be perceived out of the corner of his eye in
one
s i t u a t i o n as the c y l i n d r i c a l top of his
electric coffee grinder (he is
at
home
in
his kitchen), but as the flasher of a police
car (he
is
speeding
on
the
freeway)
in
another.
This
suggests
that
at
every
moment, some
fairly
large
swatch
of
his
knowledge
about the world somehow has found
its way
to
the
foreground
to
exert
its
influence;
as
our
robot
moves
about,
swatches must fade
in
and
out,
sometimes
coalescing,
so that at any moment, just the
right one is standing by to help guide
acts
of planning, infePence and perception.
THE C O M M O N S E N S E A L G O R I T H M
AS A BASIS FOR COMFUTER MODELS
OF HUMAN MEMORY, INFERENCE, BELIEF
AND C O N T E X T U A L LANGUAGE C O M P R E H E N S I O N
Chuck Rieger
D e p a r t m e n t of Computer Science
U n i v e r s i t y of Maryland
ABSTRACT
The
notion
of
a
commonsense
algorithm
is
presented as a basic
data s t r u c t u r e for
modeling
human
cognition.
This
data
structure
unifies many
current
ideas
about
human
memory
and
information
processing.
The
structure
is
defined
by
specifying
a
set
of
proposed cognitive primitive
links
which,
when used to build up large
structures
of
actions,
states,
statechanges
and
tendencies,
provide an a d e q u a t e
formalism
for
expressing
human
plans
and
activities,
as
well
as
general
m e c h a n i s m s and c o m p u t e r algorithms.
The c o m m o n s e n s e a l g o r i t h m is a type
of f r a m e w o r k (as Minsky has defined
the
term)
for
representing
algorithmic
processes,
hopefully
the way humans do.
I.
INTRODUCTION
AND
Marvin Minsky has captured
this
whole
idea
very
neatly
in his w i d e l y - c i r c u l a t e d
"Frames"
paper
[MI].
While
this
paper
describes
an
overall
a p p r o a c h to m o d e l i n g
human
memory,
inference
and
beliefs,
we
still
lack
any specific f o r m u l a t i o n of the
ingredients
which
make
up
the
large,
explicitly-unified
s t r u c t u r e s which seem to
u n d e r l i e many h i g h e r - l e v e l
human
cognitive
functions.
It is the purpose of this paper
to define the notion " c o m m o n s e n s e a l g o r i t h m "
(CSA)
and
to
propose the CSA as the basic
Cognitive
structure
which
underlies
the
human
processes
of planning, i n f e r e n c e and
c o n t e x t u a l i n t e r p r e t a t i o n of meaning.
I do not have a
complete
theory
yet:
the
intent
of
this
paper
is to record a
memory dump of ideas
accumulated
over
the
past
few
months
and
to show how they can
unify my past ideas on inference and memory,
as well as the ideas of others.
MOTIVATION
It is b e c o m i n g i n c r e a s i n g l y evident
to
human
intelligence
model
builders
and
t h e o r i s t s that,
in
order
to
characterize
human
k n o w l e d g e and belief as c o m p u t e r data
s t r u c t u r e s and processes, it is n e c e s s a r y to
deal
with
very
large,
explicitly unified
s t r u c t u r e s rather
than
smaller,
ununified
fragments.
The reason for this seems to be
that the o r i g i n a l e x p e r i e n c e s
which
caused
the s t r u c t u r e s and processes to exist in the
first
place
come
in
chunks
themselves;
knowledge
is
never
gained outside of some
context,
and
in
gaining
some
piece
of
knowledge
X
in
context
C, X and C become
inseparable.
This
suggests
that
it
is
meaningless
to model "a piece of k n o w l e d g e "
without regard for the larger s t r u c t u r e s
of
w h ich it is a part.
If our goal is to build
a
robot
which
behaves
and
perceives
in
manners
s i m i l a r to a human, this means that
the process by which
the
robot
selects
a
piece
of
knowledge
as being a p p l i c a b l e to
the
planning,
executory,
inferential
or
interpretive
process
at hand at the moment
is a
fu n c t i o n
not
only
of
the
specific
problem,
but
also of the larger context in
which that instance of planning,
execution,
i n f e r e n c e or i n t e r p r e t a t i o n occurs.
If, for
example, our robot sees his
friend
with
a
w r e t c h e d facial expression, the i n f e r e n c e he
makes about the
reasons
for
his
friend's
misery
will
reflect
the larger picture of
w h ich he is aware at the
time:
his
friend
has
just
returned
from a trip to p u r c h a s e
opera tickets vs.
his friend has Just eaten
the
cache
of m u s h r o o m s c o l l e c t e d y e s t e r d a y
vs . . . . .
The same p e r v a s i v e n e s s of context
exists
in
the
realm
of
the
robot's
i n t e r p r e t a t i o n s of visual
perceptions:
the
very same object (visible at eye level) will
II.
THE SCOPE
OF THE CSA'S A P P L I C A B I L I T Y
Most
of
human
knowledge
can
be
c l a s s i f i e d as either static or dvnamic.
For
example, a person's static k n o w l e d g e
of
an
automobile
tells
him
its general p h y s i c a l
shape, size,
position
of
steering
wheel,
wheels,
engine,
seats, etc.; these are the
a b s t r a c t a s p e c t s of a
car
which,
although
many
differ
in detail from car to car, are
i n h e r e n t l y unchanging.
They are in
essence
the
physical
definition
of a car.
On the
other hand, a person s dynamic k n o w l e d g e
of
a car tells him the f u n c t i o n s of the v a r i o u s
c o m p o n e n t s and how
and
why
to
coordinate
them
when
the car is a p p l i e d to some goal.
The static k n o w l e d g e tells the person
where
to
expect
the s t e e r i n g wheel to be when he
gets in; the dynamic k n o w l e d g e tells him how
to get in in the first place, and what to do
with the wheel (and why) once he is in.
For
a
robot
immersed
in
a
highly
kinematic
world -- physically,
psychologically
and
s o c i a l l y -- a very large part of his b e l i e f s
and k n o w l e d g e must relate to
dynamics:
how
he
can
effect
changes
in h i m s e l f and his
world, and how
he
perceives
other
robots
e f f e c t i n g changes.
It is the purpose of the
CSA to capture the d y n a m i c s of the w o r l d
in
belief
structures
which
are
amenable
to
computer manipulation
of
plans,
inference
and c o n t e x t u a l i n t e r p r e t a t i o n .
It should be s t r e s s e d that
the
phrase
"dynamics
of
the world" is i n t e n d e d in its
broadest
possible
sense.
As
will
be
elaborated
upon
in
a
later
section, the
18o
I
I
I
I
I
I
i
I
i
II
I
A refinement of the notion of an AND/OR
graph
introduces
the concepts of c a u s a l i t y
and
enablement,
and
actions
and
states
(statechanges);
edges
in
the
graph
are
d i s t i n g u i s h e d as either causal or
enabling,
the
nodes
are
distinguished
as
either
actions or states, and the graph
obeys
the
syntactic contraints:
phrase
is
intended
to
encompass
such
s e e m i n g l y diverse r o b o t / h u m a n a c t i v i t i e s as:
I. c o m m u n i c a t i n g
with
another
robot/human
(e.g., how to t r a n s f e r
information,
instill
wants,
convince, etc.)
