8PROCEEDINGS OF THE IRE
Steps Toward Artificial Intelligence*
MARVIN MINSKYt, MEMBER, IRE
The work toward attaining "artificial intelligence" is the center of considerable computer research, design,
and application. The field is in its starting transient, characterized by many varied and independent efforts.
Marvin Minsky has been requested to draw this work together into a coherent summary, supplement it with
appropriate explanatory or theoretical noncomputer information, and introduce his assessment of the state-ofthe-art.
This paper emphasizes the class of activities in which a general purpose computer, complete with a
library of basic programs, is further programmed to perform operations leading to ever higher-level information
processing functions such as learning and problem solving. This informative article will be of real interest to
both the general PROCEEDINGS reader and the computer specialist. The Guest Editor
Summary-The problems of heuristic programming-of making
computers solve really difficult problems-are divided into five main
areas: Search, Pattern-Recognition, Learning, Planning, and
Induction.
A computer can do, in a sense, only what it is told to do. But even
when we do not know how to solve a certain problem, we may program
a machine (computer) to Search through some large space of
solution attempts. Unfortunately, this usually leads to an enormously
inefficient process. With Pattern-Recognition techniques, efficiency
can often be improved, by restricting the application of the machine's
methods to appropriate problems. Pattern-Recognition, together
with Learning, can be used to exploit generalizations based on accumulated
experience, further reducing search. By analyzing the
situation, using Planning methods, we may obtain a fundamental
improvement by replacing the given search with a much smaller,
more appropriate exploration. To manage broad classes of problems,
machines will need to construct models of their environments, using
some scheme for Induction.
Wherever appropriate, the discussion is supported by extensive
citation of the literature and by descriptions of a few of the most
successful heuristic (problem-solving) programs constructed to date.
INTRODUCTION
A VISITOR to our planet might be puzzled about
the role of computers in our technology. On the
one hand, he would read and hear all about
wonderful "nmechaniical brains" baffling their creators
with prodigious initellectual performance. And he (or it)
would be warned that these machines must be restrained,
lest they overwhelm us by might, persuasion,
or even by the revelation of truths too terrible to be
borne. On the other hand, our visitor would find the
machines being denounced, on all sides, for their slavish
obedience, unimaginative literal interpretations, and
incapacity for innovation or initiative; in short, for
their inhuman dullness.
Our visitor might remain puzzled if he set out to find,
and judge for himself, these monsters. For he would
*
Received by the IRE, October 24, 1960. The author's work
summarized here which was done at Lincoln Lab., a center for research
operated by M.I.T. at Lexington, Mass., with the joint support
of the U. S. Army, Navy, and Air Force under Air Force Contract
AF 19(604)-5200; and at the Res. Lab. of Electronics, M.I.T.,
Cambridge, Mass., which is supported in part by the U. S. Army
Signal Corps, the Air Force Office of Scientific Res., and the ONRis
based on earlier work done by the author as a Junior Fellow of the
Society of Fellows, Harvard Univ., Cambridge.
t Dept. of Mathematics and Computation Center, Res. Lab.
of Electronics, M.I.T., Cambridge, Mass.
find only a few machines (mostly "general-purpose"
computers, programmed for the moment to behave according
to some specification) doing things that nmight
claim any real initellectual status. Some would be proving
mathematical theorems of rather undistinguished
character. A few machines might be playing certain
games, occasionally defeating their designers. Sonme
might be distinguishing between hand-printed letters.
Is this enough to justify so much interest, let alone deep
concern? I believe that it is; that we are on the threshold
of an era that will be strongly influenced, and quite
possibly dominated, by intelligent problem-solving mlachines.
But our purpose is not to guess about what the
future may bring; it is only to try to describe and explain
what seem nlow to be our first steps toward the
construction of "artificial intelligence."
Along with the development of general-purpose computers,
the past few years have seen an increase in effort
toward the discovery and mechanizationi of problermsolving
processes. Quite a number of papers have appeared
describing theories or actual comiiputer programs
concerned with game-playing, theorem-provinig, pattern-recognition,
and other domainis which would seem
to require some intelligence. The literature does not include
any general discussion of the outstanding problems
of this field.
In this article, an attempt will be made t-o separate
out, analyze, and find the relations between some of
these problems. Analysis will be supported with enough
examples from the literature to serve the introductory
function of a review article, but there remains much
relevant work not described here. This paper is highly
compressed, and therefore, cannot begin to discuss all
these matters in the available space.
There is, of course, no generally accepted theory of
"intelligence"; the analysis is our own and mnay be controversial.
We regret that we cannot give full personal
acknowledgments here suffice it to say that we have
discussed these matters with almnost every one of the
cited authors.
It is convenient to divide the problems into five main
areas: Search, Pattern-Recognition, Learning, Planning,
and Iniduction; these comprise the main divisions
January8
Minsky: Steps Toward Artificial Intelligence
of the paper. Let us summarize the entire argument
very briefly:
A computer can do, in a sense, only what it is told to
do. But even when we do not know exactly how to solve
a certain problem, we may program a machine to
Search through some large space of solution attempts.
Unfortunately, wheni we write a straightforward program
for such a search, we usually find the resulting
process to be enormously inefficient. With PatternRecognition
techniques, efficiency can be greatly improved
by restricting the machine to use its methods
only on the kind of attempts for which they are appropriate.
And with Learning, efficiency is further improved
by directing Search in accord with earlier experiences.
By actually analyzing the situation, usinig what we call
Planning methods, the machinie may obtain a really
fundanmental improvemeint by replacing the originally
given Search by a much smiialler, more appropriate exploration.
Finally, in the section on Induction, we consider
some rather more global concepts of how one
might obtain intelligenit machine behavior.
I. THE PROBLEM OF SEARCH1
Summary-If, for a given problem, we have a means for checking
a proposed solution, then we can solve the problem by testing all
possible answers. But this always takes much too long to be of
practical interest. Any device that can reduce this search may be of
value. If we can detect relative improvement, then "hill-climbing"
(Section I-B) may be feasible, but its use requires some structural
knowledge of the search space. And unless this structure meets certain
conditions, hill-climbing may do more harm than good.
\Vheni we talk of problem-solvinig in what follows we
will usually suppose that all the problems to be solved
are iniitially well defined [1]. By this we meain that with
each problem we are given somie systematic way to
decide when a proposed solution is acceptable. Most of
the experimental work discussed here is concerined with
such well-definied problems as are met in theorem-proving,
or in games with precise rules for play and scoring.
In one sense all such problems are trivial. For if there
exists a solutioni to such a problem, that solution can be
found eventually by any blind exhaustive process which
searches through all possibilities. And it is usually not
difficult to mechanize or program such a search.
But for any problem worthy of the name, the search
through all possibilities will be too inefficient for practical
use. And on the other hand, systems like chess,
or nontrivial parts of mathematics, are too complicated
for complete analysis. Without complete analysis, there
must always remaini some core of search, or "trial and
error." So we need to finid techniques through which the
1 The adjective "heuristic," as used here and widely in the literature,
means related to improving problem-solving performance; as a
noun it is also used in regard to any method or trick used to improve
the efficiency of a problem-solving system. A "heuristic program,"
to be considered successful, must work well on a variety of problems,
and may often be excused if it fails on some. We often find it worthwhile
to introduce a heuristic method which happens to cause occasional
failures, if there is an over-all improvement in performance.
But imperfect methods are nlot niecessarily heuristic, nor vice versa.
Hence "heuristic" shouild not be regarded as opposite to "foolproof";
this has caused some coniftusioni in the literature.
results of incomplete analysis can be used to mnake the
search more efficient. The necessity for this is simply
overwhelming: a search of all the paths through the
game of checkers involves some 1040 move choices [21;
in chess, some 10120 [3]. If we organized all the particles
in our galaxy into some kind of parallel computer
operating at the frequency of hard cosmic ravs, the latter
computation would still take impossibly long; we
cannot expect improvements in "hardware" aloile to
solve all our problems! Certainly we must use whatever
we know in advance to guide the trial generator. And
we must also be able to make use of results obtained
alonig the way.23
A. Relative Improvement, Hill-Climbing, and Heuristic
Connections
A problem can hardly come to interest us if we have
no backgroutnd of information about it. We usually
have some basis, however flimsy, for detecting improvement;
some trials will be judged more successful than
others. Suppose, for example, that we have a comparator
which selects as the better, one from any pair of trial
outcomes. Now the comparator cannot, alonie, serve to
make a problem well-defined. No goal is definied. But if
the comparator-defined relation betweeni trials is
"transitive" (i.e., if A dominates B and B dominates C
implies that A dominates C), then we can at least define
"progress," and ask our machine, given a time limit, to
do the best it can.
But it is essential to observe that a comparator by
itself, however shrewd, cannot alone give any imnprovement
over exhaustive search. The comparator gives us
information about partial success, to be sure. But we
need also some way of using this information to direct
the pattern of search in promising directionis; to select
new trial points which are in some senise "like," or
"similar to," or "in the same direction as" those which
have given the best previous results. To do this we need
some additional structure on the search space. This
structure need not bear much resemblance to the ordinary
spatial notion of direction, or that of distanice, but
it must somnehow tie together poinlts which are heuristically
related.
\Ve will call such a structure a heuristic connection.
We introduce this term for informal use onl--that is
why our definition is itself so informal. But we need it.
Many publications have been marred by the misuse,
2 McCarthy [11 has discussed the enumeration problem from a
recursive-function theory point of view. This incomplete but suggestive
paper proposes, among other things, that "the eniumeratioln
of partial recursive functions should give an early place to compositioIns
of functions that have already appeared."
I regard this as an important notion, especially in the light of
Shannon's results [4] on two-terminal switching circuits that the
"average" n-variable switching function requires about 2"/n contacts.
This disaster does not usually strike when we construct "interesting"
large machines, presumably because they are based on composition
of functions already found useful.
3 In 151 and especially in [61 Ashby has ani excellent discussioni of
the search problem. (However, I am not convinced of the usefulniess
of his notion of "ultrastability," which seems to be little more than
the property of a machine to search unitil somethin-g stops it.)
1961 9
PROCEEDINGS OF THE IRE
for this purpose, of precise mathematical terms, e.g.,
metric anid topological. The term "connection," with its
variety of dictionary meanings, seems just the word to
designate a relation without comnmitmiienit as to the
exact nature of the relation.
An important and simple kind of heuristic connection
is that defined when a space has coordinates (or parameters)
and there is also defined a nunmerical "successfunction"
E which is a reasonably smooth functioni of
the coordinates. Here we can use local optimization or
hill-climbing methods.
B. Hill- Climbing
Suppose that we are given a black-box machine with
inputs Xi, , XI, and an output E(X1, , XI/). W;e
wish to maximize E by adjusting the input values. But
we are not given any mathemnatical description of the
function E; hence we cannot use differentiation or related
methods. The obvious approach is to explore
locally about a point, finding the direction of steepest
ascent. One moves a certain distance in that direction
anid repeats the process until improvement ceases. If the
hill is smooth this may be done, approximately, by
estimating the gradient component dE/3Xi separately
for each coordinate Xi. There are more sophisticated
approaches (one may use noise added to each variable,
and correlate the output with each input, see Fig. 1),
but this is the general idea. It is a fundamental technique,
and we see it always in the backgrounid of far
mnore complex systems. Heuristically, its great virtue
is this: the sampling effort (for determnining the direction
of the gradient) grows, in a sense, only linearly
with the number of parameters. So if we can solve, by
such a method, a certain kind of problem inivolving
many parameters, then the addition of more paramiieters
of the same kind ought not cause an inordinate increase
in difficulty. We are particularly interested in
problem-solving methods which can be so extended to
more difficult problems. Alas, most interesting systems
which involve combinational operations usually grow
exponentially more difficult as we add variables.
A great variety of hill-climbing systems have been
studied under the names of "adaptive" or "self-optimizinig"
servomechanisms.
C. Troubles with Hill-Climbing
Obviously, the gradient-following hill-climber would
be trapped if it should reach a local peak which is not a
true or satisfactory optimum. It must then be forced to
try larger steps or changes.
It is often supposed that this false-peak problem is
the chief obstacle to machine learning by this method.
This certainly can be troublesome. But for really difficult
problems, it seems to us that usually the more
fundamental problem lies in finding any significant
peak at all. Unfortunately the known E functions for
difficult problemns often exhibit what we have called
[7] the "Mesa Phenomenon" in which a small change in
FROM OTHER U's
-8 U. - TO OTHER U's
Fig. 1 "Multiple simultaineous optimizers" search for a (local) mnaximum
value of some funiiction E(XI, - - XJ) of several parameters.
Each unit Ui independently "jitters" its parameter Xi,
perhaps randomly, by adding a variation 1i(t) to a current mean
value mi. The changes in the quiantities bi and E are correlated,
anld the result is used to (slowly) change yi. The filters are to remove
dc components. This simultaneous techniique, really a formii
of coherenit detection, usually has an advanltage ov-er methods
dealing separately and sequentially with each paramneter. (Cf. the
discussion of "informative feedback" in Wiener [11], p. 133 ff.)
a parameter usually leads to either no change in performance
or to a large change in performuaince. The
space is thus composed primarily of flat regions or
"mesas." Any tendency of the trial generator to make
small steps then results in much aimless wandering
without compensating information gainis. A profitable
search in such a space requires steps so large that hillclimbing
is essenitially ruled out. The problem-n-solver
must find other methods; hill-climbing m-light still be
feasible with a different heuristic connectioni.
Certainly, in our own intellectual behavior we rarely
solve a trickiy problem by a steady climb toward success.
I doubt that in any one simple mechanism, e.g.,
hill-climbing, will we find the mieanis to build an efficienit
and general problem-solving machine. Probably, an
intelligent machine will require a variety of differenit
mechanisms. These will be arranged in hierarchies, and
in even more comnplex, perhaps recursive, structures.
And perhapswhat amounts to straightforward hill-climbing
on one level may sonmetimes appear (onl a lower
level) as the sudden jumllps of "insight."
II. THE PROBLEM OF PATTERN RECOGNITION
Summary-In order not to try all possibilities, a resourceful
machine must classify problem situations into categories associated
with the domains of effectiveness of the machine's different methods.
These pattern-recognition methods must extract the heuristically
significant features of the objects in question. The simplest methods
simply match the objects against standards or prototypes. More
powerful "property-list" methods subject each object to a sequence
of tests, each detecting some property of heuristic importance. These
properties have to be invariant under commonly encountered forms
of distortion. Two important problems arise here-inventing new
useful properties, and combining many properties to form a recognition
system. For complex problems, such methods will have to be
augmented by facilities for subdividing complex objects and describing
the complex relations between their parts.
Any powerful heuristic program is bound to contain a
variety of different methods and techniques. At each
step of the problem-solving process the machinie will
have to decide what aspect of the problem to work on,
to Janitary
A1insky: Steps Toward Artificial Intelligence
and then which mlethod to use. A choice mnust be made,
for we usually canntiot afford to try all the possibilities.
