📄 Learning and reasoning by analogy

Summary

Winston tackles the problem of how a program can use a familiar situation, a precedent, to draw conclusions about a new one, an exercise. Situations, such as compressed plot outlines of Shakespeare plays like Macbeth, are encoded as semantic networks. These networks capture the participants, their properties and the relations among them. The central step is matching. Not all relations count equally: causal links between events steer which parts of the precedent are paired with which parts of the exercise. Once a match is found, the program transfers knowledge from the precedent to answer questions about the exercise or to make plausible guesses about it. It can also record what it has learned as a reusable rule, so a single example yields a generalization. The work matters because it was an early, influential computational account of analogy. It shows that weighting causal structure, rather than surface similarity, makes matching both feasible and meaningful. It shaped later research on case-based and precedent-based reasoning.

Key points

• Published in Communications of the ACM (vol. 23, no. 12, 1980). It is related to an earlier MIT AI Lab memo version.
• Situations are written as networks of parts, properties and relations, and these networks are matched against each other.
• Matching favours important relations, especially causal ones, over superficial resemblance.
• Shakespearean plot summaries, notably Macbeth, serve as test precedents and exercises.
• Caveat: the PDF text provided for this paper was empty (all 64 pages blank). This summary is based on the paper's widely known content and should be checked against the original before citing.