2. getting about in the world
3. building things (both physical
and
mental)
and
understanding
the
operation of things
already
built
by others
4. conceiving,
designing
.and
implementing
computer programs and
other
commonsense
algorithms
(a
special form of building)
5. i n t e r p r e t i n g
sequences
of
perceptions
(e.g.,
language
utterances) in context
6. making
contextually
meaningful
inferences from perceptions
(a) actions cause states
(b) states enable actions
Bob Abelson [AI]
was
among
the
first
to
employ
these h i s t o r i c a l l y very old concepts
in the framework
of a computer
model
of
human
belief,
and
since
then,
numerous
computer-oriented
systems
of
knowledge
reDresentation
(e.g.,
Schank's
conceptual
d e D e n d e n c y [ S 2 ] , Schmidt's models of p e r s o n a l
causation
[$4]),
as well
as
systems
of
inference (Rieger [RI], Charniak [CI])
have
found
these
four
concepts
to be vital to
meaning representation
and
inference.
In
some
sense,
enablement,
causality, states
and actions seem to be cognitive primitives.
Figure
2 is a refinement of Figure I which
makes explicit the nature of each
node
and
each
connecting
arc,
and
hence
the
u n d e r l y i n g gross conceptual structure of the
algorithm.
I am c o n v i n c e d that all
such
dynamics
of
the world can and should be e x p r e s s e d in
a uniform
CSA
formalism
built
around
a
relatively
small
number
of c o g n i t i v e l y
primitive ingredients.
III.
E V O L U T I O N OF THE CSA IDEA
While
the
inclusion
of
these
four
concepts
(and
their
resulting
syntactic
constraints) in the basic paradigm makes for
a
theoretically
more
coherent
representation,
the
scheme
is
still
too
coarse
to
capture
the
kinds
of detailed
knowledge of a l g o r i t h m s people possess.
The
following
section
proposes
an
extended
framework
of
event
types
and
event
connectors
based
on these four notions and
some
others.
These
event
types
and
connectors
will
be
regarded
as
model-primitives
which
hopefully
are
in
correspondence
with
"psychological
p r i m i t i v e s " in humans.
The next section will define a CSA as a
network-like
structure c o n s i s t i n g of events
tied together by primitive links.
Taken
as
a whole, the CSA specifies a process: how to
get something
done,
how
something
works,
etc.
A computer scientist's first reaction
to this type of s t r u c t u r e is "Oh yes, that's
an
AND/OR
problem-reduction
graph"
(see
Nilsson [NI] for example).
Figure
I shows
an
AND/OR graph for how to achieve the goal
state "a
McDonald's
hamburger
is
in
P's
stomach."
Edges
with
an
arc through them
specify AND successors of a node
(subgoals,
all
of which achieved imply the parent node
has
been
achieved);
edges
with
no
arc
through
them
specify
OR
successors
(subgoals, any one of which being sufficient
to achieve the parent goal).
IV.
D E F I N I T I O N OF THE C O M M O N S E N S E A L G O R I T H M
In the new formalism,
nodes of five types:
AND/OR graphs
have
been
demonstrated
adequate
in
practice
for
guiding various
aspects
of
problem-solving
behavior
in
existing
robots
(see
[$I]
for e x a m p l e ) .
However,
they
are
intuitively
not
theoretically
adequate
structures
for
representing
general
knowledge
of
world
dynamics: their principal d e f i c i e n c y is that
they are ad-hoc c o n s t r u c t i o n s which
express
neither
the
implicit
conceptual
r e l a t i o n s h i p s among
their
components,
nor
the
inherent
types
of
their
components.
Because of this, there is no constraint
on
their
organization, and this means that two
AND/OR graphs which a c c o m p l i s h or model
the
same
thing
might
bear
very
little
resemblance to o n e - a n o t h e r when in fact they
are
c o n c e p t u a l l y very similar.
This may be
little more than a nuisance in practice, but
it
is u n d e s i r a b l e
in principle because it
makes
learning,
reasoning
by
analogy,
sharing
of
subgoals,
etc. t e d i o u s if not
i m p o s s i b l e in a generalized problem solver.
a CSA c o n s i s t s of
I. WANTS
2. ACTIONS
3. STATES
4. S T A T E C H A N G E S
5. T E N D E N C I E S
The first four types are not new
(see
[$3]
for
example),
and will not be covered here
beyond the f o l l o w i n g brief mention.
A WANT
is
some
goal
state
which is desired by a
potential actor.
An action is something
an
(animate)
actor
does
or
can
do:
it
is
enabled
by
certain
states
(certain
conditions
which
must be true in order for
the action to begin and/or proceed), and
in
turn
causes
other
states
(discrete)
or
s t a t e c h a n g e s (continuous) to occur.
Actions
are
characterized
by
an
actor,
a
m o d e l - p r i m i t i v e action,
a time
aspect,
a
location
aspect,
and
a conceptual
case
framework
which
is
specific
to
each
model-primitive
action.
States
are
c h a r a c t e r i z e d by an object, an attribute,
a
181
value
and
a time aspect; s t a t e c h a n g e s are
c h a r a c t e r i z e d by an
object,
a continuous
state
scale
(temperature, degree of anger,
distance, etc.), a time aspect and b e g i n n i n g
and end points on the scale.
((TYPE • TENDENCY)
( R E F E R E N C E - N A M E . GROW-LONELY)
(ENABLEMENTS
((ALONE P)
iDURATION * ORDERDAYS))
(RESULTS . (WANT P (COMMUNICATE P
X))))
((TYPE . TENDENCY)
( R E F E R E N C E - N A M E . FORGET)
(ENABLEMENTS . (INHEAD ITEM P)
( ( U N R E F E R E N C E D ITEM P)
(DURATION * ORDER??)))
(RESULTS
(STATECHANGE ITEM
R E F E R E N C E - D I F F I C U L T Y X X+d))
It is the notion of a t e n d e n c y which is
new
and
which
serves
to unify a class of
problems
which
have
been
continually
experienced
in
representing
processes.
Basically,
a
tendency
is
an
actorless
action.
Tendencies
are
characterized
by
s p e c i f y i n g a set of enabling c o n d i t i o n s
,and
a
set of result states and/or statechanges.
Whenever
the
enabling
conditions
are
satisfied, the tendency, by some u n s p e c i f i e d
means, causes the
states
and
statechanges
s p e c i f i e d as the t e n d e n c y ' s results.
Hence,
a t e n d e n c y may be r e g a r d e d as a special type
of
non-purposive
action
which
must occur
w h e n e v e r all
its
enabling
conditions
are
satisfied.
Contrasting
the
notion
of
a
t e n d e n c y with the notion of an action yields
a
rather compact d e f i n i t i o n of what makes a
" v o l i t i o n a l " a c t i o n volitional: a v o l i t i o n a l
action
is
an
action
w h i c h need not occur
even
though
all
its
physical
enabling
conditions
are
met.
The reason it may not
occur is, of course, that the actor does not
desire
it to occur; t e n d e n c i e s have no such
desires.
Tendencies,
thus
characterized,
will
play
an
important
role
in
modeling
a l g o r i t h m i c p r o c e s s e s via CSA's.
In
fact,
a d o p t i n g the notion of a t e n d e n c y as a model
p r i m i t i v e points
out
a rather
ubiquitous
principle:
humans
spend
a large amount of
time in
planning
either
how
to
overcome
tendencies
which
stand in the way of their
goals, or how to harness them at the
proper
times
in place of an action (e.g., d r o p p i n g
the large rock on the coconut).
Although
a
t e n d e n c y ' s p r i m a r y use is at the edge of the
world
model,
where
things
happen
simply
because "that's the way things are", it will
p r o b a b l y be d e s i r a b l e to have the a b i l i t y to
regard
as
tendencies
things which in fact
can be explained.
Characterizing
something
as
a
tendency
even
though
it may
be
r e d u c e a b l e to further a l g o r i t h m s is p r o b a b l y
one
tactic
a human employs when c o n f r o n t e d
with the
analysis
of
very
complex,
olny
p a r t i a l l y u n d e r s t o o d processes.
Even t h o u g h
something ~
be
further
explained,
the
system
of
r e p r e s e n t a t i o n should a l l o w that
s o m e t h i n g to be treated as though it were
a
tendency.
