ai人工智能_古典AI的简要史前-CSDN博客

ai人工智能

推理与经典AI (Reasoning and Classical AI)

To talk about Reasoning, it’s important to understand how we got here. This article covers what I call the pre-history of Classical AI — those parts of the story that happened before the invention of modern computers (pre 1950s) but are crucial to understanding why we believe that AI is possible.

要谈论推理，了解我们如何到达这里很重要。 本文涵盖了我所说的古典AI的史前历史-故事的那部分发生在现代计算机发明之前(1950年代之前)，但是对于理解为什么我们相信AI可行至关重要。

This is part 2 in a series on Reasoning.

这是推理系列的第2部分。

Like most things, the very beginnings of classical AI is rooted in philosophy, and starts in the ancient world (the Greeks, Indians, and Chinese all had some early forms of logic). But, as I’m not a masochist we start in more contemporary times with two big ideas that lay the foundation for modern AI:

像大多数事物一样，古典AI的最开始就是植根于哲学，并且起源于古代世界(希腊人，印度人和中国人都有一些早期的逻辑形式)。但是，由于我不是受虐狂，所以我们从更现代的时代开始，为现代AI奠定了两个重要思想：

Mechanizing intelligence — Is there a trustworthy, automatic process that takes in some information and spits out new information? Trustworthy, in this context, means that the information produced is supported by the information taken in. Automatic means that it works without human intervention. Of course, simply discovering new information may seem like a very limited definition of intelligence but it reflects AI’s mathematical origins where it was these techniques were first applied to topics like automatic theorem proving.
机械化智能-是否存在一个可信赖的，自动的过程，该过程可以吸收一些信息并吐出新信息？在此上下文中，“ 可信赖”表示所生成的信息得到所接收信息的支持。“ 自动”表示无需人工干预即可工作。当然，简单地发现新信息似乎对智能的定义非常有限，但是它反映了AI的数学起源，这些技术最初是用于自动定理证明之类的。
Computability — Is it possible to implement this automatic process on a computer?
可计算性-是否可以在计算机上实现此自动过程？

命题逻辑 (Propositional Logic)

The development of logic was humanity’s first great attempt at mechanizing intelligence, and the basis for modern logic lies with George Boole, Charles Pierce, and Gottlob Frege. George Boole, not being satisfied with laying the basis for modern computers (hello Boolean Algebra), also nearly scuttled my fledgling graduate career (I flunked my PHIL 650 Intro to Logic course) by inventing Propositional Logic. Propositional logic represents the current state of the world by a set of facts or propositions. These propositions can be combined using boolean operators to form formulas. For example:the cat is to the left of the dog, or Whiskers is a cat and Fido is a dog and Whiskers is sitting to the left of Fido

逻辑学的发展是人类机械化智力的第一次重大尝试，而现代逻辑学的基础在于乔治·布尔，查尔斯·皮尔斯和戈特洛夫·弗雷格。乔治·布尔(George Boole)对现代计算机的基础知识不满意(您好，布尔代数)，他还通过发明命题逻辑 ( Propositional Logic )几乎使我刚起步的研究生生涯(我的PHIL 650 Intro to Logic课程失败)。命题逻辑通过一组事实或命题表示世界的当前状态。可以使用布尔运算符组合这些命题以形成公式。例如： 猫在狗的左侧 ，或者晶须是猫，而Fido是狗，晶须坐在Fido的左侧

In addition to the current state, propositional logic also has knowledge about the world stored in the form of if-then rules.if Whiskers is a cat then Whiskers has four feet and Whiskers has a tail, orif Whiskers is a cat and Fido is a dog then Fido is afraid of Whiskers

命题逻辑除了当前状态外，还具有以if-then规则形式存储的关于世界的知识。如果 晶须是猫， 那么 晶须有四英尺，晶须有尾巴 ; 如果 晶须是猫而费多是狗， 则费 多害怕晶须。

New facts are derived from currently known facts using a very important deductive tool called Modus Ponens which says that when you have a rule like so: If p then q and the fact p is true then the fact q is true.

新事实是使用一种非常重要的演绎工具Modus Ponens从当前已知事实中得出的，该演绎工具表示您具有如下规则： 如果p则q且事实p为真，则事实q为真。

Modus Ponens might seem trivial but it guarantees that when knowledge is represented in the form of trustworthy rules, new information generated from these rules using Modus Ponens is also trustworthy.

