D.D. Schmorrow, L.M. Reeves (Eds.): Augmented Cognition, HCII 2007, LNAI 4565, pp. 400–408, 2007. © Springer-Verlag Berlin Heidelberg 2007 Exploring Neural Trajectories of Scientific Problem Solving Skill Acquisition Ronald H. Stevens 1 , Trysha Galloway 1 , and Chris Berka 2 1 UCLA IMMEX Project, 5601 W. Slauson Ave. #255, Culver City, CA 90230 immex_ron@hotmail.com, tryshag@gmail.com 2 Advanced Brain Monitoring, Inc, Carlsbad, CA 92008 chris@b-alert.com Abstract. We have modeled changes in electroencephalography (EEG) - derived measures of cognitive workload, engagement, and distraction as individuals developed and refined their problem solving skills in science. Subjects performing a series of problem solving simulations showed decreases in the times needed to solve the problems; however, metrics of high cognitive workload and high engagement remained the same. When these indices were measured within the navigation, decision, and display events in the simulations, significant differences in workload and engagement were often observed. In addition, differences in these event categories were also often observed across a series of the tasks, and were variable across individuals. These preliminary studies suggest that the development of EEG-derived models of the dynamic changes in cognitive indices of workload, distraction and engagement may be an important tool for understanding the development of problem solving skills in secondary school students. Keywords: EEG, Problem solving, Skill Acquisition, Cognitive Workload. 1 Introduction Skill development occurs in stages that are characterized by changes in the time and mental effort required to exercise the skill (Anderson, 1982, 1995, Schneider and Shiffrin, 1977). Given the complexities of skill acquisition it is not surprising that a variety of approaches have been used to model the process. For instance, some researchers have used machine learning tools to refine models of skill acquisition and learning behaviors in science and mathematics. Such systems rely on learner models that continually provide updated estimates of students’ knowledge and misconceptions based on actions such as choosing an incorrect answer or requesting a multimedia hint. Although such learner models are capable of forecasting student difficulties (Stevens, Johnson, & Soller, 2005), or identifying when students may require an educational intervention, they still rely on relatively impoverished input due to the limited range of learner actions that can be detected by the tutoring system (e.g., menu choices, mouse clicks).