Figure 1. Snapshot of VirtualLabs.
interaction styles. Also, none of these past approaches have been shown to work for more than a single
ELE, whereas one of our algorithms is shown to generalize to two ELEs for chemistry and statistics education that are significantly different in their design
and interaction style with the student.
Finally, we mention works that use recognition
techniques to model students’ activities in intelligent
tutoring systems (ITSs) (VanLehn et al. 2005; Conati,
Gertner, and VanLehn 2002; Roll, Aleven, and
Koedinger 2010; Koedinger et al. 1997). Such systems
coach students during their problem solving, providing support with proven learning gains. Conati, Gertner, and VanLehn (2002) used online plan-recognition algorithms to infer students’ plans to solve a
problem in an educational software for teaching
physics by comparing their actions to a set of predefined possible plans. The number of possible plans
grow exponentially in the types of domains we consider, making it unfeasible to apply this approach.
Actions and Plans
In this section we provide the basic definitions that
are required for formalizing the plan-recognition
problems in ELEs. Throughout the article we will use
an ELE called VirtualLabs to demonstrate our
approach. VirtualLabs allows students to design and
carry out their own experiments for investigating
chemical processes (Yaron et al. 2010) by simulating
the conditions and effects that characterize scientific
inquiry in the physical laboratory. We use the fol-
lowing problem called Oracle as a running example:
Given four substances A, B, C, and D that react in a
way that is unknown, design and perform virtual lab
experiments to determine the correct reaction
between these substances.
The flexibility of VirtualLabs affords two classes of
solution strategies to this problem (and many variations within each). The first strategy mixes all four
solutions together and infers the reactants by inspect-