What is AI?

Extinguished philosophies lie about the cradle of every science as

Strangled snkes beside that of Hercules.-adapted from T. H. Huxley

　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　

What is artificial intelligence?

　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　

John McCarthy

Computer Science Department

Stanford University

Stanford, CA 94305

Jmc@cs.stanford.edu

Http://www-formal.stanford.edu/jmc/

1999 Nov 23, P.M.

Revised November 23,199 9:

　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　

1. Basic Questions

Q: What is AI?

A: artificial intelligence is the scientific and engineering implementation of intelligent machines, especially intelligent computer programs.

Technology. Artificial intelligence has some relationship with the goal of understanding human intelligence with computers, but it does not have to be used.

Biological methods.

 

Q: Oh, what is intelligence?

A: intelligence refers to the computing part of the ability to achieve the goal. Humans, many animals, and some machines have different categories.

Don't be smart to different degrees.

 

Q: Is there a general "intelligence" that doesn't depend on human intelligence?

A: No. We cannot determine what kind of computing process can be called "intelligence" in the general sense ".

We know some mechanisms of intelligence, but do not know others.

 

Q: Is intelligence isolated? For example, can I simply determine whether a machine has intelligence "?

A: No. Intelligence involves many mechanisms. AI researchers have already learned some of them and

These mechanisms can be implemented by computers, but not by other components. If a task only needs

The mechanisms that have been well understood can be well accomplished by computer programs. Such a program

There are "some intelligence.

 

Q: Is AI a simulation of human intelligence?

A: Sometimes, but not always, or, often, not. On the one hand, we can observe

The methods used to solve problems are used to understand intelligence. On the other hand, a large amount of achievements in AI are not

Through the study of human or animal activities, but through direct research of various problems that require smart solutions. Manual

Intelligent researchers can freely use methods that humans do not use and are beyond human computing power.

 

Q: What about IQ? Is there a computer program?

A: No. IQ is based on children's mental development speed. It is the score obtained by the tested person.

The ratio of age to his actual age. This measure is applicable to adults. IQ

High and low have a good relationship with success and failure in real life, but it is not true for computers.

The high scores obtained during the test have nothing to do with their actual usage. For example, a child reversely retries

The ability to describe a long string of ordered numbers can reflect other capabilities, probably because it can be measured

The child's real-time computing capability. However, a computer with very limited functions can complete its "Number

The sequence length of the range.

Of course, some problems in iqtest are also beneficial to artificial intelligence.

 

Q: What is the comparison between humans and computers about intelligence?

A: Arthur R. JASON [jen98], one of the most important researchers in Human Intelligence Research, proposed an "inspiration ".

Hypothesis: all normal people share the same smart mechanism, and the intellectual difference is that "biochemistry and Production

The quantity conditions in science are different. I regard these conditions as speed, short-term memory, and accurate and

Recoverable long-term memory capability.

There is no final conclusion about whether Jenkins's hypothesis about human intelligence is true or not; however, the current situation in artificial intelligence

On the contrary.

Computers have extremely high speed and huge capacity, but their capabilities depend on what programmers fully understand and

The intelligent mechanism implemented in the program. Some talents may have the ability to work in their teens.

Computing machines are ready, while other two-year-olds have no computers. Since cognitive science is still

Without being able to accurately define what capabilities a person has, it is more difficult for us to understand it. Generally

The composition of AI mechanisms is often different from that of human intelligence mechanisms.

As long as a person performs better than a computer on some issues, or a computer uses a lot of computing

The same performance as that of People indicates the smart mechanism that programmers need to efficiently solve this problem.

Is not fully understood.

 

Q: When did AI research begin?

A: After World War II, many people began their independent research on smart machines. English mathematician Alan Turing or

He was the first. He gave a speech on this issue in 1947. He also made his first attempt to exploit programmable computers

The method of artificial intelligence research is much better than that of direct manufacturing machines. By the end of 1950s

Most of the researchers of multiple artificial intelligence work on programmable computers.

 

Q: Is AI intended to implant human thoughts into computers?

A: Some researchers say they have such plans, but they may just use this word for an analogy. Person's

"Thinking" has many features, and I don't think someone really wants to imitate all the features.

 

Q: What is a Turing test?

A: Alan Turing 1950's article computing machines and intelligence [tur50] discusses the conditions under which machines

It can be considered intelligent. He believes that if a machine can successfully make a certain knowledge

The observer thinks that he is a human, then of course he can think that this machine is intelligent. This test standard

Most people except philosophers are acceptable. The observer in the test can pass

The telex typewriter interacts with the tested machine and one person to prevent the machine from imitating the human appearance.

And sound. The man tried to persuade the observer that he was a person, and the machine tried to deceive the observer that he was a person.

The Turing test is just a test. A machine that passed the Turing test can of course be considered wise.

