Transferable strategic meta-reasoning models

How do strategic agents make decisions? For the first time, a confluence of advances in agent design, formation of massive online data sets of social behavior, and computational techniques have allowed for researchers to construct and learn much richer models than before. My central thesis is that, when agents engaged in repeated strategic interaction undertake a reasoning or learning process, the behavior resulting from this process can be characterized by two factors: depth of reasoning over base rules and time-horizon of planning. Values for these factors can be learned effectively from interaction and are transferable to new games, producing highly effective strategic responses. The dissertation formally presents a framework for addressing the problem of predicting a population’s behavior using a meta-reasoning model containing these strategic components. To evaluate this model, I explore several experimental case studies that show how to use the framework to predict and respond to behavior using observed data, covering settings ranging from a small number of computer agents to a larger number of human participants.