Modus Ponens看起来很琐碎，但可以保证当知识以可信赖的规则形式表示时，使用Modus Ponens从这些规则生成的新信息也是可信赖的。

The mechanization process, called rule-matching, applies rules to facts. As an example, consider a situation where there is only one rule:If the cat is to the left of the dog and the dog is to the left of the human then the cat is to the left of the human. Now imagine that there are two facts you know to be true in this situation:There is a cat to the left of the dog and There is a dog to the left of the humanThe rule-matching process will match these two pieces of information to the rule and create a third fact:There is a cat to the left of the human.

机械化过程称为规则匹配 ，将规则应用于事实。例如，考虑一种情况，该规则只有一个规则：如果 猫在狗的左边，而狗在人的左边， 那么 猫在人的左边。 现在假设在这种情况下有两个事实是正确的： 这是狗的左边是猫，人的左边 是狗。规则匹配过程将匹配这两条信息遵守规则并创造第三个事实： 人的左边有只猫 。

In our example, there is only one rule that matches and it matches exactly once. But imagine having one additional fact:The human is to the left of the lamp and one additional rule:If the cat is to the left of the human and the human to the left of the lamp then the cat is to the left of the lamp Now, once that first rule has matched and the fact the cat is to the left of the human is added, our new rule will match and add a fourth fact:There is a cat to the left of the lamp.

在我们的示例中，只有一个匹配的规则，并且仅匹配一次。但是想象一下，还有一个事实： 人在灯的左边，还有另外一条规则：如果 猫在人的左边，人在灯的左边， 那么 猫在灯的左边现在，一旦第一个规则匹配并且添加了猫在人类左边的事实，我们的新规则将匹配并添加了第四个事实： 在灯的左边有一只猫 。

Applying rules to the current situation creates new information. This new information triggers another round of matching (with a possibly different set of rules) that creates additional new information. This repetitive process is how propositional logic automates the finding new information.

将规则应用于当前情况会创建新信息。此新信息触发另一轮匹配(使用可能不同的规则集)，从而创建其他新信息。命题逻辑如何自动执行发现新信息的过程。

谓词(一阶)逻辑 (Predicate (First-order) Logic)

Now we shouldn’t let George hog all the credit for the near untimely demise of my Ph.D. Charles Pierce and Gottlob Frege are just as important to this story because they invented Predicate or First-order Logic. Take the cat-leftof-dog-leftof-human example. That is not just true for cats, dogs, and humans. It’s true for any three things. With propositional logic you have to write rules for every set of three you could think of… dog to the left of cat to the left of human, human to the left of the cat to the left of the dog, cat-cat-dog, cat-human-capybara, human-tomato-birkenstocks… you get the idea. There are workarounds to represent information more succinctly but, for the most part, representing facts this way is very tedious and not very useful.

现在，我们不应该让乔治为我的博士近乎不合时宜的逝世而全力以赴。查尔斯·皮尔斯(Charles Pierce)和戈特洛布·弗雷格(Gottlob Frege)对于这个故事同样重要，因为他们发明了谓词或一阶逻辑。以cat-leftof-dog-leftof-human为例。这不仅对猫，狗和人来说是正确的。对于任何三件事都是如此。使用命题逻辑，您必须为可能想到的每三个一组编写规则……狗在猫的左边在人的左边，人在猫的左边在狗的左边，cat-cat-dog，猫-人类水豚，人类-勃艮第……你明白了。有一些变通方法可以更简洁地表示信息，但是在大多数情况下，以这种方式表示事实非常繁琐且不太有用。

With predicate logic, you can use variables to represent facts. Instead of writing: If the cat is to the left of the dog and the dog is to the left of the human, then the cat is to the left of the human you can write: For all A,B, and C: If A is to the left of B and B is to the left of C, then A is to the left of Cwhich covers a wide range of scenarios involving animals, humans, and inanimate objects, all sitting patiently next to each other. The “For all” is called a quantifier and allows predicate logic to represent variables that can stand in for all objects. There is one additional quantifier — “There exists” — that allows predicate logic to represent variables that stand in for exactly one thing. For example: There exists x: such that x is a person AND x is the president of the United States