In order to deal with a goal or a problem, that is, to
choose an appropriate method, we have to recognize
what kind of thing it is. Thus the nieed to choose
among actions compels us to provide the machine with
classification techniques, or means of evolving them.
It is of overwhelming inmportance that the machine
have classification techniques which are realistic. But
"realistic" can be defined only with respect to the environmenits
to be encounitered by the machine, and with
respect to the methods available to it. Distinctions
which caninot be exploited are not worth recognizing.
And methods are usuall worthless without classification
schemes which can help decide when they are ap-
plicable.
A. Teleological Requirements of Classification
The useful classifications are those which match the
goals and methods of the machine. The objects grouped
together in the classifications should have something of
heuristic value in common; they should be "similar" in a
useful sense; they should depend on relevant or essential
features. WVe should not be surprised, then, to find ourselves
using inverse or teleological expressions to define
the classes. \Ve really do want to have a grip on "the
class of objects which can be transformed into a result
of form Y," that is, the class of objects which will satisfy
some goal. One should be wary of the familiar injunction
against using teleological language in science.
XVhile it is true that talking of goals in some contexts
imiay dispose us towards certain kinds of animilistic explanationis,
this need not be a bad thing in the field of
problem-solving; it is hard to see how one can solve
problems without thoughts of purposes. The real difficultv
with teleological definitions is technical, not
philosophical, and arises when they have to be used and
niot just mentionied. One obviously canniot afford to use
for classification a method which actually requires waiting
for some remote outcome, if one needs the classification
precisely for deciding whether to try out that
mlethod. So, in practice, the ideal teleological definitions
often have to be replaced by practical approximnations,
usually with some risk of error; that is, the definitions
have to be made heuristically effective, or economically
usable. This is of great importanice. (\Ve can think of
"heuristic effectiveness" as conitrasted to the ordinary
mathematical notion of "effectiveniess" which distinguishes
those definitions which can be realized at all by
machine, regardless of efficiency.)
B. Patterns and Descriptions
It is usually necessary to have ways of assigning
names symbolic expressions-to the defined classes.
The structure of the names will have a crucial influence
on the mental world of the machine, for it determines
what kinds of things can be conveniiently thought about.
There are a variety of ways to assign names. The simplest
schemes use what we will call conventional (or
proper) names; here, arbitrary symbols are assigned to
classes. But we will also want to use comnplex descriptions
or computed names; these are constructed for
classes by processes which depend on the class definitions.
To be useful, these should reflect some of the structure
of the things they designate, abstracted in a mainner
relevant to the problem area. The notion of description
merges smloothly inlto the more complex notion of
model; as we think of it, a model is a sort of active description.
It is a thinlg whose form reflects some of the
structure of the thing represented, but which also has
somle of the character of a working machinie.
In Section III we will consider "learniniig" systems.
The behavior of those systems can be made to change
in reasonable ways depending on what happened to
them in the past. But by themselves, the simuple learninig
systems are useful only in recurrent situations; they
canniot cope with any significant novelty. Nontrivial
performance is obtained only when learninig systemis are
supplemenited with classification or patterni-recognitioni
methods of some inductive ability. For the variety of
objects encountered in a nionitrivial search is so eniormous
that we cannot depend oni recurrence, and the miiere
accumulation of records of past experienice canl have
onily limited value. Patterni-Recogniitioni, by providinig
a heuristic connection which links the old to the 1new,
can make learning broadly useful.
What is a "pattern"? We ofteni use the term teleologically
to mean a set of objects which cani in somie (useful)
wav be treated alike. For each problem area we nmust
ask, "What patterns would be useful for a imiachinie
workinig on such problems?"
The problems of visual pattern-recogniition have received
much attention in recenit years anid miiost of our
examples are from this area.
1) Prototype-Derived Patterns: The problem of reading
printed characters is a clear-cut instanice of a situatioin
in which the classification is based ultimlately oni a
fixed set of "prototypes" e.g., the dies fromi which the
C. Prototype-Derived Patterns
The problem of reading printed characters is a clearcut
instance of a situation in which the classificationi is
based ultimately on a fixed set of "prototypes" e.g.,
the dies from which the type font was made. The individual
marks on the printed page may show the results
of many distortions. Some distortions are rather systematic:
change in size, position, orientation. Some are
of the nature of noise: blurring, grain, low contrast, etc.
If the noise is not too severe, we may be able to manage
the identification by what we call a normalization
and template-matching process. We first remove the
differences related to size and position that is, we
normalize the input figure. One may do this, for example,
by constructing a similar figure inscribed in a certain
fixed triangle (see Fig. 2); or one may transform the figure
to obtain a certain fixed center of gravity and a unit
seconid cenitral moment. (There is an additionial problem
1961 11
PROCEEDINGS OF THE IRE
with rotational equivalence where it is not easy to avoid
all ambiguities. One does not want to equate "6" anid
"9". For that matter, one does not want to equate (0, o),
or (X, x) or the o's in xo and xo, so that there may be
context-dependency involved.) Once normalized, the
unknown figure can be compared with templates for the
prototypes and, by means of some measure of matching,
choose the best fitting template. Each "matching criterion"
will be sensitive to particular forms of noise
and distortion, and so will each normalization procedure.
The inscribing or boxing method may be sensitive
to small specks, while the momeint method will be
especially sensitive to smearing, at least for thin-line
figures, etc. The choice of a matching criterion must depend
on the kinds of noise and transformations commonly
en-countered. Still, for many problems we may get
acceptable results by using straightforward correlation
methods.
When the class of equivalence transformations is very
large, e.g., when local stretching and distortion are
present, there will be difficulty in finding a uniform
normalization method. Instead, one may have to consider
a process of adjusting locallv for best fit to the
template. (While measuring the matching, one could
"jitter" the figure locally; if an improvement were found
the process could be repeated using a slightly different
change, etc.) There is usually no practical possibility
of applying to the figure all of the admissible transformations.
And to recognize the topological equivalenice of
pairs such as those in Fig. 3 is likely beyond anly practical
kind of iterative local-improvenment or hill-climiibing
matching procedure. (Such recognitions cani be mechanized,
though, by methods which follow lines, detect
vertices, and build up a description in the formi, say, of
a vertex-connection table.)
The template matching scheme, with its niormalization
and direct comparison and matching criterion, is
just too limited in conception to be of much use in more
difficult problems. If the transformation set is large,
normalization, or "fitting," may be impractical, especially
if there is no adequate heuristic connection on
the space of transformations. Furthermore, for each
defined pattern, the system has to be presenited with a
prototype. But if one has in minid a fairly abstract
class, one may simply be unable to represent its essential
features with one or a very few concrete examples.
How could one represent with a single prototype the
class of figures which have an even number of disconnected
parts? Clearly, the template system has negligible
descriptive power. The property-list system frees
us from some of these limitations.
D. Property Lists and "Characters"
We define a property to be a two-valued function
which divides figures into two classes; a figure is said to
have or not have the property according to whether
the function's value is 1 or 0. Given a number N of distinction
properties, we could define as many as 2n subclasses
by their set intersectionis and, henice, as many as
22n patterns by combiniing the properties with AND's
and OR's. Thus, if we have three properties, rectilinear,
connected, and cyclic, there are eight subclasses (anid 256
patterns) defined by their intersectionis (see Fig. 4).
If the giveni properties are placed in a fixed order theni
we can represenit anly of these eleinenitary regions by a
vector, or string of digits. The vector so assignied to each
figure will be called the Character of that figure (with
respect to the sequenice of properties in question). (In
[9] we use the term characteristic for a property without
restriction to 2 values.) Thus a square has the Character
(1, 1, 1) and a circle the Character (0, 1, 1) for the
given sequence of properties.
For many problems oine can use such Characters as
niames for categories and as primitive eleimieints with
Fig. 2-A simple normalization techniique. If ani object is expanded
uniformly, without rotation, until it touches all three sides of a
triangle, the resulting figure will be unLique, and pattern-recognition
can proceed withotut concern abouit relative size and position.
0A A' B B
Fig. 3 The figures A, A' and B, B' are topologically equivalent
pairs. Lengths have been distorted in an arbitrary manner, but
the connectivity relations between corresponding points have
been preserved. In Sherman [81 and Haller [391 we find computer
programs which can deal with such equivalences.
RECTILINEAR
-4 ~~~~~~~~~~~~~~~~~~~4 F
z
z
(i1,f 0)
(1,0,0)
L--j
CONTAINING
A
(I, 1, )
(Ito01)
A LOOP
D
(Of 1, )
(0,0,1)
(TD
FJ
(0,1,0)
(0,0,0)
CD
Fig. 4-The eight regions represent all the possible configurations of
values of the three properties "rectilinear," "connected," "containing
a loop." Each region contains a representative figure, and
its associated binary "Character" sequence.
6 6 - l-
12 Janu,sary
rT-T-","
Of
ll6insky: Steps Toward Artificial Intelligence
which to define an adequate set of patterns. Characters
are miiore thani conventional names. They are instead
very rudimentary forms of description (having the form
of the simplest symbolic expression-the list) whose
structure provides some information about the designated
classes. This is a step, albeit a small one, beyond
the template method; the Characters are not simple
inistances of the patterns, and the properties may themselves
be very abstract. Finding a good set of properties
is the major concern of many heuristic programs.
E. Invariant Properties
One of the prime requirements of a good property
is that it be invariant under the commonly encountered
equivalence transformations. Thus for visual
Pattern-Recognition we would usually want the object
identification to be independent of uniform changes
in size anid position. In their pioneering paper Pitts
and McCulloch [10] describe a general technique for
formiiig invariant properties from noninvariant ones,
assumiiing that the transformation space has a certain
(group) structure. The idea behind their mathematical
argumenit is this: suppose that we have a function P
of figures, and suppose that for a given figure F we
define [F] = F1, F2, - - }
to be the set of all figures
equivalent to F under the given set of transformations;
further, define P[F] to be the set IP(F1), P(F2),
of values of P onl those figures. Finally, define P*[F]
to be AVERAGE (P[F]). Then we have a new property,
P* whose values are independent of the selection
of F from an equivalence class defined by the transformations.
We have to be sure that when different
representatives are chosen from a class the collection
[F] will always be the same in each case. In the case
of continuous transformation spaces, there will have to
be a measure or the equivalent associated with the set
[F] with respect to which the operation AVERAGE is
defined, say, as an integration.4
This method is proposed [10] as a neurophysiological
model for pitch-invariant hearing anid size-invarianit
visual recognition (supplemented with visual centering
mechanisms). This model is discussed also by Wiener.5
Practical applicationi is probably limited to one-diml-ensionial
groups and analog scanning devices.
In much receiit work this probleml is avoided by
using properties already invariant unider these transformations.
Thus a property might counit the number
4 In the case studied in [10] the transformation space is a group
with a uniquely defined measure: the set [F] can be computed without
repetitions by scanning through the application of all the transforms
T, to the given figure so that the invariant property can be defined by
P*(F) f P(Ta(F))d1AaGG
wNhere G is the group and IA the measure. By substitutinig TO(F) for
F in this, one can see that the result is independent of choice of d
since we obtain the same integral over GO-'=G.
I Seep. 160 ff. of [111.
of connected components in a picture-this is invariant
unider size and position. Or a property may counit the
number of vertical lines in a picture this is invariant
under size anid position (but not rotationi).
F. Generating Properties
The problem of generating useful properties has been
discussed by Selfridge [12]; we shall summarize his approach.
The machine is given, at the start, a few basic
tranisformationis A41, * *,
A,,, each of which transforms,
[Al AA2 0 O COUNT
5
Fig. 5-An arbitrary sequence of picture-transformiiations, followed
by a numerical-valued function, can be used as a property function
for pictures. AI removes all points which are niot at the edge of
a solid region. A2 leaves only vertex poinits-at which an arcsuddenly
changes direction. The functioni C sim-iply counllts the nulmber
of points remaining in the picture. All remarks in the text
could be generalized to apply to properties, like AIA2C, which
canl have more than two values.
in some significant way, each figure inlto another figure.
A,1 miight, for example, remove all poinlts not on a
boutndary of a solid region; -42 might leave only- vertex
poinlts; A3 might fill utp hollow regions, etc. (see Fig. 5).
Each sequence A j114 .2 ..
ik of these formls a niew
transformation, so that there is available anl infiniite
variety. We provide the machinie also with onie or more
"terminal" operations which conivert a picture inito a
number, so that any sequenice of the elementary tranisformations,
followed by a terminial operationi, definies a
property. (Dineeni [13] describes how these processes
were programnmiled in a digital computer.) We canl start
with a few short sequences, perhaps choseni randomly.
Selfridge describes how the miiachinie might learni niew
useful properties.
We now feed the machine A's and O's tellinig the machinie each
time which letter it is. Beside each sequence under the two letters,
the machine builds up distribution funictionis from the resuLlts of
applyinig the sequenices to the image. Now, since the sequenices were
choseni completely randomly, it may well be that most of the sequences
have very flat distribution functions; that is, they [provide]
no information, and the sequences are therefore [by definitioni]iiot
signiificant. Let it discard these and pick some others. Sooner or later,
however, some sequences will prove significant; that is, their distribution
functions will peak up somewhere. What the machine does
now is to build up new sequences like the significant onies. This is the
important point. If it merely chose sequences at ranidom it might
take a very long while indeed to find the best sequences. Buit with
some successful sequences, or partly successful ones, to gulide it, we
hope that the process will be much qtuicker. The crucial questioni
remains: how do we build up sequences "like" other sequences, but
not idcentical? As of now we think we shall merely build sequenices
from the transition frequencies of the significanit sequences. We shall
build up a matrix of transition frequencies from the significanit onies,
anid use those as transition probabilities with which to choose lew
sequtenices.
We do niot claim that this method is necessarilv a very good way
of choosinig sequences-only that it should do better than niot uisinig
1961 13
PROCEEDINGS OF THE IRE
at all the knowledge of what kind of sequences has worked. It has
seemed to us that this is the crucial point of learning.6
It would indeed be remarkable if this failed to yield
properties more useful than would be obtainied from
completely random sequenice selection. The generatinig
problem is discussed further in Minisky [14].. Newell,
Shaw, and Simon [15] describe more deliberate, less
statistical, techniques that might be used to discover
sets of properties appropriate to a given problem area.
One may think of the Selfridge proposal as a system
which uses a finiite-state laniguage to describe its properties.
Solomonoff [55], [18] proposes some techniques
for discovering common features of a set of expressions,
e.g., of the descriptions of those properties of already
established utility; the methods can then be applied
to generate new properties with the same common
features. I consider the lines of attack in [12], [15],
[18] and [55], although still incomplete, to be of the
greatest importance.