The a b s t r a c t notion of
a tendency
is
meant to be g e n e r a l - p u r p o s e , to c h a r a c t e r i z e
a wide variety of p h e n o m e n a
which
are
not
actions,
but
action-like.
Examples
of
t e n d e n c i e s are:
I. GRAVITY,
PRESSURE,
MAGNETISM,
A T O M I C - F I S S I O N , HEAT-FLOW, and the host
of
other
physical
principles.
Commonsense
GRAVITY
might be c a p t u r e d
as f o l l o w s : * *
T e n d e n c i e s have n u m e r o u s aspects
which
will
require e x p l i c i t c h a r a c t e r i z a t i o n in a
c o m p u t e r model.
Two such aspects relate
to
(I) the
inherent
rapidity
with
which
a
tendency
exerts
itself
and
(2) the
tendency's
periodicity,
if any.
That is,
how q u i c k l y
does
a person
become
hungry
(slope
of
curve), how long does it take to
forget something, how r a p i d l y does an object
accelerate,
how
fast
does
the w a t e r flow
t h r o u g h the nozzle, etc.? If the t e n d e n c y is
periodic, what are the p a r a m e t e r s d e s c r i b i n g
its p e r i o d i c i t y ?
The
primitive
CSA
links
described
in the next section will serve in
part to capture such aspects, but
they
are
not yet adequate.
((TYPE
TENDENCY)
( R E F E R E N C E - N A M E . GRAVITY)
( E N A B L E M E N T S . (UNSUPPORTED OBJ)
(LESSP (DISTANCE OBJ EARTH)
(ORDERMILES))
(RESULTS . ( S T A T E C H A N G E OBJ V E L O C I T Y X
X+d (LOC OBJ)
(LOC EARTH)))
.
2. human b i o l o g i c a l functions: a t e n d e n c y
to GROW-HUNGRY, GROW-SLEEPY, G R O W - O L D E R
(sole e n a b l i n g c o n d i t i o n is the passage
of
time!),
GROW-LARGER,
etc.
For
example:
The CS~
nrimitive
Lin~
Using
these
five
e v e n t - t y p e s as b u i l d i n g blocks (WANTS,
ACTIONS, STATES, S T A T E C H A N G E S ,
TENDENCIES),
the
goal
is
to
be
able
to
express the
dynamics of just about
anything,
be
it
a
physical
device,
a
psychological
tactic
e m p l o y e d by one person
on
another,
how
a
person
p u r c h a s e s a M c D o n a l d ' s hamburger, or
how a c o m p u t e r
program
functions
or
was
constructed.
There
are 25 p r i m i t i v e links
in the current formulation.
They will
only
be
defined
here,
leaving J u s t i f i c a i o n and
details of their use for the e x a m p l e s
which
will
follow,
and
for s u b s e q u e n t papers on
the subject.
In the f o l l o w i n g
definitions,
W,
A,
S,
SC
and
T will stand for WANT,
ACTION,
STATE,
STATECHANGE
and
TENDENCY,
respectively.
((TYPE . TENDENCY)
(REFERENCE-NAME
GROW-HUNGRY)
( E N A B L E M E N T S . (iNOT (LOC N U T R I E N T S
STOMACH))
(DURATION * O R D E R H O U R S ) ) )
(RESULTS . (WANT P (INGEST P N U T R I E N T S
MOUTH STOMACH))))
3. human
psychological
functions:
the
tendency
to
GROW-LONELY, the t e n d e n c y
to FORGET, etc.
For example:
**The
LISP
concurrent
commonsense
actually
be
report will
implementing
notation
reflects
some
thinking
on
how
a
algorithm
system
might
engineered.
A forthcoming
describe
progress
toward
the ideas in this paper.
182
I
I
!
l
I
I
i
I
i
I
i
I
I
I
I
i
I
I
I
TYPE
l
j
l
I
i
I
i
i
I: ONE-SHOT CAUSALITY
fA -r~
Action A or tendency T causes state S.
The action
or t e n d e n c y need
occur
only once;
thereafter
S will
persist
until
altered by another
action or tendency.
For
any
given
S,
there will
ordinarily
be numerous
alternative A's or T's in the
algorithmic
base which would
provide
the
one-shot
causality.
7
TYPE 2: CONTINUOUS
CAUSALITY
Action
or
tendency
A,T's
continuing
existence
continually causes state or statechange S,SC.
Whether
one-shot
or continuous
causality
is
required to maintain S or SC is both a function of S or
• SC and
its
particular
environment
in a particular
algorithm
(i.e., what other tendencies and actions are
influencing
it).
Again,
there
will
ordinarily
be
numerous
actions or tendencies in the algorithmic base
which could provide continuous causality for any
given
state or statechange.
TYPES 3,4: GATED ONE-SHOT AND CONTINUOUS CAUSALITY
A,T causes S,SC either one-shot
or continuously,
providing that all states in [S] are satisfied.
The flow of causality cannot occur
unless
states
specified
by
[S] exist.
That is, even though A,T is
occuring and there is a potential
causal
relationship
between
A,T and S,SC,
the
relationship will not be
realized until the gating states become true.
TYPE 5: ONE-SHOT ENABLEMENT
l
i
State S's one-time
tendency T to proceed.
Thereafter, A,T's
function
of S.
A,T
one-shot
enablements,
satisfied in order for
existence allows
continuation
is
will
ordinarily
in which
case,
A or T to proceed.
action
A
or
no
longer
a
have numerous
all must
be
'
State
S's continued
presence
is requisite
to
action A's or tendency T's continuance.
S's removal causes A or T to halt.
Any given A or
T will ordinarily have numerous continuous enablements,
in which case all must reamin true in order for A or T
to proceed.
TYPE 7: CAUSAL STATE COUPLING
States
$I,
S2 or statechangges
SCI,
SC2 are
causally
coupled; because of this coupling, changes in
$I or SCI are synonomous with changes
in $2 or SC2.
This
link
provides a way of capturing the relatedness
of various aspects of the same situation.
TYPE 8: GATED CAUSAL STATE COUPLING
State $2 or statechange
SC2
is synonymous
with
(causally
coupled
to)
$I or SCI, provided that all
states in [S] are true.
This link
for the
coupling.
container
synonymous
is similar to ungated state coupling,
except
existence
of factors which could disrupt the
To illustrate, the flow of a fluid
into a
(a statechange in location of the water) is
with an increase in the amount of water
in
183
[
! ~4--'~
~r
the container (another statechange), but only providing
that there is no souL~ce of exit
from
the
container's
bottom.
TYPE
9,10,11,12:
GATED/NON-GATED)
BYPRODUCT
(0NE-SHOT/CONTINUOUS,
State S or s t a t e c h a n g e SC is a causal byproduct of
action A, relative to goal state Sg or SCg.
That is, the actor of A, w i s h i n g to achieve
state
Sg
or
statechange
SCg
also
produces
state
S or
s t a t e c h a n g e SC.
The byproduct link 'is truly a
causal
link;
what
is
and is not a by product must obviously
relate to the motive of the
actor
in
performing
the
action.
Where
gated,
all
states
in
[S]
must
be
s a t i s f i e d in order for the byproduct to occur.
TYPE
13: ORIGINAL
INTENT
Want W is the o r i g i n a l desire (goal state)
of
an
actor.
W is e x t e r n a l to the CSA in that its origin is
not e x p l i c a b l e w i t h i n the CSA itself; it is the outside
directive which m o t i v a t e d the i n v o c a t i o n of some acton.
Within
an
algorithm
for
achieving
some
goal,
motivations
are explicable: every s u b a c t i o n is, by its
nature, d e s i g n e d to produce subgoal states which, taken
together, meet the original intent.
