A dynamic theory of multi-agent goal pursuit in continuous time and space

Many of the goals that people pursue in everyday life involve dynamic social interactions with others who have their own latent goals. Modeling how these goals manifest in behavior and how different goals interact with one another is central to effectively explaining and predicting the dynamics of human interactions. In this paper, we develop a theory that postulates a balance of global and local objectives driving behavior, combining elements of control theory and approach-avoidance to predict complex movement and dyadic interactions in real time. We apply this model to explain movement in a continuous control task, where participants use a joystick to accomplish a variety of direct (attack / avoid / inspect the other player) and indirect (defend a location from the other player / herd the other player to a location) goals. In the first experiment, participants interacted with a computer opponent. In the second experiment, participants interacted with other participants who had an independent goal. The goal pursuit model successfully accounted for player movement across time as people dynamically adapted to their opponent's behavior, providing insights into participants' trial-specific and person-specific priorities. In addition, the model parameters were used to identify participants' latent goals in both the single-player and two-player tasks, yielding superior performance to classifiers based on raw behavior. Put together, the model can support real-time behavioral prediction and intent inference even in complex multi-agent interactions faced by humans and social AI.