perm filename HANDOU.226[W83,JMC]1 blob sn#694769 filedate 1983-01-08 generic text, type C, neo UTF8
COMMENT āŠ—   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	handou.226[w83,jmc]	Handout for cs226 - Common sense data base, 1983 winter
C00011 ENDMK
CāŠ—;
handou.226[w83,jmc]	Handout for cs226 - Common sense data base, 1983 winter

	This year CS226 will be based on papers relevant to the problem
of constructing a general database of common sense facts.

	Imagine a general purpose problem solving program that uses
a database of common sense facts expressed in some language of mathematical
logic.  By a general purpose program, we mean that it should be
capable of attacking any problem given the facts specific to the
problem and the general facts in the database.  We imagine that
the database is to be constructed before anyone knows what problems
will be attacked.

	Constructing such a database involves problems in artificial
intelligence and epistemology - the branch of philosophy concerned with
what knowledge is.  However, neither subject has yet developed adequate
formalisms for this task.

	Philosophy and artificial intelligence approach the job with
somewhat different attitudes, and this course takes the point of view
of artificial intelligence.  A philosopher will propose a formalism,
and then some other philosopher will find counterexamples to its
general applicability, and it will be abandoned.  In artificial
intelligence, we cannot afford to abandon formalisms until we can
replace them by something better.  Therefore, we will spend considerable
time on formalisms that are known to be restricted in their generality.
We will investigate what they can do and how they can be used and
extended within their domains of applicability.  Of course, we must
also study their limitations and how to get around them.

	A key formalism will be the "situation calculus" of
(McCarthy and Hayes 1969).  This has had its ups and downs in
popularity.  It has many inadequacies, but there doesn't yet exist
a general purpose replacements, and many purported replacements
turn out to be specializations.

	A key problem of the course is to determine what is the
domain of general common sense facts.  No-one has a definitive
answer to this question either, but here is a beginning.

	. There are both situation dependent facts like the
location of an object and situation independent facts like those
of mathematics.  We need languages that can express both.

	. Events occur and cause changes in the situation dependent
facts.  We need to be able to express the laws that determine
what changes events will cause.  We need epistemologically adequate
formalisms (McCarthy and Hayes) that can express what is known.  Thus
we need to express the qualitative fact that an egg dropped on concrete
is likely to break quite separately from our knowledge of the laws
of physics that cover the same phenomenon.  We will often have neither
the initial conditions nor the computing ability to use the physics
formalism directly.  We, or people or programs whose thinking we want
to discuss, may not know the relevant physical laws.  An epistemologically
adequate formalism can express the knowledge that is actually used.

	. Some events are actions of persons (including robots).  We need to
formalize facts about the ability of persons to achieve goals.
This involves the effects of individual actions and strategies or
programs that involve both observation and action.

	. We need formalisms that express ascriptions of beliefs,
knowledge, goals and other mental qualities and attitudes to
persons (including robots, computer programs and other machines).
Actions and programs intended to obtain knowledge must be
describable, and we must be able to reason about the conditions
under which they will succeed.

	. Often the spatial character of events is important, and
we need to be able to describe objects and substances extended in
space and their changes.  Our information about objects is
partial and so is the information obtainable by machines.  Therefore,
our formalisms must be good for expressing such partial information.

	. Programs that deal with human interactions must be able
to reason about more complicated attitudes and actions such as
likes and dislikes, intentions and hopes, buying, selling, duties
and rights.

	. A major unsolved problem of common sense formalization is
expressing facts about concurrent actions.  The theory of concurrent
programs may be useful, but notice that common sense reasoning about
concurrent events rarely gives much attention to the synchronization
problems that dominate the theory of concurrent programming.

	. Common sense reasoning itself includes parts that correspond
to logical deduction, but there is more, specifically non-monotonic
reasoning of various kinds.  Recently various systems
of formalized non-monotonic reasoning have been developed.

	. The problems that philosophers have studied in defining
common sense concepts are real, and some of them have been studied
for hundreds and even thousands of years without definitive solution.
An example is causality.  For artificial intelligence, we need to
be able to express causal facts without fully understanding
causality.  Our formalisms must be "ambiguity tolerant" and
"elaboration tolerant" - concepts that will be proposed in
CS226.

	Other problems that have been identified and will be studied
are the "frame problem" and the "qualification problem".

	We will normally cover a paper a week and sometimes a paper
per class meeting.  The papers will be made available.

	An acquaintance with the language of first order logic
is a prerequisite.  We will discuss set theory axioms and their
possible AI use, and we may use the EKL interactive theorem prover.