J. McCarthy, Dartmouth College
M. L. Minsky, Harvard University
N. Rochester, I.B.M. Corporation
C.E. Shannon, Bell Telephone Laboratories
August 31, 1955
We propose that a 2 month, 10 man study of artificial intelligence becarried out during the summer of 1956 at Dartmouth College in Hanover,New Hampshire. The study is to proceed on the basis of the conjecturethat every aspect of learning or any other feature of intelligence canin principle be so precisely described that a machine can be made tosimulate it. An attempt will be made to find how to make machines uselanguage, form abstractions and concepts, solve kinds of problems nowreserved for humans, and improve themselves. We think that asignificant advance can be made in one or more of these problems if acarefully selected group of scientists work on it together for asummer.
The following are some aspects of the artificial intelligenceproblem:
1Automatic Computers
If a machine can do a job, then anautomatic calculator can be programmed to simulate the machine. Thespeeds and memory capacities of present computers may be insufficientto simulate many of the higher functions of the human brain, but themajor obstacle is not lack of machine capacity, but our inability towrite programs taking full advantage of what we have.
2.How Can a Computer be Programmed to Use a Language
It may be speculated that a large part of human thought consists ofmanipulating words according to rules of reasoning and rules ofconjecture. From this point of view, forming a generalization consistsof admitting a new word and some rules whereby sentences containing itimply and are implied by others. This idea has never been veryprecisely formulated nor have examples been worked out.
3.Neuron Nets
How can a set of (hypothetical) neurons bearranged so as to form concepts. Considerable theoretical andexperimental work has been done on this problem by Uttley, Rashevskyand his group, Farley and Clark, Pitts and McCulloch, Minsky,Rochester and Holland, and others. Partial results have been obtainedbut the problem needs more theoretical work.
4.Theory of the Size of a Calculation
If we are given a well-defined problem (one for which itis possible to test mechanically whether or not a proposed answer is avalid answer) one way of solving it is to try all possible answers inorder. This method is inefficient, and to exclude it one must havesome criterion for efficiency of calculation. Some consideration willshow that to get a measure of the efficiency of a calculation it isnecessary to have on hand a method of measuring the complexity ofcalculating devices which in turn can be done if one has a theory ofthe complexity of functions. Some partial results on this problem havebeen obtained by Shannon, and also by McCarthy.
5.Self-lmprovement
Probably a truly intelligent machine will carryout activities which may best be described as self-improvement. Someschemes for doing this have been proposed and are worth further study.It seems likely that this question can be studied abstractly as well.
6.Abstractions
A number of types of ``abstraction'' can be distinctlydefined and several others less distinctly. A direct attempt toclassify these and to describe machine methods of forming abstractionsfrom sensory and other data would seem worthwhile.
7.Randomness and Creativity
A fairly attractive and yet clearly incomplete conjectureis that the difference between creative thinking and unimaginativecompetent thinking lies in the injection of a some randomness. Therandomness must be guided by intuition to be efficient. In otherwords, the educated guess or the hunch include controlled randomnessin otherwise orderly thinking.
In addition to the above collectivelyformulated problems for study, we have asked the individuals takingpart to describe what they will work on. Statements by the fouroriginators of the project are attached.
We propose to organize thework of the group as follows.
Potential participants will be sentcopies of this proposal and asked ifthey would like to work on the artificial intelligence problem inthe group and if so what they would like to work on. The invitationswill be made by the organizing committee on the basis of its estimateof the individual's potential contribution to the work of the group.The members will circulate their previous work and their ideas for theproblems to be attacked during the months preceding the working periodof the group.
During the meeting there will be regular researchseminars and opportunity for the members to work individually and ininformal small groups.
The originators of this proposal are:
1.C. E. Shannon, Mathematician, Bell Telephone Laboratories. Shannon developedthe statistical theory of information, the application of propositionalcalculus to switching circuits, and has results on the efficientsynthesis of switching circuits, the design of machines that learn,cryptography, and the theory of Turing machines. He and J. McCarthyare co-editing an Annals of Mathematics Study on ``The Theory ofAutomata'' .