G. Combining Properties
One cannot expect easily to find a small set of properties
which will be just right for a problem area. It is
usually much easier to find a large set of properties
each of which provides a little useful information. Then
one is faced with the problem of finding a way to combine
them to make the desired distinctions. The simplest
method is to choose, for each class, a typical character
(a particular sequence of property values) and then to
use some matching procedure, e.g., counting the inumbers
of agreemenits and disagreements, to compare an
unknown with these choseni "Character prototypes."
The linear weighting scheme described just below is a
slight genieralization on this. Such methods treat the
properties as more or less independent evidenice for and
against propositions; more general procedures (about
which we have yet little practical information) niust
account also for nonlinear relations between properties,
i.e., must contain weighting terms for joint subsets of
property values.
1) "Bayes nets" for combining independent properties:
We consider a single experiment in which an object is
placed in front of a property-list machine. Each property
Et will have a value, 0 or 1. Suppose that there
has been defined some set of "object classes" Fj, and
that we want to use the outcome of this experiment to
decide in which of these classes the object belongs.
Assume that the situation is basically probabilistic,
and that we know the probability pij that, if the object
is in class Fj then the ith property Ei will have value 1.
Assume further that these properties are independent;
that is, even given Fj, knowledge of the value of Es tells
us nothing more about the value of a different Ek in
the same experiment. (This is a strong condition see
below.) Let fj be the absolute probability that an object
6 See p. 93 of [121.
is in class Fj. Finally, for this experimiienit define V to be
the particular set of i's for which the E,'s are 1. Theni
this V represents the Character of the object. Fromn the
definition of conditional probability, we have
Pr (Fj, V) = Pr (V)*Pr (Fj V) = Pr (Fj) Pr (V| Fj).
Given the Character V, we wanit to guess which Fj has
occurred (with the least chanice of being wrong -the
so-called maximum likelihood estimuate); that is, for
which j is Pr(Fj| V) the largest? Since in the above
Pr (V) does niot depend oIn j, we have onily to calculate
for whichj is Pr (Fj) Pr(
V| Fj) =Oj Pr (VI Fj) the largest.
Hence, by our indepenidence hypothesis, we have to
maxinmize
(1)-
1 pij- CqVi =
ijGV . II q j.
iEV iev icv qij all i
These "maximnum likelihood" decisions can be made
(Fig. 6) by a simple nietwork device.7
These nets resemble the genieral schematic diagrams
proposed in the "Pandemoniumn" model of Selfridge
[19] (see his Fig. 3). It is proposed there that some intellectual
processes might be carried out by a hierarchy
of simultaneously functioning submlachines suggestively
called "demonis." Each uniit is set to detect certain patFig.
6 "Net" model for maximum-likelihood decisions based on
linear weightings of property values. The input data are examined
by each "property filter" Ei. Each Ei has "O" and "1" output
channels, one of which is excited by each input. These outputs are
weighted by the corresponding pij's, as shown in the text. The resulting
signals are multiplied in the F, units, each of which "collects
evidence" for a particular figure class. (We could have used
here log(pi,), and added at the Fi units.) The final decision is
made by the topmost unit D, who merely chooses that F, with the
largest score. Note that the logarithm of the coefficient pil/qii in
the second expression of (1) can be construed as the "weight of the
evidence" of Ei in favor of F,. (See also [211 and [221.)
7At the cost of an additional network layer, we may also account
for the possible cost guk that would be incurred if we were to assign
to Fk a figure really in class F1: in this case the minimum cost decision
is given by the k for which
E g3kOjI pij H qi
j ieV icV
is the least. V is the complement set to Viqij is (1-pii).
14 January
Minsky: Steps Toward Artificial Intelligence
terns in the activity of others and the output of each
unit announices the degree of confidence of that unit
that it sees what it is looking for. Our Es units are
Selfridge's "data demons." Our units Fj are his "cognitive
demonis"; each collects from the abstracted data
evidence for a specific proposition. The topmost "decisioIn
demoni" D responds to that one in the multitude
below it whose shriek is the loudest.8
It is quite easy to add to this "Bayes network model"
a imiechaniism which will enable it to learn the optimal
conniiectioni weightinigs. Imaginie that, after each event,
the machinie is told which Fj has occurred; we could impleinent
this by sendinig back a signal along the connectionis
leading to that Fj unit. Suppose that the connectioin
for pij (or qij) containis a two-terminal device (or
"s inapse") which stores a number wij. Whenever the
joinlt evenit (Fj, E, = 1) occurs, we modify WUj by replacinig
it by (wij+t1), where 6 is a factor slightly less than
unitv. And when the joinit event (Fj, Ej=0) occurs, we
decremenit wij by replacing it with (wij)6. It is not difficult
to show that the expected values of the wij's will
become proportional to the pij's [and, in fact, approach
pij[0j(1-0)]. Hence, the machine tends to learn the
optimal weighting on the basis of experience. (One must
put in a similar mechaniism for estimating the q5j's.) The
variance of the normalized weight wjj[(1-0)/0] approaches
[(1-0)/(1 +0) ]pijqij. Thus a snmall value for 0
meanis rapid learning but is associated with a large variance,
hence, with low reliability. Choosing 0 close to
unity means slow, but reliable, learning. 0 is really a
sort of memiiory decay constant, and its choice must be
determined by the noise and stability of the enivironment-much
nioise requires long averaging times, while
a changing environment requires fast adaptation. The
two requirements are, of course, incompatible and the
decision has to be based on an economic compromise.9
2) Possibilities of using random nets for Bayes decisions:
The nets of Fig. 6 are very orderly in structure.
Is all this structure necessary? Certainily if there were
a great maiiy properties, each of which provided very
little marginal information, some of them would not be
miissed. Theni one might expect good results with a
mere sampling of all the possible connection paths wij.
And one might thus, in this special situation, use a
random connection net.
The two-layer nets here resemble those of the "perceptron"
proposal of Rosenblatt [22]. In the latter,
there is an additional level of connections coming directly
from randomly selected poiInts of a "retina."
Here the properties, the devices which abstract the
visual input data, are simple functions which add some
iniputs, subtract others, and detect whether the result
exceeds a threshold. Eq. (1), we think, illustrates what
is of value in this scheme. It does seem clear that a
maximum-likelihood type of analysis of the output of
8 See also the report in [201.
9 See also [71 anid [211.
the property functions can be hanidled by such inets. But
these nets, with their simple, ranidomnly generated, connections
can probably never achieve recognitioni of such
patterns as "the class of figures having two separated
parts," and they cannot even achieve the effect of
template recognition without size and positioin niormiialization
(unless sample figures have been presented previously
in essentially all sizes and positions). For the
chances are extremely small of findinig, by random
methods, enough properties usefully correlated with
patterns appreciably more abstract than those of the
prototype-derived kind. And these networks cani really
only separate out (by weighting) informationi in the individual
input properties; they cannot extract further
information present in nonadditive formn. The "perceptron"
class of machines have facilities iieither for
obtaining better-than-chanlce properties nor for assembling
better-than-additive comiibiniationis of those it
gets from random construction.10
For recognizing normalized printed or hanid-printed
characters, single-point properties do surprisingly well
[23]; this amounts to just "averaging" many samples.
Bledsoe and Browning [24] claim good results with
poinit-pair properties. Roberts [25] describes a series of
experiments in this general area. Doyle [26] without
normalization but with quite sophisticated properties
obtainis excellent results; his properties are already substantially
size- and position-invariant. A general review
of Doyle's work and other pattern-recognition experiments
will be found in Selfridge and Neisser [20].
For the complex discrimination, e.g., between onie
and two connected objects, the property problem is
very serious, especially for long wiggly objects such as
are handled by Kirsch [27]. Here some kind of recursive
processing is required and combinations of simple properties
would almost certainly fail eveni with large nets
and long training.
We should not leave the discussion of some decisioni
net models without noting their important limitations.
The hypothesis that, for givenj, the pij represent independent
events, is a very strong condition indeed. Without
this hypothesis we could still construct maximumlikelihood
nets, but we would need an additional layer
of cells to represent all of the joint events V; that is,
we would need to know all the Pr (Fj4 V). This gives a
general (but trivial) solution, but requires 2n cells for
n properties, which is completely impractical for large
systems. What is required is a system which computes
some sampling of all the joint conditionial probabilities,
and uses these to estimate others when needed. The
work of Uttley [281, [29], bears on this problem, but his
proposed and experimental devices do not vet clearly
show how to avoid exponential growth."
10 See also Roberts [251, Papert [211, and Hawkins [221. We cani
find nothing resembling an analysis [see (1) abvel in [221 or subsequent
publications of Rosenblatt.
11 See also Papert [211.
1961 15
PROCEEDINGS OF THE IRE
H. Articulation and Attention Limitations of
the Property-List Method
Because of its fixed size, the property-list scheme is
limited (for any given set of properties) in the detail
of the distinctions it can make. Its ability to deal with
a compound scene containing several objects is critically
weak, and its direct extensions are unwieldy and unlnatural.
If a machine can recognize a chair and a table,
it surely should be able to tell us that "there is a chair
and a table." To an extent, we can invent properties
which allow some capacity for superposition of object
Characters.12 But there is no way to escape the information
limit.
What is required is clearly 1) a list (of whatever length
is necessary) of the primitive objects in the scene and
2) a statement about the relations among them. Thus
we say of Fig. 7(a), "A rectangle (1) contains two subfigures
disposed horizontally. The part on the left is a
rectangle (2) which contains two subfigures disposed
vertically; the upper a circle (3) and the lower a triangle
(4). The part on the right .. etc." Such a description
entails an ability to separate or "articulate"
the scene into parts. (Note that in this example the
articulation is essentially recursive; the figure is first
divided into two parts; then each part is described using
the same machinery.) We can formalize this kind of description
in an expression language whose fundamental
grammatical form is a pair (R, L) whose first member R
names a relation and whose second member L is an
ordered list (X1, X2, * * , xn) of the objects or subfigures
which bear that relation to one another. We obtain the
required flexibility by allowing the members of the list
L to contain not only the names of "elementary" figures
but also "subexpressions" of the form (R, L) designating
complex subfigures. Then our scene above may be described
by the expression
[o, (E, ((, (D, (tI , (0, A)))),
(OI (O, (,(O7, O,)O))) I
(a) (b) (c)
Fig. 7-The picture 4(a) is first described verbally in the text. Then,
by introducing notation for th, relations "inside of," "to the left
of" and "above," we construct a symbolic description. Such descriptions
can be formed and manipulated by machines. By abstracting
out the complex relation between the parts of the figure
we can use the same formula to describe the related pictures 4(b)
and 4(c), changing only the list of primitive parts. It is up to the
programmer to decide at just what level of complexity a part of a
picture should be considered "primitive"; this will depend on
what the description is to be used for. We could further divide
the drawings into vertices, lines, and arcs. Obviously, for some
applications the relations would need more metrical information,
e.g., specification of lengths or angles.
where (0, (x, y)) means that y is contained in x;
(-*, (x, y)) means that y is to the right of x; ( I, (x, y))
means that y is below x, and ( 7, (x, y, z)) means that y
is to the right of x and z is undern-eath and between theml.
The symbols 0, 0, and A represent the indicated
kinds of primitive geometric objects. This expressioinpair
description language may be regarded as a simple
kind of "list-structure" language. Powerful computer
techniques have been developed, originally by Newell,
Shaw and Simon, for manipulating symbolic expressions
in such languages for purposes of heuristic programming.
(See the remarks at the end of Section IV.
If some of the members of a list are themselves lists,
they must be surrounded by exterior parentheses, aind
this accounits for the accumulation of parentheses.)
It may be desirable to construct descriptions in which
the complex relation is extracted, e.g., so that we have
an expressioni of the form FG where F is an expression
which at once denotes the composite relation betweeni
all the primitive parts listed in G. A complicatioin arises
in connection with the "binding" of variables, i.e., in
specifying the manner in which the elements of G participate
in the relation F. This can be handled in general
by the "X" notation [32] but here we can just use integers
to order the variables.
For the given example, we could describe the relational
part F by an expression
0(1, - (0(2, 1 (3, 4)),0(5, V(6, 7, 8))))
in which we now use a "functional notation": "(0,
(x, y)) " is replaced by " (i (x, y)," etc., making for better
readability. To obtain the desired description, this
expressioni has to be applied to an ordered list of primitive
objects, which in this case is (F-, D, 0, A, 0, 0,
0, 0). This composite functional form allows us to
abstract the comnposite relation. By changing only the
object list we can obtain descriptions also of the objects
in Fig. 7(b) and 7(c).
The important thing about such "articular" descriptions
is that they can be obtained by repeated application
of a fixed set of pattern-recognition techniques. Thus
we can obtain arbitrarily complex descriptions from a
fixed complexity classification-mechanism. The new element
required in the mechanism (beside the capacity
to manipulate the list-structures) is the ability to articulate
to "attend fully" to a selected part of the picture
and bring all one's resources to bear on that part. In efficient
problem-solving programs, we will not usually
complete such a description in a single operation. Instead,
the depth or detail of description will be unider
the control of other processes. These will reach deeper,
or look more carefully, only when they have to, e.g.,
when the presently available description is inadequate
for a current goal. The author, together with L. Hodes,
is working on patternl-recognition schemes using articular
descriptions. By manipulating the formal descrip-
16 January
12 Cf. Mooers' technique of Zatocoding [301, [31].
o
I&.e-l.
Minsky: Steps Toward Artificial Intelligence
tions we can deal with overlapping and incomplete figures,
and several other problems of the "Gestalt" type.
It seems likely that as machines are turned toward
more difficult problem areas, passive classification systems
will become less adequate, and we may have to
turn toward schemes which are based more on internally-generated
hypotheses, perhaps "error-controlled"
along the lines proposed by MacKay [89].
Space requires us to terminate this discussion of pattern-recognition
and description. Among the important
works not reviewed here should be mentioned those of
Bomba [33] and Grimsdale, et al. [34], which involve
elements of description, Unger [35] and Holland [36]
for parallel processing schemes, Hebb [37] who is concerned
with physiological description models, and the
work of the Gestalt psychologists, notably Kohler [38]
who have certainly raised, if not solved, a number of important
questions. Sherman [8], Haller [39] and others
have completed programs using line-tracing operations
for topological classification. The papers of Selfridge
[12], [43], have beeni a major influence on work in this
general area.
See also Kirsch, et al. [27 ], for discussion of a number
of interesting computer image-processing techniques,
and see Minot [40] and Stevens [41 ] for reviews of the
reading machine and related problems. One should also
examine some biological work, e.g., Tinbergen [42] to
see instances in which some discriminations which seem,
at first glance very complicated are explained on the
basis of a few apparently simple properties arranged in
simple decision trees.
III. LEARNING SYSTEMS
Summary-In order to solve a new problem, one should first try
using methods similar to those that have worked on similar problems.
