| Abstract: | Many current theories of human problem solving and skill acquisition assume that people work only on the unsatisfied goal that was created most recently. That is, the architecture obeys a last-in-first-out (LIFO) constraint on the selection of goals. This restriction seems necessary for the proper functioning of automatic learning mechanisms, such as production compilation and chunking. It is argued that this restriction is violated by some subjects on some tasks. In particular, 8 subjects (from a sample of 26) execute subtraction procedures in a way that violates the LIFO constraint. Although there is a great deal of between- and within-subject strategy variation in the 8 subjects' behavior, it can be simply explained by hypothesizing that (1) the goal selection is not necessarily LIFO, (2) goal selection knowledge is represented by explicit preferences, and (3) the 8 subjects have just a few preferences that are overgeneralized, overspecialized, or missing. The rest of their preferences are correct. On the other hand, LIFO-based models seem unable to explain the strategy variations in any simple way. Thus, it seems that part of the flexibility in human problem solving comes from having a choice of which goal to work on next. Fortunately, it is simple to ammend automatic learning mechanisms so that they will function correctly in a non-LIFO architecture. |
