Artificial Intelligence

"Artificial intelligence" is the ability of machines to do things that people would say require intelligence. Artificial intelligence (AI) research is an attempt to discover and describe aspects of human intelligence that can be simulated by machines. For example, at present there are machines that can do the following things: 

1. Play games of strategy (e.g., Chess, Checkers, Poker) and (in Checkers) learn to play better than people. 

2. Learn to recognize visual or auditory patterns. 

3. Find proofs for mathematical theorems. 

4. Solve certain, well-formulated kinds of problems. 

5. Process information expressed in human languages. 

The extent to which machines (usually computers) can do these things independently of people is still limited; machines currently exhibit in their behavior only rudimentary levels of intelligence. Even so, the possibility exists that machines can be made to show behavior indicative of intelligence, comparable or even superior to that of humans.' Alternatively, AI research may be viewed as an attempt to develop a mathematical theory to describe the abilities and actions of things (natural or man-made) exhibiting "intelligent" behavior, and serve as a calculus for the design of intelligent machines. As yet there is no "mathematical theory of intelligence," and researchers dispute whether there ever will be. This book serves as an introduction to research on machines that display intelligent behavior. Such machines some fimes will be called "artificial intelligence's," "intelligent machines," or "mechanical intelligence's." The inclination in this book is toward the first viewpoint of AI research, without forsaking the second. Since AI research is still in its infancy, it is therefore prudent to withhold estimation of its future. It is best to begin with a summation of present knowledge, considering such questions as: 

1. What is known .about natural intelligence? 

2. When can we justifiably call a machine intelligent?

3. How and to what extent do machines currently simulate intelligence or display intelligent behavior? 

4. How might machines eventually simulate intelligence? 

5. How can machines and their behavior be described mathematically? 

 6. What uses could be made of intelligent machines? 

Each of these questions will be explored in some detail in this book. The first and second questions are covered in this chapter. It is hoped that the six questions are covered individually in enough detail so that the reader will be guided to broader study if he is so inclined. For parts of this book, some knowledge of mathematics (especially sets, functions, and logic) is presupposed, though much of the book is understandable without it.  

TURING'S TEST 

A basic goal of AI research is to construct a machine that exhibits the behavior associated with human intelligence, that is, comparable to the intelligence of a human being. It is not required that the machine use the same underlying mechanisms (whatever they are) that are used in human cognition, nor is it required that the machine go through stages of development or learning such as those through which people progress. The classic experiment proposed for determining whether a machine possesses intelligence on a human level is known as Turing's test (after A. M. Turing, who pioneered research in computer logic, undecidability theory, and artificial intelligence). This experiment has yet to be performed seriously, since no machine yet displays enough intelligent behavior to be able to do well in the test. Still, Turing's test is the basic paradigm for much successful work and for many experiments in machine intelligence, from the Samuel's Checkers Player to "semanticinformation processing" programs such as Colby's PARRY or Raphael's.

Basically, Turing's test consists of presenting a human being, A, with a typewriter-like or TV-like terminal, which he can use to converse with two unknown (to him) sources, B and C ). The interrogator A is told that one terminal is controlled by a machine and that the other terminal is controlled by a human being whom A has never met. A is to guess which of B and C is the machine and which is the person. If A cannot distinguish one from the other with significantly better than 50% accuracy, and if this result continues to hold no matter what people are involved in the experiment, the machine is said to simulate human intelligence. 

Some comments on Turing's test are in order. First, the nature of Turing's test is such that it does not permit the interrogator A to observe the physical natures of B and C; rather, it permits him only to observe their "intellectual behavior," that is, their ability to communicate with formal symbols and to "think abstractly." So, while the test does not enable A to be prejudiced by the physical nature of either B or C, neither does it give a way to compare those aspects of an entity's behavior that reflect its ability to act non abstractly in the real world-that is, to be ·intelligent in its performance of concrete operations on objects. Can the machine, for example, fry an egg or clean a house? Second, one possible achievement of AI research would be to produce a complete description of a machine that can successfully pass Turing's test, or to find a proof that no machine can pass it. The complete description must be of a machine that can actually be constructed. A proof that there is no such constructive machine (it might say, e.g., "The number of parts in such a machine must be greater than the number of electrons in the universe.") is consequently to be regarded as a proof of the "no machine" alternative. Third, it may be that more than one type of machine can pass Turing's test. In this case, AI research has a secondary problem of creating a general description of all machines that will successfully pass Turing's test. Fourth, if a machine passes Turing's test, it means in effect that there is at least one machine that can learn to solve problems as well as a human being. This would lead to asking if a constructive machine can be described which would be capable of learning to solve not only those problems that people can usually solve, but also those that people create but can only rarely solve. That is, is it possible to build mechanical intelligence's that are superior to human intelligence? It is not yet possible to give a definite answer to any of these questions. Some evidence exists that AI research may eventually attain at least the goal of a machine that passes Turing's test. It is clear that the intellectual capabilities of a human being are directly related to the functioning of his brain, which appears to be a finite structure of cells. Moreover, people have succeeded in constructing machines that can "learn" to produce solutions to certain specific intellectual problems,· which are superior to the solutions people can produce. The most notable example is Samuel's Checkers Player, which has learned to play a better game of Checkers than its designer, and which currently plays at a championship level.