2.M. L. Minsky, Harvard Junior Fellow in Mathematics andNeurology. Minsky has built a machine for simulating learning by nervenets and has written a Princeton PhD thesis in mathematics entitled,``Neural Nets and the Brain Model Problem'' which includes results inlearning theory and the theory of random neural nets.
3.N. Rochester, Manager of Information Research, IBM Corporation, Poughkeepsie, NewYork. Rochester was concerned with the development of radar for sevenyears and computing machinery for seven years. He and another engineerwere jointly responsible for the design of the IBM Type 701 which is alarge scale automatic computer in wide use today. He worked out someof the automatic programming techniques which are in wide use todayand has been concerned with problems of how to getmachines to do tasks which previously could be done only by people.He has also worked on simulation of nerve nets with particularemphasis on using computers to test theories in neurophysiology.
4.J. McCarthy, Assistant Professor of Mathematics, DartmouthCollege. McCarthy has worked on a number of questions connected with themathematical nature of the thought process including the theory ofTuring machines, the speed of computers, the relation of a brain modelto its environment, and the use of languages by machines. Some resultsof this work are included in the forthcoming ``Annals Study'' edited byShannon and McCarthy. McCarthy's other work has been in the field ofdifferential equations.
The Rockefeller Foundation is being asked to provide financialsupport for the project on the following basis:
1. Salaries of $1200for each faculty level participant who is not being supported by hisown organization. It is expected, for example, that the participantsfrom Bell Laboratories and IBM Corporation will be supported by theseorganizations while those from Dartmouth and Harvard will requirefoundation support.
2. Salaries of $700 for up to two graduatestudents.
3. Railway fare for participants coming from a distance.
4. Rent for people who are simultaneously renting elsewhere.
5. Secretarial expenses of $650, $500 for a secretary and $150 forduplicating expenses.
6. Organization expenses of $200. (Includes expense of reproducing preliminary work by participants and travel necessary for organizationpurposes.
7. Expenses for two or threepeople visiting for a short time.
#& #Estimated Expenses6 salaries of 1200 & $72002 salaries of 700 & 14008 traveling and rent expenses averaging 300 & 2400Secretarial and organizational expense & 850Additional traveling expenses & 600Contingencies & 550&----& $13,500
I would like to devote my research to one or both of the topics listedbelow. While I hope to do so, it is possible that because of personalconsiderations I may not be able to attend for the entire two months.I, nevertheless, intend to be there for whatever time is possible.
1. Application of information theory concepts to computing machines andbrain models. A basic problem in information theory is that oftransmitting information reliably over a noisy channel. An analogousproblem in computing machines is that of reliable computing usingunreliable elements. This problem has been studies by von Neumann forSheffer stroke elements and by Shannon and Moore for relays; but thereare still many open questions. The problem for several elements, thedevelopment of concepts similar to channel capacity, the sharperanalysis of upper and lower bounds on the required redundancy, etc.are among the important issues. Another question deals with the theoryof information networks where information flows in many closed loops(as contrasted with the simple one-way channel usually considered incommunication theory). Questions of delay become very important in theclosed loop case, and a whole new approach seems necessary. This wouldprobably involve concepts such as partial entropies when a part of thepast history of a message ensemble is known.
2. The matched environment - brain model approach to automata. Ingeneral a machine or animal can only adapt to or operate in a limitedclass of environments. Even the complex human brain first adapts tothe simpler aspects of its environment, and gradually builds up to themore complex features. I propose to study the synthesis of brainmodels by the parallel development of a series of matched(theoretical) environments and corresponding brain models which adaptto them. The emphasis here is on clarifying the environmental model,and representing it as a mathematical structure. Often in discussingmechanized intelligence, we think of machines performing the mostadvanced human thought activities-proving theorems, writing music,or playing chess. I am proposing here to start at the simple and whenthe environment is neither hostile (merely indifferent) nor complex,and to work up through a series of easy stages in the direction ofthese advanced activities.
