Artificial Intelligence

views updated

Artificial intelligence

Certain tasks can be performed faster and more accurately by traditionally programmed computers than by human beings, particularly numerical computation and the storage, retrieval, and sorting of large quantities of information. However, the ability of computers to interact flexibly with the real world—their "intelligence"—remains slight. Artificial intelligence (AI) is a subfield of computer science that seeks to remedy this situation by creating software and hardware that possess some of the behavioral flexibility shown by natural intelligences (people and animals).

In the 1940s and 1950s, the first large, electronic, digital computers were designed to accomplish specific tasks (e.g., a numerical calculation set up by a human programmer) by completing a series of clearly defined steps, an algorithm . Programmers wrote algorithmic software that precisely specified both the problem and how to solve it. AI programmers, in contrast, seek to program computers not with rigid algorithms but with flexible rules for seeking solutions. An AI program may even be designed to modify the rules it is given or to develop entirely new rules.

What types of problem are appropriate for traditional, algorithmic computing and what types call out for AI? Some of the tasks that are hardest for people are, fortunately, algorithmic in nature. Teachers can describe in detail the process for multiplying numbers, and accountants can state accurately the rules for completing tax forms, yet many people have difficulty performing such tasks. Straightforward, algorithmic programs can perform them easily because they can be broken down into a series of precise procedures or steps that do not vary from case to case. On the other hand, tasks that require little thought for human beings can be hard to translate into algorithms and therefore difficult for computers to perform. For example, most people know that a pot of boiling water requires careful handling. We identify hot pots by flexible recognition of many possible signs: steam rising, radiant heat felt on the skin, a glimpse of blue flame or red coils under the pot, a rattling lid, and so forth. Once we know that the pot is boiling, we plan our actions accordingly. This process seems simple, yet describing exactly to a computer how to reliably conclude "this pot is hot" and to take appropriate action turns out to be extremely difficult. The goal of AI is to create computers that can handle such complex, flexible situations. One obstacle to this goal is uncertainty or confusion about what is intelligence.

What is intelligence?

One possible definition of intelligence is the acquisition and application of knowledge. An intelligent entity, on this view, is one that learns—acquires knowl edge—and is able to apply this knowledge to changing real-world situations. In this sense, a rat is intelligent, but most computers, despite their impressive numbercrunching capabilities, are not. To qualify as intelligent, an AI system must use knowledge (whether acquired from databases, sensory devices, trial and error, or all of the above) to make effective choices when confronted with data that are to some extent unpredictable Insofar as a computer can do this, it may be said, for the purposes of AI, to display intelligence. Note that this definition is purely functional, and that in AI the question of consciousness, though intriguing, need not be considered.

This limited characterization of intelligence would, perhaps, have been considered overcautious by some AI researchers in the early days, when optimism ran high. For example, U.S. economist, AI pioneer, and Nobel Prize winner Herbert Simon (1916–2001) predicted in 1965 that by 1985 "machines will be capable of doing any work man can do." Yet over 35 years later, despite exponential growth in memory size and processing speed, no computer even comes close to commonplace human skills like conversing, driving a car, or diapering a baby, much less "doing any work" a human being can do. Why has progress in AI been so slow?

One answer is that while intelligence, as defined above, requires knowledge, computers are only good at handling information, which is not the same thing. Knowledge is meaningful information, and "meaning" is a nonmeasurable, multivalued variable arising in the real world of things and values. Bits—"binary digits," 1s and 0s—have no meaning, as such; they are meaningful only when people assign meanings to them. Consider a single bit, a "1": its information content is one bit regardless of what it means, yet it may mean nothing or anything, including "The circuit is connected," "We surrender," "It is more likely to rain than snow," and "I like apples." The question for AI is, how can information be made meaningful to a computer? Simply adding more bits does not work, for meaning arises not from information as such, but from relationships involving the real world. In one form or another, this basic problem has been stymieing AI research for decades. Nor is it the only such problem. Another problem is that computers, even those employing "fuzzy logic" and autonomous learning , function by processing symbols (e.g., 1s and 0s) according to rules (e.g., those of Boolean algebra)—yet human beings do not usually think by processing symbols according to rules. Humans are able think this way, as when we do arithmetic , but more commonly we interpret situations, leap to conclusions, utter sentences, and plan actions using thought processes that do not involve symbolic computation at all. What our thought processes really are—that is, what our intelligence is, preicsely—and how to translate it into (or mimic it by means of) a system of computable symbols and rules is a problem that remains unsolved, in general, by AI researchers.

In 1950, British mathematician Alan Turing (1912–1954) proposed a hypothetical game to help decide the issue of whether a given machine is truly intelligent. The "imitation game," as Turing originally called it, consisted of a human questioner in a room typing questions on a keyboard. In another room, an unseen respondents—either a human or a computer—would type back answers. The questioner could pose queries to the respondent in an attempt to determine if he or she was corresponding with a human or a computer. In Turing's opinion, if the computer could fool the questioner into believing that he or she was having a dialog with a human being then the computer could be said to be truly intelligent.