To implement this "basic learning heuristic" one must
generalize on past experience, and one way to do this is to use
success-reinforced decision models. These learning systems are
shown to be averaging devices. Using devices which learn also which
events are associated with reinforcement, i.e., reward, we can build
more autonomous "secondary reinforcement" systems. In applying
such methods to complex problems, one encounters a serious difficulty-in
distributing credit for success of a complex strategy among
the many decisions that were involved. This problem can be managed
by arranging for local reinforcement of partial goals within a
hierarchy, and by grading the training sequence of problems to
parallel a process of maturation of the machine's resources.
In order to solve a new problem one uses what might
be called the basic learning heuristic-first try using
methods similar to those which have worked, in the past,
on similar problems. We want our machines, too, to
benefit from their past experience. Since we cannot expect
new situations to be precisely the same as old ones,
any useful learning will have to involve generalization
techniques. There are too many notions associated with
"learning" to justify defining the term precisely. But we
may be sure that any useful learning system will have
to use records of the past as evidence for more general
propositions; it must thus entail some commitment or
other about "inductive inference." (See Section V-B.)
Perhaps the simplest way of generalizing about a set of
entities is through constructing a new one which is an
"ideal,"? or rather, a typical member of that set; the
usual way to do this is to smooth away variation by
some sort of averaging technique. And indeed we find
that most of the simple learning devices do incorporate
some averaging technique often that of averaging
some sort of product, thus obtaining a sort of correlation.
We shall discuss this family of devices here, and
some more abstract schemes in Section V.
A. Reinforcement
A reinforcement process is one in which some aspects
of the behavior of a system are caused to become more
(or less) prominent in the future as a consequence of the
application of a "reinforcement operator" Z. This
operator is required to affect only those aspects of behavior
for which instances have actually occurred re-
cently.
The analogy is with "reward" or "extinction" (not
punishment) in animal behavior. The important thing
about this kind of process is that it is "operant" (a
term of Skinner [44]); the reinforcement operator does
not initiate behavior, but merely selects that which the
Trainer likes from that which has occurred. Such a system
must then contain a device M which generates a
variety of behavior (say, in interacting with some environment)
and a Trainer who makes critical judgments
in applying the available reinforcement operators.
(See Fig. 8.)
Let us consider a very simple reinforcement model.
Suppose that on each -presentation of a stimulus S an
animal has to make a choice, e.g., to turn left or right,
and that its probability of turning right, at the nth
trial, is pn. Suppose that we want it to turn right. Whenever
it does this we might "reward" it by applying the
operator Z+;
pn+l = Z+(pn) = f9Op. + (1 - 0) 0 < 0 < 1
RESPON SE
ENVIROME N r
STIMU'LUS_ MACHINE
Fig. 8 Parts of an "operant reinforcement" learning system. In response
to a stimulus from the environment, the machine makes
one of several possible responses. It remembers what decisions
were made in choosing this response. Shortly thereafter, the
Trainer sends to the machine positive or negative reinforcement
(reward) signal; this increases or decreases the tendency to make
the same decisions in the future. Note that the Trainer need not
know how to solve problems, but only how to detect success or
failure, or relative improvement; his function is selective. The
Trainer might be connected to observe the actual stimulus-response
activity or, in a more interesting kind of system, just some
function of the state of the environment.
1961 17
PROCEEDINGS OF THE IRE
which moves p a fractioni (1 -6) of the way towards
unjitv.3 If we dislike what it does we apply negative re-
iiforcemeint,
pn+l = Z-(pn) = Op,
mlovinig p the same fraction of the way toward 0. Somiie
theory of such "linear" learniing operators, generalized
to several stimuli and responses, will be found in Bush
and Mosteller [45]. We canl show that the learninig result
is an average weighted by- anl exponentially-decayinig
time factor: Let Z,, be + 1 accordinig to whether the
nth event is rewarded or extiniguished anid replace
pl bycv = 2p,, -1 so that - 1 < c,, < 1, as for a correlatioii
coefficienit. Theni (with co=0) we obtaini by inlduc-
tioIl
Cn= (1 -) Ej " iZi,
t.=(l
aiid siiice
11/1(1 _a) E n-i,0
we canl write this as
fn-iz.
Cn+1 ~
2 ,0n-i
If the term Zi is regarded as a product of i) how the
creature responded and ii) which kinid of reiniforcement
was given, then c,, is a kinid of correlationi funictioni
(with the decay weightinig) of the joinit behavior of these
quatntities. The ordinary, uniiformiilv-weighted average
has the same general formn but with timiie-dependenit 0:
/ 1\ 1
Cn+l=t1 - c. + f Zn. (2)
In (1) we have againi the situation described in Sectioni
II-G, 1; a small value of 0 gives fast learning, anid the
possibility of quick adaptationi to a changinig environimnenit.
A near-unity value of 6 gives slow learning, but
also smooths away unicertainities due to nioise. As nioted
in Section II-G, 1, the response distributioni comiies to
approximate the probabilities of rewards of the alterniative
responises. (The imuportanice of this phenomenon
has, I think, been overrated; it is certainily niot ani especially
rational strategy. Onie reasoniable alterniative
is that of computinig the numbers pij as inidicated, but
actually playinig at each trial the "milost likelv" choice.
Except in the presenice of a hostile opponenit, there is
usually nio reasoin to play a "mixed" strategy.14)
13 Properly, the reinforcenent functions should depend both on
the p's and on the previous reaction reward should decrease p if our
animal has just turned to the left. The notation in the literature is
also somewhat confusinig in this regard.
14 The question of just how often one should play a strategy different
from the estimated optimum, in order to gain information, is
an underlying problem in many fields. See, e.g., [851.
In Samuel's coefficient-optimiiizinig program1 [2] [see
Sectioni II I-C, 1) ], there is a most inigeniious comiipromnise
betweeni the exponietitial anid the uniiforimi averagilig
methods: the value of N in (2) above beginis at 16 aind
so remains until n= 16, theni N is 32 until n=32, anid
so oni unitil n =256. Thereafter N remiains fixed at
256. This nicely prevenits violenit fluctuationis ii1 c>, aIt
the start, approaches the uniiforimi weightling for a while,
anid finially approaches the exponentially-weighted correlation,
all in a maniner that requires ver} little COm1lpUtatioII
effort! Samuel's programil is at presenit the outstanidinig
examilple of a gamie-playing program which
maitches average huniiani ability, and its success (in real
time) is attributed to a wealth of such elegmciles, both
in heuristics anid in programiniiiig.
The problemii of extinictioni or "unilearniing" is especiallv
critical for complex, hierarchical, leariniig. For,
onice a genieralizationi about the past has been miade,
onie is likelI to build uponi it. Thus, onie may comiie to select
certaini properties as imiiportanit and begini to ulse
them in the characterizationi of experienice, perhaps storinig
onle's imemories in terimis of themii. If later it is discovered
that some other properties would serve better,
theni onie must face the probleml of tranislatinig, or abandoninig,
the records based oni the older system. This
may be a very high price to pav. Onie does niot easil\
give up anl old way of lookinig at thinigs, if the better onie
demands much effort and experienice to be useful. TIhus
the training sequtences oni which ouir miachinies will spenid
their inifancies, so to speak, mllust be choseii very
shrewdly to inisure that early abstractionis will provide a
good founidationi for later difficult problemils.
Inicidenitally, in spite of the space giveni here for
their expositioni, I amii niot coniviniced that such "inicremenital"
or "statistical" leariiing schemes should play a
cenitral role in our milodels. They will certainl\- conitiniue
to appear as componienits of our prograiams but, I thinik,
mainilv by default. The niore initelligenit onie is, the milore
ofteni he should be able to learin fromii an experienice
somilethinig rather definiite; e.g., to reject or accept a
hypothesis, or to chanige a goal. (The obvious exception
is that of a truly- statistical eniviroiinmenit in which
averaginig is iniescapable. But the heart of problemiisolvinig
is alwavs, we thinik, the combinatorial part that
gives rise to searches, anid we should usually be able to
regard the complexities caused by "nioise" as miere
annoyances, however irritatinig they may be.) In this
conlniectiotn we cani refer to the discussioni of menmory in
Miller, Galaniter anid Pribramii [46].15 This seems to be
the first imajor work in Psychology to show the inifluenice
of work in the artificial initelligenice area, anid its programmiie
is genierally quite sophisticated.
B. Secondary Reinforcement and Expectation Models
The simple reinforcement system is linmited by its depenidenice
on the Trainer. If the Trainier cani detect only
1' See especially ch. 10.
18 Janiuary
Minsky: Steps Toward Artificial Intelligence
the solution of a problem, then we may encounter
"mesa"~ phenomena which will limit performance on
difficult problems. (See Section I-C.) One way to escape
this is to have the machine learn to generalize on what
the Trainer does. Then, in difficult problems, it may be
able to give itself partial reinforcements along the way,
e.g., upon the solution of relevant subproblems. The
machine in Fig. 9 has some such ability. The new unit
Fig. 9-An additional device U gives the machine of Fig. 8 the ability
to learn which signals from the environment have been associated
with reinforcement. The primary reinforcement signals Z are
routed through U. By a Pavlovian conditioning process (not described
here), external signals come to produce reinforcement signals
like those that have frequentlv succeeded them in the past.
Such signals might be abstract, e.g., verbal encouragement. If the
"secondary reinforcement" signals are allowed, in turn, to acquire
further external associations (through, e.g., a channel Zu as
shown) the machine might come to be able to handle chains of
subproblems. But something must be done to stabilize the system
against the positive symbolic feedback loop formed by the path
Zu. The profound difficulty presented by this stabilization problem
may be reflected in the fact that, in lower animals, it is very
difficult to demonstrate such chaining effects.
U is a device that learns which external stimuli are
strongly correlated with the various reinforcement signals,
and responds to such stimuli by reproducing the
corresponding reinforcement signals. (The device U is
not itself a reinforcement learning device; it is more like
a "Pavlovian" conditioning device, treating the Z signals
as "unconditioned" stimuli and the S signals as
conditioned stimuli.) The heuristic idea is that any signal
from the environment which in the past has been
well correlated with (say) positive reinforcement is
likely to be an indication that something good has just
happened. If the training on early problems was such
that this is realistic, then the system eventually should
be able to detach itself from the Trainer, and become
autonomous. If we further permit "chaining" of the
"secondary reinforcers," e.g., by admitting the connection
shown as a dotted line in Fig. 9, the scheme becomes
quite powerful, in principle. There are obvious
pitfalls in admitting such a degree of autonomy; the
values of the system may drift to a "nonadaptive"
condition.
C. Prediction and Expectation
The evaluation unit U is supposed to acquire an
ability to tell whether a situation is good or bad. This
evaluation could be applied to imaginary situations as
well as to real ones. If we could estimate the consequences
of a proposed action (without its actual execution),
we could use U to evaluate the (estimated) resulting
situation. This could help in reducing the effort in
search, and we would have in effect a machine with
some ability to look ahead, or plan. In order to do this
we need an additional device P which, givein the descriptions
of a situation and an actioni, will predict a
description of the likely result. (We will discuss schemes
for doing this in Section IV-C.) The device P might
be conistructed along the lines of a reinforcement learning
device. In such a system the required reinforcemienit
signals would have a very attractive character. For the
machine must reinforce P positively when the actual
outtcome resembles that which was predicted accurate expectations
are rewarded. If we could further add a premium
to reinforcement of those predictions which have
a novel aspect, we might expect to discern behavior
motivated by a sort of curiosity. In the reinforcement
of mechanisms for confirmed novel expectations (or new
explanations) we may find the key to simulation of intellectual
motivation.16
Samuel's Program for Checkers: In Samuel's "generalization
learning" program for the game of checkers
[2] we find a novel heuristic technique which could be
regarded as a simple example of the "expectation reinforcement"
notion. Let us review very briefly the situation
in playing two-person board games of this kind. As
noted by Shannon [3] such games are in principle finite,
and a best strategy can be found by following out all
possible continuations-if he goes there I can go there,
or there, etc.-and then "backing-up" or "minimaxing"
from the terminal positions, won, lost, or drawni.
But in practice the full exploration of the resulting
colossal "move-tree" is out of the question. No doubt,
some exploration will always be necessary for such
games. But the tree must be pruned. We might simply
put a limit on depth of exploration-the number of
moves and replies. We might also limit the number of
alternatives explored from each position-this requires
some heuristics for selection of "plausible moves."17
Now, if the backing-up technique is still to be used
(with the incomplete move-tree) one has to substitute
for the absolute "win, lose, or draw" criterion some
other "static" way of evaluating nonterminal positions."8
(See Fig. 10.) Perhaps the simplest scheme is
to use a weighted sum of some selected set of "property"
functions of the positions-mobility, advancement,
center control, and the like. This is done in Samuel's
program, and in most of its predecessors. Associated
with this is a multiple-simultaneous-optimizer method
16 See also ch. 6 of [471.
17 See the discussion of Bernstein [481 and the more extensive review
and discussion in the very suggestive paper of Newell, Shaw, and
Simon [49]; one should not overlook the pioneering paper of Newell
[501, and Samuel's discussion of the minimaxing process in [2].
18 In some problems the backing-up process can be handled in
closed analytic form so that one may be able to use such methods as
Bellman's "Dynamic Programming" [51]. Freimer [521 gives some
examples for which limited "look-ahead" doesn't work.
1961 19
PROCEEDINGS OF THE IRE
4 z
MAX MIN MAX MIN MAX
Fig. 10-"Backing-up" the static evaluations of proposed moves in a
g-ame-tree. From the vertex at the left, representing the present
positioni in a board game, radiate three branches, representing
the player's proposed moves. Each of these might be countered
by a variety of opponent moves, and so on. According to some
program, a finite tree is generated. Then the worth to the player
of each terminal board position is estimated. (See text.) If the
opponent has the same values, he will choose to miniimize the
score, while the player will always try to maximize. The heavy
lines show how this minimaxing process backs up until a choice is
determined for the present position.
The full tree for chess has the order of 10120 branches-beyond
the reach of any man or computer. There is a fundamental heuristic
exchange between the effectiveness of the evaluation function
and the extent of the tree. A very weak evaluation (e.g., one
which just compares the players' values of pieces) would yield
a devastating game if the machine could explore all continuLations
ouit to, say, 20 levels. But only 6 levels, roughly within range of
ouir presetntly largest computers, would probably not give a brillianit
game; less exhaustive strategies, perhaps along the linles of
[49], would be more profitable.
for discovering a good coefficient assignment (usinig the
correlation techniique noted in Sectioni III-A). But the
source of reinforcement signals in [2 ] is novel. Onie
canniiot afford to play out one or more entire games for
each sinigle learning step. Samuel measures instead for
each more the difference betweeni what the evaluation
funiction yields directly of a position and what it predicts
oni the basis of an extenisive conitinuation exploratioln,
i.e., backinig-up. The signl of this error, "Delta," is
used for reiniforcement; thus the system maxy learn
somiiethinig at each move.19
D) The Basic Credit-Assignment Problem for Complex
Reinforcement Learning Systems
In playinig a complex game such as chess or checkers,
or in writing a computer program, one has a definite
29 It should be noted that [2] describes also a rather successful
checker-playing program based on recording and retrieving iniformation
about positionls encountered in the past, a less abstract way of
exploiting past experience. Samuel's work is notable in the variety of
experiments that were performed, with and without various heuristics.