TYPE
14: ACTION C O N C U R R E N C Y
Actions A I , . . . , A n must be c o n c u r r e n t l y executed.
This link will arise in the dynamics of an
actual
plan,
rather
than
be
stored
originally
in
the
a l g o r i t h m i c base explicitly.
As plans evolve
and
the
actor
learns
c o n c u r r e n c y by rote, the link will begin
to appear in the
algorithmic
base
as
well.
Action
concurrency
is
nearly
always
caused
by m u l t i p l e
e n a b l i n g states for some other
action,
all
of which
must
be
c o n t i n u a l l y present, or one-time s y n c h r o n i z e d
as a c o l l e c t i o n of one-shot enablements.
TYPE
15: DYNAMIC A N T A G O N I S M
State $I or s t a t e c h a n g e
SCI
is
antagonistic
to
state $2 or s t a t e c h a n g e SC2 along some dimension.
This link relates two states Or s t a t e c h a n g e s which
are
opposites
in some sense; t y p i c a l l y the a n t a g o n i s m
link will make explicit the final link in some sort
of
feedback
cycle
in an algorithm.
The link is hard to
describe outside the context of
an
example;
examples
will appear in the next section.
TYPE
16: M O T I V A T I N G DYNAMIC ~ N T A G O N I S M
As with o r d i n a r y dynamic antagonism,
$I,
$2
are
antagonistic
states.
Typically, $2 is required as an
e n a b l i n g state (continuous) for some action,
but
that
action,
or
some
other
action,
produces
$I
as
a
byproduct; this gives rise
to
the
need
for
another
corrective
action
A which can s u p p r e s s the byproduct,
therby p r e s e r v i n g
the
original
required
enablement.
This link is intended to capture the e x e c u t i o n dynamics
of a
situation
in which
antagonistic
states
are
expected
to
arise.
That
is,
it will
provide
a
r e p r e s e n t a t i o n wherein a n t a g o n i s m s can
be a n t i c i p a t e d
in advance
of the SCA's actual execution.
An example
of m o t i v a t i n g dynamic a n t a g o n i s m
is
included
in
the
next section.
184
II
i
I
I
i
I
!
I
i
I
I
I
i
I
I
1
I
I
I
TYPE
17: G O A L - R E A L I Z A T I O N
COUPLING
State
S
is
an
alternative
way
of
expressing
original goal W or subgoal S g .
This link s u p p l i e s a way of s p e c i f y i n g t e r m i n a t i o n
criteria
for
CSA's i n v o l v i n g repretition.
Its use is
i l l u s t r a t e d in one of the examples~
TYPE
18: C O M P O U N D
GOAL
STATE
DEFINITION
State S is a shorthand for e x p r e s i n g
the
set
of
states SI,...,Sn.
This link allows a "situation" to be c h a r a c t e r i z e d
as
a
c o l l e c t i o n of goal states.
When all goal states
are satisfied, the situation is satisfied.
An
example
of
a compound goal state would be: "get the kids ready
for the car trip", where this means
a
set
of
things
rather than one thing.
goal
TYPES
19,20,21,22:
GATED/NON-GATED)
DISENABLEMENT
(ONE-SHOT/CONTINUOUS,
Action A or t e n d e n c y T o n e - s h o t / c o n t i n u a l l y causes
S or s t a t e c h a n g e SC not to exist.
These four forms are s h o r t h a n d s for
causality
in
conjunction
with antagonism.
They will be p r i n c i p a l l y
useful for r e p r e s e n t i n g acts
of
disenabling
unwanted
tendencies.
state
TYPE
23:
REPETITION
UNTIL
THRESHOLD
Action A or t e n d e n c y
T occurs
repeatedly
until
S becomes true.
This link provides for the r e p e a t e d a p p l i c a t i o n of
an
action
or
tendency.
Normally,
the
action
or
t e n d e n c y will, d i r e c t l y or indirectly, c a u s a l l y produce
a
statechange
along some scale; this s t a t e c h a n g e will
e v e n t u a l l y t h r e s h o l d at state S.
state
TYPE
24:
INDUCEMENT
State S's or s t a t e c h a n g e
SC's
existence
induces
W in a potential actor.
Origins of wants can be e x p l i c i t l y r e p r e s e n t e d via
this
link.
Typically,
W will
be
a stabe which is
antagonistic
to
S or
SC.
For e x a m p l e ,
if
the
temperature
is
too high in the room, the want is that
the
temperature
become
lower;
if
the
tendency,
PRESSURE,
has been enabled, a l l o w i n g blood to flow out
of P's body, the induced want is that this t e n d e n c y
be
disenabled,
and
hence
that
the a n t a g o n i s m of one of
P R E S S U R E ' s e n a b l i n g states start to exist.
want
TYPE
25:
OPTIMIZATION
MARKER
State S is an enabling c o n d i t i o n for action
this r e l a t i o n s h i p makes possible an o p t i m i z a t i o n
the e x e c u t i o n of A in a p a r t i c u l a r environment.
A, and
during
When several actions arise in a plan,
they
may
share
enabling
states.
This
means that when the plans are
c o n s i d e r e d together, some of the states needed for
one
action may coincide with those needed for another.
The
optimization
marker
allows
this
phenomenon
to
be
recorded.
Its i n t e r p r e t a t i o n is: when state S becomes
true, c o n s i d e r p e r f o r m i n g acton
A~
because
action
A
also has S as an enabling state.
~ denotes a savings.
185
D
decrease
in tank water
height,
and this
decrease thresholds at point
X,
synonomous
with
the
fresh water supply valve opening.
This opening enables the tendency
pressure
to move water from the fresh water line into
the
tank;
this
is synonomous
with
an
increase
in tank water
height,
but only
providing that the
flush
valve
is closed
(this will have to wait for the movement of
waste from tank to bowl to cease).
When the
tank
water
height
finally
begins
its
increase, this increase
will
threshold
at
point
X again, this time being synchronous
with the
ball
cock
supply
valve's
being
closed,
stopping
the fresh water and hence
the tank water
height
increase.
At
this
point,
the
system
has
become
quiescent
again.
(Note:
in the actual
simulation
which will be performed, flow rates, or more
generally, rates of statechanges,
will
be
incorporated.)
These
are
the
commonsense
algorithm
primitive
links.
It is felt that they are
conceptually
independent
enough
of
one-another
so that unique algorithms will
be forced into unique, or at least
similar,
representations
under
this
formalism.
Although it is the eventual
intent
of the
theory to be able to capture all the nuances
of
intentional
human
problem-solving
behavior,
there
is no real feeling yet for
the completeness of this
set
of links
in
this
regard;
all
that
can be said now is
that they do seem to suggest
a reasonable
approach
to representing
large classes of
purposive human behavior.
The adequacy
of
these
primitives
for representing devices
and mechanisms, on the other hand, is easier
to see, at least intuitively; the links seem
to
be adequate
for
some
fairly
complex
"purposive"
mechanisms.
Accordingly,
the
first
example
of their
use will
be to
characterize
a mechanism very dear to most
of us.
EXAMPLE
2.
Sawing a
hoard
decrease its length (Figure 5)
V.
in
half
to
EXAMPLES OF COMMONSENSE ALGORITHMS
EXAMPLE
I.
Operation
toilet [Figure 3]
of
g
Figure 5 is a bare-bones representation
Of a purposive human process: sawing a board
in two
using
a
handsaw.
This
CSA
illustrates
the
concepts
of motivating
dynamic
antagonsim,
original
intent
and
byproduct
with
respect
to a goal.
The
schematic of Figure 5 is only a fragment
of
the
larger
algorithm; many enabling states
and
byproducts,
as
well
as
their
compensatory
actions have been omitted.
In
this CSA, the act of sawing for the
purpose
of decreasing
the board's length produces,
among others, the byproduct of the
board's
moving.
