From: Nan Li (annli_at_cs.washington.edu)
Date: Fri May 16 2003 - 12:05:30 PDT
Leap Before You Look / Keith Golden
This paper describes a novel approach to represent and make plans in
domains with incomplete informations.
An action can be applicable at a state only when the action's
preconditions hold in the state. This is always true. However, in typical
planner systems, in order to represent acurately the effects of an action,
some "knowledge preconditions" are added into the action's precondition
set. They are not essential for the action to be applicable, but they
constrain the effects. So, in incomplete-informed domains, some actions
will never be applicable because the planner has no idea of the truth
value of those "knowledge preconditions".
Thus, the first part of the idea is, remove those knowledge preconditions.
Even though you have no idea what are under a stone until you turn it
over, this ignorance won't stop you from turning it over and finding out.
So why don't do it? But the problem is, you can't search your plan
completely blindly. That's too expensive.
So the second part of the idea is, let's at least know what are possible
effects for an action. So those knowledge preconditions are put into
conditional effects set, and actions that can verify these preconditions
are connected with this action, by "verification links" and "observational
links". Since the main information is still maintained in the
representation, the performance of the planner won't be degraded. And the
relaxed preconditions set makes it possible to generate more flexible
plans.
The authors make it clear that they are interested in softbot domain. I
wonder how well PUCCINI would do in some other domains, e.g. a domain w/ a
large set of action schemas.
This archive was generated by hypermail 2.1.6 : Fri May 16 2003 - 12:05:32 PDT