我们提议于1956年夏季在新罕布什尔州汉诺威的达特茅斯学院开展一项为期两个月、由10人参加的人工智能研究。该研究将基于这样一个设想展开：学习的各个方面，以及智能所具有的其他任何特征，原则上都可以被如此精确地描述，以至于能够制造出一台机器对其进行模拟。 研究将尝试探索如何使机器使用语言、形成抽象概念与一般概念、解决目前只有人类才能处理的各种问题，以及实现自我改进。 我们认为，如果由一组选拔出来的科学家在一个夏季期间共同开展研究，那么在上述一个或多个问题上都可能取得具有重要意义的进展。

We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.

以下是人工智能问题的若干方面：

1．自动计算机

如果一台机器能够完成某项工作，那么就可以对自动计算机进行编程，使其模拟这台机器的工作。现有计算机的运算速度和存储容量可能不足以模拟人类大脑的许多高级功能，但目前面临的主要障碍并非机器能力不足，而是我们还无法编写出能够充分利用现有计算机能力的程序。

The following are some aspects of the artificial intelligence problem:

1. Automatic Computers

If a machine can do a job, then an automatic calculator can be programmed to simulate the machine. The speeds and memory capacities of present computers may be insufficient to simulate many of the higher functions of the human brain, but the major obstacle is not lack of machine capacity, but our inability to write programs taking full advantage of what we have.

2．如何对计算机进行编程，使其使用语言

可以推测，人类思维的很大一部分是按照推理规则和猜想规则对词语进行操作。从这一观点来看，形成一个概括，就是引入一个新的词语，并规定一些规则，使包含该词语的句子能够蕴含其他句子，同时也能够由其他句子反推得出。这一思想从未得到过十分精确的表述，也没有形成经过具体推演的实例。

2. How Can a Computer be Programmed to Use a Language

It maybe speculated that a large part of human thought consists of manipulating words according to rules of reasoning and rules of conjecture. From this point of view, forming a generalization consists of admitting a new word and some rules whereby sentences containing it imply and are implied by others. This idea has never been very precisely formulated nor have examples been worked out.

3．神经网络

如何将一组（假设的）神经元组织起来，使其能够形成概念。Uttley、Rashevsky及其研究团队、Farley和Clark、Pitts和McCulloch、Minsky、Rochester和Holland等人已经针对这一问题开展了大量理论与实验研究。目前已经取得了一些阶段性成果，但这一问题仍需要进一步的理论研究。

3. Neuron Nets

How can a set of (hypothetical) neurons be arranged so as to form concepts. Considerable theoretical and experimental work has been done on this problem by Uttley, Rashevsky and his group, Farley and Clark, Pitts and McCulloch, Minsky, Rochester and Holland, and others. Partial results have been obtained but the problem needs more theoretical work.

4．计算规模理论

如果给定一个定义明确的问题（即能够通过机械方式检验一个给定答案是否为有效答案的问题），一种解决方法是依次尝试所有可能的答案。这种方法效率低下，因此若要排除这种做法，就必须建立某种计算效率的判定标准。经过一定的考察可以发现，要衡量计算的效率，就必须首先具备一种衡量计算装置复杂程度的方法，而这又要求建立函数复杂度的理论。关于这一问题，Shannon以及McCarthy都已经取得了一些阶段性的研究成果。

4. Theory of the Size of a Calculation

If we are given a well-defined problem (one for which it is possible to test mechanically whether or not a proposed answer is a valid answer) one way of solving it is to try all possible answers in order. This method is inefficient, and to exclude it one must have some criterion for efficiency of calculation. Some consideration will show that to get a measure of the efficiency of a calculation it is necessary to have on hand a method of measuring the complexity of calculating devices which in turn can be done if one has a theory of the complexity of functions. Some partial results on this problem have been obtained by Shannon, and also by McCarthy.

5．自我改进

真正具有智能的机器，很可能会执行一些最适合被描述为“自我改进”的活动。人们已经提出了一些实现这一目标的方案，值得进一步研究。看来，这一问题同样可以从抽象层面加以研究。

5. Self-Improvement

Probably a truly intelligent machine will carry out activities which may best be described as self-improvement. Some schemes for doing this have been proposed and are worth further study. It seems likely that this question can be studied abstractly as well.

6．抽象

可以明确界定出若干种不同类型的“抽象”，此外还有一些类型则较难作出清晰的界定。直接尝试对这些抽象进行分类，并描述机器如何从感知数据及其他数据中形成抽象，似乎是一项值得开展的研究。

6. Abstractions

A number of types of ``abstraction'' can be distinctly defined and several others less distinctly. A direct attempt to classify these and to describe machine methods of forming abstractions from sensory and other data would seem worthwhile.

7．随机性与创造力

一个颇具吸引力、但显然并不完整的推测是：创造性思维与缺乏想象力但又胜任工作的思维之间的区别，在于前者引入了某种随机性。这种随机性必须受到直觉的引导，才能发挥作用。换言之，经过训练的猜测或灵感式的预感，在原本有序的思维过程中包含了一种受控的随机性。

7. Randomness and Creativity

A fairly attractive and yet clearly incomplete conjecture is that the difference between creative thinking and unimaginative competent thinking lies in the injection of a some randomness. The randomness must be guided by intuition to be efficient. In other words, the educated guess or the hunch include controlled randomness in otherwise orderly thinking.