The Turing test is obviously biased toward human language prowess, which most AI programs do not even seek to emulate. Nevertheless, it is significant that even the most advanced AI programs devoted to natural language are as far as ever from passing the Turing test. "Intelligence" has proved a far tougher nut to crack than the pioneers of AI believed, half a century ago. Computers remain unintelligent.

Even so, AI has made many gains since the 1950s. AI software is now present in many devices, such as automatic teller machines (ATMs), that are part of daily life, and is finding increasing commercial application in many industries.

Overview of AI

All AI computer programs are built on two basic elements: a knowledge base and an inferencing capability. (Inferencing means drawing conclusions based on logic and prior knowledge.) A knowledge base is made up of many discrete units of information—representing facts, concepts, theories, procedures, and relationships—all relevant to a particular task or aspect of the world. Programs are written to give the computer the ability to manipulate this information and to reason, make judgments, reach conclusions, and choose solutions to the problem at hand, such as guessing whether a series of credit-card transactions involves fraud or driving an automated rover across a rocky Martian landscape. Whereas conventional, deterministic software must follow a strictly logical series of steps to reach a conclusion, AI software uses the techniques of search and pattern matching; it may also, in some cases, modify its knowledge base or its own structure ("learn"). Pattern matching may still be algorithmic; that is, the computer must be told exactly where to look in its knowledge base and what constitutes a match. The computer searches its knowledge base for specific conditions or patterns that fit the criteria of the problem to be solved. Microchip technology has increased computational speed, allowing AI programs to quickly scan huge arrays of data. For example, computers can scan enough possible chess moves to provide a challenging opponent for even the best human players. Artificial intelligence has many other applications, including problem solving in mathematics and other fields, expert systems in medicine, natural language processing, robotics , and education.

The ability of some AI programs to solve problems based on facts rather than on a predetermined series of steps is what most closely resembles "thinking" and causes some in the AI field to argue that such devices are indeed intelligent.

General problem solving

Problem solving is thus, something AI does very well as long as the problem is narrow in focus and clearly defined. For example, mathematicians, scientists, and engineers are often called upon to prove theorems. (A theorem is a mathematical statement that is part of a larger theory or structure of ideas.) Because the formulas involved in such tasks may be large and complex, this can take an enormous amount of time, thought, and trial and error. A specially designed AI program can reduce and simplify such formulas in a fraction of the time needed by human workers.

Artificial intelligence can also assist with problems in planning. An effective step-by-step sequence of actions that has the lowest cost and fewest steps is very important in business and manufacturing operations. An AI program can be designed that includes all possible steps and outcomes. The programmer must also set some criteria with which to judge the outcome, such as whether speed is more important than cost in accomplishing the task, or if lowest cost is the desired result, regardless of how long it takes. The plan generated by this type of AI program will take less time to generate than by traditional methods.

Expert systems

The expert system is a major application of AI today. Also known as knowledge-based systems, expert systems act as intelligent assistants to human experts or serve as a resource to people who may not have access to an expert. The major difference between an expert system and a simple database containing information on a particular subject is that the database can only give the user discrete facts about the subject, whereas an expert system uses reasoning to draw conclusions from stored information. The purpose of this AI application is not to replace our human experts, but to make their knowledge and experience more widely available.

An expert system has three parts: knowledge base, inference engine, and user interface. The knowledge base contains both declarative (factual) and procedural (rules-of-usage) knowledge in a very narrow field. The inference engine runs the system by determining which procedural knowledge to access in order to obtain the appropriate declarative knowledge, then draws conclusions and decides when an applicable solution is found.

An interface is usually defined as the point where the machine and the human "touch." An interface is usually a keyboard, mouse, or similar devices. In an expert system, there are actually two different user interfaces: One is for the designer of the system (who is generally experienced with computers) the other is for the user (generally a computer novice). Because most users of an expert system will not be computer experts, it is important that system be easy for them to use. All user interfaces are bi-directional; that is, are able to receive information from the user and respond to the user with its recommendations. The designer's user interface must also be capable of adding new information to the knowledge base.

Natural language processing

Natural language is human language. Natural-language-processing programs use artificial intelligence to allow a user to communicate with a computer in the user's natural language. The computer can both understand and respond to commands given in a natural language.

Computer languages are artificial languages, invented for the sake of communicating instructions to computers and enabling them to communicate with each other. Most computer languages consist of a combination of symbols, numbers, and some words. These languages are complex and may take years to master. By programming computers (via computer languages) to respond to our natural languages, we make them easier to use.

However, there are many problems in trying to make a computer understand people. Four problems arise that can cause misunderstanding: (1) Ambiguity—confusion over what is meant due to multiple meanings of words and phrases. (2) Imprecision—thoughts are sometimes expressed in vague and inexact terms. (3) Incompleteness—the entire idea is not presented, and the listener is expected to "read between the lines." (4) Inaccuracy—spelling, punctuation, and grammar problems can obscure meaning. When we speak to one another, furthermore, we generally expect to be understood because our common language assumes all the meanings that we share as members of a specific cultural group. To a nonnative speaker, who shares fewer of our cultural background, our meaning may not always be clear. It is even more difficult for computers, which have no share at all in the real-world relationships that confer meaning upon information, to correctly interpret natural language.

