This thesis investigates the learning and decision making process of a principal agent, who must make decisions about if and from whom to obtain reputation information, and how to make decisions once that information is obtained. The thesis addresses the challenges of learning the complex behavior of individual reputation providers, designing ways to combine reputation information from multiple reputation providers, learning about the abilities of individual partner agents and the partner population, and making decisions about which reputation providers to query, while explicitly reasoning about costs associated with obtaining reputation information.

Numerous models, algorithms, and innovative techniques are used to show that a principal agent can thrive in uncertain environments in which it must learn about multiple aspects of the environment to make beneficial decisions. The thesis provides various probabilistic models, such as Bayesian networks and hierarchical models, for learning about partners, the partner population, and reputation providers. It provides a framework for increasing the accuracy of the principal's estimation of a partner's competence by combining reputation information from multiple reputation providers that may vary in their ability to provide useful information. These models are used for a formal cost-benefit analysis of the acquisition of reputation information and add a new level of reasoning in which an agent must decide if and from whom it will query for information.

The performance of these models was tested in experiments of varying complexity. These experiments established that principal agents can identify beneficial reputation providers even when there is a high ratio of prejudiced reputation providers and the providers use complex policies to provide their information. The principal agents' performances were robust to inaccurate and costly reputation information. In addition, the experiments show that the short-term loss in performance that results from buying costly reputation information leads to long-term increases in performance over using no reputation information at all. Additional experiments revealed that in certain multi-agent environments, the use of reputation information can reduce overall agent performance.