COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	common[w84,jmc]		Definition of common sense for Kay Mills
C00008 00003		The core of the knowledge concerns events that occur in
C00010 00004	What is common sense?
C00021 00005	A leftover:
C00022 ENDMK
C⊗;
common[w84,jmc]		Definition of common sense for Kay Mills

	There is common sense knowledge and common sense ability.
Common sense knowledge includes facts about events occurring in
time, about the effects of actions by the knower and others, about
physical objects and how they are perceived, and about their properties
and their relations to one another.  An example is the fact that
eggs contain a yolk and a white and a shell, how to recognize an
egg, the effects of hard boiling them and the effects of dropping them.
Common sense ability involves the use of common sense knowledge and
the observation of the world to decide what to do to achieve one's
goals.  The "common" in "common sense" refers to the fact that a
large amount of this knowledge and ability is common to all humans.
Not much of it is understood well enough to include it in computer
programs.

	For AI purposes we take common sense to include certain
knowledge about the world and a certain ability to use it to infer
addtional facts from observations and general knowledge and to use
this reasoning to achieve goals.

	The core of common sense knowledge concerns actions and other
events.  Every human knows that he can perform a variety of actions
and that the chosen actions affect the future in characteristic ways.
However, if we want computer programs to know this, we must provide
the knowledge and build into their structure the ability to use it.
Very few expert systems are provided with any such knowledge.  Instead
they have rules that say what action to take when given conditions occur.
The likely consequences of these actions were considered by the programmer
in deciding what rules to provide, but the languages for writing expert
systems provide no direct way of his telling the program such facts
or any automatic way for the program to use the information if he did.

	Humans have both.  If the car in front of us suddenly stops,
a learned rule causes us to automatically hit the brakes.  We also
have at least a verbal knowledge of what will happen if we don't.
Some of us can think fast, so that we can quickly decide whether it
is better in a particular case to hit the brakes as hard as possible
or try to swerve around the car in front.  This requires estimating
the chances of a head on collision with a car going the other way.

	It is not just an oversight that programs don't have both.
Rules are easier to write than common sense facts, and it is very
much easier to make a system that obeys rules or compiles them into
program than to make a system that acts according to reasoning from
facts or compiles facts into rules or into program.  It would be
still harder to make a system that will compile rules that take facts
into account but also provide for reverting to reasoning when this
is appropriate, i.e. when the situation is complex and there is time
to think.  Humans have, and machines need, the ability to operate
by rule and intermittently monitor and modify the operation of the rules.

	The facts that describe the consequences of actions, and
while we're at it, the facts that describe the consequences of
events the actor doesn't control, are the most important common
sense knowledge.  However, there's lots more.

	The core of the knowledge concerns events that occur in
time and the effects of events on the state of the world.  Knowledge
about the effects of actions by the knower are particularly important.
Facts about the relations between appearances and reality are
included.

Facts about goals and how to achieve them.

Facts about what actions are permissible.

To what extent are common sense abilities achievable by production
systems?

We would like a system that can find an algorithm for block stacking
given the facts about moving blocks.  The ontology of this system
should include sequences of events.  This enables us to say that
an efficient algorithm is one whose execution results in a short
(more generally cheap) sequence of actions and other events leading
to the goal.
What is common sense?

	It is now generally agreed among AI researchers that
making programs with common sense abilities is at present the
key problem facing AI research.  However, there is so far no
paper attempting a general description of what common sense
ability is.  There are many examples of specific common sense
capabilities.  The object of this paper is to attempt a general
description of common sense.

	Common sense involves certain abilities to decide what
to do to achieve goals.  These abilities are common to all
non-feeble-minded humans; indeed many very retarded people
possess common sense abilities so far lacking in any computer
program.  Many of these abilities seem to transcend specific
domains of application.

	Human activity seems to be governed by a combination of
behavior rules and reasoning based on facts.  By a behavior rule
we mean one of the form: If X is true, then do Y.  X is typically
a conjunction and included in the conjunction are usually statements
about current goals.  Moreover, both  X  and  Y  contain variables,
and the variables have to be matched against the world and the
database in order to generate a definite action.