This gives an unusual opportunity to really find out how different
heuristic methods compare. More workers should choose (other
things being equal) problems for which such variations are practi-
cable.
success criterion-the game is woin or lost. But ill the
course of play, each ultimlate success (or failure) is
associated with a vast niumiiber of initeriial decisionis. If
the run is successful, how canl we assignl credlit for the
success among the multitude of decisions? As Newell
noted,
It is extremely doubtful whether there is enotugh informationi in
"iwinl, lose, or draw" when referred to the whole play of the gamlie to
permit any learning at all over available time scales.... For learnilng
to take place, each play of the game must yield much more information.
This is . . achieved by breakiing the problem into componienits.
The uinit of success is the goal. If a goal is achieved, its subgoals are
reinforced; if not they are inhibited. (Actually, what is reinforced is
the transformation rule that provided the subgoal.) This also is
true of the other kinds of structure: every tactic that is created provides
informationi about the success or failure of tactic search rules;
every opponent's action provides information about success or failure
of likelihood inferences; and so on. The amount of informatioil relevant
to learning increases directly with the number of mechanlisms
in the chess-playing machine.20
We are in complete agreement with Newell onl this approach
to the problem.21
It is my impression that many workers in the area of
"self-organiizing" systems and "ranidomn neural niets" do
not feel the urgency of this problem. Suppose that onie
million decisions are involved in a complex task (such
as winning a chess gamiie). Could we assign to each decisioIn
element onie-millionith of the credit for the coinpleted
task? In certain special situations we caii do just
this e.g., in the imiachinies of [22], [25] anid [92], etc.,
where the connectionis being reinforced are to a sufficienit
degree inidependent. But the problemii-solvinig
ability is correspondingly weak.
For more comuplex problems, with decisionis in hierarchies
(rather thani summed oni the same level) anid
with incremenits small enough to assure probable conivergenice,
the runninig times would become fanitastic.
For complex problems we will have to define "success"
in some rich local sense. Some of the difficulty may be
evaded by usinlg carefullv graded "traininig sequenices"
as described in the following sectioni.
Friedberg's Program- Writing Program: An imnportanit
examiiple of comparative failure in this credit-assignmiiient
matter is provided by the programii of Friedberg [53],
[54] to solve programii-writinig problemns. The problemii
here is to write progranms for a (simnulated) very simiiple
digital comiiputer. A simlple probleml is assigned, e.g.,
"comiipute the AND of two bits in storage anid put the
result in ani assigned locationi." A genierating (levice
produces a random (64-inistruction) programn. The program
is runi anid its success or failure is nioted. The
success iniformation is used to reiniforce individutal instructions
(in fixed locationis) so that each success tenids
to ii1crease the chanice that the instructioins of successful
programs will appear in later trials. (We lack space
for details of how- this is done.) Thus the programii tries
20 See p. 108 of [501.
21 See also the discussion in Samuel (p. 22 of [21) oni assigning
credit for a change in "Delta."
Janit.ary270
Minsky: Steps Toward A rtificial Intelligence
to find "good" instructions, more or less independently,
for each location in program memory. The machine did
learn to solve some extremely simple problems. But it
took of the order of 1000 times longer than pure chance
would expect. In part II of [54], this failure is discussed,
and attributed in part to what wxe called (Section I-C)
the "Mesa phenomena." In changing just one instruction
at a time the machine had not taken large enough
steps in its search through program space.
The second paper goes on to discuss a sequence of
modifications in the program generator and its reinforcement
operators. With these, and with some "priming"
(starting the machine off on the right track with
some useful instructions), the system came to be only a
little worse than chance. The authors of [54] conclude
that with these improvements "the generally superior
performance of those machines with a success-number
reinforcement mechanism over those without does serve
to indicate that such a mechanism can provide a basis
for constructing a learning machine." I disagree with
this conclusion. It seems to me that each of the "improvements"
can be interpreted as serving only to increase
the step size of the search, that is, the randomness
of the mechanism; this helps to avoid the Mesa
phenomenon and thus approach chance behavior. But
it certainly does not show that the "learning mechanism"
is working-one would want at least to see
some better-than-chaiice results before arguing this
point. The trouble, it seems, is with credit-assignment.
The credit for a working program can only be assigned
to functional groups of instructionis, e.g., subroutines,
and as these operate in hierarchies we should not expect
individual instruction reinforcement to work well.22 It
seems surprising that it was not recognized in [54]
that the doubts raised earlier were probably justified!
In the last section of [54] we see some real success obtained
by breaking the problem into parts and solving
them sequentially. (This successful demonstration using
division into subproblems does not use any reinforcement
mechanism at all.) Some experiments of
similar nature are reported in [94].
It is my conviction that no scheme for learning, or
for pattern-recognition, can have very general utility
unless there are provisions for recursive, or at least hierarchical,
use of previous results. We cannot expect a
learning system to come to handle very hard problems
without preparing it with a reasonably graded sequence
of problems of growing difficulty. The first problem
must be one which can be solved in reasonable time
with the initial resources. The next must be capable of
solution in reasonable time by using reasonably simple
and accessible combinations of methods developed in
the first, and so on. The only alternatives to this use of
an adequate "training sequence" are 1) advanced resources,
given initially, or 2) the fantastic exploratory
22See the introduction to [53] for a thoughtful discussion of the
plausibility of the scheme.
processes found perhaps only in the history of organic
evolution.23 And even there, if we accept the general
view of Darlington [56] who emphasizes the heuristic
aspects of genetic systems, we mlust have developed
early (in, e.g., the phenomena of meiosis and crossingover)
quite highly specialized mechanisms providing for
the segregation of groupings related to solutions of subproblems.
Recently, much effort has been devoted to
the construction of trainiing sequences in connection
with programming "teaching machines." Naturally, the
psychological literature abounds with theories of how
complex behavior is built up from simpler. In our own
area, perhaps the work of Solomonoff [55], while overly
cryptic, shows the most thorough consideration of this
dependency on training sequences.
IV. PROBLEM-SOLVING AND PLANNING
Summary-The solution, by machine, of really complex problems
will require a variety of administration facilities. During the course
of solving a problem, one becomes involved with a large assembly of
interrelated subproblems. From these, at each stage, a very few must
be chosen for investigation. This decision must be based on 1) estimates
of relative difficulties and 2) estimates of centrality of the different
candidates for attention. Following subproblem selection (for
which several heuristic methods are proposed), one must choose
methods appropriate to the selected problems. But for really difficult
problems, even these step-by-step heuristics for reducing search
will fail, and the machine must have resources for analyzing the
problem structure in the large-in short, for "planning." A number
of schemes for planning are discussed, among them the use of
models-analogous, semantic, and abstract. Certain abstract models,
"Character Algebras," can be constructed by the machine itself, on
the basis of experience or analysis. For concreteness, the discussion
begins with a description of a simple but significant system (LT)
which encounters some of these problems.
A. The "Logic Theory" Program of Newell,
Shaw and Simon
It is not surprising that the testing grounds for early
work on mechanical problem-solving have usually
been areas of mathematics, or games, in which the rules
are defined with absolute clarity. The "Logic Theory"
machine of [57], [58], called "LT" below, was a first
attempt to prove theorems in logic, by frankly heuristic
methods. Although the program was not by human
standards a brilliant success (and did not surpass its designers),
it stands as a landmark both in heuristic programming
and also in the development of modern automatic
programming.
The problem domain here is that of discovering proofs
in the Russell-Whitehead system for the propositional
calculus. That system is given as a set of (five) axioms
and (three) rules of inference; the latter specify how
23 It should, however, be possible to construct learninig mechanisms
which can select for themselves reasonably good training sequences
(from an always complex environment) by pre-arranging a
relatively slow development (or "maturation") of the system's facilities.
This might be done by pre-arranging that the sequence of goals
attempted by the primary Trainer match reasonably well, at each
stage, the complexity of performance mechanically available to the
pattern-recognition and other parts of the system. One might be able
to do much of this by simply limiting the depth of hierarchical activity,
perhaps only later permitting limited recursive activity.
1961 21
PROCEEDINGS OF THE IRE
certain transformiiationis can be applied to produce niew
theorems fromn old theorems and axioms.
The LT program is ceintered around the idea of "workinig
backwards" to finid a proof. Giveni a theorein T to
be proved, LT searches anmonig the axioms anid previously
established theorems for onie from which T canl
be deduced by a single application of onie of three simple
"Methods" (which embody the giveni rules of iniferenice).
If onie is found, the problem is solved. Or the search
might fail completely. But finially, the search may vield
onie or more "problems" which are usually propositionls
from which T may be deduced directly. If one of
these can, in turn, be proved a theoremn the maini problem
will be solved. (The situationi is actually slightly
more complex.) Each such subproblem is adjoinied to
the "subproblem list" (after a limited preliminary attempt)
and LT works arounid to it later. The full power
of LT, such as it is, can be applied to each subproblem,
for LT cani use itself as a subroutinie in a recursive
fashion.
The heuristic technique of working backwards yields
somnething of a teleological process, anid LT is a forerunniier
of more complex systems which conistruct hierarchies
of goals and subgoals. Eveni so, the basic adminiistrative
structure of the program is Ino more than
a inested set of searches through lists in memory. We
shall first outlinie this structure anid theni menitioni a few
heuristics that were used in attemiipts to improve per-
formanice.
1) Take the inext problem froml problem list.
(If there are no more problems, EXIT with total
failure.)
2) Choose the next of the three basic Methods.
(If nio more mnethods, go to 1.)
3) Choose the next member of the list of axiomns and
previous theorems.
(If no more, go to 2.)
Then apply the Method to the problem, usinig the
chosen theorem or axionm.
If problem is solved, EXIT with complete proof.
If nio result, go to 3.
If new subproblemi arises, go to 4.
4) Try the special (substitutioni) Method on the sub-
problem.
If problem is solved, EXIT with complete proof.
If no result, put the subproblem at the end of the
problem list and go to 3.
Among the heuristics that were studied were 1) a
similarity test to reduce the work in step 4 (which includes
another search through the theorem list), 2) a
simplicity test to select apparently easier problems from
the problem list, and 3) a strong nonprovability test to
remove from the problenm list expressiotns which are
probably false atnd hence not provable. In a series of
experiments "learning"' was used to find which earlier
theorems had been most useful and should be given
priority in step 3. We cainnot review the effects of these
chaniges in detail. Of interest was the balanice between
the extra cost for adminiistrationi of certaini heuristics
and the resultanit search reduction; this balanice was
quite delicate in some cases when computer memory
became saturated. The system seemed to be quite
sensitive to the traininig sequence -the order in which
problems were given. And some heuristics which gave
nio significant over-all improvemenit did nievertheless
affect the class of solvable problems. Curiously enlough,
the genieral efficiency of LT was nlot greatly improvedl
by any or all of these devices. But all this practical experienice
is reflected in the design of the mluch imiore sophisticated
"GPS" system described brieflyt in Section
IV-D, 2).
WTang [59] has criticized the LT project oni the
grounds that there exist, as he anid others have showni,
mechanized proof methods which, for the particular
runi of problems coinsidered, use far less machinie effort
than does LT and which have the advantage that they
will ultimately find a proof for any provable propositioIn.
(LT does inot have this exhaustive "decisioni procedure"
character anid can fail ever to finid proofs for
somne theoreimis.) The authors of [58], perhaps unaware
of the existenice of eveni moderately efficienit exhaustive
methods, supported their argunmenits by comiiparison
with a particularly iniefficienit exhaustive procedeire.
Nevertheless, I feel that some of \NVang's criticismiis are
misdirected. He does niot seem to recogiiize that the
authors of LT are n1ot so much initerested in provinig
these theoremiis as they are in the genieral problenm of
solving difficult problemiis. The combitnatorial system of
Russell anid Whitehead (with which LT cleals) is far
less simple anid eleganit thani the system used byWangg.24
(Note, e.g., the emphasis in [49] and [60].) WVang's
problems, while logically equivalent, are formally mtuch
simnpler. His methods do inot iniclude any facilities for
usinig previous results (henice theyt are sure to degrade
rapidly at a certaini level of problem conmplexity), while
LT is funidamenitally- orienited around this probleml.
Finallv, because of the verv effectiveniess of Wanig's
method oni the particular set of theoremls in questiotn,
he simply did inot have to face the futndan1enital heuristic
problem of when to decide to give utp on a linle of attack.
Thus the formidable performanice of his program [59]
perhaps diverted his attenitioni fromii heuristic problems
that must againl sprinig up wheni real mathemuatics is
ultimatelyr enicountered.
This is niot meant as a rejectioii of the importanice of
Waiig's work and discussion. He aind others workinig oni
"nmechaniical mathematics" have discovered that there
are proof procedures which are much more efficienit
thani has beeni suspected. Such work will uniquestionably
help in conistructinig intelligenit machinies, anid
24 Wang's procedure [591 too, works backwards, and can be regarded
as a generalization of the method of "falsification" for deciding
truth-functional tautology. In 1931 and its unpublished sequel,
Wang introduces more powerful methods (for much more difficult
problems).
22 Januabry
Minsky: Steps Toward Artificial Intelligence
these procedures will certainly be preferred, when available,
to "unreliable heuristic methods." Wang, Davis
and Putnam, and several others are now pushing these
new techniques into the far more challenging domain
of theorem-proving in the predicate calculus (for which
exhaustive decision procedures are no longer available).
We have no space to discuss this area,25 but it seems
clear that a program to solve real mathematical problems
will have to combine the mathematical sophistication
of Wang with the heuristic sophistication of
Newell, Shaw and Simon.26
B. Heuristics for Subproblem Selection
In designing a problem-solving system, the programmer
often comes equipped with a set of more or
less distinct "Methods"-his real task is to find an
efficient way for the program to decide where and
when the different methods are to be used.
Methods which do not dispose of a problem may still
transform it to create new problems or subproblems.
Hence, during the course of solving one problem we may
become involved with a large assembly of interrelated
subproblems. A "parallel" computer, yet to be conceived,
might work on many at a time. But even the
parallel machine must have procedures to allocate its resources
because it cannot simultaneously apply all its
methods to all the problems. We shall divide this administrative
problem into two parts: the selection of
those subproblem(s) which seem most critical, attractive,
or otherwise immediate, and, in the next section,
the choice of which method to apply to the selected
problem.