It is not difficult to design a machine which exhibits the followingtype of learning. The machine is provided with input and outputchannels and an internal means of providing varied output responses toinputs in such a way that the machine may be ``trained'' by a ``trial anderror'' process to acquire one of a range of input-output functions.Such a machine, when placed in an appropriate environment and given acriterior of ``success'' or ``failure'' can be trained to exhibit``goal-seeking'' behavior. Unless the machine is provided with, or isable to develop, a way of abstracting sensory material, it canprogress through a complicated environment only through painfully slowsteps, and in general will not reach a high level of behavior.
Now let the criterion of success be not merely the appearance of a desiredactivity pattern at the output channel of the machine, but rather theperformance of a given manipulation in a given environment. Then incertain ways the motor situation appears to be a dual of the sensorysituation, and progress can be reasonably fast only if the machine isequally capable of assembling an ensemble of ``motor abstractions'' relating its output activity to changes in the environment. Such``motor abstractions'' can be valuable only if they relate to changes inthe environment which can be detected by the machine as changes in thesensory situation, i.e., if they are related, through the structureof the environrnent, to the sensory abstractions that the machine isusing.
I have been studying such systems for some time and feel thatif a machine can be designed in which the sensory and motorabstractions, as they are formed, can be made to satisfy certainrelations, a high order of behavior may result. These relationsinvolve pairing, motor abstractions with sensory abstractions in such a way as to produce new sensory situations representing thechanges in the environment that might be expected if the correspondingmotor act actually took place.
The important result that would belooked for would be that the machine would tend to build up withinitself an abstract model of the environment in which it is placed. Ifit were given a problem, it could first explore solutions within theinternal abstract model of the environment and then attempt externalexperiments. Because of this preliminary internal study, theseexternal experiments would appear to be rather clever, and thebehavior would have to be regarded as rather ``imaginative''
A verytentative proposal of how this might be done is described in mydissertation and I intend to do further work in this direction. I hopethat by summer 1956 I wi11 have a model of such a machine fairlyclose to the stage of programming in a computer.
Originality in Machine Performance
In writing a program for anautomatic calculator, one ordinarily provides the machine with a setof rules to cover each contingency which may arise and confront themachine. One expects the machine to follow this set of rules slavishlyand to exhibit no originality or common sense. Furthermore one isannoyed only at himself when the machine gets confused because therules he has provided for the machine are slightly contradictory.Finally, in writing programs for machines, one sometimes must go atproblems in a very laborious manner whereas, if the machine had just alittle intuition or could make reasonable guesses, the solution of theproblem could be quite direct. This paper describes a conjecture as tohow to make a machine behave in a somewhat more sophisticated mannerin the general area suggested above. The paper discusses a problem onwhich I have been working sporadically for about five years and whichI wish to pursue further in the Artificial Intelligence Project nextsummer.
The Process of Invention or Discovery
Living in theenvironment of our culture provides us with procedures for solvingmany problems. Just how these procedures work is not yet clear but Ishall discuss this aspect of the problem in terms of a model suggestedby Craik
. He suggeststhat mental action consists basically ofconstructing little engines inside the brain which can simulate and thuspredict abstractions relating to environment. Thus the solution of aproblem which one already understands is done as follows:
The prediction willcorrespond to the goal if living in the environment of his culture hasprovided the individual with the solution to the problem. Regarding theindividual as a stored program calculator, the program contains rulesto cover this particular contingency.
For a more complex situation therules might be more complicated. The rules might call for testing eachof a set of possible actions to determine which provided the solution.A still more complex set of rules might provide for uncertainty aboutthe environment, as for example in playing tic tac toe one must notonly consider his next move but the various possible moves of theenvironment (his opponent).
Now consider a problem for which noindividual in the culture has a solution and which has resistedefforts at solution. This might be a typical current unsolvedscientific problem. The individual might try to solve it and find thatevery reasonable action led to failure. In other words the storedprogram contains rules for the solution of this problem but the rulesare slightly wrong.
