As people are increasingly connected to other people and computer agents, forming mixed networks, collaborative teamwork offers great promise for transforming the way people perform their everyday activities and interact with computer agents. This thesis presents new representations and algorithms, developed to enable computer systems to function as effective team members in settings characterized by uncertainty and partial information.

For a collaboration to succeed in such settings, participants need to reason about the possible plans of others, to be able to adapt their plans as needed for coordination, and to support each other's activities. Reasoning on general teamwork models accordingly requires compact representations and efficient decision-theoretic mechanisms. This thesis presents Probabilistic Recipe Trees, a probabilistic representation of agents' beliefs about the probable plans of others, and decision-theoretic mechanisms that use this representation to manage helpful behavior by considering the costs and utilities of computer agents and people participating in collaborative activities. These mechanisms are shown to outperform axiomatic approaches in empirical studies.

The thesis also addresses the challenge that agents participating in a collaborative activity need efficient decision-making algorithms for evaluating the effects of their actions on the collaboration, and they need to reason about the way other participants perceive these actions. This thesis identifies structural characteristics of settings in which computer agents and people collaborate and presents decentralized decision-making algorithms that exploit this structure to achieve up to exponential savings in computation time. Empirical studies with human subjects establish that the utility values computed by this algorithm are a good indicator of human behavior, but learning can help to better understand the way these values are perceived by people.

To demonstrate the usefulness of these teamwork capabilities, the thesis describes an application of collaborative teamwork ideas to a real-world setting of ridesharing. The computational model developed for forming collaborative rideshare plans addresses the challenge of guiding self-interested people to collaboration in a dynamic setting. The empirical evaluation of the application on data collected from the real-world demonstrates the value of collaboration for individual users and environment.