![]()
|
Abstract |
|
![]() |
![]() ![]()
3. System Architecture |
|
![]() |
|
![]() Figure 1. HalVis Architecture. |
![]() |
![]()
|
|
![]() |
|
![]() Table 1. Experiment Summary. |
![]() |
![]() Experiment I showed that hypermedia visualization is more effective than a textbook for learning an algorithm. This result was successfully replicated in Experiment II using more advanced students. Experiment III compared learning from HalVis to learning from a compilation of the best algorithm descriptions (extracted from 19 textbooks published between 1974 and 1997) followed by problem solving activities. While the posttest averages favored the HalVis group, the difference between the two groups is not statistically significant in this case. This experiment indicated that interactive visualization could be as effective as learning from carefully crafted text, when problem solving followed that textual learning. Experiment IV was designed to compare and evaluate how algorithm visualization (AV) and conventional classroom lectures interact and contribute to student learning. Our hypothesis was that AV alone would assist learning more than a lecture. If so, it follows that AV used in conjunction with lecture should assist learning even more. We used a 100-minute live classroom lecture, by a computer science faculty known for his teaching ability, delivered in two consecutive class sessions. Participants were second year computer science students at Auburn University, divided into two matching groups. One group used HalVis before attending the lecture (called the visualization-lecture –VL– group) and the other group used HalVis after the lecture (called the lecture-visualization –LV– group). Having both groups attend the same lecture eliminated variations that can occur in different lecture sessions, while allowing common interactions with the teacher. A first posttest was given between the two phases, and a second posttest was given after both phases. The first posttest results (shown in the table) indicated that the VL group learned significantly more from HalVis than the LV group from the lecture. Following a session with HalVis, the VL group's average posttest score was 70% compared to 44% for the LV group. The second posttest results showed that after receiving both the lecture and the visualization, both groups performed at the same level (72%). These results suggest that interactive visualizations can have a significantly higher learning impact than a lecture, and that a combination of both is even better. Experiment V compared the effectiveness of learning from HalVis to learning from an algorithm animation typical of extant research on this topic. One of the most mature, widely reported, and publicly available algorithm animation platforms is the Tango software suite developed by Stasko (1997). The Tango software distribution contains a library of animated algorithms, including eight animations of the Shortest Path algorithm. Of these eight, we selected one that appeared to be the most complete, easiest to understand, and which most closely matched the features of the HalVis system, as a representative animation. One group of students interacted with HalVis and the other group interacted with the Tango animation. The second group was also provided with a textual description of the algorithm, to which the HalVis group did not have access. On the posttest, the HalVis group averaged 89% while the Tango group averaged 71%, indicating that a hypermedia algorithm visualization is more effective than a mere algorithm animation. With five studies demonstrating the learning benefits of HalVis, we next asked the question: which features or modules of HalVis are producing the observed learning benefits? One approach to answering this is to build different versions of HalVis by selectively eliding specific features or modules, and then to experimentally compare these versions against the original. We designed and carried out three such studies. In Experiment VI we identified three features as most likely to have an impact on learning – animation chunking, highlighting steps of the pseudocode, and questions (both ticklers and articulation points). We built three versions of the QuickSort algorithm visualization in HalVis, with each of these features removed. Four groups of students worked with the full and three elided versions. They were tested for their knowledge of the algorithm before and afterwards. As seen in the table, the HalVis group learned the most, followed by the group that did not get highlighted steps of the pseudocode, the group that did not get the questions, and the group that saw one-shot animations with no chunking. While these results did not attain statistical significance, the trend indicates that animation chunking may be the most important feature, followed by questions and highlighting steps of the algorithm in tandem with the animations. Experiments VII and VIII were designed to reveal learning contributions of the three modules of HalVis – Conceptual View, Detailed View and Populated View. In Experiment VII four matched groups of students interacted with one of four versions of HalVis illustrating the QuickSort algorithm – a complete version (CDP version: all three views) and three elided versions (DP version: Conceptual View removed; CP version: Detailed View removed; and CD version: Populated View removed). Our hypothesis was that the most important view was the most information rich view – the Detailed View. We also expected that the Populated View would follow in significance and that the contribution of the Conceptual View would be the least since it contained the least amount of algorithm-specific information. As expected, the group that received all three HalVis views performed the best. However, the groups that performed at the next level were not the ones exposed to the Detailed View but rather the groups that interacted with the Conceptual View. Perhaps the most noteworthy observation from this experiment is the effect of the Conceptual View in apparently priming the learning of information presented in subsequent views. The groups that interacted with the Conceptual View in any combination with other views performed more than twice as better as the group that lacked the Conceptual View. Experiment VIII isolated and compared each of the three views. Its procedure was identical to that of Experiment VIII, with four matched groups of participants. The CDP group worked with a full version of HalVis illustrating Dijkstra's Shortest Path algorithm. The C group worked only with the Conceptual View of this algorithm. The D group worked only with the Detailed View of this algorithm. The P group worked only with the Populated View of this algorithm. Our hypothesis was that the Detailed View would prove to be the most valuable because of the amount of information it provided. We were uncertain as to how impacts of the other views would get ranked, since neither the Populated View nor the Conceptual View contained the volume or depth of information available in the Detailed View. As expected, the CDP group outperformed the others, followed closely by the D group. Interestingly, the C group outperformed the P group by 21%. It is illuminating to note how well the C group did with the limited amount of information (only the analogy) that they received. This reinforces the important role of interactive and animated analogies in learning about algorithms from visualizations. |
![]() |
![]() |
|
![]() |
![]() |
![]()
|
![]() |