In order to solve this problem the individual willhave to do something which is unreasonable or unexpected as judged bythe heritage of wisdom accumulated by the culture. He could get such behavior by trying differentthings at random but such an approach would usually be tooinefficient. There are usually too many possible courses of action ofwhich only a tiny fraction are acceptable. The individual needs ahunch, something unexpected but not altogether reasonable. Someproblems, often those which are fairly new and have not resisted mucheffort, need just a little randomness. Others, often those which havelong resisted solution, need a really bizarre deviation fromtraditional methods. A problem whose solution requires originalitycould yield to a method of solution which involved randomness.
In terms of Craik's S model, the engine which should simulate theenvironment at first fails to simulate correctly. Therefore, it isnecessary to try various modifications of the engine until one isfound that makes it do what is needed.
Instead of describing theproblem in terms of an individual in his culture it could have beendescribed in terms of the learning of an immature individual. When theindividual is presented with a problem outside the scope of hisexperience he must surmount it in a similar manner.
So far the nearestpractical approach using this method in machine solution of problemsis an extension of the Monte Carlo method. In the usual problem whichis appropriate for Monte Carlo there is a situation which is grosslymisunderstood and which has too many possible factors and one isunable to decide which factors to ignore in working out analyticalsolution. So the mathematician has the machine making a few thousandrandom experiments. The results of these experiments provide a roughguess as to what the answer may be. The extension of the Monte CarloMethod is to use these results as a guide to determine what to neglectin order to simplify the problem enough to obtain an approximateanalytical solution.
It might be asked why the method should include randomness. Whyshouldn't the method be to try each possibility in the order of theprobability that the present state of knowledge would predict for itssuccess? For the scientist surrounded by the environment provided byhis culture, it may be that one scientist alone would be unlikely tosolve the problem in his life so the efforts of many are needed. Ifthey use randomness they could all work at once on it without completeduplication of effort. If they used system they would requireimpossibly detailed communication. For the individual maturing incompetition with other individuals the requirements of mixed strategy(using game theory terminology) favor randomness. For the machine,randomness will probably be needed to overcome the shortsightednessand prejudices of the programmer. While the necessity for randomnesshas clearly not been proven, there is much evidence in its favor.
The Machine With Randomness
In order to write a program to make anautomatic calculator use originality it will not do to introducerandomness without using forsight. If, for example, one wrote aprogram so that once in every 10,000 steps the calculator generateda random number and executed it as an instruction the result wouldprobably be chaos. Then after a certain amount of chaos the machinewould probably try something forbidden or execute a stop instructionand the experiment would be over.
Two approaches, however, appear tobe reasonable. One of these is to find how the brain manages to dothis sort of thing and copy it. The other is to take some class ofreal problems which require originality in their solution and attemptto find a way to write a program to solve them on an automaticcalculator. Either of these approaches would probably eventuallysucceed. However, it is not clear which would be quicker nor how manyyears or generations it would take. Most of my effort along theselines has so far been on the former approach because I felt that it would be best to master all relevant scientific knowledge inorder to work on such a hard problem, and I already was quite aware ofthe current state of calculators and the art of programming them.
The control mechanism of the brain is clearly very different from thecontrol mechanism in today's calculators. One symptom of thedifference is the manner of failure. A failure of a calculatorcharacteristically produces something quite unreasonable. An error inmemory or in data transmission is as likely to be in the mostsignificant digit as in the least. An error in control can do nearlyanything. It might execute the wrong instruction or operate a wronginput-output unit. On the other hand human errors in speech are apt toresult in statements which almost make sense (consider someone who isalmost asleep, slightly drunk, or slightly feverish). Perhaps themechanism of the brain is such that a slight error in reasoningintroduces randomness in just the right way. Perhaps the mechanismthat controls serial order in behavior
guides the random factor so asto improve the efficiency of imaginative processes over purerandomness.