To alleviate these problems, natural language processing programs seek to analyze syntax—the way words are put together in a sentence or phrase; semantics—the derived meaning of the phrase or sentence; and context—the meaning of distinct words within a sentence. But even this is not enough. The computer must also have access to a dictionary which contains definitions of every word and phrase it is likely to encounter, and may also use keyword analysis—a pattern-matching technique in which the program scans the text, looking for words that it has been programmed to recognize. If a keyword is found, the program responds by manipulating the text to form a reasonable response.

In its simplest form, a natural language processing program works like this: a sentence is typed in on the keyboard; if the program can derive meaning—that is, if it has a reference in its knowledge base for every word and phrase—it will respond, more or less appropriately. An example of a computer with a natural language processor is the computerized card catalog available in many public libraries. The main menu usually offers four choices for looking up information: search by author, search by title, search by subject, or search by keyword. If you want a list of books on a specific topic or subject you type in the appropriate phrase. You are asking the computer—in English—to tell you what is available on the topic. The computer usually responds in a very short time—in English—with a list of books along with call numbers so you can find what you need.

Computer vision

Computer vision is the use of a computer to analyze and evaluate visual information. A camera is used to collect visual data. The camera translates the image into a series of electrical signals. This data is analog in nature—that is, it is directly measurable and quantifiable. A digital computer, however, operates using numbers expressed directly as digits. It cannot read analog signals, so the image must be digitized using an analog-to-digital converter. The image becomes a very long series of binary numbers that can be stored and interpreted by the computer. Just how long the series is depends on how densely packed the pixels are in the visual image. To get an idea of pixels and digitized images, take a close look at a newspaper photograph. If you move the paper very close to your eyes, you will notice that the image is a sequence of black and white dots—called pixels, for picture elements—arranged in a certain pattern. When you move the picture away from your eyes, the picture becomes clearly defined and your brain is able to recognize the image.

Artificial intelligence works much the same way. Clues provided by the arrangement of pixels in the image give information as to the relative color and texture of an object, as well as the distance between objects. In this way, the computer can interpret and analyze visual images. In the field of robotics, visual analysis is very important.

Robotics

Robotics is the study of robots, which are machines that can be programmed to perform manual tasks. Most robots in use today perform various functions in an industrial setting. These robots typically are used in factory assembly lines, by the military and law enforcement agencies, or in hazardous waste facilities handling substances far too dangerous for humans to handle safely.

Most robots do not resemble the humanoid creations of popular science fiction. Instead, they usually consist of a manipulator (arm), an end effector (hand), and some kind of control device. Industrial use robots generally are programmed to perform repetitive tasks in a highly controlled environment. However, more research is being done in the field of intelligent robots that can learn from their environment and move about it autonomously. These robots use AI programming techniques to understand their environment and make appropriate decisions based on the information obtained. In order to learn about one's environment, one must have a means of sensing the environment. Artificial intelligence programs allow the robot to gather information about its surroundings by using one of the following techniques: contact sensing, in which a robot sensor physically touches another object; noncontact sensing, such as computer vision, in which the robot sensor does not physically touch the object but uses a camera to obtain and record information; and environmental sensing, in which the robot can sense external changes in the environment, such as temperature or radiation .

The most recent robotics research centers around mobile robots that can cope with environments that are hostile to humans, such as damaged nuclear-reactor cores, active volcanic craters, or the surfaces of other planets.

Computer-assisted instruction

Intelligent computer-assisted instruction (ICAI) has three basic components: problem-solving expertise, student model, and tutoring module. The student using this type of program is presented with some information from the problem-solving expertise component. This is the knowledge base of this type of AI program. The student responds in some way to the material that was presented, either by answering questions or otherwise demonstrating his or her understanding. The student model analyzes the student's responses and decides on a course of action. Typically this involves either presenting some review material or allowing the student to advance to the next level of knowledge presentation. The tutoring module may or may not be employed at this point, depending on the student's level of mastery of the material. The system does not allow the student to advance further than his or her own level of mastery.

Most ICAI programs in use today operate in a set sequence of presentation of new material, evaluation of student response, and employment of tutorial (if necessary). However, researchers at Yale University have created software that uses a more Socratic way of teaching. These programs encourage discovery and often will not respond directly to a student's questions about a specific topic. The basic premise of this type of computer-assisted learning is to present new material only when a student needs it. This is when the brain is most ready to accept and retain the information. This is exactly the scenario most teachers hope for: students who become adroit self-educators, enthusiastically seeking the wisdom and truth that is meaningful to them. The cost of these programs, however, can be far beyond the means of many school districts. For this reason, these types of ICAI are used mostly in corporate training settings.