Artificial Intelligence

History

The study of artificial intelligence has a long history, dating back to the work of British mathematician Charles Babbage (1791–1871) who developed a special-purpose "Difference Engine" for mechanically computing the values of certain polynomial functions. Similar work was also done by German mathematician Gottfried Wilhem von Leibniz (1646–1716), who introduced the first system of formal logic and constructed machines for automating calculation. George Boole, Ada Byron King, Countess of Lovelace, Gottlob Frege, and Alfred Tarski have all significantly contributed to the advancement of the field of artificial intelligence.

Knowledge Representation

It has long been recognized that the language and models used to represent reality profoundly impact one's understanding of reality itself. When humans think about a particular system, they form a mental model of that system and then proceed to discover truths about the system. These truths lead to the ability to make predictions or general statements about the system. However, when a model does not sufficiently match the actual problem, the discovery of truths and the ability to make predictions becomes exceedingly difficult.

A classic example of this is the pre-Copernican model in which the Sun and planets revolved around the Earth. In such a model, it was prohibitively difficult to predict the position of planets. However, in the Copernican revolution this Earth-centric model was replaced with a model where the Earth and other planets revolved around the Sun. This new model dramatically increased the ability of astronomers to predict celestial events.

Arithmetic with Roman numerals provides a second example of how knowledge representation can severely limit the ability to manipulate that knowledge. Both of these examples stress the important relationship between knowledge representation and thought.

In AI, a significant effort has gone into the development of languages that can be used to represent knowledge appropriately. Languages such as LISP, which is based on the lambda calculus, and Prolog, which is based on formal logic, are widely used for knowledge representation. Variations of predicate calculus are also common languages used by automated reasoning systems. These languages have well-defined semantics and provide a very general framework for representing and manipulating knowledge.

Knowledge Manipulation

Many problems that humans are confronted with are not fully understood. This partial understanding is reflected in the fact that a rigid algorithmic solution—a routine and predetermined number of computational steps— cannot be applied. Rather, the concept of search is used to solve such problems. When search is used to explore the entire solution space, it is said to be exhaustive. Exhaustive search is not typically a successful approach to problem solving because most interesting problems have search spaces that are simply too large to be dealt with in this manner, even by the fastest computers. Therefore, if one hopes to find a solution (or a reasonably good approximation of a solution) to such a problem, one must selectively explore the problem's search space.

The difficulty here is that if part of the search space is not explored, one runs the risk that the solution one seeks will be missed. Thus, in order to ignore a portion of a search space, some guiding knowledge or insight must exist so that the solution will not be overlooked. Heuristics is a major area of AI that concerns itself with how to limit effectively the exploration of a search space. Chess is a classic example where humans routinely employ sophisticated heuristics in a search space. A chess player will typically search through a small number of possible moves before selecting a move to play. Not every possible move and countermove sequence is explored. Only reasonable sequences are examined. A large part of the intelligence of chess players resides in the heuristics they employ.

A heuristic-based search results from the application of domain or problem-specific knowledge to a universal search function. The success of heuristics has led to focusing the application of general AI techniques to specific problem domains. This has led to the development of expert systems capable of sophisticated reasoning in narrowly defined domains within fields such as medicine, mathematics, chemistry, robotics, and aviation.

Another area that is profoundly dependent on domain-specific knowledge is natural language processing. The ability to understand a natural language such as English is one of the most fundamental aspects of human intelligence, and presents one of the core challenges for the AI community. Small children routinely engage in natural language processing, yet it appears to be almost beyond the reach of mechanized computation. Over the years, significant progress has been made in the ability to parse text to discover its syntactic structure. However, much of the meaning in natural language is context-dependent as well as culture-dependent, and capturing such dependencies has proved highly resistant to automation.

The Turing Test

At what point does the behavior of a machine display intelligence? The answer to this question has raised considerable debate over the definition of intelligence itself. Is a computer capable of beating the world chess champion considered intelligent? Fifty years ago, the answer to this question would most likely have been yes. Today, it is disputed whether or not the behavior of such a machine is intelligent. One reason for this shift in the definition of intelligence is the massive increase in computational power that has occurred over the past fifty years, allowing the chess problem space to be searched in an almost exhaustive manner.

Two key ingredients are seen as essential to intelligent behavior: the ability to learn and thereby change one's behavior over time, and synergy, or the idea that the whole is somehow greater than the sum of its parts.

In 1950 British mathematician Alan Turing proposed a test for intelligence that has, to some extent, withstood the test of time and still serves as a litmus test for intelligent behavior. Turing proposed that the behavior of a machine could be considered intelligent if it was indistinguishable from the behavior of a human. In this imitation game, a human interrogator would hold a dialogue via a terminal with both a human and a computer. If, based solely on the content of the dialogue, the interrogator could not distinguish between the human and the computer, Turing argued that the behavior of the computer could be assumed to be intelligent.

Opponents of this definition of intelligence argue that the Turing Test defines intelligence solely in terms of human intelligence. For example, the ability to carry out complex numerical computation correctly and quickly is something that a computer can do easily but a human cannot. Given that, is it reasonable to use this ability to distinguish between the behavior of a human and a computer and conclude that the computer is not intelligent?

SEE ALSO ASSISTIVE COMPUTER TECHNOLOGY FOR PERSONS WITH DISABILITIES; LISP; OPTICAL CHARACTER RECOGNITION; ROBOTICS; ROBOTS.

Victor L. Winter

Bibliography

Luger, George F., and William A. Stubblefield. Artificial Intelligence: Structures and Strategies for Complex Problem Solving. Redwood City, CA: Benjamin/Cummings Publishing Company, 1993.

Mueller, Robert A., and Rex L. Page. Symbolic Computing with LISP and Prolog. New York: Wiley and Sons, 1988.

Russel, Stuart J., and Peter Norvig. Artificial Intelligence: A Modern Approach. Englewood Cliffs, NJ: Prentice Hall, 1994.