EXTINGUISHED Philosophies Lie About The Cradle of Every Science As The CRADLE OF EVERY
STRANGED SNAKES BESide That of Hercules. - Adapted from T. H. Huxley
What is artificial intelligence?
John McCarthy
Computer science department departments
Stanford University
Stanford, CA 94305
JMC@cs.stanford.edu
http://www-formal.stanford.edu/jmc/
1999 Nov 23, 8:39 p.m.
Revised November 23, 1999:
1. fundamental issue
Q: What is artificial intelligence?
A: Artificial intelligence is a scientific and engineering that achieves smart machines, especially with intelligent computer programs
technology. Artificial intelligence has some relationship with the goal of understanding human intelligence with computer, but it does not have to use
Biological methods.
Q: Oh, what is intelligent?
A: Intelligence refers to the calculation section of the ability to achieve the target. People, many animals and some machines have different classes
Don't vary from degrees of intelligence.
Q: Is there a "intelligence" in general, it does not depend on people's intelligence
A: No. We can't determine how the calculation process in the general sense can be called "intelligence".
We know some of the mechanisms of intelligence, and don't understand others.
Q: Is the intelligence isolated? For example, can it easily determine if a machine is intelligent?
A: No. Intelligence contains many mechanisms, and artificially intelligent researchers have already learned some parts of them, and
These mechanisms can be implemented with a computer, and other parts can not be used. If a task only needs now
Those mechanisms that have been depleted, then the computer program is quite well completed. Such a program
There is "some intelligence".
Q: Is artificial intelligence to simulate human intelligence?
A: Sometimes, but not always, or it is often not. On the one hand, we can observe humans.
The way to solve the problem is to understand the intelligence; on the other hand, the large number of artificial intelligence is not pass
Through the activities of people or animals, it is obtained by directly studying the needs of intelligent solutions. Artificial
Intelligent researchers can freely adopt methods of mankind and how to exceed human calculations.
Q: So IQ? Is there any computer program?
A: No. IQ is based on children's intelligence development. It corresponds to the score obtained by the test.
The age of age with his actual age. This measurement is promoted to adults in a relatively appropriate manner. IQ
High and low and success failures in actual life have a good relationship, but for the computer, otherwise, in IQ test
The high scores obtained in the test have nothing to do with the size of the actual use. For example, a child is reversed
The ability of a long series of sequence numbers can reflect his ability in other aspects, which may be because this can be measured.
The real-time computing power of this child. However, a very limited computer can complete its "number"
The sequence length of the range.
Of course, some of IQ test is also a beneficial challenge for artificial intelligence.
Q: What is more than other than intelligence between human and computers?
A: The most important researcher of human intelligence research Arthur R. Jensen [Jen98] proposed a "inspiration
False ": All normal people have the same intelligent mechanism, and the difference in intelligence is due to the difference between" number of biochemistry and physiology ". I think these conditions as speed, short time memory and formation accurate
The ability to recover long-term memory.
Jensen's hypothesis correct or not, no conclusion; however, now in artificial intelligence
But the opposite.
The computer has extremely high speeds and huge capacity, but its ability depends on the programmer's full understanding and
The intelligent mechanism implemented in the program. Some people who may have the ability to have the ability to have the ability to have the ability to have the ability to be teenagers.
The computer already has, while others have the ability to computers. Due to understanding science
What is unable to precisely define people's ability, we are more difficult to understand it. Generally speaking, people
The composition of intelligence mechanisms of workers is often different from the composition of the intelligent mechanism of people.
As long as people perform better than the computer, or the computer can only be able to
Make the same performance as people, explain the intelligent mechanism required by the program for efficient resolution.
The understanding is not sufficient.
Q: When is the study of artificial intelligence begin?
A: After World War II, many people started their own research on intelligent machines. British mathematician Alan Turing or
Xu is the first. He made a speech on this issue in 1947. He also confirmed the use of programmable computer for the first time
Studying artificial intelligence is much better than direct manufacturing machines. By the end of the 1950s, it has already been
Multi-artificial intelligence researchers, most of them work on programmable computers.
Q: Is artificial intelligence intend to implant people's "thinking" into your computer?
A: Some researchers say they have such a plan, but they may only use this word to make an alteration. Human
"Thought" has a lot of features, and I don't think there will be people really want to imitate all of the features.
Q: What is Tulex test?
A: Alan Turing 1950's article "Computing Machine and Intelligence" [TUR 50] Discussed what conditions under the machine
It can be considered to be intelligent. He believes that if a machine can successfully make a certain knowledge
The observer believes that he is a person, and it is of course we can think that this machine is intelligent. This test standard is
In addition to most people outside the philosophers, they are acceptable. The observer in the test can pass one
The teleprend typewriter interacts with the test machine and a person, which is to avoid the external table of the machine.
