COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	THE STRUCTURE OF THE ARTIFICIAL INTELLIGENCE PROBLEM
C00006 ENDMK
C⊗;
THE STRUCTURE OF THE ARTIFICIAL INTELLIGENCE PROBLEM


	The object  of this paper  is to  discuss how the  artificial
intelligence  problem as it  exists today is  composed of subproblems
and to discuss the relation of the work being done by  various groups
to this structure.  Naturally, its purpose is to influence workers to
move towards problems that seem to be important. 

	The  object  of  AI  research  is  to  acquire the  knowledge
necessary to build  intelligent machines.   To this  end, we want  to
understand the  various mechanisms and other  aspects of intelligence
[1].    We  could   consider  the  artificial  intelligence   problem
substantially solved if we could write a  computer program that could
understand any human produced texts and could improve its own problem
solving ability better than we can improve it ourselves. 

	We shall proceed by trying  to divide the problem into  parts
and analyze the parts separately. 

	1.  First  consider the  problem  of  generalizations.   This
problems has four subproblems:

	a. What kinds of generalizations are there? 

	b. How are generalizations to be represented in the memory of
a computer? 

	c. How are generalizations to be used? 

	d. How are new generalizations to be obtained? 

We shall start our discussion  with the last of these problems.   The
ability  to make  general statements  on the  basis of  experience or
deduction from previous knowledge  is one of  the highest aspects  of
intelligence.  Even confirming generalizations already conjectured is
difficult.    The learning  of generalizations  from  experience that
computer programs have so  far been programmed to  carry out is on  a
very elementary  level.  So  far it  amounts to hill-climbing  on the
values of some parameters.  Therefore, we suggest that the ability to
make powerful generalizations will  be one of the later  successes of
artificial  intelligence research.   Once we  understand how  to make
computers carrry out other difficult intellectual tasks, then we  may
be ready to try to program them to  make powerful generalizations.  I
don't say that nothing  can be done now, but I do assert that we will
have to understand intelligence a  lot better before we will be  able
to program computers to improve  their intellectual capabilities in a
general way.