通过谓词逻辑，您可以使用变量来表示事实。而不是写：如果 猫在狗的左边，而狗在人的左边， 那么 猫在人的左边，您可以写： 对于所有A，B和C： 如果 A在B的左侧，B在C的左侧， 然后 A在C的左侧 ，涵盖了涉及动物，人类和无生命物体的各种场景，所有场景都耐心地彼此相邻。 “ For all ”称为量词，它允许谓词逻辑表示可以代表所有对象的变量。还有一个附加的量词(“ exists ”)，它允许谓词逻辑表示恰好代表一件事的变量。例如： 存在x：使得x是一个人而x是美国的总统

Predicate (and propositional) logic can be written in a shorter, easier-to-read form:LeftOf(A,B) & LeftOf(B,C) → LeftOf(A,C). Here, LeftOf is called a predicate, and A and B are the variables that stand in for propositions. Predicate logic uses the same vocabulary of propositional logic — propositions, the boolean operators, an inference rule (usually Modus Ponens) — with the addition of variables and the two quantifiers. These additions make predicate logic much more powerful at representing rules but comes at the cost of making it more complicated to run (due to the need to match variables).

谓词(和命题)逻辑可以用更短，更易读的形式编写： LeftOf(A，B)＆LeftOf(B，C)→LeftOf(A，C) 。在这里， LeftOf被称为谓词，而A和B是代表命题的变量。谓词逻辑使用相同的命题逻辑词汇-命题，布尔运算符，推理规则(通常是Modus Ponens)-加上变量和两个量词。这些添加使谓词逻辑在表示规则时功能更强大，但代价是使其运行起来更加复杂(由于需要匹配变量)。

There is also second-order predicate logic where you can even use a variable for LeftOf because what applies to Left also applies to other similar predicates. So, instead of LeftOf(A,B) & LeftOf(B,C) → LeftOf(A,C), we can write D(A,B) & D(B,C) → D(A,C). Where D could be left, right, top, bottom, in front, behind and so on. Second-order logics make rule-matching much harder and I’ve never seen one used in the wild but I flunked my Intro to Logic course so what do I know.

还有二阶谓词逻辑，您甚至可以为LeftOf使用变量，因为适用于Left的内容也适用于其他类似谓词。因此，我们可以写D(A，B)＆D(B，C)→D(A，C)来代替LeftOf(A，B)＆LeftOf(B，C)→LeftOf (A，C) 。 D可以位于左，右，上，下，前，后等位置。二阶逻辑使规则匹配变得更加困难，而且我从未见过在野外使用规则，但是我不了解Intro to Logic课程，所以我知道什么。

From propositional and predicate logic, we can distill that mechanizing intelligence means three requirements:

从命题和谓词逻辑中，我们可以得出结论，机械化智能意味着三个要求：

A way to represent your knowledge (Rule for e.g.)
表示您的知识的一种方式(例如规则)
A way to represent the current situation (propositions or predicates)
表示当前情况(命题或谓词)的方式
A way to automatically keep applying knowledge (from 1) to the current situation (2) to generate/infer/deduce trustworthy new information. (Modus Ponens and rule-matching)
一种自动将知识(从1)应用于当前情况(2)以生成/推断/推断可信赖的新信息的方法。 (方法Ponens和规则匹配)

These requirements are not particularly onerous. Neural networks satisfy most of these conditions — The weight matrix represents knowledge, the input matrix is the current sitch, and forward pass is the way to apply knowledge to discover new information. The only place where neural networks fail is that they do not keep applying the process to new information (recurrent networks being the exception). Neural networks are not logical systems obviously. A real logical system has stricter requirements and definitions which we will encounter later. Our current purpose is to introduce the idea of mechanizing intelligence, i.e., identifying a process where intelligence (narrowly defined as discovering new information) is automatic.

这些要求并不是特别繁重。神经网络满足大多数这些条件-权重矩阵代表知识，输入矩阵是当前势力，而前向传递是应用知识发现新信息的方式。神经网络唯一失败的地方是它们不会继续将过程应用于新信息(循环网络除外)。神经网络显然不是逻辑系统。真正的逻辑系统具有更严格的要求和定义，我们将在以后遇到。我们当前的目的是引入机械化情报的想法，即，识别情报(狭义地定义为发现新信息)是自动的过程。

偏离完整性和完整性。 (A digression into soundness and completeness.)