Some work has been done on simulating neuron nets on ourautomatic calculator. One purpose was to see if it would be therebypossible to introduce randomness in an appropriate fashion. It seemsto have turned out that there are too many unknown links between theactivity of neurons and problem solving for this approach to workquite yet. The results have cast some light on the behavior of netsand neurons, but have not yielded a way to solve problems requiringoriginality.
An important aspect of this work has been an effort to make themachine form and manipulate concepts, abstractions, generalizations,and names. An attempt was made to test a theory
of how the brain doesit. The first set of experiments occasioned a revision of certaindetails of the theory. The second set of experiments is now inprogress. By next summer this work will be finished and a final reportwill have been written.
My program is to try next to write a program to solve problems which are members of some limited class of problemsthat require originality in their solution. It is too early to predictjust what stage I will be in next summer, or just; how I will thendefine the immediate problem. However, the underlying problem which isdescribed in this paper is what I intend to pursue. In a singlesentence the problem is: how can I make a machine which will exhibitoriginality in its solution of problems?
1. K.J.W. Craik, The Nature of Explanation, CambridgeUniversity Press, 1943 (reprinted 1952), p. 92.
2. K.S. Lashley, ``The Problem of Serial Order in Behavior'', inCerebral Mechanism in Behavior, the Hixon Symposium, editedby L.A. Jeffress, John Wiley & Sons, New York, pp. 112-146, 1951.
3. D. O. Hebb, The Organization of Behavior, John Wiley & Sons, NewYork, 1949
During next year and during the Summer Research Project on ArtificialIntelligence, I propose to study the relation of language tointelligence. It seems clear that the direct application of trial anderror methods to the relation between sensory data and motor activitywill not lead to any very complicated behavior. Rather it isnecessary for the trial and error methods to be applied at a higherlevel of abstraction. The human mind apparently uses language as itsmeans of handling complicated phenomena. The trial and error processesat a higher level frequently take the form of formulating conjecturesand testing them. The English language has a number of properties whichevery formal language described so far lacks.
1. Arguments in English supplemented by informal mathematics can beconcise.
2. English isuniversal in the sense that it can set up any other language withinEnglish and then use that language where it is appropriate.
3. Theuser of English can refer to himself in it and formulate statementsregarding his progress in solving the problem he is working on.
4. Inaddition to rules of proof, English if completely formulated wouldhave rules of conjecture.
The logical languages so far formulated haveeither been instruction lists to make computers carry out calculationsspecified in advance or else formalization of parts of mathematics.The latter have been constructed so as:
1. to be easily described ininformal mathematics,
2. to allow translation of statements frominformal mathematics into the language,
3. to make it easy to argue about whether proofs of (???)
No attempt has been made to make proofs in artificial languagesas short as informal proofs. It therefore seems to be desirable toattempt to construct an artificial language which a computer can beprogrammed to use on problems requiring conjecture and self-reference.It should correspond to English in the sense that short Englishstatements about the given subject matter should have shortcorrespondents in the language and so should short arguments orconjectural arguments. I hope to try to formulate a language havingthese properties and in addition to contain the notions of physicalobject, event, etc., with the hope that using this language it willbe possible to program a machine to learn to play games well and doother tasks .
The purpose of the list is to let those on it know who is interestedin receiving documents on the problem. The people on the 1ist wlllreceive copies of the report of the Dartmouth Summer Project onArtificial Intelligence. [1996 note: There was no report.]
The list consists of people who particlpated in or visited theDartmouth Summer Research Project on Artificlal Intelligence, or whoare known to be interested in the subject. It is being sent to thepeople on the 1ist and to a few others.
For the present purpose the artificial intelligence problem is takento be that of making a machine behave in ways that would be calledintelligent if a human were so behaving.
A revised list will be issued soon, so that anyone else interested ingetting on the list or anyone who wishes to change his address on itshould write to:
1996 note: Not all of these people came to theDartmouth conference. They were people we thoughtmight be interested in Artificial Intelligence.