And sound. That person tried to convince the observer he is a personal; that machine is trying to deceive the observer. It is a person.
Tulex test is just a test. A machine that passes the Tulex test can of course be considered to have a wisdom.
Yes, but a machine that knows not much to humans and can't be very good, can also be considered to have a wisdom.
can. "Brainchildren [Den98]) written by Daniel Dennett
Tulex testing and various partial graphic tests (the knowledge and inquiries of observed observer in the test
The content is restricted) a very exciting discussion. The results show that some programs that have poor function will be subject to some people.
It is easy to think that it is intelligent.
Q: Is the goal of artificial intelligence to achieve the intelligence of human level?
A: Yes. The ultimate effort is to make computer programs can solve problems in the world like people.
the goal. However, many people do not have such a sense in the special research field.
Q: How far is the intelligent intelligence of manual intelligence to achieve human level? When will it become a reality?
A: There are a few people think that the intelligence of human level can be obtained in this way: is based on people now used
Writing and accumulating a large number of languages that can represent knowledge, use it to write a lot of programs,
To achieve the intelligence of human level.
However, most artificially intelligent researchers believe that some new basic ideas are needed. Therefore, it is still not possible to predict when a human level is intelligent.
Q: Is the computer used to generate a suitable machine?
A: The computer can be programmed to simulate any machine. Many researchers invented some non-mechanical machines
I hope they can have more intelligence than the computer in some respects. However, they usually in the computer
After simulating the machines they invented, they suspect whether this new machine is worth implementing. Because
Number of funds have enabled the speed of the computer faster and faster, and the performance of the actual implementation of the machine must be
It is better to achieve more performance than its simulation on your computer.
Q: Is it fast enough to get a smart, the speed of the current computer is enough?
A: Some people think that there is also a faster computer while needing new ideas. My personal point is that as long as
We know how to program intelligence, 30 years ago, the computer is already enough. Of course,
The computer will remain getting faster and faster outside the manual intelligent researcher.
Q: What about parallel computers?
A: The speed of the computer with multiple processors is much faster than the speed of the single processor computer. Parallel
There is no advantage, and it is difficult to use in parallel. Use only when you need a very high speed, use
This difficult machine is required.
Q: What is the "child-like machine" that can be self-improved by reading and learning from experience in experience?
A: This idea has been proposed many times since the 1940s. Finally, it will be manufactured.
However, the current artificial smart program cannot be effectively learned from natural experience as children. Current program
It is not good to understand the language, let alone learn from reading.
Q: A artificial intelligence system can continue to improve his wisdom by thinking about artificial intelligence.
Can it be horizontal?
A: I think it is ok. However, artificial intelligence has not yet reached the level of such a process.
Q: Can you talk about chess?
A: A Russian artificial intelligence researcher Alexander Kronrod believes "Chess is artificial intelligence research
Drosophila ". He has played more than one, and the fruit fly is often used by genetics.
There is a need for a smart mechanism, but there is no need for some other mechanisms. The program of chess can now reach
The master level, but compared with the human player, it only uses a very simple intelligent mechanism, which is used in large quantities.
Calculations instead of understanding. Once we have a better understanding of the smart mechanism of chess, we can do it.
A lot of chess programs that are less than the current calculation is now, but also with human chess.
Unfortunately, the manufacturing chess program is now more used as a competitive means of business, not
As a research method in the scientific field. This is like 1910, after the organization of Drosophila competition, genetics
They have concentrated force to breed fruit flies as winning the game.
Q: Do you talk about Go?
A: China and Japanese Go are also in one chessboard with two players. Go exposes us to us.
Insufficient understanding of the intelligent mechanism of human playback. Despite the considerable research (no in chess
There are many studies, but the playful chess game is still very poor. This problem may be due to a chess in Go.
The bureau can be divided into some sub -sso bureaus, then the independent analysis of each sub-chess office again
Inter-contact. People can use such a method while playing chess, but the Go program is not available, it can only put the whole
The chessboard is considered as a whole. The play chess program can only use a lot of calculations to make up for the lack of this intelligent mechanism.
Loss. This calculation is usually more than thousands of people, and for deep blue, it is millions of times.
There will always be one day, and artificially intelligent researchers can overcome this disadvantage. Q: Isn't some people think that artificial intelligence is a harmful idea?
A: The philosopher John Searle believes that non-biological machines will have intelligent ideas that are not logical. philosophy
Hubert Dreyfus believes that artificial intelligence is impossible. Computer scientist Joseph Weizenbaum
It is believed that artificial intelligence is an anti-human. Various people insist on since now artificial intelligence
It didn't reach the intelligence of human level, it must be impossible. There are still some people invested in them.
The company is disappointed.
Q: What is the key role in calculating theory and computational complexity? Is there a key role in artificial intelligence? [Note: Calculation theory
And computational complexity Theory is a quite skillful branch of mathematical logic and computer science, and solve it.