Yes, but a machine that does not know much about humans and cannot imitate humans can also be considered smart.

Yes. The book brainchildren [den98] by Daniel Dennett

The Turing Test and various part of the Turing Test (knowledge and inquiry about the testing observer in AI)

. The results show that some programs with poor functionality will be

It is easy to think that it is intelligent.

 

Q: Is AI designed to achieve human intelligence?

A: Yes. The final effort is to make computer programs solve problems in the world and achieve

Target. However, many people who work in special research areas do not have such ambition.

 

Q: How far is Ai from reaching the human level? When will it become a reality?

A: A few people think that human intelligence can be achieved in this way:

The fact that a language that has compiled and accumulated a large amount of knowledge can express knowledge is used to write a large number of programs,

To achieve human intelligence.

However, most AI researchers believe that new basic ideas are still needed. Therefore

It cannot predict when intelligence at the human level can be achieved.

 

Q: Are computers suitable for generating intelligence?

A: A computer can simulate any machine by programming. Many researchers have invented some non-computer computers.

In some ways, we hope they can have more intelligence than computers. However, they are usually

After simulating the machines they invented, they doubt whether the new machines are worth implementing in reality. Because it is not counted

The number of capital investment has made the computer faster and faster, and the actual performance of another machine must be

It is much better than the simulation on the computer.

 

Q: Is the current computer fast enough for intelligence?

A: Some people think that a faster computer is needed while new ideas are needed. In my opinion, as long

We know how to program and implement intelligence. Thirty years ago, the computer speed was enough. Of course,

In addition to the ambitious goals of AI researchers, computers will continue to grow faster and faster.

 

Q: How about parallel computers?

A: computers with multiple processors are much faster than single-processor computers. Parallelism itself

It does not provide any advantages, and it is difficult to use parallel machines. Only when high speed is required

This difficult-to-use machine is necessary.

 

Q: How can I create a "child-like machine" that can improve myself by reading and learning from experiences?

A: This idea has been raised many times since 1940s. In the end, it will be manufactured.

However, at present, artificial intelligence programs cannot learn from natural experiences as effectively as children do. Current Program

I still cannot understand the language well, not to mention learning from reading.

 

Q: an AI system can constantly improve its own intelligence by thinking about AI.

Level?

A: I think so. However, artificial intelligence has not yet reached the level where such a process can be achieved.

 

Q: Can I talk about chess?

A: Alexander kronrod, a Russian AI researcher, believes that "Chess is an AI research

The fruit fly in the study ". He made an example here. Fruit fly is often used by genetic scientists to study genetics. Chess

Some smart mechanisms are required, but not others. The chess program can now be used

Master level, but compared with human players, it only uses a very simple smart mechanism, that is, a large number

Instead of understanding. Once we have a better understanding of the smart mechanism of playing chess, we can do it.

There is much less computing than we need now, but there is also a chess program at the level of human chess.

Unfortunately, the creation of chess programs is now more used as a means of competition for business, rather

As a scientific research method. This is like a genetic engineer who started organizing the fruit fly competition in 1910.

They concentrate on raising fruit fly to win the competition.

 

Q: Can I talk about go?

A: Go games in China and Japan are also played by two players on the same board. Go exposes us

Lack of understanding of human intelligence mechanisms for playing chess. Despite a considerable amount of research (not on chess)

But the go program's chess skills are still poor. This problem may be caused by a game in go.

The Board can be naturally divided into sub-board games, and then each sub-board is analyzed independently before the sub-board

. People can use this method when playing chess, but the go program cannot.

As a whole. Chess programs can only use a large amount of computing to make up for the lack of such smart mechanisms

Loss. This type of computing is usually several thousand times more than humans, and millions of times more than deep blue.

One day, artificial intelligence researchers will be able to overcome this objectionable shortcoming.

 

Q: aren't some people thinking that artificial intelligence is harmful?

A: The philosopher John Searle believes that the idea that a non-biological machine will possess intelligence is not logical. Philosophy

Hubert Dreyfus believes that artificial intelligence is impossible. Computer scientist Joseph weizenbaum

Artificial intelligence is an immoral and indecent idea. All kinds of people insist that

It cannot be achieved without reaching the human level of intelligence. Some others invest in them.

The company broke down.

 

Q: Does Computing Theory and computing complexity theory play a key role in artificial intelligence? [Note: Computing Theory

Computing complexity theory is a branch of mathematical logic and computer science that is quite skillful and solves

The problem also requires considerable skills.]

A: No. Those theories are related to artificial intelligence, but they cannot solve the basic problems of artificial intelligence.