The list consists of:
Adelson, Marvin
Hughes Aircraft Company
Airport Station, Los Angeles, CA
Ashby, W. R.
Barnwood House
Gloucester, England
Backus, John
IBM Corporation
590 Madison Avenue
New York, NY
Bernstein, Alex
IBM Corporation
590 Madison Avenue
New York, NY
Bigelow, J. H.
Institute for Advanced Studies
Princeton, NJ
Elias, Peter
R. L. E., MIT
Cambridge, MA
Duda, W. L.
IBM Research Laboratory
Poughkeepsie, NY
Davies, Paul M.
1317 C. 18th Street
Los Angeles, CA.
Fano, R. M.
R. L. E., MIT
Cambridge, MA
Farley, B. G.
324 Park Avenue
Arlington, MA.
Galanter, E. H.
University of Pennsylvania
Philadelphia, PA
Gelernter, Herbert
IBM Research
Poughkeepsie, NY
Glashow, Harvey A.
1102 Olivia Street
Ann Arbor, MI.
Goertzal, Herbert
330 West 11th Street
New York, New York
Hagelbarger, D.
Bell Telephone Laboratories
Murray Hill, NJ
Miller, George A.
Memorial Hall
Harvard University
Cambridge, MA.
Harmon, Leon D.
Bell Telephone Laboratories
Murray Hill, NJ
Holland, John H.
E. R. I.
University of Michigan
Ann Arbor, MI
Holt, Anatol
7358 Rural Lane
Philadelphia, PA
Kautz, William H.
Stanford Research Institute
Menlo Park, CA
Luce, R. D.
427 West 117th Street
New York, NY
MacKay, Donald
Department of Physics
University of London
London, WC2, England
McCarthy, John
Dartmouth College
Hanover, NH
McCulloch, Warren S.
R.L.E., M.I.T.
Cambridge, MA
Melzak, Z. A.
Mathematics Department
University of Michigan
Ann Arbor, MI
Minsky, M. L.
112 Newbury Street
Boston, MA
More, Trenchard
Department of Electrical Engineering
MIT
Cambridge, MA
Nash, John
Institute for Advanced Studies
Princeton, NJ
Newell, Allen
Department of Industrial Administration
Carnegie Institute of Technology
Pittsburgh, PA
Robinson, Abraham
Department of Mathematics
University of Toronto
Toronto, Ontario, Canada
Rochester, Nathaniel
Engineering Research Laboratory
IBM Corporation
Poughkeepsie, NY
Rogers, Hartley, Jr.
Department of Mathematics
MIT
Cambridge, MA.
Rosenblith, Walter
R.L.E., M.I.T.
Cambridge, MA.
Rothstein, Jerome
21 East Bergen Place
Red Bank, NJ
Sayre, David
IBM Corporation
590 Madison Avenue
New York, NY
Schorr-Kon, J.J.
C-380 Lincoln Laboratory, MIT
Lexington, MA
Shapley, L.
Rand Corporation
1700 Main Street
Santa Monica, CA
Schutzenberger, M.P.
R.L.E., M.I.T.
Cambridge, MA
Selfridge, O. G.
Lincoln Laboratory, M.I.T.
Lexington, MA
Shannon, C. E.
R.L.E., M.I.T.
Cambridge, MA
Shapiro, Norman
Rand Corporation
1700 Main Street
Santa Monica, CA
Simon, Herbert A.
Department of Industrial Administration
Carnegie Institute of Technology
Pittsburgh, PA
Solomonoff, Raymond J.
Technical Research Group
17 Union Square West
New York, NY
Steele, J. E., Capt. USAF
Area B., Box 8698
Wright-Patterson AFB
Ohio
Webster, Frederick
62 Coolidge Avenue
Cambridge, MA
Moore, E. F.
Bell Telephone Laboratory
Murray Hill, NJ
Kemeny, John G.
Dartmouth College
Hanover, NH
John McCarthy
Wed Apr 3 19:48:31 PST 1996
[8]ページ先頭