Since a stationary board is a gate
condition on the flow of causality from
the
sawing
action
to the statechange
in cut
depth,
the
two
states
joined
by
the
motivating
dynamic
antagonsim link form an
antagonistic pair, indicating in advance
of
actual
execution
that it will be necessary
to perform a compensatory action:
hold
the
wood down.
If we were to illustrate more of
this
algorithm,
it might
be
found
that
holding
the wood
down would require more
hands
than were
available.
This
would
provide
another
dynamic
antagonsim
which
would motivate
the
engagement
of another
compensatory
action,
such
as
"call
for
help," "go to a vise," etc.
reverse-trap
As a first
test
of the
theory,
the
reverse-trap
toilet
is
a
relatively
demanding
mechanism.
It is
a
complex
feedback
mechanism
which is the product of
some
rather
sophisticated
human
problem-solving.
It
is
therefore
interesting both in its own right and
as a
tangible
manifestation
of human-concocted
causality and
enablement.
By one
simple
action,
a complex sequence of tendencies is
unleasehed;
the
sequence
not
only
stops
itself,
but
restores
the
system
to its
initial state, and does something useful
in
the process.
The
English
description
of
the
schematic
of Figure
4 is as follows: The
trip handle is pushed down, one-shot causing
the
flush-ball
to be raised; this one-shot
enables
the
tendency
to float,
in turn
continually causing the float ball to remain
raised.
The float ball's
being
raised
is
synonomous
with the flush valve being open,
and this openness continuously
e n a b l e s the
tendency
of gravity to move water from the
tank to the bowl beneath (as long as water
remains
in the
tank,
of course.)
This
movement of water
is synonomous
with
two
other
state
changes:
a decrease of water
height in the tank, and an increase of water
height
in the bowl.
The increase of bowl
water
height
thresholds
when
the
water
reaches
waste
channel
lip level, at which
time
it
begins
providing
continuous
enablement
for gravity
to move the water
into
the waste
channel;
this
movement
thresholds when the channel fills, providing
the beginning of continuous
enablement
of
the
tendency
capillary
action.
This
tendency, in turn,
sustains
the
flow
of
water
from
the
bowl to the waste channel,
continually
moving
waste
water
into
the
sewer.
This
action
ceases
when the bowl
becomes
empty.
Meanwhile,
the
tendency
gravity is continually moving water from the
tank to the bowl.
This is synonomous with a
It should again
be pointed
out
that
points
of antagonism could alternatively be
detected at the execution time
of the
CSA
and
compensatory
solutions
dynamically
fabricated.
This would
likely
occur
via
some
sort
of interrupt mechanism.
But the
antagonsim link allows
for
planning
ahead
(e.g.
when
two
arbitrary
algorithms are
selected
to accomplish
a
task,
their
coexistence
will
probably
not
always
be
without
antagonism -- this
allows
the
planning
mechanism
to anticipate and solve
such antagonisms before
execution).
Also,
after
a
successful
plan
involving
antagonisms has actually been executed, this
link
provides a means of recording once and
for all the compensating actions which
were
performed.
186
I
I
I
I
I
I
I
EXAMPLE
3.
Vicious cycles
<Figure 6)
~WOODI,
Consider tendencies such
as
fire
and
forgetfulness.
Both
roughly
follow
the
paradigm:
a tendency
has
state
S as
a
continuous enablement, and produces the same
state
as
continuous
causality.
Once
started,
such
a system is self-sustaining.
In the case
of
fire,
a one-shot
causing
action
causes
a statechange in t e m p e r a t u r e
which
thresholds
at
the
point
of
the
material's
combustion
temperature;
this
enables the tendency to burn, which in
turn
produces
as
a continual
byproduct
heat,
causing a vicious cycle.
In forgetting, the
tendency
to
forget
X is enabled
by not
r e f e r e n c i n g X for periods of time; but as
X
grows
more
forgotten,
it
becomes
less
referenceable.
Here,
dynamic
antagonism
lies at the root of the vicious cycle.
EXAMPLE 4.
Description
(synthesis)
computer a l g o r i t h m (Figure 7)
-
WOOD2~
_L_
T
oooJOINED
,
~/
In this
type
of
situation,
the
state
coupling
concept
is required at this level
to stop the r e p r e s e n t a t i o n of some
sort
of
inexplicable
"micro-causality"
when
it
transcends
the
model's
knowledge
of
the
world.
VI.
A L T E R N A T I V E ACTION SELECTION
In
looking
at
devices
and
simple
processes
such
as
sawing a board in half,
there have been few choices;
the
causality
and
enablement are in a sense already built
in
or
strongly
prescribed.
In a
real
planning
environment
on
the
other
hand,
there
will
ordinarily
be
numerous
alternative
actions
which
could
causally
produce some desired goal
state,
providing
all
gating
conditions
could
be met.
For
example, if the goal
of
a planner
is
to
produce
a
statechange
in
his location to
some s p e c i f i e d point, the
various
subplans
of walking, driving a car, hitching a ride,
bicycling, taking a plane, etc. all
suggest
themselves
as
potentially
relevant,
some
more
than
others.
The
one
the
planner
actually
selects will be a function of more
than
Just
the
relative
costs
of
each
alternative;
the s e l e c t i o n will also relate
to
the
inherent
applicability,
or
reasonableness
of
the
plan,
based on the
sDecif%cs
of
where
his
destination
is
relative
to his
current location, w e a t h e r
conditions,
etc.
Of
course,
all
the
relevant
factors
could
eventually
be
discovered by
simulating
each
alternative
plan
before
choosing,
watching
out
for
undesirable
or
suboptimal
events.
For
example,
in
simulating
the
walking,
h i t c h h i k i n g or b i c y c l i n g plans, the
planner
finds
himself
outside for p o t e n t i a l l y long
durations.
Hence, if
it
is
raining,
the
cost
is
judged
high.
If the d i s t a n c e is
less than a mile, or is indoors,
simulation
of
the a i r p l a n e plan leads to some a b s u r d l y
high
costs
and
perhaps
some
unsolvable
antagonisms.
Certainly,
a degree of such
forward s i m u l a t i o n must occur
in
planning;
however,
it
seems
that
the
process
of
selecting
among
alternative
actions
is,
intuitively,
more
unified
than
just
a
c o l l e c t i o n of forward simulations.
of
Suppose the
goal
is
to
compute
the
average
of
a
table
of
numbers,
TABLE(1),...,TABLE(n).
Figure 7 shows
both
how to conceive of the a l g o r i t h m and how the
a l g o r i t h m will a c t u a l l y run.
As a computer
algorithm,
this is not as fully explicit as
might
be
desired:
it
lacks
explicit
iteration and explicit t e r m i n a t i o n c r i t e r i o n
testing.
These will have to be worked
out
before
the
theory
adequately
handles
repetition.
i
Causal gating seems to play
a central
role
in
this
sort
of computer algorithm.
Intuitively,
this
is
the
case
because,
though
a computer i n s t r u c t i o n t y p i c a l l y has
no physical enabling conditions (it could be
issued
at any time), desired effects can be
achieved
only
by
tying
the
syntax
of
instruction
causality
to
the semantics of
logical causality.
For example, the flow of
causality
from
the
action "fetch location
SUM to ACI" to the
logical
semantic
state
"partial
sum in ACI" can take place only if
location SUM logically contains
the
actual
partial
sum
at
that
point!
Otherwise,
garbage is fetched.
The r e l a t i o n s h i p s of certain
types
of
causal
gating and state coupling (e.g.
the
valve closing because the float has risen in
the toilet tank) are not c o m p l e t e l y a p p a r e n t
yet.
Perhaps state coupling is a s h o r t h a n d
for
an
implicit
sequence
of
gated
c a u s a l i t i e s between
two
statechanges.