除上述由全体成员共同提出的研究问题之外，我们还要求参与者分别说明各自计划开展的研究工作。项目四位发起人对此所作的说明附于文后。

我们拟按照以下方式组织研究小组的工作。

我们将向潜在参与者寄送本提案的副本，并询问他们是否愿意作为小组成员参与人工智能问题的研究；如果愿意，还需说明各自希望开展的研究方向。邀请名单将由组织委员会根据其对个人可能为小组研究工作作出贡献程度的评估来确定。在小组正式开展研究前的数月内，各成员将相互交流此前的研究成果，以及对于研究期间拟解决问题的设想。

会议期间将定期举行研究讨论会，同时为成员开展个人研究以及进行非正式小组讨论提供机会。

In addition to the above collectively formulated problems for study, we have asked the individuals taking part to describe what they will work on. Statements by the four originators of the project are attached.

We propose to organize the work of the group as follows.

Potential participants will be sent copies of this proposal and asked if they would like to work on the artificial intelligence problem in the group and if so what they would like to work on. The invitations will be made by the organizing committee on the basis of its estimate of the individual's potential contribution to the work of the group. The members will circulate their previous work and their ideas for the problems to be attacked during the months preceding the working period of the group.

During the meeting there will be regular research seminars and opportunity for the members to work individually and in informal small groups.

本提案的发起人如下：

1．C. E. 香农（C. E. Shannon） ，数学家，贝尔电话实验室。香农创立了信息的统计理论，研究了命题演算在开关电路中的应用，并在开关电路的高效综合、学习机器的设计、密码学以及图灵机理论等方面取得了研究成果。他与J. 麦卡锡共同担任《自动机理论》（“The Theory of Automata”）《数学年刊研究专论》（Annals of Mathematics Studies）的主编。

The originators of this proposal are:

1. C.E. Shannon, Mathematician, Bell Telephone Laboratories. Shannon developed the statistical theory of information, the application of propositional calculus to switching circuits, and has results on the efficient synthesis of switching circuits, the design of machines that learn, cryptography, and the theory of Turing machines. He and J. McCarthy are co-editing an Annals of Mathematics Study on ``The Theory of Automata'' .

2．M. L. 明斯基（M. L. Minsky） ，哈佛大学数学与神经学初级研究员。明斯基曾研制出一台用于模拟神经网络学习过程的机器，并在普林斯顿大学完成数学博士论文《神经网络与脑模型问题》（“Neural Nets and the Brain Model Problem”），其中包含关于学习理论和随机神经网络理论的研究成果。

2. M.L. Minsky , Harvard Junior Fellow in Mathematics and Neurology. Minsky has built a machine for simulating learning by nerve nets and has written a Princeton PhD thesis in mathematics entitled, ``Neural Nets and the Brain Model Problem ''which includes results in learning theory and the theory of random neural nets.

3．N. 罗切斯特（N. Rochester） ，IBM公司信息研究部经理，任职于纽约州波基普西。罗切斯特曾从事雷达研发七年，之后又从事计算机及计算设备研发七年。他与另一位工程师共同负责设计了IBM 701型计算机，这是一种大型自动计算机，如今已得到广泛应用。他还研究并发展了一些目前广泛使用的自动编程技术，并长期致力于研究如何使机器完成过去只有人类才能完成的任务。此外，他还开展了神经网络模拟方面的研究，尤其关注利用计算机检验神经生理学理论。

3. N. Rochester , Manager of Information Research, IBM Corporation, Poughkeepsie, New York. Rochester was concerned with the development of radar for seven years and computing machinery for seven years. He and another engineer were jointly responsible for the design of the IBM Type701 which is a large scale automatic computer in wide use today. He worked out some of the automatic programming techniques which are in wide use today and has been concerned with problems of how to get machines to do tasks which previously could be done only by people. He has also worked on simulation of nerve nets with particular emphasis on using computers to test theories in neurophysiology.

4．J. 麦卡锡（J. McCarthy） ，达特茅斯学院数学系助理教授。麦卡锡曾研究一系列与思维过程的数学本质有关的问题，包括图灵机理论、计算机的运行速度、脑模型与其环境之间的关系，以及机器对语言的使用等。其中部分研究成果收录于即将出版、由香农和麦卡锡共同编辑的《数学年刊研究专论》。麦卡锡的其他研究则集中于微分方程领域。

4. J. McCarthy , Assistant Professor of Mathematics, Dartmouth College. McCarthy has worked on a number of questions connected with the mathematical nature of the thought process including the theory of Turing machines, the speed of computers, the relation of a brain model to its environment, and the use of languages by machines. Some results of this work are included in the forthcoming ``Annals Study'' edited by Shannon and McCarthy. McCarthy's other work has been in the field of differential equations.

我希望将我的研究投入以下列出的一个或两个课题。尽管我希望能够如此开展研究，但由于个人原因，我可能无法参加完整的两个月研究活动。不过，无论实际情况如何，我都计划在可能的时间内尽量参加。

I would like to devote my research to one or both of the topics listed below.

While I hope to do so, it is possible that because of personal considerations I may not be able to attend for the entire two months. I, nevertheless, intend to be there for whatever time is possible.

1．信息论概念在计算机与脑模型中的应用。

信息论中的一个基本问题，是如何在有噪信道中可靠地传输信息。与之类似，计算机领域中的一个问题是，如何利用不可靠的元件实现可靠计算。冯·诺依曼曾针对谢弗函数元件研究过这一问题，香农和摩尔也曾对继电器进行过相关研究；但这一领域仍存在许多有待解决的问题。其中，多个元件情况下的问题、发展类似于信道容量的概念，以及对所需冗余度的上下界进行更精确的分析等，都是重要的研究课题。另一个问题涉及信息网络理论，即研究 信息如何在多个闭环中流动 ，而不是像通信理论通常研究的那样，仅在单向信道中传输。在闭环情况下，延迟问题变得尤为重要，因此似乎有必要采用一种全新的研究方法。这可能需要引入诸如“ 部分熵 ”之类的概念，即在已知消息集合部分过去历史的情况下，对其信息特征进行描述。

1.Application of information theory concepts to computing machines and brain models.

