Page 236 - Socially Intelligent Agents Creating Relationships with Computers and Robots

Socially Intelligent Agents in Educational Games 219

players’ emotions can be detected, based on current research on how to measure
emotional reactionsthrough bodilyexpressions suchas facial expressions, vocal
intonation, galvanic skin response and heart rate [15].

3.3 Action Generators

The action generator for each SIA in the game relies on a decision-theoretic
model of decision-making predicting that agents act so as to maximize the
expected utility of their actions [9]. Other researchers have started adopting
a decision theoretic approach to regulate the behavior of interactive desktop
assistants [8] and of an intelligent tutor to support coached problem solving [5].
In our architecture, the function representing an agent’s preferences in terms
of utility values depends on the role of the agent in the game. So, for instance,
the Collaboration Manager will act so as to maximize students’ learning as well
as their collaborative behavior. A Help Agent will act to maximize the stu-
dent’s understanding of a specific activity, while an agent in charge of eliciting
a specific meta-cognitive skill will select actions that maximize this specific
outcome. All the agents will also include in their utility functions the goal of
maintaining the student’s level of fun and engagement above a given threshold,
although the threshold may vary with the agent’s role. The action generators’
decision-theoretic models can be represented as influence diagrams [9], an ex-
tension of Bayesian networks devised to model rational decision making under
uncertainty. By using influence diagrams, we can compactly specify how each
SIA’s action influences the relevant elements in the Bayesian student model,
such as the player’s cognitive and emotional states. We can also encode the
agent’s utility function in terms of these states, thus providing each agent with
a normative theory of how to intervene in the students’ game playing to achieve
the best trade-off between engagement and learning.

4. Conclusions
We have presented a preliminary architecture to improve the effectiveness
of collaborative educational games. The architecture relies on the usage of
socially intelligent agents that calibrate their interventions by taking into ac-
count not only the students’ cognitive states, but also their emotional states
and the unfolding of collaborative interactions within the game. We propose to
rely on Bayesian networks and influence diagrams to provide our agents with
a principled framework for making informed decisions on the most effective
interventions under the multiple sources of uncertainty involved in modelling
interaction and learning in multi-player, multi-activity educational game.