Artificial Intelligence

Questions on this topic? Just ask!

Artificial Intelligence

Artificial Intelligence (AI) is a subfield of computer science that is concerned with the representation, study, and automation of knowledge and intelligent reasoning. One of the main goals of AI is to create computer programs that exhibit "intelligent" behavior. The AI field is a vast one, encompassing a variety of problem domains, methodologies, and applications. Some subjects that currently fall under the rubric of AI include automated planning, robotics, computer vision, natural language understanding, speech and pattern recognition, neural networks, machine learning, rule-based expert systems, knowledge representation, and automatic theorem proving.

The birth of AI is said to have occurred at a conference held at Dartmouth college in the summer of 1956. The conference was organized by Marvin Minsky, who later helped found the AI laboratory at the Masschusetts Institute of Technology (MIT) and who is currently at the MIT Media Laboratory. Also present was John McCarthy, creator of the LISP programming language that is still an important tool for AI research and development. McCarthy himself at that time named the field "artificial intelligence." Herbert Simon and Allen Newell, who had already implemented an automatic theorem-proving program for the Rand corporation called the Logic Theorist, were also present. These four--Minsky, McCarthy, Newell, and Simon--are considered the "fathers" of AI. McCarthy went on to found an AI lab at Stanford while Simon and Newell did the same at Carnegie Mellon. At the Dartmouth conference were also assembled a number of other fledgling researchers who had all written programs exhibiting the rudiments of intelligent behavior, and it was the collective efforts of these individuals that gave impetus to the young field of AI.

An important early contribution to the philosophy of AI was made by the brilliant British mathematician Alan Turing in his 1950 paper "Computing Machinery and Intelligence." Here Turing proposed the now-famous Turing Test, which stands to this day as an ideal measure of intelligent behavior: a machine is "intelligent" if it can succeed in an imitation game in which a human being and the machine itself both communicate with a human interrogator in a separate room only through written questions and responses. If the interrogator cannot, through these written questions and answers, tell the human apart from the machine, then the machine must be considered "intelligent." Of course, no computer built or even conceived can actually pass the Turing test; computers are very stupid conversationalists. (Today an annual contest is held to see which AI system can fail the Turing test least badly.) The Turing Test is a philosophical construct, not a workaday benchmark for measuring system performance. However, the concept of the Test has provoked much discussion, some of it helpful, of just what is meant by the word "intelligence."

Much modern AI research has been concerned with formally characterizing or describing the reasoning processes humans use to solve problems, and in creating very general architectures or theoretical frameworks to encode such reasoning. One approach is to view all problems as essentially search problems that must progress through a state-space of partial solutions. An AI system's task is thus to move from some initial state of affairs (initial partial solution) to a goal, or final complete solution. The process of moving from partial solution to partial solution is driven by a set of rules which can be thought of as a set of if-then statements. That is, if a partial solution satisfies a certain hypothesis (the condition of the if-statement), then that the given partial solution is accepted as one that is closer to the goal. This type of system is called a rule-based or production system--"production" because each rule produces a different partial solution--and this type of reasoning is known as forward chaining or deduction. Production systems can also support a different type of reasoning known as backward chaining, in which the various rules are applied backwards through the state-space from the goal toward the initial partial solution.

A number of problems can be solved using this rule-based framework, provided that the representation of knowledge in the encoding of partial solutions, and the rules for moving through the state-space of partial solutions, are chosen judiciously. One of the first examples of such a rule-based system was a simple software robot capable of moving around blocks in an imaginary block world. The state of the block world was encoded by a list of which blocks were stacked on each other and which block, if any, the robot arm had in its possession. The goal would a prespecified configuration of blocks, and rules were given to allow the robot arm to rearrange blocks from any start configuration to achieve the final configuration (without toppling the blocks).

A more realistic application is the well-known system Mycin, which is capable of diagnosing bacterial infections. The input to Mycin is the set of symptoms a patient experiences. Based on a set of rules that are triggered by each symptom, Mycin suggests the particular type of bacteria that could be the cause of the problem. An AI system that has a limited domain of expertise, in this case bacterial infections, is known as an expert system. Although expert systems could not pass the Turing test because of their limited domain of expertise, they have nevertheless become an important part of AI research because of their immediate commercial value. Another aspect of AI whose development has received a large impetus from work on rule-based and expert systems is the field of knowledge representation, which is concerned with the creation of data structures to efficiently encode the knowledge that is relevant to solving specific problems. Useful conceptual tools developed in this field include the notions of frames and inheritance, which apply to the organization of data structures and their internal relationships.

Whereas rule-based systems attempt to formally capture logical reasoning, their underlying architecture bears very little resemblance to the human brain--an intricate network of relatively simple processing units (nerve cells) called neurons. Another aspect of AI, known as parallel distributed algorithms, seeks to create intelligent systems whose architecture mimics the distributed nature of the human brain. Neural networks have been successfully created and trained to accomplish a host of pattern recognition tasks. The mimicking of nature in AI does not stop there. The field of genetic algorithms (GAs) mimics evolution by using the process of natural selection to solve complicated optimization and search problems. A GA maintains a whole population of candidate solutions to the optimization problem at any given time. By picking from among the best solutions in the population at any given time according to a prespecified fitness function and by combining or mating these fittest members to produce new and hopefully better offspring, the GA algorithm efficiently explores the search space and yields optimized solutions after many iterations or generations. These biologically inspired methods form an interesting complement to the more formal methods initially explored in AI.

AI is an interdisciplinary subject that draws on the methodologies and expertise of many different fields, including computer science, biology, cognitive science, mathematics, linguistics, philosophy, and physics. It is likely that knowledge from all of these fields will be useful, even essential, in searching for the holy grail of AI: a machine that can pass the Turing test. In the last 50 or so years, AI has produced a wealth of intelligent applications that can achieve many tasks from diagnosing disease to beating grandmasters in chess. It is worth noting that progress in AI has been much slower than was anticipated early in its history. Herbert Simon predicted in 1957 that "a digital computer [would] be the world's chess champion" within 10 years, but it was not until 1997--30 years behind schedule--that a computer beat Gary Kasparov, reigning world chess champion (with the help of a team of human operators). Nevertheless, the insights gained from AI research have lead to advances in the understanding of human reasoning and intelligence, and no doubt many valuable applications will continue to be forthcoming.