On
the
other
hand, state coupling between two
states, as opposed to statechanges, seems to
be a concept which is independent of gated
causality.
To illustrate; "a nail
through
two
pieces
of wood"
(state
I) has to be
regarded as s t a t e - c o u p l e d to "the pieces
of
wood
are joined" (state 2, a d e s c r i p t i o n of
the
same
situation,
but
at
a different
level):
For this reason,
the model
incorporates
the
notion
of a
denoted by the c o n s t r u c t i o n :
187
of
CSA's
selector,
VIII.
SEL is a place where heuristics, as well
as
forward
s i m u l a t i o n control can reside,
The
heuristics
test
relevant
dimensions
(e.g. distance, w e a t h e r conditions, etc.) of
the
context
in
which
the
state
or
statechange
is
being
sought
(either
for
execution
of
some
larger
plan,
or
for
interpreting
what
another might do in some
context).
Based on
the
outcomes
of
such
tests,
the SELector chooses one a l t e r n a t i v e
action as most reasonable.
Currently,
the
selector
function
is
imagined
to
exist
"outside"
the
CAS
formalism
as
'an
u n r e s t r i c t e d program which runs and decides.
Eventually, since it is one goal of the
CSA
formalism
to be able to represent a r b i t r a r y
decision processes (these
are,
after
all,
just
other
algorithms),
the
SELector
function should simply reference other CSA's
which
carry
out the heuristic testing.
In
other words,
defer
the
"intelligence"
in
selecting
an
alternative
at this level to
u n i n t e l l i g e n t CSA's at the next
level,
and
SO
A later
version
of
this
paper
will
contain
more examples of the CSA, i n c l u d i n g
its use in language context
problems.
The
theory
is
by
no
means
complete;
to
illustrate:
(I) Is there
such
a thing
as
gated
enablement?
The answer seems to be
"yes", since it seems r e a s o n a b l e to
regard
enablement
as a flow which
can be cut off in much the same way
as
causality.
Perhaps an example
of gated
enablement i s
when
the
horses
begin
their
race
at
the
racetrack: the start
gate's
being
open
is
a one-shot enablement for
the horse to run, but only
if
the
horse
is in the box to start with!
If he's not in the box, the
gate's
position
isn't
relevant
as
an
enablement
to
run;
its
flow
is
severed.
on.
VII.
LEVELS OF R E S O L U T I O N S
THE THEORY HAS ONLY JUST BEGUN
(2) What kinds of time
and
sequencing
i n f o r m a t i o n need to be i n c o r p o r a t e d
in
the
formalism?
For
example,
causality
can
be either abrupt or
gradual:
taking
medicine
for
an
ulcer
provides
a
conceptually
gradual
statechange
in
the
stomach's
condition,
whereas
surgery
provides
a
conceptually
abrupt cure! This suggests the need
for
classifying
statechanges
on
some
discrete
conceptual
scale.
Another i n a d e q u a c y of
the
present
model
is
its i n a b i l i t y to s p e c i f y
time sequencing; a d o p t i o n
of
some
t r a d i t i o n a l flowchart concepts will
p r o b a b l y prove a d e q u a t e for this.
IN CSA'S
The a l g o r i t h m i c content of a CSA can be
described
at
many
different
levels
of
resolution.
For example, the "action" "take
a
plane
to
San
Francisco" is quite a bit
higher in level and more abstract
than
the
action
"grasp
a saw".
In the former, the
act of
taking
a plane
somewhere
is
not
really
an
action
at
all,
but
rather
a
d e s c r i p t i o n of an
entire
set
of
actions,
themselves
related
in a CSA; "take a plane
to San Francisco" is a high level
surrogate
for
a low level c o l l e c t i o n of true actions
in the sense of a c t u a l l y p e r f o r m i n g physical
movements,
etc.
in the real world (things
like g r a s p i n g a saw,
reaching
into
pants
pocket for some money, and so on).
(3) There is no c o n v e n i e n t way to model
decision-making
processes
on
the
part of the planner of a CSA.
This
will have to be developed.
Another
example
of
resolution
level
differences
relates
to how enabling states
for actions are c h a r a c t e r i z e d .
For example,
in (A2) Abelson employs the primitive (OKFOR
object
application),
as
in
(OKFOR
AUTO
TRAVEL).
The q u e s t i o n here is, what is the
relationship
between
this
high
level
description
of
OKness and the specifics of
what OKFOR means for any given object?
That
is,
for
a car, OKFOR means "gas in tank",
"tires
inflated",
"battery
charged",...,
whereas
(OKFOR TOILET FLUSHING) means quite
a different set of things.
The basic
issue
is:
should the m e m o r y plan and interpret in
the
abstract
realm
of
OKFORedness,
then
instantiate
with details later, or must the
details serve as the primary planning basis,
with
the
abstract ideas being reserved for
other
higher
level
processes
such
as
reasoning
by analogy, g e n e r a l i z a t i o n and so
forth? There is
probably
no
cut-and-dried
answer;
however,
the
tendency
in a CSA
system would be to favor
the
details
over
the abstract.
But the CSA r e p r e s e n t a t i o n is
intended to be flexible enough to a c c o m o d a t e
both
the
abstract
and
the concrete.
The
idea of state coupling is an i l l u s t r a t i o n of
this.
IX.
A P P L I C A T I O N S OF THE CSA
On the b r i g h t e r side, the CSA
provides
a u n i f i e d basis for p r o b l e m - s o l v i n g - r e l a t e d
cognitive models.
Specifically,
I believe
it
shores
up,
under
one
basic
data
structure, the ideas
presented
in
my
own
past
research
in
conceptual
memory
and
i n f e r e n c e (RI,R2) and in c o n c e p t u a l o v e r l a y s
(R3)
which
suggests
a meaning
context
m e c h a n i s m for language
comprehension
based
around CSA's.
I want to conclude by listing
anticipated
applicaions
of
CSA's.
The
applications
have
been
divided
into
two
categories: general (those which are c e n t r a l
to some major t h e o r e t i c a l issues in language
understanding
and
problem-solving),
and
specific
(those
which
provide
some local
i n s i g h t s into m e m o r y o r g a n i z a t i o n ) .
General A p P l i c a t i o n s
I. As the basis
for active
Rroblem-solv~ng
The CSA s u p p l i e s an a l g o r i t h m i c
format
wherein
plans can be conceived, s y n t h e s i z e d
and
executed.
One
immediate
goal
of
188
I
I
I
I
I
I
I
!
I
i
I
i
I
research
should
be
to
construct
a
commonsense
algorithm1
interpreter
which
could
"execute" the contents of portions of
its
own
CSA memory
in order
to
effect
actions
of moving about, communicating, and
so forth.
2. As the basis for conceptual
sense, r e p r e s e n t i n g such m e c h a n i s m s in terms
of
causality,
enablement,
byproducts,
thresholds, etc.
is quite meaningful.
5. As a basis for m o d e l i n g dynamic
m e a n i n g context i__nnlanguage
c o m p r e h e n s i o n and general perception
inference
(R3)
describes
an
expectancy-based
system
called
"conceptual
overlays" which
can
impose
high-level,
contextual
interpretations
on
sentences by consulting
its a l g o r i t h m i c
base.
In that
paradigm,
some stimuli (i.e.
meaning graphs resulting
from
a conceptual
parser " which
receives
language
utterances
as
input)
activate
action overlays,
while
other
stimuli
fit
into
previously
a c t i v a t e d action overlays.
Since an overlay is a collection of pointers
to
CSA's in the a l g o r i t h m i c base which have
been predicted as likely to occur
next,
to
"fit
into"
is to identify subsequent input
as steps in the
various
algorithms
actors
have been p r e d i c t e d to engage.
