and Knowledge Representation LPKR 97
Danford's Inn, Port Jefferson, Long
Island, Oct. 16
This ILPS'97 postconference workshop
was the third in a series organized by J. Dix, L.M. Pereira and T. Przymusinski since
1994. It is obvious that intelligent machines need tremendous amounts of knowledge for
solving difficult tasks like diagnosis, planning, configuration, decision support and many
others. Logic programming's declarative nature as well as its amenability to
implementation make it a very good candidate for knowledge representation purposes.
The papers presented this year's
workshop were centered around four major topics: updates, abduction, priorities, and
Updates, that is program modifications
reflecting changes in the actual world, have received considerable interest in the logic
programming community within the last few years. In particular, specifications of updates
through so-called update programs have been investigated. The paper presented by J. Leite
and L.M. Pereira shows that in order to model update in a logic programming context
adequately it is often insufficient just to take the models of programs into account. The
program rules themselves give additional information about why a certain belief is
accepted, e.g. because of a causal relationship. Consequently, updates should be conceived
as program transformations, not as operations on models. The authors describe how such
transformations can be defined both for normal and for extended logic programs.
The paper by A. Yahya (which was
actually presented by D. Seipel) investigates the issue of adding a clause to,
respectively deleting it from, a disjunctive deductive database. Several possibilities for
accomplishing this in different classes of theories are presented. In each case minimality
of change is measured in terms of its effect on the minimal model structure of the theory.
Two presentations were given in the
abductive logic programming session. R. Li, L.M. Pereira and V. Dahl showed how the
framework of abductive logic programming can be used to model the incorporation of test
results into incomplete action domain descriptions. The approach leads to a stepwise
refinement of descriptions in a high level action language and solves a problem suggested
by V. Lifschitz.
A system for learning abductive logic
programs was presented by E. Lamma, M. Milano and F. Riguzzi. The underlying algorithm is
based on a top-down algorithm which takes into account abducibles and integrity
constraints. Instead of the standard deductive proof procedure abductive inference is used
establish coverage of positive and negative examples in the learning process.
In the priority section I myself
presented a new approach to preference handling in extended logic programs under answer
set semantics. A given preference relation on program rules is extended to a preference
relation on answer sets, and the meaning of a prioritized program is determined through
its most preferred answer sets. The approach is based on a program reduction which is dual
to the standard Gelfond/Lifschitz reduction. It provably satisfies reasonable principles
for preference handling in rule based systems.
Instead of modifying the semantics, M.
Gelfond and T.C. Son proposed to use a certain set of domain
independent axioms describing declaratively how conclusions are to be drawn from
defeasible rules with additional preference information. To make this possible the
original rules are reified and a number of meta-predicates (like holds_by_default,
may_hold, defeated etc.) are introduced. The authors demonstrated that the standard
examples from the literature can be handled adequately this way.
The last and biggest session on
semantics contained four presentations. H. Decker developed a three-
valued semantics for integrity constraints based on the notion of a sustained model.
Existing two-valued semantics for integrity constraints turn out to be biased either
towards satisfaction or towards violation of constraints. Using a three-valued semantics
such problems can be avoided.
In the paper by S. Greco, N. Leone and
F. Scarcello disjunctive datalog is extended by nested rules, i.e. rules whose heads can
be composed of rules. The authors argue that this extension leads to more natural
formulations of complex reasoning problems. They presented a semantics for their extension
and discussed its complexity.
D. Seipel introduced so-called partial
evidential stable models and showed that such models always exist for disjunctive
deductive databases. These models are a subset of the minimal models, they coincide with
partial stable models whenever the latter exist. The definition of partial evidential
stable models is based on a preference relation which captures the idea of minimizing
reasoning by contradiction.
Finally, an introspective logic of
belief was presented by L.-Y. Yuan and J.-H. You. The logic is a modal logic extended by a
new negative introspection rule. It turns out that this new logic is able to characterize
almost all nonmonotonic semantics. The authors also argue that their logic has
considerable advantages when it comes to implementation.
In conclusion, this year's workshop -
as the two workshops held in 94 and 96 - demonstrated important progress in the field.
Although semantical issues still play an important role in the discussion, more and more
solutions for hard knowledge representation problems based on logic programming technology
seem to be emerging. I take this as a clear indication that investigations of the kind
discussed at the workshop are on the right track.
Institut fuer Informatik Augustusplatz
10/11 04109 Leipzig Germany