Modeling Agent by Designer`s Social Preference

Human social preferences have been shown to play an important role in many areas of decision-making; e.g., interaction in labor markets, bilateral or small-group bargaining, social welfare considerations.  There is evidence from the social science literature that human preferences in interpersonal interactions depend partly on a stable, measurable personality trait called a human`s Social Value Orientation (SVO). In multi-agent systems, agents are often written by humans to serve as their delegates when interacting with other agents.  Thus, one might expect an agent`s behavior to be influenced by the SVO of its human designer. The purpose of this paper is to explore this hypothesis.

There are many methods to gauge human social preferences or even personalities, as it is well studied in social psychology. Several measurement methods for quantifying variations in SVO across individuals have been developed. To obtain the social preferences of computer agents, we needed to have some measurement methods.  Quantifying the social preferences of computer agents presents several challenges.

We collected a set of students` agents and conducted psychological SVO based evaluation to get the corresponding SVO value of the human who constructed the agent.  We estimated the social preference of computer agents by the proposed methods, and studied the correlation. The results show that the SVO of human designer is *highly correlated* with the social preference of the corresponding agent. We also show the value of having the SVO of the designer of other agents by presenting an application of using that information on agent modeling.

Next, we extend the SVO model by developing a behavioral signature, a model of how an agent`s behavior over time will be affected by both its own SVO and the other agent`s SVO.   We provide a way to measure an agent`s behavioral signature, and methods for using behavioral signatures to predict agents` performance, and present experimental results using agents that students wrote to compete in repeated-game tournaments. The experimental results show that our predictions are highly correlated with the agents` actual performance in tournament settings.  This shows that our proposed model is an effective way to generalize SVO to situations where agents interact repeatedly.