What makes propositional and predicate (and other forms of) logics interesting is not simply what you can do with them but what you can prove about them; particularly three properties called soundness, completeness, and decidability. If you don’t already know about these concepts then you should pay close attention because it’s probably the coolest thing you will read today:

使命题和谓词(以及其他形式的)逻辑有趣的原因不仅在于您可以对它们做些什么，还可以证明它们。特别是三个特性，称为健全性 ， 完整性和可判定性 。如果您还不了解这些概念，那么您应该密切注意，因为这可能是您今天将阅读的最酷的东西：

A logical system is said to be sound if any information derived from that system is true. In a sound logical system, if you start off with a set of facts and run the logic machine, all information that is produced is logically provable, i.e., you can trust the machine to produce only correct information.

如果从该逻辑系统派生的任何信息为真，则该逻辑系统被称为健全系统。在合理的逻辑系统中，如果从一组事实开始并运行逻辑机器，则所产生的所有信息在逻辑上都是可证明的，即，您可以信任机器仅产生正确的信息。

Completeness is sort of the opposite side of the coin. A logical system (initial facts + process) is said to be complete if all information that is true in that system can be derived from the starting set of facts using the logical process, i.e., if there is new information that logically follows what the system knows, the system will eventually find it.

完整性有点像硬币的反面。如果可以使用逻辑过程从起始事实集合中导出该系统中所有正确的信息，即，如果逻辑上遵循该系统的新信息，则该逻辑系统(初始事实+过程)被认为是完整的知道，系统最终会找到它。

Together, soundness and completeness allow for a logical system to be dependable because you can be sure that everything that comes out of it is valid and that it will find anything that is logically valid. To be truly useful, logical systems also have to be decidable. That is, there needs to be an effective method (a method that runs in a finite though possibly very long amount of time) for determining if a statement is true or false in that logical system.

健全性和完整性共同保证了逻辑系统的可靠性，因为您可以确保从逻辑系统中得到的一切都是有效的，并且可以找到在逻辑上有效的任何内容。为了真正有用，逻辑系统也必须是可决定的。也就是说，需要一种有效的方法(这种方法可以在有限的时间内运行，尽管可能会花费很长的时间)，以确定该逻辑系统中语句是对还是错。

Propositional logics are sound, complete, and decidable. First-order predicate logics are sound and complete but not decidable for non-trivial systems. (P. S. You can have a 5 credit graduate-level course that is just about proving the above statements)

命题逻辑是合理，完整和可判定的。一阶谓词逻辑是健全且完整的，但对于非平凡的系统则无法确定。 (附言：您可以开设5个学分的研究生课程，证明上述陈述是正确的)

Together soundness, completeness, and decidability confer respectability on our choice of mechanistic process because it gives us theoretical assurances about its effectiveness.

健全性，完整性和可判定性共同赋予我们对机械过程选择的尊重，因为它为我们提供了对其有效性的理论保证。

Anyway, there are three takeaways here for anyone still reading.

无论如何，这里还有三点可供任何还在阅读的人参考。

Boole, Pierce, and Frege laid the basis for logic, reasoning, and artificial intelligence
Boole，Pierce和Frege为逻辑，推理和人工智能奠定了基础
If you are in computer science, take the Intro to Logic course offered by the Math department not the Philosophy department. You’ll thank me later.
如果您是计算机科学专业的学生，请选择数学系而非哲学系开设的逻辑入门课程。待会儿我会谢谢你的。
Suck it Boole, Pierce, and Frege. I got my Ph.D.
吸吮Boole，Pierce和Frege。我获得了博士学位

One final note about logic. Logic can trip anyone up. Bertrand Russell spent an entire lifetime trying to build a logical understanding for mathematics. He wanted to start from a few assumptions (four or five) and show that all of mathematics can be derived from these assumptions using the machinery of logic. His attempt was ultimately futile just like my attempt to pass my Intro to Logic course. He did manage to accomplish some other stuff while he was failing at finding a logical foundation of mathematics — wrote extensively on a broad range of subjects, created entire new disciplines, revolutionized the understanding of mathematics and logic, won a Nobel prize. What a slacker. Jokes aside, Russell is one of the titans of philosophy and liberalism of the 20th Century. He saw through the veneer of Bolshevism, made exceptions to his pacifist beliefs with the rise of Hitler, chaired the India league, and went to jail at age 89 for taking part in an anti-nuclear demonstration. Truly, someone worth learning from.

