Page 76 - Socially Intelligent Agents Creating Relationships with Computers and Robots

Adapting to Affect and Personality 59

5. Conclusions

We described a research area aimed at producing more realistic behavior
in socially intelligent agents, namely the recognition of, and adaptation to,
a user’s affective state. We developed an adaptive methodology and demon-
strated its effectiveness by implementing a prototype Affect and Belief Adap-
tive Interface System (ABAIS). ABAIS assessed the user affect and belief
states using a knowledge-based approach and information from a variety of
sources, predicted the effects of this state within the constrained context of
the demonstration task, and suggested and implemented specific GUI adapta-
tion strategies based on the pilot’s individual information presentation prefer-
ences. The preliminary results indicate the general feasibility of the approach,
raise a number of further research questions, and provide information about
the specific requirements for a successful, operational affective adaptive inter-
face. Although the initial prototype was developed within a military domain,
we believe that the results are applicable to a broad variety of non-military
application areas, as outlined below.

Requirements for Adaptation. A number of requirements were identified
as necessary for affective adaptive interface system implementation. These
include: Limiting the number, type, and resolution of affective states, and using
multiple methods and data sources for affective state assessment; providing
individualized user data and user-customized knowledge-bases; implementing
‘benign’ adaptations that at best enhance and at worst maintain current level of
situation awareness (i.e., never limit access to existing information).

Key Issues to Address. A number of issues must be addressed to further
validate this approach and to provide robust affect adaptive systems. These in-
clude: an empirical evaluation; multiple-method affect assessment; and demon-
stration of the ABAIS methodology across multiple task contexts.

Future Work. Possible future work in the broad area of user affect and
modeling is limitless at this point, as the endeavor is in its infancy. Key ques-
tions include issues such as: What emotions should and can be addressed in
adaptive systems? When should an agent attempt to enhance the user’s af-
fective state, adapt to the user’s affective state, or attempt to counteract it?
Cañamero offers an excellent summary of some of the affect-related issues
that must be addressed by the SIA architecture research community [2].
Individually, both modeling and recognition of affective and beliefs states,
and personality traits provide a powerful enhancement to agent architectures,
by enabling socially intelligent, adaptive behavior. The coordinated integra-
tion of these two enhancements within a single agent architecture promises
even further benefits, by enhancing the realism and effectiveness of human-