|
Introductory Talk
Architecture
Home
|
ModPlan - Modern Action Planning
Grounding
Grounding is the process of finding (supersets of) reachable actions,
facts and fluents by instantiating operators, predicates and functions
with the objects that come with the problem description. Most current
planners perform some form of grounding to apply planning state space
exploration.
Knowledge Acquisition
One of the most efficient methods to infer a
grounded representation is fact space exploration
that uses a fact queue, in which the initial
state is enqueued and where one fact is dequeued and processed at a
time. For all operators that include this fact in the precondition
list this fact is marked. If all preconditions are marked, the
operator is fired, enqueuing all its add effects.
Constant predicates are detected and removed and constant fluents
are substituted to simplify numerical expressions.
Quantification, conditional effects, domain
axioms in form of derived predicates, numeric conditions and temporal
operators have been integrated into this process.
Knowledge Engineering
The in- and output format of the instantiation process itself is valid
PDDL, so that the grounded representation can be fed into every
existing planner by the domain expert.
Even though the intermediate results might become
large, it is important for the expert to
access the simplified state and operator descriptors
to understand the working of the plan engines.
To perform the grounding process in our workbench we adapted three
different technologies:
-
translate: the instantiation process that is provided with
the planner Fast-Downward
-
adl2strips : the domain translation that comes
with the planner FF
- ground :
the preprocessing result
that is implicit in the planner MIPS.
|
|