In 1930s, the mathematical logic experts represented by Kurt Godel and Alan Turing confirmed

No algorithm can determine all problems in some important mathematical fields. True or false propositions in level 1 Logic

The determination is an example; whether a polynomial equation with multiple variables has an integer solution is another example. But human 1

Directly solve problems in these fields. This result provides such a viewpoint (usually exaggerated ),

That is to say, computers cannot do things that humans can do in essence. However, people cannot guarantee

Humans can solve any problem in these fields.

In 1960s, computer scientists represented by Steve cook and Richard Karp

The theory of NP completeness is established. An NP-complete problem can be solved, but it is likely to be necessary and the scale of the problem.

An exponential time. The satisfiable Problem of a proposition in Proposition calculation is a basic example of a complete NP problem. Person

Class can often be used to solve some special problems much shorter than the general algorithm.

Yes.

What is important in artificial intelligence is to obtain algorithms that can solve problems like humans. The confirmed algorithm is

It is important to have other fields, but the solutions to many AI problems are not related to this field.

The theory of difficulty about all kinds of problems is called computational complexity theory. So far, this

The theory has not interacted with artificial intelligence as expected. Problems with people and AI programs

The attribute of the problem on which the success of the solution depends and the method of solving the problem, as if it were a complex investigator and person.

AI researchers cannot be completely determined.

Complexity of algorithms independently developed by solomonoff, Kolmogorov, and chaitin

The same is true. It defines the complexity of a symbolic object as the length of the shortest program that can generate it. Already

It has been proved that the problem of finding a program with the shortest or proximity to the shortest is unsolvable, but the short one can be used.

Programs that generate these symbols are often enlightening. Of course, you cannot prove this program.

Is the shortest.

2 branch of Artificial Intelligence

Q: What are the branches of AI?

A: Some of them are listed here, but some branches are not listed here, because no one can confirm

They, some of which are regarded as concepts or themes, rather than full branches.

 

Logical artificial intelligence

Generally, the knowledge of the world that a program knows is a fact in a specific situation,

In this case, the program must take actions, and its goal is to use all these words in mathematical logic.

Statement. The program determines what to do by reasoning the appropriate behavior to achieve the goal. Process this

The earliest article about the problem was [mcc59], and [mcc89] was a recent summary. [mcc96] listed some

[Sha97] is an important document in the concept of logical artificial intelligence.

 

Search

Artificial intelligence programs often need to search for a large number of possibilities, such as moving or passing Theorem in chess games.

Prove the program to reason. In many different fields, people are continuing to discover various methods to improve the program's

Efficiency.

 

Pattern Recognition

When a program observes something, it is usually designed to compare what you see with the pattern. For example,

To find a face, a visual system tries to match the eye and nose pattern in a scene. More

Complex patterns, such as text in a natural language, a location in a chess game, or something to be studied

Component history. These more complex models are more different than the vast majority of simple models that have been studied.

Method.

 

Indicates

Knowledge about the world must be expressed in some way. It generally uses a mathematical logic language.

 

Reasoning

Some facts can be used to introduce other facts. For some purposes, mathematical logic deduction is enough,

However, since 1970s, a method called non-monotonic reasoning has been developed.

 

Common sense knowledge and common sense reasoning

This field is the farthest distance from artificial intelligence to human beings, although it has been a non-

Frequently active research areas. Although considerable progress has been made, such as non-monotonous reasoning systems and Behaviors

The development of theory, but we need more new ideas. The Cyc system contains a large number of but unreliable

A set of facts.

 

Learn from experience

The program must do this. Connection-based and neural network processing for AI is dedicated to this. Also

This includes learning the rules expressed in logic. [Mit97] is an undergraduate course on machine learning. Program

You can only learn the facts and actions that can be expressed in their forms, but unfortunately

It is based on a very limited ability to represent information.

 

Planning

The planning procedure is expressed in regard to the world (especially the facts about the results of the Act) or with respect to special circumstances and objectives

. Then, they generate a policy to achieve this goal. In many common cases,

This policy is a series of actions.

 

Cognition Theory

This is a study of the types of knowledge we need to address the problems in the world.

 

Ontology

Ontology studies the types of things that exist. In AI, programs and statements need to process a large number of objects, while I

We want to study the types of these objects and their basic properties. The focus on Ontology began in 1990s.

 

Heuristic

Heuristic is a method that tries to discover ideas and other things in a program. In AI, this technique

It is widely used. Heuristic functions are used to process searches. They are used to measure the number of nodes in a search tree.

How far is the mark node. A heuristic predicate is used to compare two nodes in a search tree.

A good one is to make progress towards the goal. According to my opinion, heuristic predicates are more useful.

 

Genetic Programming

Genetic Programming is a technique that selects the most appropriate from thousands of generations by means of a lisp program of hybridization randomization

Suitable program to complete the task. The group led by John Koza developed the genetic program design with a copy attached here

Guide. (Http://www.genetic-programming.com/gpanimatedtutorial.html ).

Q: Where is the application of AI?

A: Some of them are listed here.

