraphplan. Second, identifying where exactly TGraphplan works and where
it fails might give more insight into the problem.
One could also analyze how to extend these ideas to the case of
partially observable domains.
This archive was generated by hypermail 2.1.6 : Tue Apr 29 2003 - 11:03:09 PDT