ModPlan - Modern Action Planning
Goal Ordering
Let (O,I,G) be a planning instance and p,q in
G and let (p,not q) be a planning state for
which p has just been reached and in which q does not hold. We
define the forced ordering p <=(f) q if for all (p,not q)
there is no plan P
with q is the result of applying P to
(p, not q).
Moreover, the reasonable ordering p <=(r) q is given
if for all (p, not q)
there is no plan P with q is the result of applying
P to (p, not q).
Here we restrict to actions that have
p not in their delete list. Unfortunately, the decision
problems for p <=(f) q and
p <=(r) q are both PSPACE hard.
Knowledge Acquisition
We have implemented approximation <=(h) of <=(r)
as an individual static analysis option. The algorithm prompts
the outcome of this phase to the domain expert so that he can refine
the induced goal agenda. If the agenda is fixed, a sequence of
PDDL files is generated wrapping any selected planning
module.
Knowledge Enginering
As finding the best goal ordering is hard,
we leave it to the domain expert to adjust the approximated one.
The inference I(i+1) given I(i) is transparent
to the expert, as it simulates plan execution within the
validator VAL.
Using our extension to flush state sequences according to plan
happenings, we
apply VAL to I(i) and write the result I(i+1)
together with goal G(i+1) back to disk.
|