For example,
knowing what the sentence "John
asked
Mary
for
the keys" means c o n t e x t u a l l y is quite a
bit more simply u n d e r s t a n d i n g the
"picture"
this
utterance
elicits
(its
conceptual
dependency representation).
If we know that
John was hungry:
In (RI), which describes
a theory
of
conceptual
memory
and
inference,
sixteen
classes
of
conceptual
inference
were
identified
as
the
logical foundation of'a
l a n g u a g e - b a s e d meaning c o m p r e h e n s i o n system.
Interestingly
enough
(but not surprising),
nine of those inference
classes
correspond
directly
to
traversals
of
CAS
primitive
links.
In
the
t h e o r y of
(RI),
every
language stimulus, represented in c o n c e p t u a l
form
via
Schank's
conceptual
dependency
notation
(S2),
was
subjected
to
a
spontaneous expansion in
"inference
space"
along
the
sixteen dimensions c o r r e s p o n d i n g
to the sixteen inference classes.
Making an
inference
in
that
model
corresponds
to
i d e n t i f y i n g each perception as a step in one
or more
CSA's, then e x p a n d i n g outward from
those
points
along
the
CSA
links
breadth-first.
A l t h o u g h there is c e r t a i n l y
a class
of more
goal-directed
conceptual
inference,
this
kind
of
spontaneous
expansion
seems
necessary
to
general
comprehension,
and
the
CSA
is a natural
formalism
to
use.
The
nine
classes
of
inference which relate directly to CSA links
are:
I.
2.
3.
4.
5.
6.
7.
8.
9.
John hadn't eaten in days.
John asked Mary for the car keys.
we a c t i v a t e an overlay
which
expects
that
John
will engage CSA's which will a l l e v i a t e
his inferred hunger; needing car
keys
fits
nicely as a c o n t i n u o u s enablement in several
of these algorithms.
Of course, the
virtue
of
such
a
system
is
that
it allows the
h i g h - l e v e l i n t e r p r e t a t i o n of a sentence
to
change
as
a
function
of
its
contextual
environment:
causative
resultative
motivational
enablement
function
enablement-prediction
missing enablement
intervention
action-prediction
John had some h a m b u r g e r stuck
in his teeth.
John asked Mary for the car keys.
3. As the basis for the c o n c e p t u a l
r e p r e s e n t a t i o n o_~f language.
Change
the
interpretation
A very large percentage of what
people
communicate
deals
with algorithms, the how
and why of their a c t i v i t i e s
in the
world.
Schank's
conceptual
dependency
framework
does
a good
job
at
representing
rather
complex
utterances
which
reference
u n d e r l y i n g actions, states and statechanges.
This
theory of CSA's extends this f r a m e w o r k
to a c c o m o d a t e larger
chunks
of
experience
and
language
to
begin
dealing
with
p a r a g r a p h s and stories instead
of
isolated
sentences.
expectancies,
changes!
and
the
6. As ~he basis for
c o m p u t e r a l ~ o r l t h m synthesis
~nd
Since
a computer
algorithm
is
a
relatively
direct
reflection
of
a
programmer's
internal
model
of
an
algorithmic
process,
it
seems
reasonable
that both the
processes
of
synthesis
and
final
implementation
be represented in the
same
terms
as
his
internal
model.
The
present theory only suggests an approach; it
is not yet
adequate
for
general
computer
algorithms.
But it seems that the idea of a
CSA
might
be
very
relevant
to
recent
research
in
the
area
of
automatic
programming,
at
least
as
a
basis
of
representation.
4. As the basis for m o d e l i n ~ m e c h a n i s m s
Every m a n - m a d e mechanism,
as well
as
every n a t u r a l l y - o c c u r r i n g biological system,
is
rich
in
algorithmic
content.
As
i l l u s t r a t e d in a previous example, CSA's can
do a r e s p e c t a b l e
job at
characterizing
complex
servo- and feedback mechanisms.
It
is not hard to envision the CSA as
a basis
for
physiological
models
in
such
an
a p p l i c a t i o n as medical diagnosis.
Since all
biological
systems
are
purposively
c o n s t r u c t e d m e c h a n i s m s in
the
evolutionary
7. As ~ basis o__ffg self-model
If a CSA
interpreter
can
indeed
be
defined,
and
if
indeed
the
CSA
can
e v e n t u a l l y capture any
computer
algorithm,
then
creating
a
self-model
amounts
to
189
interfere
with
the
picnic
plans
that
afternoon.
How
do
such
situations
get
judged
"anomalous",
and
how
does
the
perceiver
try to explain or cope with them?
The
answer
undoubtedly
relates
to
expectancies
and
a knowledge of a l g o r i t h m s
for putting things on
one-another,
getting
s o m e w h e r e in a hurry and a n t a g o n i s t i c states
when eating outdoors.
By playing e x p e r i e n c e
against CSA's we d i s c o v e r things which would
not o t h e r w i s e be discovered.
s p e c i f y i n g the CSA i n t e r p r e t e r in terms
of
CSA's.
For example, an act of c o m m u n i c a t i o n
amounts
to
the
communication
of
enough
referential
information
(features
of
objects,
times,
etc.)
to
enable
the
comprehender
to identify, in his own model,
the
concepts
being
communicated.
The
how-to-communicate
algorithm
which the CSA
i n t e r p r e t e r employs could itself be a CSA.
8. As a basis for i n v e s t i g a t i o n
of a l g o r i t h m learning
4. For filling in missing i n f o r m a t i o n
If we posit the existence
of
a small
set
of
primitive
CSA
links
and make the
a s s u m p t i o n that these are either part of the
brain's
hardware,
or learned i m p l i c i t l y as
soon as the intellect begins perceiving,
we
have a basis from which to study how a child
learns world dynamics.
For
example,
how,
and
at
what
point,
does the toddler know
that he must grasp the
cup
in an
act
of
continuous
enablement before he can lift it
to his mouth, and how does he know
it must
be at
his mouth before he can s u c c e s s f u l l y
drink?
Perhaps
algorithmic
knowledge
develops
from random e x p e r i m e n t a t i o n within
the syntactic c o n s t r a i n t s imposed by the set
of CSA primitive links.
If a person is p e r c e i v i n g in a noisy or
incomplete
environment,
having
CSA's
a v a i l a b l e to guide
his
interpretations
of
p e r c e p t i o n s provides enough m o m e n t u m to fill
in m i s s i n g details, scarcely n o t i c i n g
their
absence.
If
John is h a m m e r i n g a nail into
the wall with his hand on the backswing, but
the
object
in
his
hand
is
occluded, it
requires very little effort to surmise
that
it
is a hammer.
If we believe that Mary is
going to M c D o n a l d ' s to buy a hamburger,
but
she
comes
back
into
the house saying "It
won't start", we have
a pretty
good
idea
"it" refers to the car.
This a p p l i c a t i o n of
CSA's
corresponds
to
the
notion
of
a
s p e c i f i c a t i o n i n f e r e n c e in (RI).
X.
Specific
CONCLUSIONS
CSA A p p l i c a t i o n s
I. For r e p r e s e n t i n g
Instead of a conclusion, I will
simply
state
the order in which research along CSA
lines should,
and
hopefully
w i l l at
the
U n i v e r s i t y of Maryland, progress:
the functions of objects
As with mechanisms, any m a n - m a d e object
is made for a purpose.
T r a n s l a t e d to CSA's,
this
means
that
part
of
every
purposively-constructed
object's d e f i n i t i o n
is a set of pointers into the a l g o r i t h m base
to
CSA's
in which the object occurs.
This
is true for all
objects
from
pencils,
to
furnaces,
to w i n d o w
shades,
to
a bauble
which p r o v i d e d its c o n s t r u c t o r amusement, to
newspapers.
