Why Example Fading Works: A Qualitative Analysis Using Cascade

Abstract

Why Example Fading Works: A Qualitative Analysis Using Cascade Eric S. Fleischman (esfleisc@colby.edu) Randolph M. Jones * (rjones@colby.edu) Department of Computer Science, Colby College, 5830 Mayflower Hill Drive Waterville, ME 04901 − 8858 USA Abstract “Faded” examples are example problems that provide a solution, but first require students to generate a portion of the solution themselves. Empirical studies have shown that such examples can be more effective teaching aids than completely worked examples that require no work from the student. Cascade is a model of problem-solving skill acquisition that was originally developed to explain other empirical regularities associated with human problem solving and learning, most notably the self-explanation effect. Past research demonstrated that Cascade might also explain the mechanisms underlying the effectiveness of example fading. This paper analyzes new protocol data, and finds that it is consistent with predictions derived from Cascade. Overview Renkl, Atkinson, and Maier (2000) empirically demonstrated the qualitative result that, when learning problem-solving skills, students studying a series of “faded” examples show improved post-test performance over students studying only completely worked examples. Jones and Fleischman (2001) argue that this result can be explained by Cascade (VanLehn, Jones, & Chi, 1991), a computational model of problem-solving skill acquisition. Cascade was originally developed to understand the mechanisms of the self-explanation effect (Chi, Bassok, Lewis, Reimann, & Glaser, 1989; Pirolli & Anderson, Jones and Fleischman demonstrated that the mechanisms underlying self-explanation might also explain the effectiveness of studying faded examples. Although they showed that Cascade is consistent with the fading result, the explanation involved assumptions that had not yet been tested empirically. Therefore Jones and Fleischman (2001) finished with a small set of predictions and suggestions for new experiments to confirm or dispute Cascade’s account. Since that time, Renkl, Atkinson, and their colleagues have run additional experiments, collecting detailed transcripts of subjects studying two types of faded sequences of problems. Although the experiments are not yet complete, we have been able to perform a qualitative analysis of the protocol data for eight of the subjects. Additionally, we have fine-tuned Cascade’s knowledge base (but not its underlying mechanisms) to more faithfully model the current data. This paper reports the result of using Cascade to develop a qualitative analysis of the eight subjects. The primary result is that the findings remain The second author is also affiliated with Soar Technology, Inc. consistent with Cascade’s account of example fading, as well as the predictions made by Jones and Fleischman Background Years of research have demonstrated effective techniques for teaching students problem-solving skills in a variety of task domains. In particular, a number of studies show that students benefit from being given a series of completely worked example problems, followed by a series of unworked practice problems (e.g., Chi et al., 1989; Pirolli & Anderson, 1985; Renkl, 1997, VanLehn, 1996). Other studies show that the effectiveness of such a curriculum depends in part on the willingness of the students to explain the worked examples to themselves in detail, rather than simply giving the examples a superficial read (Chi et al., 1989; Fergusson-Hessler & de Jong, 1990; Pirolli & Bielaczyc, 1989). VanLehn and Jones (1993a, 1993b; VanLehn et al., 1991) developed Cascade in order to determine the cognitive mechanisms behind this self- explanation effect. In essence, Cascade suggests that thorough study of worked examples help students consciously expose and patch gaps in their task knowledge. In addition, self-explanation provides contextual memories that can guide future problem solving by analogy to familiar examples. Subsequent experiments by Renkl et al. (2000) suggest that student learning can improve even further by fading a curriculum from fully worked examples to partially worked examples. The partially worked examples provide a complete solution to the problem (as with fully worked examples), but first require students to derive one or more steps on their own. This in turn requires the students to understand the rest of the example in at least enough detail to be able to attempt a solution. Jones and Fleischman (2001) argue that the reason faded examples improve learning is that they retain much of the guidance provided by the context of a solved example, but they force the students to work on particular parts of the problem, in turn possibly forcing them to expose and patch knowledge gaps. This is in contrast to studying completely worked examples, where it is basically up to the students to decide whether they are going to put any effort into understanding the examples (because the students are not required to produce any answers in that case). This argument came directly from the assumption that Cascade is