Steps Toward Artificial Intelligence

What the article set out to do

"Steps Toward Artificial Intelligence" is a survey article by Marvin Minsky, published in Proceedings of the IRE, volume 49, issue 1, January 1961, pages 8 to 30, with DOI 10.1109/JRPROC.1961.287775 ^[1]. The IRE received it on 24 October 1960 ^[2]. Minsky's affiliation is given as the Department of Mathematics and Computation Center, Research Laboratory of Electronics, MIT ^[2].

The starting point is a plain observation about computers: a computer can do only what it is told to do ^[2]. That seems to rule out writing a program for a problem whose solution method we do not know. The article's answer is that even when we do not know how to solve a problem, we may still program a machine to search a large space of solution attempts ^[2]. Heuristic programming, in this framing, is the work of making such searches succeed.

Five areas

The article divides the problems of heuristic programming into five main areas: search, pattern recognition, learning, planning and induction ^[2]. These headings set the structure of the discussion, and each of the last four is treated as a way of keeping an otherwise unmanageable search under control.

Search and its cost

Search is both the basic mechanism and the basic difficulty. Simply programming a machine to try candidate solutions usually leads to an enormously inefficient process ^[2]. Something has to reduce the space that the machine has to look through, or the approach does not scale beyond small problems.

Pattern recognition and learning

Pattern recognition is the first of the reductions. Techniques for recognizing patterns can improve efficiency by restricting the application of the machine's methods to appropriate problems ^[2]: rather than applying every method to every problem, the machine recognizes what kind of situation it is in and applies the methods that fit it.

Learning reaches further. Pattern recognition together with learning can exploit generalisations based on accumulated experience and further reduce search ^[2]. Experience is not kept as a list of individual cases but in general form, so that what was learned in one situation narrows the search in the next.

Planning and induction

Planning attacks the search from another angle. By analysing the situation with planning methods, the article argues, one may replace the given search with a much smaller and more appropriate exploration ^[2]. Instead of working through the original problem step by step, the machine first works at a level where the exploration is cheaper and better directed.

Induction is the most ambitious of the five. To manage broad classes of problems, the article concludes, machines will need to construct models of their environments using some scheme for induction ^[2]. A machine that handles only the situation it was explicitly programmed for stays narrow; a machine that builds models of its environment can carry what it has learned into situations its programmer did not anticipate.

Where the work was done

The work summarised in the article was done at Lincoln Laboratory and at the Research Laboratory of Electronics at MIT, with support from the US Army, Navy and Air Force ^[2].

Why it matters

The article is best read as a statement of a problem rather than a report of a result. It does not announce a program that solves these questions; it sets out what would have to be done, and why each part is needed. Its contribution is the map: intelligence treated as the management of search, with recognition, learning, planning and induction as the means by which search is kept affordable. That is a compact description of what the field took itself to be doing at the start of the 1960s, and the five headings give a reader a clear way to see how the parts of the problem connect — recognition and learning cutting down the space, planning replacing it with a smaller exploration, and induction supplying the models that make any of this apply beyond the case at hand.