Understanding mixed human-agent societies.

摘要

The coupling of trends in agent-based technologies with increasing numbers of processors pervading the human environment will assuredly lead to mixed societies composed of both humans and agents. Presently, it is unknown how a person might fare in mixed societies, or how a person's behavior might be altered when participating in such societies. The objective of the research described in this dissertation is to explore human and agent mixed societies in order to characterize human behavior toward agents. The approach is to perform human-subject experiments and from the results develop human-inspired agents as tools to further investigate models of human behavior.;The experiments used the dictator game to explore direct contributing behavior of humans toward agents. We show that there were insignificant differences between how people behaved towards agents versus towards humans with respect to direct contributing behavior, despite both the participants' social value orientation (i.e., either altruist or egoist) and the context of the interaction (i.e., either in public or in private). We then considered human behavior toward agents in the ultimatum game, a simplified negotiation context. In this case, we did observe significant differences in how people accept offers from agents versus humans. Specifically, we observed that altruists are more accepting of very low offers from agents, while they unilaterally sanctioned these same offers when they were made by humans.;Further steps in the understanding of human-agent interaction were taken by presenting (1) models, within a limited domain, for agents that behave like humans; and (2) results of simulated interactions between the human-like agents and a variety of purely rational agents. Modeling human behaviors as we have done presents a means for further exploring both purely human and human-agent interactions. A detailed equation-based analysis of agent performance during the simulated dictator game and an extension of this game called the indirect reciprocator game is presented. An example of the utility of these human-behavior based agents is presented as we model dynamic indirect reciprocity and use simulation based experimentation to explore our model's potential in enhancing the survivability of altruists in a society.