A basic problem in information theory is that of transmitting information reliably over a noisy channel. An analogous problem in computing machines is that of reliable computing using unreliable elements. This problem has been studies by von Neumann for Sheffer stroke elements and by Shannon and Moore for relays; but there are still many open questions. The problem for several elements, the development of concepts similar to channel capacity, the sharper analysis of upper and lower bounds on the required redundancy, etc. are among the important issues. Another question deals with the theory of information networks where information flows in many closed loops (as contrasted with the simple one-way channel usually considered in communication theory). Questions of delay become very important in the closed loop case, and a whole new approach seems necessary. This would probably involve concepts such as partial entropies when a part of the past history of a message ensemble is known.

2．匹配环境——脑模型的自动机研究方法。

一般而言，一台机器或一种动物只能适应或运行于有限类别的环境之中。即使是复杂的人类大脑，也必须首先适应环境中较为简单的方面，然后逐步发展到处理更为复杂的特征。我拟通过平行构建一系列相互匹配的（理论）环境及相应的脑模型来研究脑模型的综合，即让每一个脑模型都能够适应与之对应的环境。这里的重点在于明确环境模型，并将其表示为一种数学结构。在讨论机器智能时，我们往往会想到让机器执行人类最高级的思维活动——证明定理、创作音乐或下棋。而我在此提出的方法，是从简单的问题入手，在环境既不具有敌意（仅仅处于中性状态）又不复杂的条件下开始研究，然后通过一系列由易到难的阶段，逐步向这些高级活动的方向推进。

2. The matched environment - brain model approach to automata.

In general a machine or animal can only adapt to or operate in a limited class of environments. Even the complex human brain first adapts to the simpler aspects of its environment, and gradually builds up to the more complex features. I propose to study the synthesis of brain models by the parallel development of a series of matched(theoretical) environments and corresponding brain models which adapt to them. The emphasis here is on clarifying the environmental model, and representing it as a mathematical structure. Often in discussing mechanized intelligence, we think of machines performing the most advanced human thought activities-proving theorems, writing music, or playing chess. I am proposing here to start at the simple and when the environment is neither hostile (merely indifferent) nor complex, and to work up through a series of easy stages in the direction of these advanced activities.

设计一种能够表现出以下类型学习能力的机器并不困难。机器配备输入和输出通道，并具有一种内部机制，可以针对输入产生多种不同的输出响应，使机器能够通过“试错”过程进行“训练”，从而获得一系列 输入—输出函数 中的某一种。当这样的机器被置于适当的环境中，并被赋予某种“成功”或“失败”的判定标准后，就可以通过训练使其表现出“目标导向”的行为。除非为机器提供一种能够对感知材料进行抽象的方法，或者机器能够自行发展出这种方法，否则它在复杂环境中的进步只能通过缓慢而艰难的步骤逐步实现，而且通常无法达到较高层次的行为水平。

It is not difficult to design a machine which exhibits the following type of learning. The machine is provided with input and output channels and an internal means of providing varied output responses to inputs in such a way that the machine may be ``trained'' by a ``trial and error'' process to acquire one of a range of input-output functions . Such a machine, when placed in an appropriate environment and given a criterior of ``success'' or ``failure'' can be trained to exhibit ``goal-seeking'' behavior. Unless the machine is provided with, or is able to develop, a way of abstracting sensory material, it can progress through a complicated environment only through painfully slow steps, and in general will not reach a high level of behavior.

现在，假定成功的判定标准不再仅仅是机器输出通道中出现某种预期的活动模式，而是在特定环境中完成某种给定的操作。那么，从某些方面来看，运动情境似乎是感觉情境的一个对应面；而只有当机器同样能够建立一组“运动抽象”，将其输出活动与环境的变化联系起来时，其进展才可能较为迅速。这类“运动抽象”只有在它们所涉及的环境变化能够被机器感知为感觉情境的变化时才具有价值；也就是说，它们必须通过环境的结构，与机器正在使用的感觉抽象建立联系。

Now let the criterion of success be not merely the appearance of a desired activity pattern at the output channel of the machine, but rather the performance of a given manipulation in a given environment. Then in certain ways the motor situation appears to be a dual of the sensory situation, and progress can be reasonably fast only if the machine is equally 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 in the environment which can be detected by the machine as changes in the sensory situation, i.e., if they are related, through the structure of the environment, to the sensory abstractions that the machine is using.

我已经研究这类系统一段时间，并认为，如果能够设计出一种机器，使其在形成感觉抽象和运动抽象的过程中，能够使二者满足某些特定关系，那么机器便可能产生较高层次的行为。这些关系包括将运动抽象与感觉抽象相互配对，使机器能够产生新的感觉情境，而这些新的感觉情境代表了：如果相应的运动行为实际发生，环境可能随之产生的变化。

I have been studying such systems for some time and feel that if a machine can be designed in which the sensory and motor abstractions, as they are formed, can be made to satisfy certain relations, a high order of behavior may result. These relations involve pairing, motor abstractions with sensory abstractions in such a way as to produce new sensory situations representing the changes in the environment that might be expected if the corresponding motor act actually took place.