关于逻辑的最后一点。逻辑可以使任何人绊倒。伯特兰·罗素(Bertrand Russell)毕生致力于建立对数学的逻辑理解。他想从几个假设(四个或五个)开始，并证明所有数学都可以使用逻辑机制从这些假设中推导出来。就像我尝试通过逻辑入门课程一样，他的尝试最终是徒劳的。当他未能找到数学的逻辑基础时，他确实设法完成了其他工作-在广泛的学科上广泛写作，创建了全新的学科，彻底改变了对数学和逻辑的理解，并获得了诺贝尔奖。真懒。除了开玩笑，罗素是20世纪哲学和自由主义的巨人之一。他看穿了布尔什维克主义的表面饰面，随着希特勒的崛起，打破了和平主义信仰的例外，主持了印度联赛，并因参加反核示威活动而享年89岁。确实，值得向某人学习。

教会转向论题 (The Church-Turing Thesis)

Logical Systems (propositional and predicate logic) revolutionized our understanding of how a mechanistic process can produce new information. The Church-Turing thesis convinces us that it is possible to build a machine that actually implements such processes.

逻辑系统(命题和谓词逻辑)彻底改变了我们对机械过程如何产生新信息的理解。 Church-Turing的论点使我们相信，可以构建一台实际执行此类过程的机器。

Alonzo Church, an American mathematician, invented a system of mathematical logic called the Lambda-calculus as a framework for solving math problems. The expressive power of Lambda calculus allows it to represent a whole variety of logics. Alan Turing (he of the Turing Test fame) is, of course, the inventor of the Turing machine, a machine for computation. Church and Turing realized that anything represented using the Lambda calculus can be calculated using a Turing machine and vice versa. And since the Lambda calculus covers a wide range of processes including first-order logic, the Church-Turing thesis allows for the possibility of computational machines that can demonstrate inference and intelligence.

美国数学家阿隆佐·丘奇(Alonzo Church)发明了一种称为Lambda演算的数学逻辑系统，作为解决数学问题的框架。 Lambda演算的表达能力使其可以代表各种逻辑。当然，Alan Turing(图灵测试界的著名人物)是图灵机(一种用于计算的机器)的发明者。丘奇和图灵意识到，可以使用图灵机来计算使用Lambda微积分表示的任何事物，反之亦然。而且由于Lambda演算涵盖了包括一阶逻辑在内的广泛过程，因此Church-Turing论点允许出现可以证明推理和智能的计算机器。

The Church-Turing hypothesis is also the reason why we refer to our brains as computers. It’s not because our brains and computers are similar in the way they compute. We refer to brains as computers because

Church-Turing假设也是我们将大脑称为计算机的原因。这不是因为我们的大脑和计算机的计算方式相似。我们称大脑为计算机是因为

to the best of our knowledge, the brain is implementing a computable function,
据我们所知，大脑正在执行可计算的功能，
a Turing machine can implement any computable function, and
图灵机可以实现任何可计算功能，并且
computers are Turing machines.
计算机是图灵机。

Ergo the brains-as-computers analogy.

像计算机一样类比。

We don’t know what that computable function is or how it is implemented. But this distinction between computable function and its implementation is reflected in the divide between Classical AI and probabilistic counterparts like Neural Networks. Classical AI has a good story on what this computable function looks like but they are less convincing in their implementation. Neural networks have a really good story about what implementations can look like but are less convincing about exactly what computable function they are implementing.

我们不知道该可计算函数是什么或如何实现。但是，可计算功能与其实现之间的区别体现在经典AI与概率对等物(如神经网络)之间的鸿沟。古典AI很好地讲述了此可计算函数的外观，但它们在实现方面的说服力较弱。神经网络对于实现看起来是一个很好的故事，但对于实现的可计算功能却不太有说服力。

Finally, it is quite possible that there are many different but equivalent such computable functions and many different ways to implement them. The history of AI is the story of trying to find one such function and implementation (while the history of Cognitive Science is of trying to identify the specific computable function and implementation that underlies human intelligence).

最后，很可能存在许多不同但等效的可计算函数，以及实现它们的许多不同方式。 AI的历史就是试图找到一种这样的功能和实现的故事(而认知科学的历史就是试图找出构成人类智能基础的特定可计算功能和实现)。

本系列的下一步是什么？ (What’s next in this series?)

I originally intended this article to be a brief history of classical AI but I soon realized that I’d have to break it up into two articles — this one that addresses the pre-history and the next one in the series that will focus on developments since the 1950s.

我原本打算将这篇文章作为经典AI的简要历史，但是我很快意识到，我必须将其分解为两篇文章-这篇文章讲述了史前史，而下一篇文章着重于发展自1950年代以来。