In the basic program for LT (Section IV-A), subproblem
selection is very simple. New problems are
examined briefly and (if not solved at once) are placed
at the end of the (linear) problem list. The main program
proceeds along this list (step 1), attacking the
problems in the order of their generation. More powerful
systems will have to be more judicious (both in
generation and selection of problems) for only thus can
excessive branching be restrained.27 In more complex
systems we can expect to consider for each subproblem,
at least these two aspects: 1) its apparent "centrality"
how will its solution promote the main goal, and
2) its apparent "difficulty"-how much effort is it liable
to consume. We need heuristic methods to estimate
each of these quantities and, further, to select accordingly
one of the problems and allocate to it some rea-
25 See [611 and [931.
2' All these efforts are directed toward the reduction of search
effort. In that sense they are all heuristic programs. Since practically
no one still uses "heuristic" in a sense opposed to "algorithmic,"
serious workers might do well to avoid pointless argument on this
score. The real problem is to find methods which significantly delay
the apparently inevitable exponential growth of search trees.
27 Note that the simple scheme of LT has the property that each
generated problem will eventually get attention, even if several are
created in a step 3. If one were to turnfull attention to each problem,
as generated, one might never rettirn to alternate branches.
sonable quantity of effort.2'8 Little enough is known
about these matters, and so it is not entirely for lack of
space that the following remarks are somewhat cryptic.
Imagine that the problems and their relations are
arranged to form some kind of directed-graph structure
[14], [57], [62]. The main problem is to establish a
"valid" path between two initially distinguished nodes.
Generation of new problems is represented by the addition
of new, not-yet-valid paths, or by the insertion of
new nodes in old paths. Then problems are represented
by not-yet-valid paths, and "centrality" by location in
the structure. Associate with each connection, quantities
describing its current validity state (solved, plausible,
doubtful, etc.) and its current estimated difficulty.
1) Global Methods: The most general problem-selection
methods are "global" at each step they look over
the entire structure. There is one such simple scheme
which works well on at least one rather degenerate interpretation
of our problem graph. This is based on an
electrical analogy suggested to us by a machine designed
by Shannon (related to one described in [63]
which describes quite a variety of interesting gameplaying
and learning machities) to play a variant of the
game marketed as "Hex" (and known amonig mathematicians
as "Nash"). The initial board position can
be represented as a certain network of resistors. (See
Fig. 11.) One player's goal is to construct a shortcircuit
path between two given boundaries; the opponent
tries to open the circuit between them. Each
move consists of shorting (or opening), irreversibly, one
of the remaining resistors. Shannon's machinie applies a
potential between the boundaries and selects that
resistor which carries the largest current. Very roughly
speaking, this resistor is likely to be most critical because
changing it will have the largest effect on the ret
4 7
3 6 9
Fig. 11-This board game (due to C. E. Shannon) is played on a nletwork
of equial resistors. The first player's goal is to open the circuit
between the endpoints; the second player's goal is to short the circuit.
A move consists of opening or shorting a resistor. If the first
player begins by opening resistor 1, the second player might
counter by shorting resistor 4, following the strategy described in
the text. The remaining move pairs (if both players use that
strategy) would be (5, 8) (9, 13) (6, 3) (12, 10 or 2) (2 or 10 win).
In this game the first player should be able to force a win, and the
maximum-current strategy seems always to do so, even on larger
networks.
28 One will want to see if the considered problem is the same as
one already considered, or very similar. See the discussion in [621.
This problem might be handled more generally by simply remembering
the (Characters of) problems that have been attacked, and
checking new ones against this memory, e.g., by methods of [311,
looking more closely if there seems to be a match.
1961 23
PROCEEDINGS OF THE IRE
sistance of the net and, hence, in the goal direction of
shorting (or openinig) the circuit. And although this argument
is not perfect, nor is this a perfect model of the
real combinatorial situation, the machine does play extremely
well. (It can make unsound moves in certain
artificial situationis, but no one seems to have been able
to force this during a game.)
The use of such a global method for problem-selection
requires that the available "difficulty estimates"
for related subproblems be arranged to combine in
roughly the manner of resistanice values. Also, we could
regard this machine as using an "analog model" for
"'planninig." (See Sectioni IV-D.) 29
2) Local, and "Ilereditary," Methods: The prospect of
havinig to study at each step the whole problem structure
is discouraginig, especially since the structure usually
changes only slightly after each attempt. One naturally
looks for methods which merely update or modify
a small fragmenit of the stored record. Between the extremes
of the "first-come-first-served" problem-list
method aind the full global-survey methods, lie a variety
of compromise techniques. Perhaps the most attractive
of these are what we will call the Inheritance methods
-essentially recursive devices.
In ani Inheritanice method, the effort assigned to a
subproblem is determinled only by its immediate ancestry;
at the tinme each problem is created it is assignied
a certaini total quanitity Q of time or effort. When
a problem is later split inito subproblems, such quantities
are assigiied to them by some local process which
depends only on their relative merits and on what remains
of Q. Thus the cenitrality problem is maniaged implicitly.
Such schemnes are quite easy to program, especially with
the niew programminig systems such as IPL [64] anid
LISP [32 ] (which are thetmiselves based on certain
hereditary or recursive operationis). Special cases of the
iniheritanice miiethod arise wheni onie canl get along with a
simiple all-or-nionie Q, e.g., a "stop condition" this
yields the exploratorynmethod called "back-trackinig"
by Golumb [65 ]. The decoding procedure of Wozenicraft
[66] is another imiiportant variety of Inheritanice
method.
Iln the complex exploration process proposed for chess
by Newell, Shaw, and Simloni [49] we have a form of
Inheritanice method with a non-numerical stop-condition.
Here, the subprobleimis iniherit sets of goals to be
achieved. This teleological conitrol has to be admiinistered
by ani additionial goal-selectioni system and is further
comiiplicated by a global (but reasoniably simple) stop
rule of the backing-up variety [Section III-C]. (Note:
we are identifying here the move-tree-limitation problemii
with that of problem-selection.) Eveni though extenisive
experimental results are not yet available, we
feel that the scheme of [49] deserves careful study by
29 A variety of combinatorial methods will be matched against
the network-analogy opponient ini a program being completed by R.
Silver, Lincoln Lab., M.I.T., Lexinigton, Mass.
anyone planning serious work in this area. It shows
only the beginning of the complexity sure to come iti
our developmenit of intelligent machinies.30
C. "Character-Method" Machines
Once a problem is selected, we must decide which
method to try first. This depends oni our ability to
classify or characterize problems. We first compute the
Character of our problem (by usinig some patterni recognition
techniique) and theni conisult a "CharacterM\lethod"
table or other device which is supposed to tell
us which method(s) are most effective oni problemns of
that Character. This informationi might be built up
from experienice, given initially by the programmiiiier, deduced
from "advice" [70], or obtainied as the solutioni
to some other problem, as suggested in the GPS proposal
[68]. In any case, this part of the machinie's behavior,
regarded from the outside, canl be treated as a
sort of stimulus-responise, or "table look-up," activity.
If the Characters (or descriptionis) have too wide a
variety of values, there will be a serious problem of fillinig
a Character-Method table. One miiight theni have to
reduce the detail of iniforn1atioin, e.g., by usinig only a
few importanit properties. Thus the Differences of GPS
[see Sectioni IV-D, 2) ] describe ino mnore thani is iiecessary
to define a single goal, aind a priority schem-iie selects
just olle of these to characterize the situationi.
Gelernter and Rochester [62] suggest usinig a propertyweightinig
schemne, a special case of the "Bayes niet"
described in Section II-G.
D. Planning
Ordinarily onie can solve a complicated problemii only
by dividing it into a number of parts, each of which cani
be attacked by a smaller search (or be further divided).
Generally speaking, a successful divisioni will reduce the
search time not by a mere fractioni, but by a Jfractional
exponenit. In a graph with 10 branches descenidinig from
each niode, a 20-step search might inivolve 1020 trials,
which is out of the questioni, while the inisertioni of just
four lemmas or sequential subgoals might reduce the
search to only 5.104 trials, which is withini reasoni for
machitne explorationi. Thus it will be worth a relativelxenormous
effort to find such "islanids" in the solutioni of
complex problems.3' Note that eveni if onie enicouniterecd,
say, 106 failures of such procedures before success, onie
would still have gainied a factor of perhaps 1010 in overall
trial reductioni! Thus practically any ability at all to
"plan," or "analyze," a problem will be profitable, if the
problem is difficult. It is safe to say that all simple, uniitary,
notionis of how to build anl intelligenit machiine wNill
fail, rather sharply, for somle modest level of problemii
difficulty. Onlxy schemes which actively pursue anl anialysis
toward obtainiing a set of sequential goals cani be
3 Some further discussioll of this qutestioni may be fotntd ill
Slagle [671.31 See section 10 of [61.
24 January
Minsky: Steps Toward Artificial Intelligence
expected to extend smoothly into increasingly complex
problem domains.
Perhaps the most straightforward concept of planning
is that of using a simplified model of the problem
situation. Suppose that there is available, for a given
problem, some other problem of "essentially the same
character" but with less detail and complexity. Then
we could proceed first to solve the simpler problem. Suppose,
also, that this is done using a second set of methods,
which are also simpler, but in some correspondence
with those for the original. The solution to the simpler
problem can then be used as a "plan" for the harder one.
Perhaps each step will have to be expanded in detail.
But the multiple searches will add, not multiply, in the
total search time. The situation would be ideal if the
model were, mathematically, a homomorphism of the
original. But even without such perfection the model
solution should be a valuable guide. In mathematics
one's proof procedures usually run along these lines;
one first assunmes, e.g., that integrals and limits always
converge, in the planning stage. Once the outline is
completed, in this simple-minded model of mathematics,
then one goes back to try to "make rigorous"
the steps of the proof, i.e., to replace them by chains of
argument using genuine rules of inference. And even if
the plan fails, it may be possible to patch it by replacing
just a few of its steps.
Another aid to planning is the semantic, as opposed
to the homomorphic, model [14], [9]. Here we may
have an interpretation of the current problem within
another system, not necessarily simpler, but with which
we are more familiar and have already more powerful
methods. Thus, in connection with a plan for the proof
of a theorem, we will want to know whether the proposed
lemmas, or islands in the proof, are actually true;
if not, the plan will surely fail. We can often easily tell
if a proposition is true by looking at an interpretation.
Thus the truth of a proposition from plane geometry
can be supposed, at least with great reliability, by
actual measurement of a few constructed drawings (or
the analytic geometry equivalent). The geometry machine
of Gelernter and Rochester [62], [69] uses such
a semantic model with excellent results; it follows
closely the lines proposed in [14].
1) The "Character-Algebra" Model: Planning with the
aid of a model is of the greatest value in reducing search.
Can we construct machines which find their own
models? I believe the following will provide a general,
straightforward way to construct certain kinds of useful,
abstract models. The critical requirement is that
we be able to compile a "Character-Method Matrix"
(in addition to the simple Character-Method table in
Section IV-C). The CM matrix is an array of entries
which predict with some reliability what will happen when
methods are applied to problems. Both of the matrix
dimensions are indexed by problem Characters; if there
is a method which usually transforms problems of character
Ci into problems of character Cj then let the matrix
entry C1i be the name of that method (or a list of such
methods). If there is no such method the corresponding
entry is null.
Now suppose that there is no entry for Cij-meaning
that we have no direct way to transform a problem of
type Ci into one of type C1. Multiply the matrix by itself.
If the new matrix has a non-null (i, j) entry then
there must be a sequence of two methods which effects
the desired transformation. If that fails, we may try
higher powers. Note that [if we put uniity for the (i, i)
terms] we can reach the 2n matrix power with just n
multiplications. We don't need to define the symbolic
multiplication operation; one may instead use arithmetic
entries putting unity for any non-null entry and
zero for any null entry in the original matrix. This yields
a simple connection, or flow diagram, matrix, and its
nth power tells us something about its set of paths of
length 28.32 (Once a non-null entry is discovered, there
exist efficient ways to find the correspondinig sequenices
of methods. The problem is really just that of findinig
paths through a maze, and the method of Moore [71]
would be quite efficient. Almost any problenm can be
converted into a problem of finding a chaini between
two terminal expressions in some formal system.) If
the Characters are taken to be abstract representations
of the problem expressions, this "Character-Algebra"
model can be as abstract as are the available patternrecognition
facilities. See [14] and [9].
The critical problem in using the Character-Algebra
model for planning is, of course, the prediction-reliability
of the matrix entries. One cannot expect the Character
of a result to be strictly determined by the Character
of the original and the method used. And the reliability
of the predictions will, in any case, deteriorate rapidly
as the matrix power is raised. But, as we have noted,
any plan at all is so much better than nonie that the system
should do very much better than exhaustive search,
even with quite poor prediction quality.
This matrix formulation is obviously onily a special
case of the character planning idea. M\Iore generally,
one will have descriptions, rather than fixed characters,
and one must then have more general methods to calculate
from a description what is likely to happeni when a
method is applied.
2) Characters and Differences: In the GPS (General
Problem Solver) proposal of Newell, Shaw, and Simon
[68], [15], we find a slightly different framework: they
use a notion of Difference between two problems (or expressions)
where we speak of the Character of a sinigle
problem. These views are equivalent if we take our
problems to be links or connections betweeni expressions.
But this notion of Difference (as the Character of a
pair) does lend itself more smoothly to teleological reasoning.
For what is the goal defined by a problemn but
to reduce the "difference" between the present state and the
32See, e.g., [881.
1961 25,
PROCEEDINGS OF TIHE IRE
desired state? The underlying structure of GPS is precisely
what we have called a "Character-Method Machinie"
in which each kind of Differenice is associated in
a table with onie or more methods which are known to
"reduce" that Differenice. Since the characterization
here depends always on 1) the current problem expressiotn
and 2) the desired end result, it is reasonable to
think, as its authors suggest, of GPS as usinig "meainsenid"
analysis.
To illustrate the use of Differences, we shall review
an example [15]. The problem, in elementary propositional
calculus, is to prove that from SA(-PDQ) we
canl deduce (QVP) AS. The program looks at both of
these expressions with a recursive matching process
which braniches out from the main conn1ectives. The
first Difference it enicouniters is that S occurs oII differenit
sides of the main coinnective "A". It therefore
looks in the Difference-Method table under the headinig
"chanige position." It discovers there a method
which uses the theorem (A AB) (B AA) which is
obviously useful for removinig, or "reducing," differenices
of position. GPS applies this method, obtainintg
(-PDQ) AS. GPS now asks what is the Differel1ce betwreeni
this niew expressioni anid the goal. This time the
matchinig procedure gets downi inito the conniectives inside
the left-hand members and finids a Difference betw-eeni
the coinnectives "D" anid "V". It n1ow looks in
the CM table unider the headinig "Chanige Conniiective"
and discovers the appropriate method usling (-ADB)
=(A VB). It applies this miiethod, obtaininlg (PVQ) AS.
