COMMENT ⊗   VALID 00004 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "memo2.pub" source
C00003 00003	.cb INTRODUCTION
C00016 00004	.skip 1
C00020 ENDMK
C⊗;
.require "memo2.pub" source;
.turn off "{"
.skip 20
.cb EPISTEMOLOGICAL PROBLEMS OF COGNOLOGY

.SKIP 15
Abstract: This paper attempts to identify the epistemological problems
of %2artificial intelligence%1, or, as we might rename the science,
%2cognology%1.  An admittedly incomplete classification is given,
and some partial results on a two problems are given.  These are:
formalization of concepts and the reasoning involved in jumping to
conclusions.

.skip 5
.cb INTRODUCTION

	One's opinion on how long it will take to achieve human-level
artificial intelligence largely depends
on what problems one can see between the present state of
our science and this goal.  Unless one is convinced that the problems he
can see are all the important problems, one can only give lower bounds on
the time required to overcome them.  Since 1958 (see McCarthy 1959), I
have seen epistemological problems as major obstacles to artificial
intelligence, and I have always been disappointed at how few people work
seriously on them.  For example, when Erik Sandewall and I announced a
special issue of the AI Journal devoted to these problems, we received too
few relevant papers to justify the special issue.

	The neglect of epistemology seems related to
the fact that 95α% of the work in artificial intelligence is engineering
(trying to make a program with specified performance), and only 5α% is
science (studying a natural phenomenon).  I think it should be about
75-25.  For this reason, and also to get a name more parallel to those of
other sciences, I suggest renaming our subject %2cognology%1 - %3the
science of intellectual processes and their relation to the problems they
solve%1.

	%2Epistemological%1 problems of cognology are clarified by trying
to separate them from the %2heuristic%1 problems.  %2Epistemology%1 studies
what kinds of facts are available for attaining goals, and %2heuristics%1
studies how to obtain goal-relevant facts
and how to use them to achieve the goals.
Epistemology is also concerned with
the valid and useful modes of reasoning that allow concluding
that a problem has a certain solution.  Admittedly this separation
can only be partial, because facts about heuristics may be needed to
solve the subproblem of what heuristics to use (whenever finding
heuristics reaches the level of a subproblem),
and the heuristic methods for dealing with facts
depend on what kinds of facts there are.
Moreover, the problem of how to represent facts has both epistemological and
heuristic aspects.
Nevertheless, we can identify many specifically epistemological problems
that lie between us and human-level artificial intelligence.

	Our first observation about the information available for
solving problems is that it is usually partial knowledge.
Suppose, for example, that a person or a program wants to persuade
someone to allow it access to a certain file of information by
arguing that mutual benefits will result.  The would-be persuader
doesn't know the state of the persuadee's mind, he doesn't know
precisely what is in the file and why it is valuable, and he doesn't
know precise laws specifying how people are affected by argument.
Contrast this with chess wherein positions are known,
and so are the laws giving the effects of moves.

	A formalism that is suitable for representing complete
information about a class of circumstances and for representing
the laws of motion of completely described circumstances may be
entirely unsuitable for representing the information actually
available to a person or program with given opportunities to
observe and compute.  For example, consider a cup of coffee
spilled on a table with papers on it.
Suppose that a person or robot who sees the coffee spilled wishes
to save the papers.  The behavior of the coffee is determined
by the Navier-Stokes equations of mathematical physics.  Given
a controlled experiment, six months to program, and a good computer,
it should be possible to compute rather precisely which papers
are endangered.  However, a human can make a quick estimate and in
less than a second can perform a fairly good triage among the papers.  
We may call the information used %2common sense hydrodynamics%1 and
try to formalize it in a way that can be used by a computer program.
While scientific hydrodynamics may not be useful in a coffee
spilling emergency, its importance is not mainly for predicting
the outcome of particular situations, however well measured.
Rather it is used for establishing general laws useful for designing
boats, airplanes and waterways.  In fact if we wanted to develop
our ability to rescue papers from coffee, our research might
well use scientific hydrodynamics to develop improved rules of thumb
that we would incorporate into reflexes by training.

	More generally, the epistemological problem of cognology
is that of discovering what kinds of information will be available
in particular situations, what are the general facts about
phenomena, and what laws determine the effects of actions.

	Here is a list of some of the common sense phenomena which
will have to be formalized.
.skip 1
.begin verbatim
John McCarthy
Artificial Intelligence Laboratory
Computer Science Department
Stanford University
Stanford, California 94305

ARPANET: MCCARTHY@SU-AI
.end

.turn on "{"
%7This draft of
IJCAI[E77,JMC]
PUBbed at {time} on {date}.%1