Answer Set Programming for Information Agents

1
Answer Set Programming for
Information Agents
Thomas Eiter and Michael Fink
Vienna University of Technology Knowledge-Based
Systems Group
2
Answer Set Programming

• Recent development in nonmonotonic reasoning
• Use (non-monotonic) logic programs for
  declarative problem solving
• Method
• Represent the problem, P, by an (extended) logic
  program,P.
• Compute some / all answer sets AS(P) of P
• Extract some / all solutions S of P from some /
  all sets A in AS(P).
• Transformations P -gt P , A -gt S efficient

3
ASP (2)

• Related Problem reduction to SAT solvers
  (chaff,).
• Difference to Prolog
• purely declarative rule order, subgoal order
  irrelevant
• nonmonotonic negation (also unstratified)
• nondeterminism
• Efficient ASP solvers are available (Smodels,
  DLV, ASSAT)
• Useful tools / reasoning engines for
  domain-specific KR formalisms and problems

4
Information Agents

• Crucial in emerging (global) information systems.
• Cooperation within societies of agents
• Some kinds of agents
• Facilitators control sub-agents and coordinate
  services
• Brokers Match between data sources / services
  and user requests
• Mediators Exploit meta-knowledge about provider
  agents to create higher-level services

5
Information Agents (2)

• Further infrastructure
• Yellow pages info matchmaking
• Blackboards
• Further auxiliary agents
• Web-Crawlers
• Info-Raiders
• ...

6
Information Agents (3)

• Desire Intelligent Agents
• Need Rational Capabilities
• inferences (deduction, abduction, )
• plausible conclusions
• deal with incomplete / unsure / unreliable
  information
• Exploit nonmonotonic formalisms logics
• Build task-tailored reasoning components

7
Information Agents (4)

• Prototypical Architecture

8
Information Agents (5)

• Problems Challenges
• decompose query requests
• integrate query request user profile
• select information source
• create execute a query plan
• compose / merge query answers
• data cleaning
• data integration detect / resolve
  inconsistencies
• All using
• (rich) domain knowledge
• meta-knowledge about sources

9
Example Site Selection

• Task Given a user query, select most relevant
  source.
• Requires background knowledge (about application
  domain, information sources),
• nonmonotonic reasoning due to incomplete
  information,
• declarative semantics diserable.
• ? use ASP Eiter et al., KR02.

10
Example
Movie Domain information sources s1, s2,
s3. Which movies are directed by Alfred Hitchcock?

• FUNCTION HCMovies(MovieDB"movie.dtd")
• CONSTRUCT ltMovieListgt WHERE
• ltMovieDBgt ltMoviegt ltTitlegt t lt/Titlegt
• ltDirectorgt ltPersonaliagt
• ltFirstNamegt "Alfred" lt/FirstNamegt
• ltLastNamegt "Hitchcock" lt/LastNamegt
• lt/Personaliagt lt/Directorgt
• lt/Moviegt lt/MovieDBgt IN source(MovieDB)
• CONSTRUCT ltMoviegt t lt/Moviegt
• lt/MovieListgt

Known s1 good for directors, s2 for person data,
s3 not reliable. Expected select s1.
11
General Architecture
Selected Site
?Q, lt
parsing
?sd
?dom
R(Q) ? ?qd
Q
?sel, ltu

• Query Description ?qd Abstract representation of
  query.
• Domain Theory ?dom Domain specific background
  knowledge.
• Site Description ?sd Information about the
  sources.
• Site-Selection Program ?sel Qualitative and
  quantitative selection rules and constraints
  user preferences.

12
Abstract Query Description

• Based on a general view of a query consisting of
  a construct part, a where part, and a source
  part.
• Generated from a set of elementary facts R(Q) by
  application of program ?qd.
• Relevant items identified with context-reference
  pairs (C,P), e.g., access(O,C,P,Q).
• High-level description predicates
  query(Q), access(O,C,P,Q), occurs(O,V),
  selects(O,C,V), constructs(O,C,P),
  joins(O1,O2,C).

13
Example Program

• A site-selection program ?sel for the movie
• domain
• Core rules ?sel
• r1 query_site(s2,Q) ? default_object(O,Person
  ,Q)
• r2 query_site(s1,Q) ? selects(O,equal,Hitchcoc
  k),
  
• 
  access(O,Director,Personalia/LastName,Q)
• r3 query_site(S,Q) ? default_path(O,LastName
  ,Q),
• 
  default_object(O,T,Q), accurate(S,T,high)

c
14
Example Program (ctd.)
aux

• Auxiliary rules ?sel
• r4 high_acc(T,Q) ? access(O,T,P,Q),
  accurate(S,T,high)
• r5 high_cov(T,Q) ? access(O,T,P,Q),
  covers(S,T,high)
• Optimization rules ?sel
• c1 ? query_site(S,Q),
  high_acc(T,Q),
• not
  accurate(S,T,high) 101
  
• c2 ? query_site(S,Q),
  high_cov(T,Q),
• not
  covers(S,T,high)
  51
• User preferences ltu
• nr1(Q,_) ltu
  nr3(Q,_,_,_).

o
15
Application

• Implemented on top of dlv Eiter et al. 1998 and
  its front end plp Delgrande, Schaub, Tompits
  2001.
• Agentized in IMPACT Subrahmanian et al. 2000.
• Experimental site selection environment movie
  domain
• Modeled DTD from a set of relevant movie concepts
  captured by the Open Directory Project.
• Wrapped parts of the Internet Movie Database
  (IMDb) and the EachMovie Database to XML created
  6 different databases.

16
Application (ctd.)

• Movie databases
• RandomMovies (RM),
• RandomPersons (RP),
• EachMovie (EM),
• Hitchcock (HC),
• KellyGrant (KG),
• Horror60 (H60).
• Site descriptions contents, quality, cost,
  reliability,etc.
• Domain knowledge simple ontological knowledge
  from the DTD some background knowledge.
• Site selection program (several pages of code).

17
Experimental Queries

• Formulated a number of natural user queries,
  including
• Q1 Which movies were directed by Alfred
  Hitchcock?
• Q2 In which movies directed by Josef von
  Sternberg acted
• Marlene Dietrich?
• Q3 In which year has the movie Arsenic and Old
  Lace been
• released?
• Q4 In which movies directed by Alfred Hitchcock
  acted Marlene
• Dietrich?
• Q5 In which film noirs did Marilyn Monroe act?
• Modified selection base adding/removing
  databases and
• their descriptions.

18
Results

• Selection results satisfactory and explainable
• Specific core site selection rules trigger.
• Domain knowledge identifies irrelevant or
  specific sites by
• ontological reasoning, reasoning over genre info.
  etc,
• Quantitative selection in case of equally
  preferred AS.

19
Conclusion

• ASP is a new problem solving paradigm
• applicable for many problems,
• useful and promising for information agents.
• IMPACT agents reasonable status set semantics
  answer set semantics.
• Problem specific reasoning components on top of
  ASP, e.g., site selection.
• Future work
• Tackle further problems of information agents (or
  related fields) where ASP approaches are
  promising.
• Coupling of approaches with existing tools, e.g.,
  for learning, planning, ontological reasoning,
  etc.