The problem also requires quite skills. ]
A: No. Those theoretical and artificial intelligence related, but it is not possible to solve the basic problems of artificial intelligence.
In the 1930s, the number of mathematics represented by Kurt Godel and Alan Turing was indeed
There is no algorithm to determine all the problems in certain important mathematics. First-order logic in the true and false
The determination is an example; a multi-variable polynomial equation is an integer solution is another example. Human one
Directly solve problems in these areas. This result gives such a view (usually exaggerated),
That is, the computer is inherently unable to do some people can do. However, people cannot guarantee
Humans can solve any of these areas.
In the 1960s, computer scientists represented by Steve Cook and Richard Karp
The theory of NP completeness was established. A complete problem with an NP is solved, but it is very likely that the needs and problem scale
Exponential time. The satisfaction of proposition in proposition calculations is a basic example of an NP full problem. people
Categories often solve certain special problems with a short time than the general algorithm, but for general problems
Is it unable to solve quickly.
It is important in artificial intelligence to get an algorithm that can solve problems like people. Confirmed algorithm is
No existence is important, but the solution of many artificial intelligence issues is not related to this area.
The theory of the difficulty of general various problems is called computational complexity theory. So far, this
Theory has not been interacting with artificial intelligence as once they expect. People and artificial intelligence programs in the problem
The methods and problems of the problem dependent on the success of the solution, it seems to be a complex researcher and person
Indecent researchers cannot be fully determined.
Complexity of algorithm developed by Solomonoff, Kolmogorov and Chaitin
Theory is also the same. It defines the complexity of a symbol object to generate the length of its shortest program. Already
It is impossible to find a shortest or close to the shortest program, but the shortest can be
The program that generates these symbols is often very inspiring, of course, you can't prove this program.
It is the shortest.
2 Hand-intelligent branch
Q: What is the branch of artificial intelligence?
A: Here is some, but some branches have not appeared here, because no one can confirm
They, some of which is considered conceptual or subject, not a complete branch.
Logical manual intelligence
In general, a program knows that the world's knowledge is in a specific situation,
In this case, the program must have actions, and its goal is to put all of these manager logic.
The sentence is expressed. The program decides what to do by reasoning the appropriate behavior of the goal. Treat this
The earliest article is [MCC59], and [MCC89] is a recent summary, [MCC96] lists some included
In the concept of logical manual intelligence, [sha97] is an important file.
search for
Artificial intelligence programs often search for a large number of possibilities, for example, movement in chess games or by theorem
Problem proceedings. In many different areas, people are continuing to find various methods to improve the efficiency of the program.
Pattern recognition
When a program observes a certain type of thing, it is usually designed to compare the subject matter and model. E.g,
A vision in order to find a face, it will try to match the mode of the eyes and nose in a scene. more
Complex mode, such as a natural language, a location in the chess game, or what is studied
History of pieces. These more complex patterns need more different than most simple models that have been studied.
method.
Indicate
Knowledge about the world must be expressed in some way. Usually used is the language of the proactive logic.
reasoning
From some facts, you can launch another fact. For some purposes, the number of logical interpretation is enough.
However, since the 1970s, a method called non-monoculated reasoning has been developed.
Common knowledge and common sense reasoning
This area is the farthest level of artificial intelligence, although this has been a non-in-law since the 1950s
Often active research areas. Although some considerable progress, such as non-monotonic reasoning systems and behaviors
The development of theory, but we need more new ideas. The CYC system contains a lot of but unreliable
A collection of facts.
Learn from experience
The program must do this. Connection and neural network processing of artificial intelligence is to do this. also
There is a study of laws that include logical expression. [Mit97] is a undergraduate textbook for machine learning. program
Can only learn the facts and behaviors that can be expressed by their form, but unfortunately, learning systems are
It is based on very limited ability to represent information.
planning
Planning procedures with respect to the world (especially the facts of behavior) or representation of special circumstances and objectives
The general fact begins. Then they generate a strategy to achieve this goal. In a lot of common situations,
This strategy is a series of behaviors.
Cognitive
This is a study on the types of knowledge we need to solve the problem in the world.
Presence
There is a type of thing that exists. In artificial intelligence, programs and statements have to handle a lot of objects, and I
We have to study the types of these objects and their basic properties. The focus of presence theory begins in the 1990s.
Heuristic
The heuristic is a way to try to find thoughts and other things in the program. In artificial intelligence, this technique
Quotes widely used. The heuristic function is used in the search for search, which is measured from the node in a search tree.
How far is the standard node. Heuristic predicates are used to compare one of two nodes in a search tree than another
A good, that is, the progress of the leading target. In accordance with my point of view, heuristic predicates are more useful.
