clarke@utcsri.UUCP (Jim Clarke) (01/27/86)
(SF = Sandford Fleming Building, 10 King's College Road)
ARTIFICIAL INTELLIGENCE SEMINAR,
Tuesday, February 4, 3 pm, SF 1105
Professor Jaime Carbonell
Carnegie-Mellon University
"Analogical Reasoning:
Some New Results and Directions"
Analogical reasoning is a significant cognitive process that has
hereto not been modelled by AI researchers in a computationally tractable
manner. Recently, several new approaches have shown significant promise
using analogical processes for problem solving and learning. Among the
first and most comprehensive, the ARIES project demonstrated that analogi-
cal problem solving is a computationally tractable means of exploiting past
experience to solve new problems of increasing complexity. Two methods
were developed: transformation analogy where solutions to related problems
are incrementally transformed into the solution of the new problem, and
derivational analogy where the problem solving strategies, rather than the
resultant solutions, are transferred across like problems.
However, many interesting questions remained unanswered, such as: How
closely related should problems be prior to analogical transfer? How does
one measure similarity? What is the relation of analogical transfer to
human problem solving capabilities? What role does analogy play in learn-
ing? Should analogical reasoning be considered an integral aspect of any
unified problem-solving architecture striving to model human cognition?
After a glimpse into the basic ARIES model, partial answers to these ques-
tions will be discussed, based on recent work. Some of these new direc-
tions are rooted in concrete computational and psychological results; oth-
ers are of a more speculative nature.
ARTIFICIAL INTELLIGENCE SEMINAR
Thursday, February 6, 11 am, SF 1105
Professor Dave Smith
Stanford University
"Controlling Backward Inference"
Effective control of inference is a critical problem in Artificial
Intelligence. Expert systems have made use of powerful domain-dependent
control information to beat the combinatorics of inference. However, it is
not always feasible or convenient to provide all of the domain-dependent
control that may be needed, especially for systems that must handle a wide
variety of inference problems, or must function in a changing environment.
In this talk a powerful domain-independent means of controlling inference
is proposed. The basic approach is to compute expected cost and probabil-
ity of success for different backward inference strategies. This informa-
tion is used to select between inference steps and to compute the best
order for processing conjuncts. The necessary expected cost and probabil-
ity calculations rely on simple information about the contents of the prob-
lem solvers database, such as the number of facts of each given form and
the domain sizes for the predicates and relations involved.
--
Jim Clarke -- Dept. of Computer Science, Univ. of Toronto, Canada M5S 1A4
(416) 978-4058
{allegra,cornell,decvax,ihnp4,linus,utzoo}!utcsri!clarke