In the final c-cle, the differenice-evaluatinig procedure
discovers the nieed for a "change position1" inside the
left member, anid applies a method using (A VB)
-(BVA). This completes the solutioni of the problem.33
EvidentlI, the success of this "imeanis-enid" anialI-sis in
reducinig general search will depenid on1 the degree of
specificity that cani be writteni in1to the DiffereniceMethod
table-basically the same requiremenit for an
effective Character-Algebra.
It may be possible to plan usinig Differences, as well.3
Onie might imagine a "Differenice-Algebra" in which the
predictions have the form D=D'D". One must con1struct
accordinigly a difference-factorization algebra for
discoverinig loniger chainis D=Di . . Dn anid correspondinig
method planis. WN\e should note that onie cannot
expect to use such planning methods with such
primitive Differenices as are discussed in [15]; for these
canniiot form ani adequate Difference-Algebra (or Character
Algebra). Unless the characterizing expression-s
33 Compare this with the "matching" process described in [571.
The notions of "Character," "Character-Algebra," etc., originate in
[14] but seem useful in describing parts of the "GPS" system of [57]
and [15]. Reference [15] contains much additional material we cannot
survey here. Essentially, GPS is to be self-applied to the problem of
discovering sets of Differences appropriate for given problem areas.
This notion of "bootstrapping"-applying a problem-solving system
to the task of improving some of its own methods-is old and familiar,
but in [15] we find perhaps the first specific proposal about how
such an advance might be realized.
have many levels of descriptive detail, the miiatrix powers
will too swiftly become degenierate. This degetieracy
will ultimately limit the capacity of any formcal plannlling
device.
Otne may thinik of the genieral planniinig heuristic as
enmbodied in a recursive process of the followinig formii.
Suppose we have a problemii P:
1) Formi a plan- for problemii P.
2) Select first (niext) step of the plan.
(If nio more steps, exit with "success.")
3) Try the suggested miiethod(s):
Success: returni to b), i.e., try niext step in the
plani.
Failure: returni to a), i.e., form niew plani, or
perhaps chanige currenit plani to avoid this
step.
Problemii judged too difficult: A-pply this entire
proceditre to the problem of the cuirrent step.
Observe that such a program scheina is essentially recursive;
it uses itself as a subroutine (explicitly, in the
last step) in such a way- that its currenit state has to be
stored, anid restored wheni it returnis conitrol to itself.34
\[/iller, Galaniter anid Pribram33 discuss possible analogies
betweeni humani problemii-solvinig anid somiie
heuristic planniinig schemes. It seemiis certaini that, for
at least a few vears, there will be a close associationi betweeni
theories of humiiani behavior aind attempts to inicrease
the initellectual capacities of miiachinies. But, in
the lonig runi, we inust be prepared to discover profitable
linies of heuristic programmiiiinig rhich (lo niot deliberatelv
imitate humani characteristics.36
34 This violates, for example, the restrictions on "DO loops' in
programming systems such as FORTRAN. Conivenient techniques
for programniing such processes were developed by Newell, Shaw,
and Simon [64]; the program state-variables are stored in "pushdown
lists" and both the program and the data are stored in the forn
of "list-structures." Gelernter [69] extended FORTRAN to manage
some of this. McCarthy has extended these notions in LISP [32] to
pernmit explicit recursive definitions of programs in a language based
on recursive functions of symbolic expressions; here the manage-
3enit of program-state variables is fully automatic. See also OrchardHays'
article in this issue.
35 See chs. 12 and 13 of [461.
36 Limitations of space preclude detailed discussion here of
theories of self-organizing neural nets, and other models based on
brain analogies. (Several of these are described or cited in [C] and
[D].) This omission is not too serious, I feel, in connection with the
subject of heuristic programming, because the motivation and methods
of the two areas seem so different. Up to the present time, at
least, research on neural-net models has been concerned mainly with
the attempt to show that certain rather simple heuristic processes,
e.g., reinforcement learning, or property-list pattern-recognition,
can be realized or evolved by collections of simple elements without
very highly organized interconnections. Work on heuristic programiiming
is characterized quite differently by the search for new, more
powerful heuristics for solving very complex problems, and by very
little concern for what hardware (neuronal or otherwise) would
minimally suffice for its realization. In short, the work on "nets" is
concerned with how far one can get with a small initial endowmnent;
the work on "artificial intelligence" is concerned with using all we
know to build the most powerful systems that we can. It is my expectation
that, in problem-solving power, the (allegedly brain-like)
minimal-structure systems will never threaten to compete with their
more deliberately designed contemporaries; nevertheless, their study
should prove profitable in the development of component elements
and subsystems to be used in the construction of the more systematically
conceived machines.
26 Janitary
Minsky: Steps Toward Artificial Intelligence
V. INDUCTION AND MODELS
A. Intelligence
In all of this discussion we have not come to grips
with anything we can isolate as "intelligence." We
have discussed only heuristics, shortcuts, and classification
techniques. Is there something missing? I am confident
that sooner or later we will be able to assemble
programs of great problem-solving ability from complex
combiniations of heuristic devices multiple optimizers,
pattern-recognitioni tricks, planning algebras,
recursive administration procedures, and the like. In
no one of these will we find the seat of intelligence.
Should we ask what intelligence "really is"? My owni
view is that this is more of an esthetic question, or one
of senise of dignity, than a technical nmatter! To me "intelligenice"
seems to denote little more than the complex
of performances which we happen to respect, but do
not understand. So it is, usually, with the question of
"depth" in mathematics. Once the proof of a theorem
is really understood its content seems to become trivial.
(Still, there may remaini a sense of wonder about how
the proof was discovered.)
Programmers, too, know that there is never any
"heart" in a program. There are high-level routines in
each program, but all they do is dictate that "if suchand-such,
then transfer to such-and-such a subroutine."
And when we look at the low-level subroutines,
which "actually do the work," we find senseless loops
and sequences of trivial operations, merely carrying out
the dictates of their superiors. The intelligence in such
a system seems to be as intangible as becomes the meaning
of a single common word when it is thoughtfully
pron-ounced over and over again.
But we should not let our inability to discern a locus
of intelligence lead us to conclude that programmed
computers therefore cannot think. For it may be so
with man, as with machine, that, when we understand
finally the structure and program, the feeling of mystery
(and self-approbation) will weaken.37 We find similar
views concerning "creativity" in [60 ]. The view expressed
by Rosenbloom [73] that minds (or brains) can transcend
machines is based, apparently, on an erroneous
interpretation of the meaning of the "unsolvability
theorems" of Godel.38
B. Inductive Inference
Let us pose now for our machines, a variety of problems
more challenging than any ordinary game or
mathematical puzzle. Suppose that we want a machine
which, when embedded for a time in a complex
environment or "universe," will essay to produce a
description of that world-to discover its regularities
or laws of nature. We might ask it to predict what will
happen next. We might ask it to predict what would be
the likely consequences of a certain action or experiment.
Or we might ask it to formulate the laws governing
some class of events. In any case, our task is to
equip our machine with inductive ability-with methods
which it can use to construct genieral statements
about events beyond its recorded experienice. Now,
there can be no system for inductive inference that will
work well in all possible universes. But given a uniiverse,
or an ensemble of universes, and a criterion of success,
this (epistemological) problem for machines becomes
technical rather than philosophical. There is quite a
literature concerning this subject, but we shall discuss
only one approach which currently seems to us the most
promising; this is what we might call the "grammatical
induction" schemes of Solomonoff [55], [16], [17],
based partly on work of Chomsky and Miller [80],
[81].
We will take language to mean the set of expressionis
formed from some given set of primitive symbols or
expressions, by the repeated application of some given
set of rules; the primitive expressions plus the rules is
the grammar of the language. Most induction problems
can be framed as problems in the discovery of grammars.
Suppose, for instance, that a machine's prior experience
is summarized by a large collection of statements,
some labelled "good" and some "bad" by some
critical device. How could we generate selectively more
good statements? The trick is to find some relatively
simple (formal) language in which the good statements
are grammatical, and in which the bad ones are not.
Given such a language, we can use it to generate more
statements, and presumably these will tend to be more
like the good ones. The heuristic argumenit is that if we
can find a relatively simple way to separate the two sets,
the discovered rule is likely to be useful beyonid the immediate
experience. If the extension fails to be consisteint
with new data, one might be able to make small chaniges
in the rules and, generally, one may be able to use nmany
ordinary problem-solving methods for this task.
The problem of fiinding an efficient grammar is much
the same as that of finding efficient encodings, or pro-
37 See [141 and [9].
38 On problems of volition we are in general agreement with McCulloch
[75] that our freedom of will "presumably means no more
than that we can distinguish between what we intend [i.e., our plan],
and some intervention in our action." See also MacKay (176] and its
references); we are, however, unconvinced by his eulogization of
"analogue" devices. Concerning the "mind-brain" problem, one
should consider the arguments of Craik [771, Hayek [781 and Pask
[791. Among the active leaders in modern heuristic programming,
perhaps only Samuel [91] has taken a strong position against the
idea of machines thinking. His argument, based on the fact that reliable
computers do only that which they are instructed to do, has a
basic flaw; it does not follow that the programmer therefore has full
knowledge (and therefore full responsibility and credit for) what will
ensue. For certainly the programmer may set up an evolutionary
system whose limitations are for him unclear and possibly incomprehensible.
No better does the mathematician know all the consequences
of a proposed set of axioms. Surely a machine has to be in
order to perform. But we cannot assign all the credit to its programmer
if the operation of a system comes to reveal structures not recognizable
or anticipated by the programmer. While we have not yet
seen much in the way of intelligent activity in machines, Samuel's
arguments in [91] (circular in that they are based on the presumption
that machines do not have minds) do not assure us against this.
Turing [72] gives a very knowledgeable discussion of such matters.
1961 27
PROCEEDINGS OF THE IRE
grams, for machinies; in. each case, onie needs to discover
the important regularities in the data, anid exploit the
regularities by- making shrewd abbreviations. The possible
importanice of Solomoioff's work [18] is that, despite
some obvious defects, it may point the way toward
sN stematic mathematical ways to explore this discovery
problem. He considers the class of all programls (for a
giveni general-purpose computer) which will produce a
certain given output (the body of data in question-).
MVlost such programs, if allowed to continue, will add to
that body of data. By properly weighting these programs,
perhaps by lenigth, we can obtaini correspotndinig
weights for the differenit possible continiuations, anid thus
a basis for prediction. If this prediction is to be of any
interest, it will be necessary to show some inidependence
of the giveni computer; it is Inot yet clear precisely
wNhat form such a result will take.
C. Miodels of Oneself
If a creature can answer a questioni about a hypothetical
experiment, without actually performinig that experimenit,
then the answer must have beeni obtainied
from some submachinie inside the creature. The output
of that submachinie (representinig a correct aniswer)
as well as the iniput (represeniting the question) must be
coded descriptions of the corresponidiing external evenits
or event classes. Seen through this pair of enicoding and
decoding chainnels, the initerinal submachine acts like
the environmenit, anid so it has the character of a
"model." The inductive iniference probleml may theni be
regarded as the problem of constructing such a miiodel.
To the extenit that the creature's actions affect the
enivironment, this internial model of the world will nieed
to iniclude some representation of the creature itself. If
onie asks the creature "why did you decide to do such
and such" (or if it asks this of itself), any aniswer must
come from the internial model. Thus the evidence of
initrospection itself is liable to be based ultimately on
the processes used in conistructinig onie's imlage of onie's
self. Speculatioin on the form of such a model leads to
the amusinig prediction that initelligenit machines mlav
be reluctant to believe that they are just machinies. The
argument is this: our owIn self-models have a substantiallv
"dual" character; there is a part concernied with
the physical or mechanical enivironimenit-with the behavior
of inanimate objects anid there is a part conicernied
with social and psychological matters. It is precisely
because we have not yet developed a satisfactory
mechaniical theory of mental activity that we have to
keep these areas apart. We could nlot give up this division1
even if we wished to until we find a unified model
to replace it. Now, wheii we ask such a creature what
sort of beinig it is, it canniiot simply answNer "directlv;"
it must inispect its model(s). Anid it miust aniswer by
sayinig that it seemiis to be a dual thinig which appears
to have two parts a "iminiid" and a "body." Thus, eveni
the robot, unless equipped with a satisfactory theory- of
artificial initelligence, would have to maintami a dualistic
opitnioIn on1 this matter.39
CONCLUSION
In attempting to combinie a survey of work on "artificial
initelligence" with a summary- of our own7 views,
we could nlot mentioni every relevaInt project and publication.
Some important omIissionls are inl the area of
"braini models"; the earlv work of Parley anid Clark
[92] (also Farley's paper in [D], ofteni unkniowiniglx
duplicated, and the work of Rochester [82] and Mjilner
[D].) The work of Lettvini, et al. [83] is related to the
theories in [19]. We did not touch at all oni the problemiis
of logic and language, anid of ilnfornmationi retrieval,
which must be faced when actioni is to be based oni the
contents of large memories; see, e.g., \'IcCarthy [70].
We have niot discussed the basic results inl mathematical
logic which bear on the questioni of what canl be donie by
machinies. There are enltire literatures we have hardly7
even sampled---the bold pionieerinig of Rashevsky (c.
1929) anid his later co-workers [95 ]; Theories of Learlinig,
e.g., Gorni [84]; Theory of Gamiies, e.g., Shubik
[85]; anid Psychology, e.g., Brunier, et al. [86]. Anld
everyonie should kniow the wA,ork of Polva [87] oni how
to solve problems. We canl hope only to have tranismitted
the flavor of somne of the more amiibitious projects
directly conicernied with gettinig n1achiines to take over
a larger portioni of problenm-solvinig tasks.
One last remiiark: we have discussed here onlv work
conicerned with more or less self-contaMied problemiisolving
programs. But as this is writteni, we are at last
beginning to see vigorous activity in the directioni of
conistructinig usable time-sharing or multiprogramming
computinlg systems. With these systems, it will at last
becomne econiomical to match humiiani beinigs in realI
tinme with really large mlachinies. This miieanls that we
canl work toward programminig what will be, in effect,
"thinkinig aids." In the years to comiie, N-e expect that
these man-machine systemls will share, anid perhaps for
a time be domlliinant, ill our advanice toward the developmenit
of "artificial itntelligel1ce."
39 There is a certaii problem of ifiinite regression in the niotionlof a
machine having a good model of itself: of course, the nested models
must lose detail and finally vanish. But the argumenit, e.g., of Hayek
(See 8.69 and 8.79 of [781) that we cannot "fully comprehend the uniitary
order" (of our own minds) ignores the power of recursive description
as well as Turing's demonstration that (with sufficient external
writing space) a "general-purpose" machine can answer any
question about a description of itself that any larger machine could
answer.
28 Janutary
Minsky: Steps Toward Artificial Intelligence
BIBLITOGRAPHY4
[1] J. McCarthy, "The inversion of functions defined by Turing
machines," [B].
[2] A. L. Samuel, "Some studies in machine learning using the game
of checkers," IBMJ. Res. &Dev.,vol. 3,pp. 211-219;July, 1959.
