From: Kevin Sikorski (kws_at_cs.washington.edu)
Date: Fri May 16 2003 - 11:28:16 PDT
Leap Before You Look
Keith Golden
The author presents a partial order planner that uses verification links
and secondary preconditions for handling incomplete information.
I think the biggest contribution from this paper is the ability of the
planner to assume information it does not have, and then later verify that
this information was true via verification links. This has the potential
to vastly increase the search space - you still have to decide what you
are going to assume is true, and then plan how to determine the truth or
falshood of that assertion. However, it does provide a powerful technique
for planning with incomplete information.
The concept of secondary preconditions is also innovative, and is a
prerequisite for the use of verification links. Generally, the planner
assumes that a secondary precondition is true, and verifies it later.
One interesting facet of the experimental results is the runtimes of the
VL version versus that of the NO version. It is unclear which is faster
in general. Some extra discussion of the problems where NO was faster
would be nice - it may be that there is a clear division of the problems
where NO is faster. If so, then this division could be exploited. Of
course, one would have to be careful, because NO doesn't always solve the
problem when VL does. It really bothers me that restricing where the
producer of a verification link is located can improve the runtime of the
planner.
I wonder if it is possible to add some simple heurisitic for deciding
which secondary preconditions are best to assume true first. That is, if
we can prove that a precondition must be false, then we shouldn't waste
any time deciding the truth value of any related precondition. I'm sure
that the best heuristics would be domain-specific, but there may be a
simple, general approach that might work, similar to that of HSPr.
I also like the fact that the author mentions that interleaved planning
and execution could be applied to the planner. If we can identify a set
of preconditions that we want to be true, we can find a subplan that will
allow the agent to observe their truth-values, and then execute that
subplan. Then, we can plan again, from our current position in the state
space to the goal, now that we have more, and more useful, information.
This would likely be preferable to finding a long plan from the start, and
then building contingencies to handle cases where the secondary
preconditions were not all true.
This archive was generated by hypermail 2.1.6 : Fri May 16 2003 - 11:28:18 PDT