ModPlan - Modern Action Planning
Domain Encoding
A pure propositional encoding can have efficiency drawbacks during the
exploration. A multivariate representation for the atom
set is often preferable. In the SAS+ encoding, groups of mutually
exclusive atoms are generated. This encoding serves as an
optional input for existing planners that can exploit this facility.
The output file format for this domain analysis step is a
Lisp-like representation of the set of reachable atoms and their
partitioning into SAS+ variable domains.
In planning with pattern databases, the
automated selection of possible abstraction functions to yield
informative pattern databases is a hard combinatorial task. This is
especially true for the creation of disjoint
databases, in which operator projections are
void in all but one abstraction. In plannning, pattern database
abstraction are most effective if they consider SAS+ groups in
common.
Knowledge Acquisition
There are different bin-packing approximation
algorithms that
infer a partitioning of variable groups before constructing the
databases. The maximum size of a pattern database is bounded by
the multiplication of the cardinalities of the selected variable domains.
Knowledge Engineering
Even though we are currently working at a genetic algorithm to improve
the first partition of groups, domain expert guidance in this optimization
process is crucial. The expert is asked to refine the approximated,
disjoint partitioning into planning pattern databases proposed in the
inference module, by modifying the SAS+ partition
for the different pattern databases in an XML frontend.
|