From: Parag (parag_at_cs.washington.edu)
Date: Tue May 20 2003 - 10:53:48 PDT
A Knowledge-Based Approach to Planning with Incomplete Information and
Sensing
by :- Ronald P.A. Petrick and Fahiem Bacchus
The paper introduces a new approach to the problem of
planning with incomplete information which is based
on a higher level representation of planner's world
knowledge and world axioms.
The key ideas in the paper are:
Since, while planning in a domain with incomplete information,
an agent can plan based on what it knows about
the world and how its action change the state of
its knowledge, one should be able to plan based
entirely on this kind of information, rather than
keeping track of all possible world states. This
approach is expected to abstract away a number of
irrelevant details.
The second idea is how to represent the agent's
knowledge in the form of a set of databases to facilitate
the above approach. The knowledge stored in these databases
could be translated to a set of first order logic formulas on
which inferencing could be done. Also, paper talks about how
to represent planning problems in this new framework and how to
incorporate the change in state caused by taking actions.
One of the problems, as the authors point out is that
the inference procedure that they use is not complete.
And also, they say that there is no formal characterization
when the inference procedure would be complete. Another problem
that I have is with the branching actions which replicate the DB.
They propose to have a separate DB for each possible branch, which
means that the number would grow exponentially with
number of such branching actions taken. In that case,
aren't they faced by the similar problem of representing
extremely large number of possible world states as in
classical approaches? Am I missing something here?
The experimental section could have been better if
they had given figures for comparison with other planners.
As for future work, it would be interesting to characterize
the domains where above approach would give better results
versus the domains where it might miss some of the plans
which would have been detected by a classical planner.
This archive was generated by hypermail 2.1.6 : Tue May 20 2003 - 10:53:49 PDT