[mod.ai] Seminar - Explicit Contextual Knowledge in Learning

KEDAR-CABELLI@RED.RUTGERS.EDU (Smadar) (11/09/86)

Reminder:   Dissertation Defense for Rich Keller
Time and Place: Thursday, Nov. 13, 1:30 p.m., Hill 423

Committee:  Tom Mitchell (chair)
            Thorne McCarty
            Lou Steinberg
            Jack Mostow

Abstract:

             The Role of Explicit Contextual Knowledge in
               Learning Concepts to Improve Performance

                          Richard M. Keller
                       (KELLER@RED.RUTGERS.EDU)


    This  dissertation  addresses some of the difficulties encountered
when using artificial intelligence-based, inductive  concept  learning
methods  to  improve an existing system's performance.  The underlying
problem is that inductive methods are insensitive to  changes  in  the
system  being  improved by learning.  This insensitivity is due to the
manner in which contextual knowledge is represented  in  an  inductive
system.   Contextual knowledge consists of knowledge about the context
in which concept learning takes place, including knowledge  about  the
desired form and content of concept descriptions to be learned (target
concept knowledge), and knowledge about the system to be  improved  by
learning  and  the  type  of  improvement  desired (performance system
knowledge).    A  considerable  amount  of  contextual  knowledge   is
"compiled" by an inductive system's designers into its data structures
and procedures.  Unfortunately, in this compiled form, it is difficult
for  the  learning  system  to  modify  its  contextual  knowledge  to
accommodate changes in the learning context over time.

    This research investigates the  advantages  of  making  contextual
knowledge  explicit  in a concept learning system by representing that
knowledge directly, in terms of express declarative structures.    The
thesis  of this research is that aside from facilitating adaptation to
change, explicit contextual knowledge  is  useful  in  addressing  two
additional  problems  with  inductive systems.  First, most  inductive 
systems are unable to learn  approximate  concept  descriptions,  even  
when approximation is necessary or desirable  to improve  performance.
Second, the capability of  a  learning  system  to  generate  its  own
concept  learning  tasks  appears  to  be outside the scope of current
inductive systems.

    To investigate the thesis, this study  introduces  an  alternative
concept learning framework -- the concept operationalization framework
-- that requires various types of  contextual  knowledge  as  explicit
inputs.    To  test  this new framework, an existing inductive concept
learning system (the LEX system [Mitchell et al. 81]) was rewritten as
a  concept  operationalization  system  (the  MetaLEX  system).   This
dissertation  describes the design of MetaLEX and reports  results  of
several experiments performed to test the system.  Results confirm the
utility  of  explicit  contextual  knowledge,  and  suggest   possible
improvements in the representations and methods used by the system.
-------