||June 9, 2006
||Dr. Gyorgy Turan, University of Illinois at Chicago, Chicago, USA and Hungarian Academy of Sciences
||Remarks on learning and commonsense reasoning
|| Providing agents with commonsense reasoning capability is a fundamental
task of artificial intelligence, and it is also important for applications
as user interfaces and natural language processing. Commonsense reasoning
is a huge area with interesting mathematical theories. It is often pointed
out that the existing frameworks should be extended to include learning.
We give a brief introduction to this direction of research, formulating some
A point of entry into the many-faceted area of commonsense reasoning
and learning is belief revision, the study of how to revise a knowledge base
if new information is received that may be inconsistent with what is known.
Here one usually begins with postulates required of a rational revision
such as the AGM postulates, aimed at formalizing the requirement of
minimal change. There are representation results, constructions (akin to
learning algorithms) and connections to probabilistic reasoning. It seems
to be a challenging general question whether successful learning and
revision can be combined. So far, this has been considered mostly in
inductive inference, but it is also discussed in machine learning.