[3] C. E. Shannon, "Programming a digital computer for playing
chess," [K].
[4] C. E. Shannon, "Synthesis of two-terminal switching networks,"
Bell Sys. Tech. J., vol. 28, pp. 59-98; 1949.
[5] W. R. Ashby, "Design for a Brain," John Wiley and Sons, Inc.,
New York, N. Y.; 1952.
[6] W. R. Ashby, "Design for an intelligence amplifier," [B].
[7] M. L. Minsky and 0. G. Selfridge, "Learning in random nets,"
[H].
[8] H. Sherman, "A quasi-topological method for machine recognition
of line patterns," [E].
[9] M. L. Minsky, "Some aspects of heuristic programming and
artificial intelligence," [C].
[10] W. Pitts and W. S. McCulloch, "How we know universals,"
Bull. Math. Biophys., vol. 9, pp. 127-147; 1947.
[11] N. Wiener, "Cybernetics," John Wiley and Sons, Inc., New
York, N. Y.; 1948.
[12] 0. G. Selfridge, "Pattern recognition and modern computers,"
[A].
[13] G. P. Dinneen, "Programming pattern recognition," [A].
[14] M. L. Minsky, "Heuristic Aspects of the Artificial Intelligence
Problem," Lincoln Lab., M.I.T., Lexington, Mass., Group
Rept. 34-55, ASTIA Doc. No. 236885; December, 1956.
(M.I.T. Hayden Library No. H-58.)
[15] A. Newell, J. C. Shaw, and H. A. Simon, "A variety of intelligence
learning in a general problem solver," [D].
[16] R. J. Solomonoff, "The Mechanization of Linguistic Learning,"
Zator Co., Cambridge, Mass., Zator Tech. Bull. No. 125, presented
at the Second Internatl. Congress on Cybernetics,
Namur, Belgium; September, 1958.
[17] R. J. Solomonoff, "A new method for discovering the grammars
of phrase structure languages," [E].
[18] R. J. Solomonoff, "A Preliminary Report on a General Theory
of Inductive Inference," Zator Co., Cambridge, Mass., Zator
Tech. Bull. V-131; February, 1960.
[19] 0. G. Selfridge, "Pandemonium: a paradigm for learning," [C].
[20] 0. G. Selfridge and U. Neisser, "Pattern recognition by
machine," Sci. Am., vol. 203, pp. 60-68; August, 1960.
[21] S. Papert, "Some mathematical models of learning," [H].
[22] F. Rosenblatt, "The Perceptron," Cornell Aeronautical Lab.,
Inc., Ithaca, N. Y., Rept. No. VG-1196-G-1; January, 1958. See
also the article of Hawkins in this issue.
[23] W. H. Highleynman and L. A. Kamentsky, "Comments on a
character recognition method of Bledsoe and Browning" (Correspondence),
IRE TRANS. ON ELECTRONIC COMPUTERS, vol.
EC-9, p. 263; June, 1960.
[24] W. W. Bledsoe and I. Browning, "Pattern recognition and reading
by machine," [F].
[25] L. G. Roberts, "Pattern recognition with an adaptive network,"
1960 IRE INTERNATIONAL CONVENTION RECORD, pt. 2, pp. 66-
70.
[26] W. Doyle, "Recognition of Sloppy, Hand-Printed Characters,"
Lincoln Lab., M.I.T., Lexington, Mass., Group Rept. 54-12;
December, 1959.
[27] R. A. Kirsch, C. Ray, L. Cahn, and G. H. Urban, "Experiments
in Processing Pictorial Information with a Digital Computer,"
Proc. EJCC, PROC. IRE, pp. 221-229; December, 1957.
[28] A. M. Uttley, "Conditional probability machines," and "Temporal
and spatial patterns in a conditional probability machine,"
[B].
[29] A. M. Uttley, "Conditional probability computing in a nervous
system," [C].
[30] C. N. Mooers, "Information retrieval on structured content,"
[J].
40 Bibliographic note: Work in this area seems to be currently
prominent in the following periodicals:
1) IBM J.Res.&Dev.
2) Information and Control.
3) Proc. EJCC and WJCC (EASTERN AND WESTERN JOINT
COMPUTER CONFS.)
4) IRE NATIONAL CONVENTION RECORD.
5) J. Assoc. Comp. Mach. (J. A CM).
6) Trans. Assoc. Comp. Mach.
7) IRE TRANS. ON INFORMATION THEORY.
A more informative bibliography, compiled by the present author,
should appear shortly in the IRE TRANS. ON HUMAN FACTORS IN
ELECTRONICS.
[311 C. N. Mooers, "Zatocoding and developments in information
retrieval," Aslib Proc., vol. 8, pp. 3-22; February, 1956.
[32] J. McCarthy, "Recursive functions of symbolic expressions," [G].
[331 J. S. Bomba, "Alpha-numeric character recognition using local
operations," [F].
[34] R. L. Grimsdale, et al., "A system for the automatic recognition
of patterns," Proc. IEE, vol. 106, pt. B; March, 1959.
[35] S. H. Unger, "Pattern detection and recognition," PROC. IRE,
vol. 47, pp. 1737-1752; October, 1959.
[36] J. H. Holland, "On iterative circuit computers constructed of
microelectronic components and systems," Proc. WJCC, pp.
259-265; 1960.
[37] D. 0. Hebb, "The Organization of Behavior," John Wiley and
Sons, Inc., New York, N. Y.; 1949.
[38] W. Kohler, "Gestalt Psychology," Mentor, No. MD 279; 1947.
[39] N. Haller, "Line Tracing for Character Recognition," M.S.E.E,
thesis, M.I.T., Cambridge, Mass.; 1959.
[40] 0. N. Minot, "Automatic Devices for Recognition of Visible
Two-dimensional Patterns: A Survey of the Field," U. S. Naval
Electronics Lab., San Diego, Calif., Tech. Memo. 364; June 25,
1959.
[41] M. E. Stevens, "A Survey of Automatic Reading Techniques,"
NBS, U. S. Dept. of Commerce, Washington, D. C., Rept. 5643;
August, 1957.
[42] N. Tinbergen, "The Study of Instinct," Oxford University
Press, New York, N. Y.; 1951.
[43] 0. G. Selfridge, "Pattern recognition and learning," [J].
[44] B. F. Skinner, "Science and Human Behavior," The Macmillan
Co., New York, N. Y.; 1953.
[45] R. R. Bush and F. Mosteller, "Stochastic Models For Learning,"
John Wiley and Sons, Inc., New York, N. Y.; 1955.
[46] G. A. Miller, E. Galanter, and K. H. Pribram, "Plans and the
Structure of Behavior," Henrv Holt and Co., Inc., New York,
N. Y.; 1960.
[47] M. L. Minsky, "Neural Nets and the Brain Model Problem,"
Ph.D. dissertation, Princeton Univ., Princeton, N. J.; 1954.
(University Microfilms, Ann Arbor.)
[48] A. Bernstein, et al., "A chess playing program for the IBM 704,"
Proc. WJCC, pp. 157-159; 1958.
[49] A. Newell, J. C. Shaw, and H. A. Simon, "Chess-playing programs
and the problem of complexity," IBM J. Res. & Dev.,
vol. 2, p. 320 ff.; October, 1958.
[50] A. Newell, "The chess machine," [A].
[51] R. Bellman, "Dynamic Programming," Princeton University
Press, Princeton, N. J.; 1957.
[52] M. Freimer, "Topics in Dynamic Programming II," Lincoln
Lab., M.I.T., Lexington, Mass., Rept. 52-G-0020; April, 1960.
(M.I.T. Hayden Library No. H-82). See especially sec. I-E.
[53] R. M. Friedberg, "A learning machine, part I," IBM J. Res. &
Dev., vol. 2, pp. 2-13; January, 1958.
[54] R. M. Friedberg, B. Dunham, and J. H. North, "A learning machine,
part II," IBM J. Res. & Dev., vol. 3, pp. 282-287; July,
1959.
[55] R. J. Solomonoff, "An inductive inference machine," 1957 IRE
NATIONAL CONVENTION RECORD, pt. 2, pp. 56-62.
[56] C. D. Darlington, "The Evolution of Genetics," Basic Books,
Inc., New York, N. Y.; 1958.
[57] A. Newell and H. A. Simon, "The logic theory machine," [L].
[58] A. Newell, J. C. Shaw, and H. A. Simon, "Empirical explorations
of the logic theory machine," Proc. WJCC, pp. 218-230;
1957.
[59] H. Wang, "Toward mechanical mathematics," IBM J. Res. &
Dev., vol. 4, pp. 2-22; January, 1960.
[60] A. Newell, J. C. Shaw and, H. A. Simon, "Elements of a theory
of human problem solving," Psych. Rev. vol. 65, p. 151; March,
1958.
[61] M. Davis and H. Putnam, "A computing procedure for quantification
theorv," J. ACM, vol. 7 pp. 201-215; July, 1960.
[62] H. Gelernter and N. Rochester, "Intelligent behavior in prob
lem-solving machines," IBM J. Res. & Dez., vol. 2, p. 336 ff.;
October, 1958.
[63] C. E. Shannon, "Game-playing machines," J. Franklin Inst.,
vol. 206, pp. 447-453; December, 1955.
[64] A. Newell and F. Tonge, "Introduction to IPL-V," Commun.
A CM, vol. 3; April, 1960.
[65] S. Golumb, "A mathematical theory of discrete classification,"
[H].
[66] J. Wozencraft and M. Horstein, "Coding for two-way channels,"
[H].
[67] J. Slagle, "A computer program for solving integration problems
in 'Freshman Calculus'," thesis in preparation, M.I.T., Cambridge,
Mass.
[681 A. Newell, J. C. Shaw, and H. A. Simoin, "Report on a general
problem-solving program," [E].
1961 29
PROCEEDINGS OF THE IRE
[691 H. L. Gelernter, "Realizationi of a geometry-proving machinie,"
[El.
[701 J. McCarthy, "Programs with common sense, ' [C].
[71] E. F. Moore, "On the shortest path through a mlaze," Proc.
Internati. Symp. on the Theory of Switching, Harvard tUniv., Cambridge,
Mass.; 1959.
[72] A. M. TuLrinig, "Cani a machine think?," [K].
[73] P. Roseuibloom, "Elements of Mathematical Logic," Dover
Publications, New York, N. Y.; 1951.
[74] H. Rogers, "Review of 'Godel's Proof' by Newmani anid Nagel,"
Am. Math. Mfonthly, vol. 67, p. 98; January, 1960.
[75] WV. S. McCulloch, "Through the den of the metaphysician,'
Brit. J. Phil. Science, vol. 5, pp. 18-34; 1954.
176] D. M. MacKay, "Operational aspects of intellect," [C].
[77] K. J. XW. Craik, "The Nature of Explanation," Cambridge tU-lliV.
Press, Cambridge, Eng.; 1952. Preface dated 1943.
[78] F. A. Hayek, "The Sensorv Order," Routledge and Kegan PauLl,
London, Eng.; 1952.
[791 G. Pask, "Physical analogues to the growth of a concept, ' [C3.
[80] A. N. Chomisky, "Syntactic Structures," Mouton, The HaguLe;
1957.
[81] N. Chomsky and G. A. Miller, "Finite state languages," Inform.
and Control, vol. 1, pp. 91-112; May, 1958.
[82] N. Rochester, et al., "Tests on a cell assembly theory of the
action of the brain, using a large digital computer," [L].
[83] J. Y. Lettvin, H. Matuirania, W. S. McCulloch, and XV. Pitts,
"What the frog's eye tells the frog's brain," PROC. IRE, vol. 47,
pp. 1940-1951; November, 1959.
[84] S. Gorn, "On the mechanical simulation of learning and habitforming,"
Inform. and Control, vol. 2, pp. 226-259; September,
1959.
[85] M. Shubik, "Games, decisions and industrial organization,"
M1anagement Science vol. 6, pp. 455-474; July, 1960.
[86] J. S. Bruner, J. Goodnow, and G. Austin, "A Study of Thinking,"
John WViley and Sons, Inc., New York, N. Y.; 1956.
[871 G. Polya, "How to Solve It," Princeton Univ. Press, Princetonl,
N. J.; 1945. Also, "Induction and Analogy in Mathematics,"
and "Patterns of Plausible Inference," 2 vols., Priniceton Univ.
Press, Princeton, N. J.; 1954. (Available in paperback.)
[88] F. E. Hohn, S. Seshut, ancd D. D. Auifenikamp, "TIhe theory of
nets," IRE TRANS. ON ELECTRONIC COMPUTERS, vol. EC-6, pp.
154-161; September, 1957.
[89] D. M. MacKay, "The epistemological problem for autotmata,"
[B].
[901 1. J. Good, "Weight of evidence and false target probabilities,"
[H].
[91] A. Samuel, Letter to the Editor, Science, vol. 132, No. 3429;
September 16, 1960. (Incorrectly labelled vol. 131 oni cover.)
[92] B. G. Farley and WV. A. Clark, "Simiulation of self-organizinig
systemiis by digital computer," IRE TRANS. ON INFORMATION
rHEORY, vol. IT-4, pp. 76-84; September, 1954.
[93] H. XVaug, "Proving theorems by patteru recognitioni, I," [G].
[941 T. Kilburli, R. L. Grimsdale, and F. H. Sumllner, "Experimilenits
inl mnachine thinking and learninlg, " [El.
[95] N. Rashevskv, "Mathematical Biophysics," Dover Publicationis,
Inc., New York, N. Y., vol. 2; 1960.
Proceedinigs and collectionis containinlg more thani onie of the above
references:
[A] Proc. WJCC; March, 1955.
[B] "AAutomlata Studies," C. E. Shannon and J. McCarthy. Eds.
Princeton Univ. Press, Princeton, N. J.; 1956.
[C] Proc. Symp. on Mechanization of Thought Processes, Her Majestv's
Stationery Office, London, Eng.; 1959.
[Dl "Self-Organizing Systems," M. T. Yovitts and S. Canmeroni, Eds.,
Pergamoni Press, New York, N. Y.; 1960.
[E] Proc. Internatl. Conf. on Information Processing, UNESCO
House, Paris, France; 1959.
[Fl Proc. EJCC; December, 1959.
[G] Commun. A CI, vol. 3; April, 1960. (Preprinits of Conf. on
Symbol Manipulation Programs.)
[H] Fourth London Symp. on Information Theory, C. Cherry, Ed.,
to be published.
[J] Third London Symp. on Information Theory, C. Cherry, Ed.,
Acadlemiiic Press, Inc., New York, N. Y.; 1956.
[K] "The World of Mathematics,' 7Newnmani, Ed., Similoii anid
SchuLster, Ilnc., New York, N. Y., vol. 4; 1956.
[L] IRE TRANS. ON INFORMATION THrvORYx Vol. IF-2; Septeml1ber
1956.
January30