 

Games

You can spend a few hundred dollars to buy a master-level chess game machine. There is

Artificial intelligence, but they primarily rely on blind computation-watching thousands of locations to confront people. Pass blind

To defeat a world champion, the heuristic algorithm needs to observe 0.2 billion positions per second.

 

Speech Recognition

In 1990s, Computer Speech Recognition has reached a practical level for some specific purposes. Although possible

Guides some computers to use voice, but for convenience, people will continue to use the keyboard and mouse.

 

Natural language understanding

It is not enough to input only one character sequence to the computer. Sentence Parsing is not enough. The computer must

Can Understand fields related to the text, and this is only possible for very limited fields.

 

Computer Vision

The world is composed of three-dimensional objects, but the images entered into human eyes and computer TV cameras are

Two-dimensional. Some useful programs can work independently in two dimensions, but the complete computer vision needs

3D information is divided into more than two-dimensional images. Currently, there are only limited ways to directly represent three-dimensional messages.

And it is not as good as humans.

 

Expert System

A "Knowledge engineer" tries to put their knowledge in computer programs by interviewing experts in a certain field

To complete a task. How does this work depend on the intellectual mechanism required to complete the task?

No is included in the current state of artificial intelligence. If this is not proved to be the case, there will be many

Disappointing results. One of the earliest expert systems was mycin in 1974, which was used to diagnose whether the blood was

Infected with bacteria and how to treat them. It works better than medical school students and doctors who have already worked.

Good, even though people have discovered its flaws. As its name implies, its existence includes bacteria, symptoms, and governance.

Second, it does not include patients, doctors, hospitals, deaths, rehabilitation, and real-time events. Communication dependency

In a simple patient to be considered. Because the experts interviewed by Knowledge engineers are familiar with patients, doctors, and deaths

Therefore, knowledge engineers must input the information that experts tell them into

Pre-determined framework. From the current situation of artificial intelligence, this is indeed the case. Currently

The usefulness of the home system depends on users with common knowledge.

 

Heuristic Classification

One of the most feasible expert systems is to input information to a fixed

Use a collection of original information types. For example, whether to accept the proposed credit card transaction is recommended. About letter

The information such as the card owner, payment record, purchased items, and the company that purchased the items should be available.

(For example, whether the company has had credit card fraud before ).

: How AI research is conducted.

A: The research on artificial intelligence involves both theory and experiment, while there are also basic and theoretical points in experiment.

There are two main clues to AI research. One is biological, based on the following ideas:

However, if humans have intelligence, artificial intelligence should learn humans and simulate their psychology and physiology. Another

It is a phenomenon. Based on common knowledge about the world and issues raised to achieve the goal

Research and formalization. These two aspects interact with each other to some extent, and both of them will eventually take

. It was a race, but two contestants seemed to be walking.

 

Q: What do I have to learn before learning artificial intelligence?

A: learn mathematics, especially mathematical logic. Generally, the more scientific knowledge the better.

Biological processing of artificial intelligence requires learning about psychology and the physiology of the nervous system. Learn several programs

Design languages: C, lisp, and Prolog. It is also a good idea to learn a basic machine language.

Further work may depend on familiarity with popular languages. The end of 1990s, including C ++ and Java.

 

Q: What are some good materials about AI?

A: Stuart Russell and Peter norvig wrote the article "artificial intelligence" published by Prentice Hall.

Intelligence is the most common teaching material of 1997, and some general ideas expressed here are not consistent with this article. Nils

Written by Nilsson, artificial intelligence: a new comprehensive possibility published by Morgan Kaufman

It is easier to read.

 

Q: What organizations and publications are related to AI?

A: aaai (http://www.aaai.org), European AI Cooperation Committee member

Society (eccai) (http://www.eccai.org/) and AI and Behavior Simulation Association (aisb)

(Http://www.cogs.susx.ac.uk/aisb) is a scientific association related to artificial intelligence research. Computer

The Association (ACM) has an interesting group related to artificial intelligence, called sigart.

(Http://www.acm.org/sigart ).

 

The International Federation of Artificial Intelligence (ijcai) (http://www.ijcai.org) is the main international conference.

Electronic Journal of Artificial Intelligence electronic transactions on artificial intelligence

(Http://www.ida.liu.se/ext/etai/), "artificial intelligence" Artificial Intelligence

(Http://www.elsevier.nl/locate/artint/), Journal of artificial intelligence research

Artificial Intelligence Research (http://www.jair.org/) and IEEE Pattern Analysis

And machine intelligence journal IEEE Transactions on Pattern Analysis and Machine

Intelligence (http://computer.org/tpami/) is the master of four artificial intelligence research papers

A magazine. All the information I can find is included here.

 

Home positive reviews (http://www.cs.utexas.edu/users/vl/ppr/) lists one

Some papers that experts think are important ..