所要寻求的重要结果，是使机器能够在自身内部逐渐建立起一个关于其所处环境的抽象模型。如果向机器提出一个问题，它可以首先在这一内部环境抽象模型中探索解决方案，然后再尝试进行外部实验。由于事先进行了这种内部探索，这些外部实验将会显得相当巧妙，而这种行为也不得不被视为具有相当程度的“ 想象力 ”。

The important result that would be looked for would be that the machine would tend to build up within itself an abstract model of the environment in which it is placed. If it were given a problem, it could first explore solutions within the internal abstract model of the environment and then attempt external experiments. Because of this preliminary internal study, these external experiments would appear to be rather clever, and the behavior would have to be regarded as rather `` imaginative ''

关于如何实现这一目标，我的博士论文中提出了一种非常初步的设想，我计划沿着这一方向继续开展研究。我希望到1956年夏季，能够建立起这样一种机器的模型，并使其基本达到可以在计算机上进行编程实现的阶段。

A very tentative proposal of how this might be done is described in my dissertation and I intend to do further work in this direction. I hope that by summer 1956 Iwi11 have a model of such a machine fairly close to the stage of programming in a computer.

机器执行中的原创性

在为自动计算机编写程序时，通常需要为机器制定一套规则，以应对可能出现并需要机器处理的各种情况。人们通常期望机器严格按照这套规则执行，不表现出任何原创性或常识。此外，当机器因为所提供的规则之间存在轻微矛盾而陷入混乱时，人们往往只能责怪自己。最后，在为机器编写程序时，有时不得不采用非常繁琐的方法来解决问题；而如果机器具备一点直觉，或者能够作出合理的猜测，那么问题的解决过程就可能直接得多。本文提出了一种设想，探讨如何使机器在上述领域中以一种更为复杂的方式运行。本文讨论的是我断断续续研究了约五年的一个问题，我希望在明年夏季的人工智能项目中继续深入研究这一问题。

Originality in Machine Performance

Inwriting a program for an automatic calculator, one ordinarily provides the machine with a set of rules to cover each contingency which may arise and confront the machine. One expects the machine to follow this set of rules slavishly and to exhibit no originality or common sense. Furthermore one is annoyed only at himself when the machine gets confused because the rules he has provided for the machine are slightly contradictory. Finally, in writing programs for machines, one sometimes must go at problems in a very laborious manner whereas, if the machine had just a little intuition or could make reasonable guesses, the solution of the problem could be quite direct. This paper describes a conjecture as to how to make a machine behave in a somewhat more sophisticated manner in the general area suggested above. The paper discusses a problem on which I have been working sporadically for about five years and which I wish to pursue further in the Artificial Intelligence Project next summer.

发明或发现的过程

生活在我们自身的文化环境中，使我们掌握了许多解决问题的方法。这些方法究竟是如何发挥作用的，目前尚不清楚，但我将借助 Craik* 提出的一种模型来讨论这一问题的这一方面。他认为，心理活动的基本过程，是在大脑内部构建一些小型的“机器”，使其能够模拟并预测与环境相关的抽象过程。因此，对于一个我们已经理解的问题，其解决过程如下：

The Process of Invention or Discovery

Living in the environment of our culture provides us with procedures for solving many problems. Just how these procedures work is not yet clear but I shall discuss this aspect of the problem in terms of a model suggested by Craik*. He suggests that mental action consists basically of constructing little engines inside the brain which can simulate and thus predict abstractions relating to environment. Thus the solution of a problem which one already understands is done as follows:

*K. J. W. Craik，《解释的本质》（The Nature of Explanation），剑桥大学出版社，1943年（1952年重印），第92页。

*K.J.W. Craik, The Nature of Explanation, Cambridge University Press, 1943 (reprinted 1952) p92.

1．环境提供数据，并据此形成某些抽象。

2．这些抽象与某些内部习惯或驱动力共同提供：

2.1 以未来需要实现的状态，即目标，来界定问题。

2.2 为解决问题而提出的行动方案。

2.3 在大脑中激活与这一情境相对应的“机器”。

3．随后，该“机器”开始运行，预测这一环境情境以及拟采取的反应将导致何种结果。

4．如果预测结果与目标相符，个体便按照所指示的方式采取行动。

1. The environment provides data from which certain abstractions are formed.

2. The abstractions together with certain internal habits or drives provide:

2.1 A definition of a problem in terms of desired condition to be achieved in the future, a goal.

2.2 A suggested action to solve the problem.

2.3 Stimulation to arouse in the brain the engine which corresponds to this situation.

3. Then the engine operates to predict what this environmental situation and the proposed reaction will lead to.

4. If the prediction corresponds to the goal the individual proceeds to act as indicated.

如果个体在其所处的文化环境中生活时，已经获得了解决该问题的方法，那么预测结果就会与目标相符。若将个体视为一台存储程序计算机，那么其程序中便包含了应对这一特定情形的规则。

The prediction will correspond to the goal if living in the environment of his culture has provided the individual with the solution to the problem. Regarding the individual as a stored program calculator, the program contains rules to cover this particular contingency.

对于更加复杂的情境，其规则也可能更加复杂。这些规则可能要求对一组可能采取的行动逐一进行检验，以确定其中哪一种能够解决问题。更为复杂的一组规则还可能需要处理对环境状态的不确定性。例如，在玩井字棋时，不仅要考虑自己的下一步行动，还必须考虑环境（即对手）可能采取的各种行动。

For amore complex situation the rules might be more complicated. The rules might call for testing each of a set of possible actions to determine which provided the solution. A still more complex set of rules might provide for uncertainty about the environment, as for example in playing tic tac toe one must not only consider his next move but the various possible moves of the environment (his opponent).

