next up previous
Next: General references
Up: Main references
Previous: Some history

The original source

A program for the solution of a class of geometric-analogy intelligence-test questions by Thomas G. Evans, in Semantic Information Processing, edited by M. Minsky (1968, MIT Press). This is in both the Psychology library and the RSL (19976 d 21).

Winston gives all the information you need about the principles. Finals certainly won't require a account as detailed as Evans' own. But it's always a good idea to look at the original sources, partly to check peoples' motivation, partly to discover reporting errors.

In this case, the source is good for (a) extra background on history, the AI culture of the time, and the programming problems faced in real AI; (b) why Evans himself says he wanted to do this work; (c) how the program really worked (remember I said Winston's account is simplified); (d) some points about how the IQ test pictures are split up.

I said above that you should read the introduction to, and section 5.3.3 of, Evans' original writeup. The introduction puts Evans' work in context and describes his motivation; section 5.3.3 talks about how complex figures were separated into their components, which is something Winston ignores. Most of the rest of the article is a technical account of the program. I have marked below sections that I think are particularly important.

  1. 5.1.1 - 5.1.4. Introduction.
  2. 5.1.4 in particular. Talks about the importance of finding a good representation (Evans calls it a ``description''): for pictures and solving problems on them this was almost untried then. There are some thoughts on why Samuel's program's description of checkers will never enable it to learn. Page 282 distinguishes the three levels of representation in Evans' program. Page 283 views the program as working by theory formation.
  3. 5.1.5 A three-page summary of past work in AI
  4. 5.2.1 A summary of how the program works. Notice the need for punched cards as intermediate storage, because the computer had so little memory.
  5. 5.3.3 Towards the end, Evans tells how he dealt with overlapping figures, and how he had to take context into account to do so.
  6. 5.4.1 General remarks on the analogy process.
  7. 5.4.3 Details on how the best solution was selected.
  8. 5.4.4 How the program might be extended to ``higher-order'' problems.
  9. 5.5.1,2 Sample runs and discussion.
  10. 5.5.3, the subsection on Comparison with Human performance.
  11. 5.6.1,2 The fundamental role of analogy in problem solving.
  12. 5.6.3 Problems with the ``property-list'' approach to pattern-matching. Property-list methods were being tried extensively in computer vision, especially for recognising letters, and they eventually failed (as described briefly in Chapter 2 (?) of Boden's The Computer and the Mind.
  13. 5.6.4

next up previous
Next: General references
Up: Main references
Previous: Some history

Jocelyn Paine
Tue Jun 3 11:33:37 BST 1997