The topic of predicting the behavior of goal-driven agents has been studied from many angles in numerous fields. There is no doubt it is an important task with a plethora of applications. In the machine learning community, the literature has focused primarily on the single-agent setting. Inverse optimal control techniques designed for this setting assume that the agent aims to maximizing its utility. That is, the agent is optimally solving an unknown control problem. This assumption allows them to both leverage off-the-shelf planners and attain strong statistical guarantees.
In this thesis, we will tackle behavior prediction in multi-agent settings. Here, unlike in the single-agent case, one may not myopically maximize its reward–it must also reason about the other agents' decisions. As such, we must instead employ a game-theoretic equilibrium concept in place of the utility-maximizing optimality criterion. We will consider this challenge from both statistical and computational perspectives. That is, we propose new methods for both structured and unstructured settings that are computationally efficient and require fewer observations than standard machine learning techniques.
Drew Bagnell (Chair)
Michael Littman (Brown University)