When encountering a new domain or when solving a novel problem in which no previously acquired skills are directly applicable, students must draw upon other sources of knowledge to overcome problem-solving impasses. In classroom settings, multiple resources exist for acquiring such knowledge. These include, for example, instructional materials (e.g., textbooks), teachers, peers, or analogies from prior knowledge. The degree to which students can properly interpret and apply this knowledge will affect their success in overcoming problem-solving impasses.
In this article, we concentrate on the declarative origins of knowledge used in initial skill acquisition. In particular, we focus on the strategies that students bring to bear when explaining instructional materials to themselves prior to problem solving. These learning strategies, hereafter called self-explanation, appear to significantly affect students' initial understanding and their subsequent problem-solving performance (Chi, Bassok, Lewis, Reimann, & Glasser, 1989; Ferguson-Hessler & de Jong, 1990; Pirolli & Recker, 1994). In a recent study, Pirolli and Recker (1994) investigated how students explained typical, school-like instruction on programming recursion to themselves before they solved recusion problems with the CMU Lisp Tutor, an intelligent tutoring system for Lisp (J. Anderson, Boyle, Corbett & Lewis, 1990). The results of the study showed that skill acquisition in the recursion lesson was correlated with the quantity and kinds of self-explanations that subjects made while initially trying to understand the instruction. In particular, the better performing subjects (a) focused more on the content of the instruction about recursion, (b) attempted to relate the instructional text to relevant prior knowledge, and (c) seemed to exhibit higher degrees of metacognition.
One interpretation of these results is that, to improve student learning, the elaborations produced by effective learners should be directly included in the instructional materials. This way, all students would read quality elaborations and thus acquire equally rich domain models. Unfortunately, the story is not quite that simple. Prior research investigating this hypotheses found that, in some cases, embedding elaborations within instruction actually hindered learning (Reder, Charney, & Morgan, 1986). Moreover, it is unclear whether externally provided elaborations are effective as self-generated explanations. Memory research has identified a generation effect, in which self-generated items are better remembered than experimenter-provided items (Hirschman & Bjork, 1988). Last, consider the case in which a learner has already successfully generated a useful elaboration. It is really necessary for him or her to spend the extra time and effort trying to understand a text-provided elaboration?
The research we have discussed so far involved the use of standard instruction, which students read in a serial fashion. Therefore, an alternative possibility is to structure the instruction in a nonlinear form, such as hypertext. In this way, students would have a choice over how they moved through the instruction and the option of viewing explanatory elaborations when they were unable to generate them on their own. Such a structure would allow low self-explainers to read quality explanations, whereas the high self-explainers could avoid having to read redundant, possibly confusing explanations.
To investigate this hypothesis, we designed a hypertext-based instructional system, called the explanation environment (EE), which contained instruction on programming recursion in Lisp. In this environment, instruction examples were annotated with explanatory elaborations that students could choose to view. Although the design of this system did not directly investigate the relative merits of self- versus text-provided elaborations, it did investigate the hypothesis that providing extra, optional elaborations to students who were unable to produce them on their own might reduce the overall variance between students.
The instructional environments also provided an important, secondary benefit. They were implemented such that they collected a detailed log of subjects' trajectories through their systems. These data provided additional means for understanding subjects' learning and browsing strategies and their range of individual differences.
In the first part of this article, we describe a study that investigates the impact of the EE on students' learning strategies and their subsequent problem-solving performance. In this study, subjects went through five programming lessons in Lisp, including the topic of recursion. Each lesson had two parts: (a) studying instructional material contained in a booklet (knowledge acquisition) followed by (b) programming using the CMU Lisp Tutor (problem solving). For the target lesson, the lesson on recursion, two sets of computer-based instruction were developed. Subjects were randomly assigned to one of the two environments and were asked to provide verbal protocol as they learned about and explained to themselves the concepts of recursion. The first environment was the EE. The second environment, the instructional environment, served primarily as a control condition to the EE. Although it was also computer-based, its structure mirrored more standard, linear instruction. All subjects then programmed a set of recursive functions using the CMU Lisp Tutor.
In the second part of this article, we present a computational model, called Strategies for Understanding Recursive Functions (SURF), which accounts for results from the empirical study. In addition to explaining results from the study, a particular focus of SURF was on modeling the learning activities of individual subjects as they are interacted with and studied the materials contained in the instructional environments.
The general modeling strategy within SURF involved coupling simulations of individual subjects with models of the interfaces of the instructional environments to capture the interactions observed in the empirical data. By setting parameters in the model, SURF exactly simulated the temporal sequence of individual subjects' interactions in terms of their mouse action and self-explanations. In particular, in modeling individual subjects, SURF was bound by two criteria. The first criterion required that every mouse clicking action by all subjects be modeled in the exact order of occurrence. The second criterion required that, for each subject, every domain-related self-explanation be modeled in the exact order of occurrence. The SURF model was implemented within the Soar architecture (Laird, Rosenbloom, & Newell, 1987).
Choosing environmental opportunities, learner strategy, and resulting interactions as the units of analysis provided tangible means for understanding and analyzing the contributions of each in shaping the learning process. As we will discuss, analyses of the simulations of the instructional environment revealed differences in their cognitive complexity. These differences may have had differential effects on subjects' learning and browsing strategies. On the basis of analyses, we argue that individual subjects' approaches to the task reflected one of several classes of general learning strategies. Furthermore, these strategies were, in some sense, attempting to be adaptive to the task, with varying degrees of success.
The remainder of this article is organized as follows. First, the theoretical framework underlying this research is presented. Second, the instructional environments and the empirical study are described. Third, results from the empirical study form the baseline data for the computational model. Fourth, several approaches for evaluating the fit of individual simulations to subject data are given. Last, we close wit discussions of some of the limitations of the current modeling approach and compare it to related research.