LCW-Based Agent Planning for the Semantic Web

advertisement
LCW-Based Agent Planning for
the Semantic Web
Jeff Heflin, Hector Muñoz-Avila
presented by Axel Polleres
cf. http://www.cse.lehigh.edu/~heflin/pubs/lcw-aaai02.pdf
1
Overview
• Add LCW to DAML+OIL and SHOE
• Combine enriched ontologies with OTD.
• Semantic Description of the involved
Sources/Services is not an issue.
2
Local Closed World
Assumption
• An expression of the form LCW() means
stating that any substitution of  which is true
in the real world is also represented in the
agents knowledge, i.e. stating "local
complete knowledge" on .
3
Semantic Web and LCW
• Why is this interesting here? The open world
causes often unbounded search (therefore,
e.g. classical planning is not directly
applicable).
• Suggestion: Extend fixed KB by a “Semantic
Web Mediator” to gather more information.
4
DAML-LCW:
State that a resource has complete information on a particular class.
5
DAML-LCW
• Limited suitability for properties:
Can express LCW(prop(X,c)), but not LCW(prop(c,X)), nor LCW(prop(x,y))
6
LCW within DAML+OIL
• States that C is the class of r1, …, rn and nothing else!
• i.e. DAML-LCW is not more expressive than DAML+OIL
itself.
Problem: very naïve, since r1, …, rn have to be
enumerated explicitely.
7
SHOE-LCW
states LCW(flight(X)  destination(X,Y)  USCity(Y))
•
•
LCW restricted to pos. conjunctions.
more flexible than DAML-LCW (variables), but LCW not expressible within the
language itself (complement/disjointness not expressible)
8
SHOE-LCW
• LCW adds implicit negation to SHOE. How to handle this? Nonmonotonic reasoning? Unsolved.
• In the paper they assume “well-formed LCW information”, but
can this be achieved in the Web?
• further remark: LCW cannot be defined on sets of resources
(neither in DAML-LCW, nor in SHOE-LCW)
9
Rest of the paper
OTD enhanced by Semantic Web mediator:
What is OTD?
• method M(h,P,ST)
...
compound task definition with
head h, preconditions P and a
set of subtasks ST.
• matches(h,t,S) ...
a task matches a method with head h
if hΘ=t and S satisfies PΘ where
Θ is a substitution.
• operator O(h,P,aL,dL) ...
primitive action, with
preconditions
add-list and delete-list (like in STRIPS)
OTD: match tasks to subtasks recursively to an ordered list of
operators p, binding all parameters.
10
Architecture:
Fixed KB contains information gathered so far (including LCW information)
and domain theory.
Other information sources queried via Semantic Web mediator.
(Problem: offline, actions are not “executed” on the source before planning
is completed)
11
LCW in planning:
• LCW information can arise in two forms in
Planning:
– LCW provided by information sources
– LCW inferred as result of an action (example:
UNIX command "ls" whereafter we know all files
in a directory and we know that these are all files
there).
12
Use LCW for precondition evaluation:
13
Querying the SWeb Mediator:
Bound by resource-constraints such as time limits or max number
of sources to be queried.
14
LCW yielding operators:
After execution with f  355 and date 1/2/2002, we can add
LCW(SeatFree(355, 1/2/2002, s))
15
Conclusions
• LCW can be udes to reduce/cut-off search
space in search/planning.
• LCW info stored explicitly or yielded by action
execution
• Unclear how to deal with inconsistent LCW
information.
16
Download