Genetic programming
Genetic programming is a technology that selects the most in combination from thousands of generations through hybridized LISP procedures.
Appropriate procedures to complete tasks. John Koza Leading Group has developed a genetic programming, here comes with a copy
guide. (http://www.genetic-programming.com/gpanimatedtutorial.html).
Q: Where is the application of artificial intelligence?
A: More here is listed here
game
You can spend a few hundred dollars to buy a machine with master level chess game. There is
Artificial intelligence, but they mainly play with people through blind calculations - watching thousands of locations. Brilliant
Location calculation To defeat a world champion, the heuristic is required to observe 200 million locations per second.
Speech Recognition
In the 1990s, computer speech recognition has reached a practical level for some specific purposes. Although it is possible
Guide some computers to use voice, but for convenience, people will continue to use keyboards and mice.
Natural language understanding
Just enter a character sequence to the computer is not enough. Sentence analysis is not enough. Computer must
It can understand the field-related fields, and this is currently only possible to very limited fields.
Computer vision
The world is composed of three-dimensional objects, but the image in the eye and computer TV cameras are entered.
2D. Some useful programs work independently in twoD, but complete computer vision requires some three-dimensional information and more than some two-dimensional images. At present, there is only a limited way to directly represent the three-dimensional letter.
This is also good for human beings.
expert system
A "knowledge engineer" through an interview with experts in a field, trying to put their knowledge in computer program
It is reflected in a task. This works how to detriment to complete the intelligence mechanism required by the task is
No in the current state of manual intelligence. If this is proven to be like this, then there will be many letters.
The result of disappointing people. One of the earliest expert systems is Mycin that appeared in 1974. It used to diagnose blood?
Infected by bacterial, and it is recommended to treat. It works more than the medical school students and the doctors who have already worked
It is better, although people also discover its defects. As the name suggests, its existence includes bacterial, symptoms and treatment
Treatment, two do not include patients, doctors, hospitals, death, rehabilitation and real-time events. Its communication dependence
A simple patient considered. Because experts interviewed by knowledge engineers are familiar with patients, doctors, death
In the case of death, rehabilitation, etc., there must be knowledge engineers tell their experts to enter one.
In advanced frame. From the state of artificial intelligence, it is true. Currently
Applicability of home systems depends on users with common sense.
Heuristic classification
Using the current knowledge of artificial intelligence, one of the most feasible experts is to input the information to a fixed make
Use some of the types of original information. For example, it is recommended to accept the proposed credit card transaction. About the letter
Information with card owners, payment records, purchased items, and companies who purchase items should be available
(For example, information about whether this company has had credit card fraud).
: How to study is made by artificial intelligence.
A: The research of artificial intelligence includes two aspects of theory and experiment, and there are also foundations and theoretical points in the experiment.
Artificial intelligence research mainly has two clues. One is biological, based on such ideas:
However, human beings have intelligence, then artificial intelligence should learn human and simulate their psychology and physiology. another
The strip is phenomenological, based on the facts about the world's common sense and to achieve the goals
Research and formation. These two aspects are somewhat interacting, and the event will take
Successful. This is a racing, but two entrants seem to be walking.
Q: What must I learn before learning artificial intelligence?
A: Learn Mathematics, especially the logic. Generally speaking, the more scientific knowledge learns, the better.
Artificial intelligence with biological treatment must learn psychology and nervous system physiology. Society a few procedures
Design language, at least C, LISP, and ProLog. Learning a basic machine language is also a good idea.
Further work may depends on familiar with the current popular languages. At the end of the 1990s, C and Java were included.
Q: Is there those good textbooks about artificial intelligence?
A: Stuart Russell and Peter Norvig Written, Prentice Hall Press Publishing "Artificial
Intelligent "is the most common textbook 1997, and some general views and this article are inconsistent. NILS
"Artificial Intelligence: A New Comprehensive" published by Morgan Kaufman Press
More readily read.
Q: What are the organizations and publications related to artificial intelligence?
A: American Artificial Intelligence Association (AAAI) (http://www.aaai.org), European Artificial Intelligence Cooperation
ECCAI (http://www.eccai.org/) and Artificial Intelligence 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 a special and interesting team related to artificial intelligence, called Sigart
(http://www.acm.org/sigart). International Artificial Intelligence Federation (IJCAI) (http://www.ijcai.org) is the most important international conference.
"Electronic Intelligence Electronics" 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" 人杂
Artificial Intelligence Research (http://www.jair.org/) and "IEEE mode analysis
And Machine Intelligence Journal of IEEE Transactions on Pattern Analysis and Machine
Intelligence (http://computer.org/tpami/) is the most master of four published artificial intelligence research papers
Magazine must be. All information I can find is already included here.
Home Positive Reviews (http://www.cs.utexas.edu/Users/vl/ppr/) lists one
Some experts think important papers.
.
