摘要：

Previous research has motivated the idea of automatically determining when programmers are having difficulty, provided an initial algorithm (unimplemented in an actual system), and performed a small student-based evaluation to justify the viability of this concept. We have taken the next step in this line of research by designing and developing two-different systems that incorporate variations of the algorithm, implementing a tool that allows independent observers to code recorded sessions, and performing studies involving both student and industrial programmers. Our work shows that (a) it is possible to develop an efficient and reusable architecture for predicting programmer status, (b) the previous technique can be improved through aggregation of predicted status, (c) the improved technique correlates more with programmers' perception of whether they are stuck than that of observers manually watching the programmers, (d) the observers are quicker than the developers to conclude that programmers are stuck, (e) with appropriate training, the tool can be used to predict even the observers' perceptions, and (f) a group training model offers more accuracy than an individual one when the training and test exercises are the same and carried over a small time frame.

展开