Page 234 - Socially Intelligent Agents Creating Relationships with Computers and Robots

Socially Intelligent Agents in Educational Games 217

reaction) and updates the corresponding elements in the student model for that
player.
A Game Actions Interpreter, for instance, processes all the student’s game
actions within a specific activity, to infer information on the student’s cognitive
and meta-cognitive skills. A Meta-Cognitive Behavior Interpreter tracks all
the additional student’s actions that can indicate meta-cognitive activity, (e.g.,
utterances and eye or mouse movements) and passes them to the student model
as further evidence on the student’s meta-cognitive skills. The agent’s action
generator then uses the student model and the expertise encoded in the agent’s
knowledge base (which depend on the agent’s pedagogical role) to generate
actions that help the student learn better from the current activity.
The agents in the architecture include a Game Manager, the Collaboration
Manager and agents related to specific game activities (like Help Agent for
activity A and Peer Agent for activity K in Figure 1). The Game Manager
knows about the structure of the game and guides the students through its
activities. The Collaboration Manager is in charge of orchestrating effective
collaborative behavior. As shown in Figure 1, its Behavior Interpreter captures
and decodes all those students’ actions that can indicate collaboration or lack
thereof, along with the related emotional reactions. The actions that pertain
to the Collaboration Manager include selecting an adequate collaboration role
and partners for a student within a particular activity. The pool of partners from
which the Collaboration Manager can select includes both the other players or
the artificial agents (e.g., the Peer Agent selected for Student N in activity K in
Figure 1), to deal with situations in which no other player can currently be an
adequate partner for a student, because of incompatible cognitive or emotional
states.
The artificial agents related to each game activity have expertise that allow
them to play specific roles within that activity. So, for instance, a Help Agent
(like Help Agent for activity A in Figure 1) has expert knowledge on a given
activity, on the emotional states that can influence the benefits of providing help
and on how to provide this help effectively. Peer agents, on the other hand, will
have game and domain knowledge that is incomplete in different ways, so that
they can be selected by the Collaboration Manager to play specific collaborative
roles in the activity (e.g., that of a more or less skilled learning companion).

3.2 Student Models
The student models in our architecture are based on the probabilistic rea-
soning framework of Bayesian networks [10] that allows performing reasoning
under uncertainty by relying on the sound foundations of probability theory.
One of the main objections to the use of Bayesian networks is the difficulty
of assigning accurate network parameters (i.e. prior and conditional proba-