现在，考虑这样一个问题：在某种文化环境中，没有任何人能够找到解决方案，而且该问题一直未能得到解决。这可能是一个典型的、目前尚未解决的科学问题。个体可能尝试解决这一问题，却发现每一种合理的行动最终都会导致失败。换言之，存储程序中虽然包含了解决这一问题的规则，但这些规则存在轻微的错误。

Now consider a problem for which no individual in the culture has a solution and which has resisted efforts at solution. This might be a typical current unsolved scientific problem. The individual might try to solve it and find that every reasonable action led to failure. In other words the stored program contains rules for the solution of this problem but the rules are slightly wrong.

为了解决这一问题，个体必须采取某种按照该文化积累下来的智慧传统来看并不合理或出人意料的做法。他可以通过随机尝试不同的方法来获得这种行为，但这种方式通常效率过低。通常存在太多可能的行动路径，而其中只有极少一部分是可行的。个体需要一种直觉，一种出乎意料但又并非完全不合理的判断。有些问题，尤其是那些较新且尚未经历大量尝试的问题，只需要引入少量随机性；另一些问题，则往往由于长期未能得到解决，需要对传统方法作出真正异乎寻常的偏离。那些需要原创性的解题问题，或许可以通过一种包含随机性的求解方法得到解决。

In order to solve this problem the individual will have to do something which is unreasonable or unexpected as judged by the heritage of wisdom accumulated by the culture. He could get such behavior by trying different things at random but such an approach would usually be too inefficient. There are usually too many possible courses of action of which only a tiny fraction are acceptable. The individual needs a hunch, something unexpected but not altogether reasonable. Some problems, often those which are fairly new and have not resisted much effort, need just a little randomness. Others, often those which have long resisted solution, need a really bizarre deviation from traditional methods. A problem whose solution requires originality could yield to a method of solution which involved randomness.

按照Craik的模型，原本应该用于模拟环境的“机器”最初无法正确地完成模拟。因此，有必要对这一“机器”进行各种修改，直到找到一种能够使其完成所需任务的形式。

不把问题描述为文化环境中的个体，也可以将其描述为一个尚未成熟的个体的学习过程。当个体面对一个超出自身经验范围的问题时，他也必须以类似的方式克服这一问题。

迄今为止，在利用这种方法通过机器解决问题方面，最接近实际应用的方法，是对蒙特卡罗方法的一种扩展。在通常适合采用蒙特卡罗方法的问题中，存在一种被严重误解的情境，其中包含过多可能影响结果的因素，以至于人们无法判断在进行解析求解时应该忽略哪些因素。因此，数学家让机器进行数千次随机实验。这些实验的结果可以为可能的答案提供一个粗略的估计。对蒙特卡罗方法的扩展，则是利用这些结果作为判断依据，确定哪些因素可以忽略，从而将问题简化到足以获得一个近似解析解的程度。

可能有人会问，为什么这种方法必须包含随机性。

为什么不能按照当前知识状态所预测的成功概率，从高到低依次尝试各种可能性？对于一个置身于自身文化环境中的科学家而言，可能一个人穷其一生也不太可能解决某个问题，因此需要许多人共同努力。如果采用随机方法，他们就可以同时开展研究，而不至于完全重复彼此的工作。如果采用系统化的方法，则需要进行极其详细的信息交流，这在实际上难以实现。对于一个在与其他个体竞争中逐渐成熟的个体而言，混合策略（借用博弈论的术语）也会使随机性成为一种有利因素。对于机器而言，随机性可能是克服程序员短视性和偏见所必需的。虽然随机性的必要性显然尚未得到证明，但已有大量证据支持这一观点。

In terms of Craik's model, the engine which should simulate the environment at first fails to simulate correctly. Therefore, it is necessary to try various modifications of the engine until one is found that makes it do what is needed.

Instead of describing the problem in terms of an individual in his culture it could have been described in terms of the learning of an immature individual. When the individual is presented with a problem outside the scope of his experience he must surmount it in a similar manner.

So far the nearest practical approach using this method in machine solution of problems is an extension of the Monte Carlo method. In the usual problem which is appropriate for Monte Carlo there is a situation which is grossly misunderstood and which has too many possible factors and one is unable to decide which factors to ignore in working out analytical solution. So the mathematician has the machine making a few thousand random experiments. The results of these experiments provide a rough guess as to what the answer may be. The extension of the Monte Carlo Method is to use these results as a guide to determine what to neglect in order to simplify the problem enough to obtain an approximate analytical solution.

It might be asked why the method should include randomness.

Why shouldn't the method be to try each possibility in the order of the probability that the present state of knowledge would predict for its success? For the scientist surrounded by the environment provided by his culture, it may be that one scientist alone would be unlikely to solve the problem in his life so the efforts of many are needed. If they use randomness they could all work at once on it without complete duplication of effort. If they used system they would require impossibly detailed communication. For the individual maturing in competition 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 shortsightedness and prejudices of the programmer. While the necessity for randomness has clearly not been proven, there is much evidence in its favor.

具有随机性的机器

为了编写一个程序，使自动计算机能够表现出 原创性 ，仅仅引入随机性而不加以预见性地利用，是行不通的。例如，如果编写一个程序，使计算机每运行10,000步就产生一个随机数，并将其作为一条指令执行，那么结果很可能会是一片混乱。经过一定程度的混乱之后，机器很可能会尝试执行某项被禁止的操作，或者执行停止指令，从而使实验就此结束。

The Machine With Randomness

In order to write a program to make an automatic calculator use originality it will not do to introduce randomness without using for sight. If, for example, one wrote a program so that once in every 10,000 steps the calculator generated a random number and executed it as an instruction the result would probably be chaos. Then after a certain amount of chaos the machine would probably try something forbidden or execute a stop instruction and the experiment would be over.

