Probabilistic modeling and reasoning currently underlie many real-world applications in diverse areas such as the world wide web, medical informatics, robotics, bioinformatics, and information security. This project aims at significantly improving the scale and utility of probabilistic reasoning systems in these application areas, where success has become increasingly dependent on the availability of efficient and accurate probabilistic reasoning systems. The project is focused on a particular class of probabilistic models, known as Bayesian and Markov networks; these are among the most successful models studied by computer scientists and statisticians. This project is concerned with attaining the highest accuracy of reasoning that is feasible under real-world constraints on computational resources. The project is based on new, fundamental discoveries by the PI's group, showing that the efficiency and accuracy of reasoning can be finely controlled by approximating model dependencies in a dynamic fashion driven by user queries. These discoveries have formed the basis of a new semantics, and a concrete realization, of one of the most influential theories of probabilistic reasoning during the last decade, known as generalized belief propagation (GBP). In addition to pursuing the theoretical and practical implications of the new semantics of GBP, the project also aims at producing a comprehensive software system that embodies this novel and practical realization of GBP, with the intent of making it publicly available to the broad scientific community on a web site. It is anticipated that the developed system, with its surrounding theory and practice, will significantly advance the state of the art in probabilistic reasoning, to the point of both allowing new applications to be handled efficiently, and also increasing the scale and scope of existing applications.