	However, besides rules, humans also accumulate facts.  Here
are some examples:

	1. London is in the south of England.

	2. When objects collide they usually make a noise.

	3. The Stanford computer science comprehensive examination
contains a section on numerical analysis.

	4. When a person fires a handgun a chemical residue is left
on his hands.

	Such facts may be used to answer geography questions, to
conclude that a side trip to London while visiting Edinburgh may
take more time than is available, to conclude that a screech of
brakes was not actually followed by the collision of the cars,
to decide to take one's graduate work in computer science
elsewhere than in Stanford or to decide to have someone else
shoot one's desired victim.  The writer of a rule-based expert
system will take such facts into account in writing his rules.
However, the present expert system technology has only a limited
ability to include the facts themselves and insure that they
will be taken into account in as wide a variety of situations
as a human would do.

	We may call the human ability to take facts into account
common sense reasoning ability.  In any given case, it may be
difficult to ascertain whether an action rule is being invoked
or whether reasoning with facts is being done.  The distinction
is not always clear.  First a system that reasons with facts may
itself be built up from rules.  Second the information is available
in redundant forms.

	Consider a person whose car needs fixing.  He may decide
that he will be without his car for several days and hence should
shop for several days food in advance, because it will inconvenient
to do so without the car.  It seems to me that for some people this
is an example of fact based reasoning, but it may become "compiled"
into a rule after some experience.  Some smart people will compile it into
a rule from just contemplating the possible situation before any
experience.

	Rule based systems are usually considered "brittle".  By
comparison with humans, especially the experts whose abilities
they are designed to incorporate, they cannot take new facts into
account unless they fit the format of their databases.  For example,
it might be difficult to tell Mycin about AIDS without rewriting
it extensively.  Presumably, if AIDS is a possible diagnosis, the
probabilities of the different bacterial infections are drastically
altered, because some bacteria that the human body can ordinarily
handle without help are now dangerous.

	Often facts enter generally rule-based behavior as an
exception handling mechanism.  One is carrying out some routine
activity in pursuit of a customary goal when suddenly an event
not previously imagined occurs and one says to oneself, "I'd
better think".  Thinking often involves going into a fact-dominated
mode.
  Thinking is often unsuccessful in the sense that
one later concludes that one could have handled a situation much
better if only one had thought of the answer in time.  A good
robot will presumably have many of the same characteristic.  It
has rules, but when exceptions are encountered, it has to reason,
and this reason is only sometimes capable of coming up with
the right answer even when the relevant facts are all in its
database.

	The rest of this essay is devoted to two questions.
First, what are the main common sense facts?  Second, how are
these facts used in common sense reasoning in conjunction with
rules.  Only very preliminary answers are possible in the present
state of AI.


THE FACTS OF THE COMMON SENSE WORLD

	It seems that the most important class of facts are those
that permit the prediction of the effects of events including
actions.  Basic to this is the fact that the world does change
and that actions and other events have effects.  However, this
fact is so basic that its explicit representation may not play
an important role.  Namely, it is built into the human mechanism
rather than used as a fact by a reasoning mechanism.  Such
treatment may also be appropriate in designing computer programs.
An analogous example may be helpful.

	Consider the regress involving modus ponens proposed
by Lewis Carroll and which Douglas Hofstadter has made much of.
Achilles infers ⊗q from ⊗p and ⊗p_⊃_q.  The tortoise asks why,
and Achilles cites the rule, "If ⊗p is true and ⊗p_⊃_q is true,
then ⊗q is also true".  The tortoise asks why, and the regression
continues.  Achilles should have stopped the regress at the first
stage.  When asked why he thinks so, he should have answered, "Because
that's the way I'm constructed.  I'm built to allow modus ponens".
The higher level modus ponens's aren't higher level principles,
they are simply more examples of the same principle.

	In any case, the idea that the future lies ahead and can
be affected is most likely built in.  The fact that we can formulate
it as a fact is of secondary importance in daily life, although it
is important when thinking about philosophy or about designing robots.

	There is an important difference between common sense ability
to predict the consequences of physical actions and the laws of
physics.  Namely, common sense physics allows only incomplete
partial prediction.  When an object is thrown, we can predict where
it will land very approximately, but this information is insufficient
to predict where it will go next.
A leftover:
The system that accumulates facts and takes them into account
may itself be organized as a rule based system.  Moreover, a
system organized primarily as rule a based system, i.e. doing
do most of its work by executing rules, may go into a mode that
uses facts when exceptions are detected.  A person may say, "This
has gotten complicated.  Now I have to think."