ROLE OF SPECIFIC INSTANCES IN CONTROLLING A DYNAMIC SYSTEM

This article examines the claim that the learning of a dynamic control task is mediated by a lookup table consisting of previously successful trials on the task. Consistent with the predictions of a lookup table, in 2 experiments participants tended to give the same response to situations in which they had previously been successful rather than unsuccessful. Further, in both experiments, participants knowledge did not generalize to new dissimilar situations, unless the dynamic control task was governed by a highly salient rule. A version of G. Logan's (1988) instance theory, which assumes that participants store each successful response as a separate instance linking the situation to the response, was able to quantitatively match a range of measures of participants' performance with one free parameter, except in the case in which the control task was governed by a salient rule. In a complementary way, an alternative rule-based model could only match participants' performance when the control task was governed by a highly salient rule.