Cooperation and competition : modeling intention and behavior in dual-agent interactions

Abstract

A major goal of artificial intelligence research today is to build something that can cooperate with humans in an intelligent manner. In order to do so, we first must understand the mental mechanisms human use when solving problems of cooperation in dual-agent interactions, or between two people. We used reinforcement learning and Bayesian modeling to create a mathematical representation of this mental model. Our model is comprised of a high-level planner that understands abstract social intentions, and it employs two low-level planners that perform cooperative and competitive planning. To validate the model, we ran two experiments via Amazon Mechanical Turk to capture how humans attribute other players' behaviors and how they themselves behave in problems of cooperation such as the prisoner's dilemma. We compared our model against lesioned models and found that our model, which used both cooperative and competitive planning strategies, was the most representative of the data collected from both experiments.