Dario D. Salvucci
John R. Anderson


In this chapter we propose an approach to developing models of analogical reasoning within production system architectures. Though existing theories of analogy have succeeded in capturing many aspects of analogical behavior, they have several limitations. First, the theories use empirical support that focuses almost exclusively on high-level data that illustrate the results of analogy, ignoring low-level data that illustrate step-by-step processes during analogy. Second, the theories cannot fully account for variability prevalent in analogical behavior, nor can they account for the adaptation of analogical strategies in learning. Third, some of the theories cannot be readily incorporated into a unified theory of cognition. We show how production rule models of analogy can address and to some extent overcome these limitations. As our exemplar, we describe empirical and modeling results for a task in which subjects solved simple physics problems by analogy. The empirical results, which include both high- and low-level data, give evidence that subjects use multiple analogical strategies and shift between strategies. The modeling results show that both a declarative and procedural ACT-R model of the task can account for much of subjects' observed behavior. We also present a model for a similar analogical task involving picture analogies (Sternberg, 1977) to illustrate how the simple physics model can generalize to other tasks.


Simple Physics Task

People-Piece Task