What We Mean by “Machine Intelligence”

We use the term “machine intelligence” to refer to the broad vision articulated in Alan Turing’s 1948 paper entitled “Intelligent Machinery” and his 1950 paper entitled “Computing Machinery and Intelligence.”

In the 1950s, the terms “machine intelligence,” “artificial intelligence,” and “machine learning” all referred to the goal of getting “machines to exhibit behavior, which if done by humans, would be assumed to involve the use of intelligence” (to quote Arthur Samuel, 1983).

However, in the intervening five decades, the terms “artificial intelligence” and “machine learning” progressively diverged from their original goal-oriented meaning. These terms are now primarily associated with particular methodologies for attempting to achieve the goal of getting computers to automatically solve problems. Thus, the term “artificial intelligence” is today primarily associated with attempts to get computers to solve problems using methods that rely on knowledge, logic, and various analytical and mathematical methods. The term “machine learning” is today primarily associated with attempts to get computers to solve problems that use a particular small and somewhat arbitrarily chosen set of methodologies (many of which are statistical in nature). The narrowing of these terms is in marked contrast to the broad field envisioned by Samuel at the time when he coined the term “machine learning” in the 1950s, the broad charter originally articulated by the founders of the field of artificial intelligence, and the broad vision encompassed by Turing’s term “machine intelligence.”

Of course, the shift in focus from broad goals to narrow methodologies is an all-too-common sociological phenomenon in academic research.

Turing’s term “machine intelligence” did not undergo this arteriosclerosis because, by accident of history, it was never appropriated and never became monopolized by any group of academic researchers whose primary dedication was to a particular methodological approach. Thus, Turing’s term remains catholic today. For this reason, we prefer to use Turing’s term “machine intelligence” to describe genetic programming because Turing’s term still communicates the broad goal of getting computers to automatically solve problems in a human-like way.

In his 1948 paper, Turing identified three broad approaches by which human-competitive machine intelligence might be achieved.

The first approach was a logic-driven search. Turing’s interest in this approach is not surprising in light of Turing’s own pioneering work in the 1930s on the logical foundations of computing.

The second approach for achieving machine intelligence was what he called a “cultural search” in which previously acquired knowledge is accumulated, stored in libraries, and brought to bear in solving a problemľthe approach taken by modern knowledge-based expert systems.

Turing’s first two approaches have been pursued over the past 50 years by the vast majority of researchers using the methodologies that are today primarily associated with the term “artificial intelligence.”

However, Turing also identified a third approach to machine intelligence in his 1948 paper entitled “Intelligent Machinery” (Turing 1948, page 12; Ince 1992, page 127; Meltzer and Michie 1969, page 23), saying:

“There is the genetical or evolutionary search by which a combination of genes is looked for, the criterion being the survival value.” (Emphasis added).

Turing did not specify in 1948 how to conduct the “genetical or evolutionary search” for solutions to problems. In particular, did not mention the concept of a population or recombination. However, he did point out in his 1950 paper “Computing Machinery and Intelligence” (Turing 1950, page 456; Ince 1992, page 156):

“We cannot expect to find a good child-machine at the first attempt. One must experiment with teaching one such machine and see how well it learns. One can then try another and see if it is better or worse. There is an obvious connection between this process and evolution, by the identifications

“Structure of the child machine = Hereditary material

“Changes of the child machine = Mutations

“Natural selection = Judgment of the experimenter”

Thus, Turing correctly perceived in 1948 and 1950 that machine intelligence might be achieved by an evolutionary process in which a description of a computer program (the hereditary material) undergoes progressive modification (mutation) under the guidance of natural selection (i.e., selective pressure in the form of what is now usually called “fitness” by practitioners of genetic and evolutionary computation).

Of course, the measurement of fitness in modern-day genetic and evolutionary computation is usually performed by automated means (as opposed to a human passing judgment on each candidate individual, as suggested by Turing). In addition, modern work generally employs a population (i.e., not just a point-to-point evolutionary progression) and sexual recombination—two key aspects of John Holland’s seminal work on genetic algorithms, Adaptation in Natural and Artificial Systems (Holland 1975).

· The home page of Genetic Programming Inc. at www.genetic-programming.com.

· For information about the field of genetic programming in general, visit www.genetic-programming.org

· The home page of John R. Koza at Genetic Programming Inc. (including online versions of most papers) and the home page of John R. Koza at Stanford University

· Information about the 1992 book Genetic Programming: On the Programming of Computers by Means of Natural Selection, the 1994 book Genetic Programming II: Automatic Discovery of Reusable Programs, the 1999 book Genetic Programming III: Darwinian Invention and Problem Solving, and the 2003 book Genetic Programming IV: Routine Human-Competitive Machine Intelligence. Click here to read chapter 1 of Genetic Programming IV book in PDF format.

· For information on 3,198 papers (many on-line) on genetic programming (as of June 27, 2003) by over 900 authors, see William Langdon’s bibliography on genetic programming.

· For information on the Genetic Programming and Evolvable Machines journal published by Kluwer Academic Publishers

· For information on the Genetic Programming book series from Kluwer Academic Publishers, see the Call For Book Proposals

· For information about the annual Genetic and Evolutionary Computation (GECCO) conference (which includes the annual GP conference) to be held on June 26–30, 2004 (Saturday – Wednesday) in Seattle and its sponsoring organization, the International Society for Genetic and Evolutionary Computation (ISGEC). For information about the annual NASA/DoD Conference on Evolvable Hardware Conference (EH) to be held on June 24-26 (Thursday-Saturday), 2004 in Seattle. For information about the annual Euro-Genetic-Programming Conference to be held on April 5-7, 2004 (Monday – Wednesday) at the University of Coimbra in Coimbra Portugal.

Last updated on August 27, 2003