In two experiments, we demonstrated two types of strategies (rule-based and memory-based) and strategy transitions within the same binary classification task. The strategy that dominated later in practice depended on the difficulty of the operative classification rule and on the salience of the cue for that rule. Accuracy increased over practice trials, and response times were faster for the dominant strategy, either rule or memory. Rule retention was superior to stimulus item retention, so that, even for participants who preferred a memory-based strategy, a rule-based strategy dominated at least temporarily after a 1-week interval. Strategy use over trials and the retention interval reflected a given task's affordance of a shift between rule- and memory-based processes.
Strategies adopted by participants during skill acquisition do not necessarily remain stable over training or practice trials, but rather often shift in the direction of speeding up responses and easing the cognitive load (Schunn & Reder, 2001). A number of formal models, notably by Logan (1988) and Rickard (1997), have been proposed to account for strategy shifts. These models have focused on a shift from performance based on a rule (algorithm, calculation) to performance based on the retrieval of previously experienced items from memory. In these models, retrieval from long-term memory is viewed as a singlestep, possibly automatic solution procedure for problems that might otherwise require multiple rule-based computations. In tasks to which these models have been applied, participants are first instructed in the rule or algorithm that governs responses. Thus, although there are certain fundamental differences between them, both the Logan and Rickard models assume that performance is under control of an algorithm-based strategy at the outset and that, at some point in practice, there is a transition in control from a strategy based on the algorithm to a strategy based on instance-memory retrieval.
The Logan (1988) and Rickard (1997) models may be limited by the tasks to which they have been applied (e.g., alphabet arithmetic). In these tasks, only two strategies are possible. Participants are given the rule (or algorithm) in advance and told to use the rule-based strategy at the outset. Under these conditions, guessing is unlikely and a transition from memory retrieval to rule use is virtually impossible. In a binary classification task, using a simple self-report procedure to identify strategies, Bourne, Healy, Parker, and Rickard (1999) found that strategy shifts are not always in the rule-to-memory direction. When participants were not preinstructed in the rule or algorithm for a task, most of them reported guessing on early trials. After some significant number of repetitions of individual stimuli, most participants adopted a preferred strategy, which might or might not be memory based. Specifically, most participants preferred a rule-based strategy hi what was called a natural task condition. This task was called natural because it involved a linguistic rule for pronunciation of English words (i.e., participants had to choose one of two pronunciations, "thee" or "thuh," for the definite article the, depending on a given noun or adjective; see Raymond, Fisher, & Healy, 2002, for evidence that college students are not fully knowledgeable about this rule despite its presence in all English dictionaries). In contrast, most participants eventually preferred a memory-based strategy in an artificial task condition (in which participants had to base their response on whether meaningless letter strings conformed to an alphabetical sequence). Overall measures (accuracy and response time [RT]) showed that performance was better under the control of the preferred strategy-either rule or memory-than under the nonpreferred strategy.
The findings by Bourne et al. (1999) suggest that extant models probably do not satisfy all tasks in which strategy use and strategy shifts occur. …