不过，有两种途径似乎是可行的。其中一种是研究大脑究竟如何完成这类事情，并对其进行模仿；另一种则是选取某一类现实问题，这类问题要求在解决过程中具有原创性，然后尝试找到一种方法，编写程序使自动计算机能够解决这些问题。无论采用哪一种途径，最终都有可能取得成功。然而，目前尚不清楚哪一种途径会更快，也不清楚需要经过多少年或多少代人的努力才能实现这一目标。到目前为止，我在这方面的大部分工作都集中于前一种途径，因为我认为，要解决这样一个困难的问题，最好先掌握所有相关的科学知识；与此同时，我对当时计算机的现状以及计算机编程技术的发展情况已经相当了解。

Two approaches, however, appear to be reasonable. One of these is to find how the brain manages to do this sort of thing and copy it. The other is to take some class of real problems which require originality in their solution and attempt to find a way to write a program to solve them on an automatic calculator. Either of these approaches would probably eventually succeed. However, it is not clear which would be quicker nor how many years or generations it would take. Most of my effort along these lines has so far been on the former approach because I felt that it would be best to master all relevant scientific knowledge in order to work on such a hard problem, and I already was quite aware of the current state of calculators and the art of programming them.

大脑的控制机制显然与当今计算机中的控制机制有着很大的不同。这种差异的一个表现，就是二者出错的方式不同。计算机一旦发生故障，其结果通常会产生一些完全不合理的东西。存储器或数据传输中的错误，出现在最高位和最低位上的可能性是相同的；而控制方面的错误则几乎可能导致任何结果，它可能执行错误的指令，也可能操作错误的输入—输出设备。另一方面，人类在说话时出现的错误，却往往会产生一些几乎说得通的语句（例如，一个人处于即将入睡、微醺或轻微发热的状态时）。或许，大脑的机制正是如此：推理过程中的轻微错误，会以恰到好处的方式引入随机性。又或许，控制行为序列的机制*会对这一随机因素加以引导，使想象过程的效率高于纯粹的随机过程。

The control mechanism of the brain is clearly very different from the control mechanism in today's calculators. One symptom of the difference is the manner of failure. A failure of a calculator characteristically produces something quite unreasonable. An error in memory or in data transmission is as likely to be in the most significant digit as in the least. An error in control can do nearly anything. It might execute the wrong instruction or operate a wrong input-output unit. On the other hand human errors in speech are apt to result in statements which almost make sense (consider someone who is almost asleep, slightly drunk, or slightly feverish). Perhaps the mechanism of the brain is such that a slight error in reasoning introduces randomness in just the right way. Perhaps the mechanism that controls serial order in behavior* guides the random factor so as to improve the efficiency of imaginative processes over pure randomness.

我们已经开展了一些利用自动计算机模拟神经网络的工作。其中一个目的，是考察能否通过这种方式以适当的形式引入随机性。结果似乎表明，目前神经元活动与问题求解之间仍存在太多尚未明确的联系，因此这一途径尚不足以发挥作用。这些研究结果虽然使我们对神经网络和神经元的行为有了进一步认识，但尚未找到一种能够解决需要原创性的问题的方法。

Some work has been done on simulating neuron nets on our automatic calculator. One purpose was to see if it would be thereby possible to introduce randomness in an appropriate fashion. It seems to have turned out that there are too many unknown links between the activity of neurons and problem solving for this approach to work quite yet. The results have cast some light on the behavior of nets and neurons, but have not yielded a way to solve problems requiring originality.

*K. S. 拉什利，《行为中的序列顺序问题》，载于 L. A. 杰弗里斯（L. A. Jeffress）编《行为的脑机制：希克森研讨会论文集》，约翰·威利父子出版公司，纽约，第112—146页，1951年。

*K.S. Lashley, ”The Problem of Serial Order in Behavior, in Cerebral Mechanisms in Behavior, the Hixon Symposium, Edited by L.A. Jeffress, John Wiley & Sons, New York, pp112-146;1951.

这项研究的一个重要方面，是尝试使机器形成并运用概念、抽象概念、概括以及名称。研究过程中曾尝试检验一种关于大脑如何完成这一过程的理论*。第一组实验促使我们对该理论的某些细节进行了修订。第二组实验目前正在进行之中。到明年夏季，这项工作将完成，并将撰写最终报告。

An important aspect of this work has been an effort to make the machine form and manipulate concepts,abstractions,generalizations, and names. An attempt was made to test a theory* of how the brain does it. The first set of experiments occasioned a revision of certain details of the theory. The second set of experiments is now in progress. By next summer this work will be finished and a final report will have been written.

我的下一步计划，是尝试编写一种程序，使其能够解决某一有限类别中的问题，而这类问题的解决需要具有原创性。现在还无法预测到明年夏季时我将达到什么阶段，也无法确定届时我会如何界定眼前需要解决的具体问题。不过，本文所描述的这一根本问题，正是我计划继续研究的方向。用一句话概括，这个问题就是：我怎样才能制造出一台机器，使其在解决问题时表现出原创性？

My program is to try next to write a program to solve problems which are members of some limited class of problems that require originality in their solution. It is too early to predict just what stage I will be in next summer, or just how I will then define the immediate problem. However, the underlying problem which is described in this paper is what I intend to pursue. In a single sentence the problem is: how can I make a machine which will exhibit originality in its solution of problems?

