Human social interaction is complex. Rarely are our interactions independent, nor do they fall into narrow categories of full cooperation or competition. People may share some goals but not others, may cooperate at some times and compete at others. More fundamentally, their behaviors may lead to conflict. To navigate this complexity, we rely on forming beliefs about the goals and behavior of others, what is called a "theory of mind." These beliefs inform not only our decisions about what to do, but also what to believe. In particular, whether we believe a message depends not only on the content of a message but also our model of (and relationship to) the communicator.

There is a range of applications where rich models of human social behavior are important. We can use them to to support the computational study of social and psychological interactions. To develop a better understanding of the causes and remedies of school bullying, for example, we could use agent models of the students to create simulations of classroom social interactions. We could also use such models as a backend to interactive drama applications, modeling the behavior of the characters in the drama. In particular, one might imagine an interactive pedagogical drama or a social training environment where a young teacher could train for handling unruly students by playing the role of a teacher in a virtual classroom populated with virtual students (e.g., see VICTEC). The teacher could explore in the safety of a virtual world the same sorts of situations and dilemmas he is likely to face in the real world, with interaction between human and virtual characters dynamically unfolding based on choices they make in their various roles.

Modeling the social interaction between a bully, his victim, other students and school authorities must take into account a range of factors. What are the bully's goals? What personal or social benefits does a bully envision from acts of bullying? Is there some imagined affront by the victim, or does the bully feel bullying makes him more powerful or even popular in the eyes of other students? How does he or she model school authorities? Why are threats of punishment not influencing behavior? Are they not credible for some reason? Is the cost of being punished outweighed by benefits? In what ways can the bully be coerced, influenced or persuaded not to victimize another student?

We have tackled such issues in the context of creating a social simulation tool, called PsychSim, designed to explore how individuals and groups interact and how those interactions can be influenced. PsychSim allows an end-user to quickly construct a social scenario, where a diverse set of entities, either groups or individuals, interact and communicate among themselves. Each entity has its own goals, relationships (e.g., friendship, hostility, authority) with other entities, private beliefs and mental models about other entities. The simulation tool generates the behavior for these entities and provides explanations of the result in terms of each entity's goals and beliefs. The richness of the entity models allows one to examine the potential consequences of minor variations on the scenario. A user can play different roles by specifying actions or messages for any entity to perform. Alternatively, the simulation itself can perturb the scenario to provide a range of possible behaviors that can identify critical sensitivities of the behavior to deviations (e.g., modified goals, relationships, or mental models).

A unique aspect of the PsychSim design is that agents have fully specified models of other agents.Representation of such recursive models gives PsychSim a powerful mechanism to model a range of factors in a principled way. For instance, we exploit this recursive modeling to:

• allow agents to form complex attributions about others.
• enrich the messages between agents to include the beliefs and goals of other agents.
• give agents the ability to reason about message credibility in a human-like fashion
• model the influence such recursive models have on an agent's own behavior and communication.
• model the influence observations of another's behavior have on the agent's recursive model of that other.
• enrich the explanations provided to the user.

Research