Page 235 - Socially Intelligent Agents Creating Relationships with Computers and Robots

218 Socially Intelligent Agents

bilities). However, even when the parameters cannot be reliably specified by
experts or learned from data, providing estimates for them allows the designer
to clearly define the assumptions the model must rely upon and to revise the
assumptions by trial and error on the model performance. Thus, we believe that
Bayesian networks provide an appropriate formalism to model and integrate in
a principled way the multiple sources of uncertainty involved in monitoring a
student’s cognitive and emotional states, and the unfolding of a collaborative
interaction.
Modeling cognitive and meta-cognitive skills. Bayesian networks have
been extensively used to build user models representing user’s knowledge and
goals [11]. In [3], we have described how to automatically specify the structure
and conditional probabilities of a Bayesian network that models the relations
between a user’s problem solving behavior and her domain knowledge. In [7],
we have extended this work to model learning of instructional material through
the meta-cognitive skill known as self-explanation. We plan to adapt this ap-
proach to formalize the probabilistic relationships between player’s behavior,
meta-cognitive skills and learning in the student models for SIAs in educational
games.
Modeling collaboration. A preliminary Bayesian model of effective collab-
orative interaction has been proposed in [13]. The model attempts to trace the
progress of group members through different collaborative roles (e.g., leader,
observer, critic) by monitoring the actions that they perform on an interface
especially designed to reify these roles. We also adopt a role-based approach
to model effective collaboration, but we cannot structure and constrain the
game interface as in [13], because this kind of highly constrained interaction
could compromise the level of fun and engagement that students experience
with Avalanche. Hence, we need to devise alternative ways to capture the col-
laborative roles that students adopt during the interaction. We plan to start
by making the adoption of different collaborative roles one of the mandatory
game activities, orchestrated by the Collaboration Manager. This will reduce
the collaboration-monitoring problem to the problem of verifying that students
effectively perform the role they have been assigned. However, as the research
proceeds, we hope to also achieve a better understanding of how to monitor and
support less constrained collaboration.
Modeling emotions. Since emotional engagement is the element that makes
educational games attractive to learners, it is fundamental that this variable be
accurately monitored and taken into account by SIAs for these games. Starting
from existing research on the structure of emotions [1], we are working on a
general Bayesian student model to represent relevant emotional states (such as
frustration, boredom and excitement) and their dynamics, as they are influenced
by the interaction with an educational game, by the SIAs interventions and by
the player’s personality [2]. The formalization includes a theory of how the