*D. O. 赫布，《行为的组织》（The Organization of Behavior），约翰·威利父子出版公司，纽约，1949年。

*D.O. Hebb, The Organization of Behavior, John Wiley & Sons, New York,1949

在下一年度以及人工智能夏季研究项目期间，我拟研究 语言与智能之间的关系 。显然，直接将试错方法应用于感觉数据与运动活动之间的关系，并不能产生十分复杂的行为。相反，试错方法必须应用于更高层次的抽象。在处理复杂现象时，人类思维显然以语言作为其工具。在更高层次上，试错过程通常表现为提出猜想并对其进行检验。英语具有若干迄今所描述的形式语言均不具备的性质。

During next year and during the Summer Research Project on Artificial Intelligence, I propose to study the relation of language to intelligence. It seems clear that the direct application of trial and error methods to the relation between sensory data and motor activity will not lead to any very complicated behavior. Rather it is necessary for the trial and error methods to be applied at a higher level of abstraction. The human mind apparently uses language as its means of handling complicated phenomena. The trial and error processes at a higher level frequently take the form of formulating conjectures and testing them. The English language has a number of properties which every formal language described so far lacks.

1. Arguments in English supplemented by informal mathematics can be concise.

2. English is universal in the sense that it can set up any other language within English and then use that language where it is appropriate.

3. The user of English can refer to himself in it and formulate statements regarding his progress in solving the problem he is working on.

4. In addition to rules of proof, English if completely formulated would have rules of conjecture .

迄今为止已经构建的逻辑语言，要么是用于向计算机下达指令、使其执行预先规定的计算的指令表，要么是对数学某些部分进行形式化的语言。后者的构建方式是为了：

The logical languages so far formulated have either been instruction lists to make computers carry out calculations specified in advance or else formalization of parts of mathematics. The latter have been constructed so as:

1. to be easily described in informal mathematics,

2. to allow translation of statements from informal mathematics into the language,

3. to make it easy to argue about whether proofs of certain classes of propositions exist.

迄今为止，还没有人尝试使 人工语言 中的证明像非形式化证明那样简短。因此，似乎有必要尝试构造一种人工语言，使计算机能够被编程为利用这种语言处理需要 猜想和自我指涉 的问题。这种语言应当与英语相对应，即关于给定主题的简短英语陈述，在这种语言中也应有相应的简短表达；简短的论证或猜想性论证也应如此。我希望尝试构建一种具有这些性质的语言，并且在其中进一步包含物理对象、事件等概念，以期利用这种语言编写程序，使机器能够学会出色地下棋以及完成其他任务。

No attempt has been made to make proofs in artificial languages as short as informal proofs. It therefore seems to be desirable to attempt to construct an artificial language which a computer can be programmed to use on problems requiring conjecture and self-reference . It should correspond to English in the sense that short English statements about the given subject matter should have short correspondents in the language and so should short arguments or conjectural arguments. I hope to try to formulate a language having these properties and in addition to contain the notions of physical object, event, etc., with the hope that using this language it will be possible to program a machine to learn to play games well and do other tasks.

1956年3月6日

达特茅斯人工智能夏季研究项目计划

A. 纽厄尔（A. Newell）

H. A. 西蒙（H. A. Simon）

March 6, 1956

Plans for the Dartmouth summer research project on artificial intelligence

A. Newell and H.A. Simon

由兰德公司（RAND Corporation）的 A. 纽厄尔（A. Newell）和 J. C. 肖（J. C. Shaw），以及卡内基理工学院（Carnegie Institute of Technology）的 H. A. 西蒙（H. A. Simon）组成的研究小组，目前正在对复杂信息处理领域开展一系列探索性研究。其中相当一部分工作属于人工智能研究的范畴。目前，我们正在探索四个方面。

A research group composed of A. Newell and J.C. Shaw of the RAND Corporation，and H.A. Simon of Carnegie Institute of Technology is currently engaged in a series of forays into the area of complex information processing. A large part of this activity comes under the heading of artificial intelligence. At present we are exploring in four areas.

1.国际象棋机器
能够下棋并学习下棋的机器，正在迅速成为理解复杂信息处理的“经典”问题。我们已经为研究国际象棋机器奠定了相当多的基础：我们已经形成了一些理论上的构想；为了探索编程问题，我们已经对约翰尼亚克（Johnniac）进行了编程，使其能够执行符合规则的着法。约翰尼亚克是兰德公司的一台高速计算机，属于普林斯顿型机器。最后，我们目前正在为约翰尼亚克开发一种“国际象棋语言”（更准确地说，是一种解释性代码），使国际象棋程序能够被方便地编写，并成为迈向机器学习和自我编程的第一步。我们目前正在编写这一解释系统，但还无法确定到今年夏季能够取得多大进展。

1. Chess Machine.
Machines that play and learn chess are rapidly becoming the “classical” problem for understanding complex information processing. We have laid a considerable amount of groundwork for the study of chess machines: we have some theoretical formulation; in order to explore the programming problem, we have Programmed the Johnniac (one of RAND’s high speed computers, a Princeton-type machine) for legal moves. Finally, we are now developing a “chess language” for Johniac (more technically, an interpretive code) which permits easy coding of chess programs and is a first step towards learning and self-programming. We are currently coding this interpretive scheme, but do not know what progress we will have trade by summer.