An
object
in memory
can be
c o m p l e t e l y c h a r a c t e r i z e d (in the
abstract)
by a set of intrinsic features (shape, size,
color, etc) and
this
set
of
pointers
to
CSA's.
2. For r e p r e s e n t i n g
people's
I. R e i m p l e m e n t a t i o n
of
the
conceptual
overlays
p r o t o t y p e system d e s c r i b e d in (R3)
to reflect the new CSA ideas and replace the
ad-hoc
AND/OR
graph
a p p r o a c h d e s c r i b e d in
that report.
2. I m p l e m e n t a t i o n of a m e c h a n i s m
simulator
which
could
accept,
in
CSA
terms,
the
definition
of
a
complex
mechanism
(electronic circuit or toilet), s i m u l a t e it,
respond
to
artificially-induced
m a l f u n c t i o n s , and answer q u e s t i o n s about the
m e c h a n i s m ' s cause and effect structure.
professions
3. E n g i n e e r i n g of
a new
total
conceptual
memory,
along the lines of the o r i g i n a l one
of (RI), but i n c o r p o r a t n g CSA's and the
new
idea
of a
tendency.
This
would involve
r e i m p l e m e n t i n g the i n f e r e n c e
mechanism
and
various searchers.
To say (ISA JOHNI PLUMBER) skirts
what
it
means
to be a plumber.
Rather, to be a
plumber means to engage plumbing
algorithms
as
a
principal
source of income.
Thus, a
profession
can
be
defined
by a set
of
pointers
to
the
CSA's
which
are
characteristic
of
that
profession.
This
makes it possible to observe someone at work
and
identify
his
profession,
to compare
professions,
etc.;
these
would
not
be
p o s s i b l e if CSA's
were
not
the
basis
of
representation.
4. D e v e l o p m e n t of a
CSA
interpreter
which
could
not only use CSA's as data s t r u c t u r e s
in the various c o g n i t i v e processes, but also
could execute them to drive itself.
5. A p p l y i n g CSA's to m e d i c a l
a u t o m a t i c programming.
3. For d e t e c t i n g and explaining
a n o m a l o u s s i t u a t i o n s and p o t e n t i a l l y
a n t a g o n i s t i c states
diagnosis
and
6. I n v e s t i g a t i n g
the
problem
of
story
comprehension
via
conceptual
o v e r l a y s and
CSA's.
Perhaps
also
investigating
generation
of
stories
(e.g.
the story of
the trip to McDonald's) or the g e n e r a t i o n of
a
description
of
a complex
electronic
circuit,
encoded
as
a CSA,
in
layman's
terms.
A person notices a license plate yearly
sticker on upside down; a person notices two
fire engines
approaching
an
intersection,
rushing
to a fire; at the intersection, one
turns left, the other turns right; a person
notices
that
the
rain
that
m o r n i n g will
190
XI.
ACKNOWLEDGEMENTS
My thanks to the members
of
the
Commonsense
Algorithm Study G r o u p at the
University of Maryland: Bob Eberlein,
Milt
Grinberg,
Bob Kirby,
Phil London and Tom
Skillman.
They have provided considerable
intellectual
stimulation.
We
hope to
continue as a group and eventually
issue a
working paper and computer system.
REFERENCES
(AI) Abelson, R., "The Structure of Belief
Systems,"
in Schank and Colby (eds.),
Computer
Models
of
Thought
and
Language, W.H.
Freeman, 1973
(A2) Abelson, R., "Frames for Understanding
Social Actions,"
Paper
for Carbonell
Conference, Pajarro Dunes CA, May 1974.
(CI) Charniak,
E.,
"Toward a
Model
of
Children's
Story
Comprehension,"
Doctoral
dissertation,
M.I.T.,
AI
TR-266, 1972.
(MI) Minsky,
M.,
"A
Framework
for
Representing
Knowledge,"
M.I.T.
AI
TR-306, 1974.
(NI) Nilsson, N., Probelm Solving Methods i__nn
Artificial
Intelligence,
McGraw Hill,
1971.
(RI) Rieger,
C.,
"Conceptual
Memory:
A
Theory
and
Computer
Program
for
Processing
the Meaning
Content
of
Natural Language Utterances," Doctoral
dissertation, Stanford Univ.
AI Memo
233, 1974.
(R2) Rieger,
C.,
"Understanding
by
Conceptual Inference," American Journal
of
Computational
Linguistics
(in
press), 1975.
(Also available as Univ.
of Maryland Technical Report #353)
(R3) Rieger,
C.,
"Conceptual
Overlays:
A
Mechanism for the Interpretation
of
Sentence Meaning in Context," to appear
in
Proceedings
4IJCAI
1975.
(Also
available
as
Univ.
of
Maryland
Technical Report #354)
($I) Sacerdoti, E., "Planning in a Hierarchy
of Abstraction Spaces," in Proceedings
3IJCAI 1973.
($2) Schank,
R.,
"Identifications
of
Conceptualizations
Underlying
Natural
Language",
in
Schank
and
Colby,
Computer
Models
of
Thought
and
Language, W.H.
Freeman, 1973.
($3) Schank, R., Goldman,
N., Rieger,
C.,
and Riesbeck,
C., "Primitive Concepts
Underlying Verbs of Thought,"
Stanford
Univ.
AI Memo 162, 1972.
($4) Schmidt,
C., and D'Addamio,
J.,
"A
Model of the Common-Sense
Theory of
Intention and
Personal
Causation,"
Proceedings 3IJCAI, 1973.
Figure !
Unrestricted AND/OR graph for getting a
McDonald's hamburger into stomach.
e.,~\e.
Figure 2
Hamburger algorithm, with actions, states,
causality and enablement explicit.
191
I
,
tank
from
~j~
/ ~ J
_ ~ ~ _ ~ ,
i
m,
,
I
itrip~
f'~,refi I l
~ / /
_ # float
- trip
%]
oat~i/i
~anole~
lf~
arm
supply II1 disvalve Ill harge
channel
III~ pe
plpe~_
I
"~ Over.fl ow
p,pe
c
• ~flush ball
ValeV~
supply|~ flush
stewOer
ll!ft
' ~'-')
TANK
FIGURE3
A reverse-trap toilet.
K
Fuu~~Ft ~ b
~Bk~6c~c~ sutp6y
6~6~r o~ ?uo#T~
V~#~.
~.
w ~,~J~
wWr~ F~
IvPPI.Y
LtN~- "r~ TA~V~
L~t. ~A~If,Q.:
v~Lv~
6J~TF.~ H¢16.~
o~rr~ L~P ~F wksl~
15owu
G,'Z~W,T7
Bo~L. 1"o w~rs~n~
FIGURE4
O~erationof the reverse-trap toilet.
192
I
I
I
OF I,~o0
I (oe,,.,,~.
(Co~t~UouS
Jr~
I
I
I
I
I
I
I
I
I
,
WooO
_ _
'Figure 5
_
_
.
+
Sawing a board in half to decrease its length.
193
Ik / ~T ~ , ~ •, ~ ' o~ . ~ - I - -~- - - "
II T;~~,.
,T ~ ~ .
'
|
I
I
I
COMBUSTION
,
I
~~'"
~
I
|
!
!
FORGETFULNESS
FIGURE 6
I
!
!
Vicious Cycles.
I
194
!
DJVII>(
,4c~. J ~ c Z.
ts ~j /tc
2,£ I ~
~F-.O
Nf.x.T
P~
t~E~T TA~LF_
l
llll
FIGURE 7
Computer algorithm to compute the
average of TABLE(1),...,TABLE(N)
expressed as a commonsense algorithm.
(NOTE: Initialization has not been shown. The assumptions are
that AC3 begins with zero, that ACI begins with zero,
and that N and TABLE(1),...,TABLE(N) exist in core.)