Design of Social Agents - Foundations on Natural and Artificial Computation - page 191

Design of Social Agents

Roman Gorbunov, Emilia Barakova, and Matthias Rauterberg

Eindhoven University of Technology, Den Dolech 2, Postbus 513, 5600 MB
Eindhoven, The Netherlands

Abstract. Social behavior, as compared to the egoistic and rational be-
havior, is known to be more beneficial to groups of subjects and even
to individual members of a group. For this reason, social norms natu-
rally emerge as a product of evolution in human and animal populations.
The benefit of the social behavior makes it also an interesting subject in
the field of artificial agents. Social interactions implemented in computer
agents can improve their personal and group performance. In this study
we formulate design principles of social agents and use them to create so-
cial computer agents. To construct social agents we take two approaches.
First, we construct social computer agents based on our understanding
of social norms. Second, we use an evolutionary approach to create social
agents. The social agents are shown to outperform agents that do not
utilize social behavior.

1 Introduction

Collaboration, trust and other social norms are crucial for good operation of
teams. Our long term goal is to create a model of social norms that will enable
prediction of behavior and interactions within a team. On the basis of such a
model we want to test hypotheses of how these interactions within a group will
develop.
Human and animal models of behavior inspired studies on social learning
[4,19,18,17,21] and models of social norms such as trust and collaboration
[19,18,17,14]. Collaboration has also been studied within economics, sociology,
and game theory. In economics, models of a rational actor who collaborates to
improve organizational effectiveness have been developed [15,11,9]. In sociology
[5], the non-rational and social aspects of collaboration has been explored. Gor-
bunov at al. [8] have proposed a model that accounts for the deviations from the
rational behavior.
Games, in particular are an establishing field that makes possible realistic
social interactions to be exposed, monitored and trained [2,1,3,12,13,16]. Hennes
[12] Voynarovskaya et al.[20] and Gorbunov at al. [8] have developed a game-
based method for monitoring the social relations in small groups of individuals.
It is based on the records of a negotiation game that monitors of the collab-
orative behavior of the players. The Colored Trails game framework [7] in a
three-player negotiation variation [6,20] was used for this study. The conclu-
sions of the study have been restricted by the limited amount of data and the