*A. 纽厄尔，《国际象棋机器：通过适应处理复杂任务的一个实例》，载于《西部联合计算机会议论文集》，1955年3月。

*A. Newell, “The chess machine: an example of dealing with a complex task by adaptation” Proc. Western Joint Computer Conference, March, 1955

2.数学机器
数学活动无疑属于复杂信息处理的领域。其中最为有趣的部分，远非一般的数学计算，而是定理的证明以及那些尚无已知算法的问题的求解。当然，现有的计算机最初就是为了执行数学家在“进行数学活动”过程中所发现的各种算法而构建的。我们已经在设计一种能够发现命题演算证明的机器方面取得了一定程度的成功。这种机器并不使用真值表或典范形式，而是采用搜索程序，其工作方式至少在总体上与人类处理同类问题时所采用的方法相似。这一机器的设计方案已经足够简单，以至于可以由人来模拟其运行过程。我们还构建了一种与“国际象棋语言”类似的“逻辑语言”或解释性代码，目前正在为约翰尼亚克编写相应的程序。我们的计划是将这一工作进一步扩展到学习、自我编程，以及自主选择需要证明的定理（即创新）。同样，我们目前还无法确定到今年夏季能够取得多大进展。

2.Mathematics Machines.
doing mathematics is certainly an area of complex information processing. By far the most interesting part concerns the proving of theorems and the solving of problems where no algorithms is known. Current computers, of course, were built to carry out all the algorithms that mathematicians have found by “doing mathematics”. We have achieved a degree of success in specifying a machine that will discover proofs in the propositional calculus, not by using truth tables or canonical forms, but by employing search procedures the appear at least grossly similar to those used by humans in dealing with the same problems. The specification of this machine has proven simple enough to allow it to be simulated by a human. We have also constructed a “logic language” or interpretive code similar to the chess language, and are currently programming this for the Johnniac. It is our intention to extend this effort to learning, self-programming and the self-selection of theorems to be proved (innovation). Again, we do not know what progress we will make by summer.

3.学习理论
显然，任何能够执行人类功能的机器，只需改变观察问题的角度，就可以被视为人类行为的一个模型。我们一直在研究心理学实验中关于人类和动物学习的大量数据，希望据此设计出一种能够以相同方式表现的机器。尽管我们已经能够明确这种机器必须执行的一些过程，但与国际象棋机器和逻辑机器相比，这项研究目前尚未取得同样程度的进展。

3.Learning Theory.
It is clear that any machine that can perform human functions can be turned into a model of human behavior simply by a change in viewpoint. We have been working with the mass of data available from psychological experiments on human and animal learning with a view to designing a machine that would behave in the same manner. Although we have been able to specify some of the processes such a machine will have to perform, we are not as far along on this project as we are on the chess and logic machines.

4.简单理论
前面提到的三类机器，以及我们正在努力设计的这些机器，从严格意义上说都是 模型 ：它们能够表现出其所模拟对象的行为（能够下国际象棋、证明定理或进行学习）。我们可以将这类模型与理论区分开来——理论是对行为某些方面的描述，以及关于这些方面相互关系的论述。例如，我们可以提出一种问题求解器或逻辑机器的理论；而问题求解器或逻辑机器本身则能够实际解决问题。

4. Simple Theories.
The three types of machines that have been mentioned, and which we are striving to design, are models in the strict sense of the word: they exhibit the behavior of what they model (they play chess or prove theorems or learn). We may distinguish these from theories——descriptions of certain aspects of behavior, and assertions about their interrelationship. A theory of problem-solver or logic machine; the problem-solver or logic machine will solve problems.

这个研究小组的基本研究策略，是同时开展模型与理论两方面的研究。在复杂信息处理系统这一科学领域，我们距离甚至能够明确应该提出什么问题、应该从哪些方面进行抽象以建立理论，都还相当遥远。目前的需要，是建立一个数量足够庞大的具体系统群体，并对这些系统进行充分理解，从而为归纳研究提供基础。综合构建模型是获得这类系统的适当方法，因为自然形成的这类系统本身往往难以处理。然而，即使是综合构建出来的系统，其自身也具有极高的复杂性；如果要理解它们的行为，就必须建立相应的理论。尽管我们认识到，对这些机器的直接实践经验的增长必然会存在相当大的滞后，我们仍一直在朝着建立简单搜索理论*的方向进行初步尝试。

The fundamental research strategy of this group contemplates work on both models and theories. In the science of very complex information processing systems, we are a long way from even knowing what questions to ask or what aspects to abstract for theory. The present need is for a large population of concrete systems that are completely understood and thereby provide a base for induction. Synthesis of models provides the appropriate technique for providing such systems, since the systems that naturally are so intractable. However, even the synthesize systems are extraordinarily complex in their own right, and will require the construction of theories if their behavior is to be understood. We have been making tentative attempts in the direction of simple search theories*, even though we recognize thar there is bound to be a sizable lag in the growth of our Direct experience with these machines.

*H. A. 西蒙，《理性选择与环境结构》，《心理学评论》，1956年4月。

**H.A. Simon, “Rational choice and the structure of the environment”, Psychological Review, April 1956.

我们无法参加整个夏季研究项目，但预计会在项目初期参加一段时间。我们认为，该项目为我们提供了一个机会，使我们能够在前述各个领域进一步推进研究，并深入探索我们的工作与其他参与者的研究之间如何相互联系。

We will not be able to attend the entire summer research project, but do expect to be there for some time during the early part. We consider the project an opportunity to push a little further in the various areas mentioned, and to explore extensively the way in which our efforts interrelate with those of the other participants.