QUAIL '97 (Question of the Day) |
|
QUESTION OF THE DAY Friday, December 6
==============================================================================
Briefly summarize the differences between the various kinds of approaches
to nonmonotonic reasoning (e.g., closed-world assumption, minimal models,
consistency, epistemic, conditional).
------------------------------------------------------------------------------
Since I got no answers for this one, I've come up with a slightly-more
detailed set than I was looking for. :)
- Patrick
CLOSED-WORLD ASSUMPTION: Assumes all relevant information has been
specified. Any positive fact not specified is false. Most obviously
employed in database theory. Reiter's (1978) formalization employs a
definition of closure
CLOSURE(DB) = DB U {!P(t) | DB !|= P(t), where P is an n-ary
predicate symbol of DB and t is an n-tuple
of ground terms formed using the function
symbols of DB}
This can generate inconsistent closures in general, though that problem
is avoided in Horn databases. Will produce incorrect results in
databases with null values.
An alternative formalization by Clark (1978) has to do with Prolog
programs. His assumption is that sufficient constraints on predicates
are also sufficient. This lacks generality; it only works for
universally quantified clauses; is sensitive to syntactic form.
CONSISTENCY-BASED APPROACHES: Interpret the absence of information
prohibiting A to mean that it can consistently be assumed.
Nonmonotonic Logic: McDermott and Doyle (1980). Uses a modal operator
M, with intended semantics MA meaning "A is consistent." Uses a
fixed-point operator to define the theorems of a theory A as the
intersection of all fixed points T for which
T = Th(A U {Mw | !w not in T}.
Neither the fixed points nor the theory is recursively enumerable. A
proof theory is only known for propositional logic.
Default Logic: Reiter (1980) treats default conditions as rules of
inference of the form a(x):b(x)/c(x), meaning "if a(x) is true, and b(x)
is consistent, infer c(x)." Default rules generate extensions of the
theory, also using fixed-points. E is an extension of a default theory
(D, W) if E = G(E), where G(S) is the smallest set for which
1. W c G(S).
2. G(S) is closed under FOL consequence.
3. If a:b/c is a default rule of D and a is in G(S) and !b not in S,
then c in G(S).
Note that extensions are comparable to fixed-points T in nonmonotonic
logic, but Reiter's interpretation is that they are all possible belief
sets for an agent; he does not intersect them.
This logic has a proof theory when rules are normal, but the extensions
are not recursively enumerable, the semantics are unclear, and since
defaults are expressed as rules, they cannot be reasoned about.
MINIMAL MODEL APPROACHES: Theorems are sentences that are true in
distinguished models of the theory.
Circumscription: Don't get me started. Predicates P are circumscribed
with respect to other predicates using a second-order logic axiom, so
that any minimal theory is one in which the predicates P have the
narrowest possible definition.
Broadest formalism; has decision theories for versions with constraints.
EPISTEMIC APPROACHES: Use belief or knowledge to model nonmonotonic
reasoning.
Autoepistemic Logic: Based upon nonmonotonic logic, Moore (1984, 1985)
adds an additional modal operator B for belief. A set of formulae T is
defined as a stable expansion of a set A if
T = Th(A U {Bw | w in T} U {!Bw | w not in T}).
Levesque extends this to first-order case; adds operator O where Ow means
"w is all that is believed." Konolige proves that there is an exact
correspondence between the extensions of default logic and strongly
grounded expansions of autoepistemic logic.
CONDITIONAL LOGICS: Use subjunctive conditionals ("If A were the case,
then B would be the case") to formalize nonmonotonic reasoning.
Not successful approach; such logics are monotonic, also very weak, and
rely upon pragmatics to achieve nonmonotnic effects, leaving them without
principled semantics.
Back to the Question of the Day
Page