From: Parag (parag_at_cs.washington.edu)
Date: Thu May 15 2003 - 21:19:48 PDT
Leap Before You Look : Information Gathering in the PUCCINI planner
Keith Golden
The paper describes the problem faced by classical planners when
an agent may not have sufficient information to execute an action
which would probably lead to a perfectly valid plan and proposes
a way to solve this problem.
The main ideas in the paper are:
1. Why it might be important to execute an action even when
the an agent does not have sufficient information to be
sure about the outcome of the action.
2. Formalize the above problem, and give a solution in terms
of what the author rightly coins as "verification" links. The
idea is to execute an action assuming certain conditions and
then verify them later.
3. How simply the standard partial order planners could be
modified to incorporate the idea of verification links.
I have one question regarding the approach the author takes.
In the experimental section, the author says that SADL
encoding makes it necessary to use verification links
for certain problems, whereas they could be solved without
using the verification links in some different encodings.
Is it that the encoding itself makes the use of
some kind of verification links unavoidable. In other words,
given a specific domain, is it the case that one could always
come up with encodings where one won't need the verification
links at all. Or there are problems where no matter what encoding
scheme you use, you have to use the idea of verification links to get
a valid plan.
An interesting direction of further research would be
to see if the idea of verification links could be useful
in domains other than the softbot world. Also, as the author
suggests, one could look at how to combine some of the ideas
from contingency planning to solve above problem.
This archive was generated by hypermail 2.1.6 : Thu May 15 2003 - 21:19:49 PDT